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;
830 /* Portion of address space to scan in MB */
831 unsigned int sysctl_numa_balancing_scan_size
= 256;
833 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
834 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
836 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
838 unsigned long rss
= 0;
839 unsigned long nr_scan_pages
;
842 * Calculations based on RSS as non-present and empty pages are skipped
843 * by the PTE scanner and NUMA hinting faults should be trapped based
846 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
847 rss
= get_mm_rss(p
->mm
);
851 rss
= round_up(rss
, nr_scan_pages
);
852 return rss
/ nr_scan_pages
;
855 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
856 #define MAX_SCAN_WINDOW 2560
858 static unsigned int task_scan_min(struct task_struct
*p
)
860 unsigned int scan
, floor
;
861 unsigned int windows
= 1;
863 if (sysctl_numa_balancing_scan_size
< MAX_SCAN_WINDOW
)
864 windows
= MAX_SCAN_WINDOW
/ sysctl_numa_balancing_scan_size
;
865 floor
= 1000 / windows
;
867 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
868 return max_t(unsigned int, floor
, scan
);
871 static unsigned int task_scan_max(struct task_struct
*p
)
873 unsigned int smin
= task_scan_min(p
);
876 /* Watch for min being lower than max due to floor calculations */
877 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
878 return max(smin
, smax
);
882 * Once a preferred node is selected the scheduler balancer will prefer moving
883 * a task to that node for sysctl_numa_balancing_settle_count number of PTE
884 * scans. This will give the process the chance to accumulate more faults on
885 * the preferred node but still allow the scheduler to move the task again if
886 * the nodes CPUs are overloaded.
888 unsigned int sysctl_numa_balancing_settle_count __read_mostly
= 4;
890 static void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
892 rq
->nr_numa_running
+= (p
->numa_preferred_nid
!= -1);
893 rq
->nr_preferred_running
+= (p
->numa_preferred_nid
== task_node(p
));
896 static void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
898 rq
->nr_numa_running
-= (p
->numa_preferred_nid
!= -1);
899 rq
->nr_preferred_running
-= (p
->numa_preferred_nid
== task_node(p
));
905 spinlock_t lock
; /* nr_tasks, tasks */
908 struct list_head task_list
;
911 atomic_long_t total_faults
;
912 atomic_long_t faults
[0];
915 pid_t
task_numa_group_id(struct task_struct
*p
)
917 return p
->numa_group
? p
->numa_group
->gid
: 0;
920 static inline int task_faults_idx(int nid
, int priv
)
922 return 2 * nid
+ priv
;
925 static inline unsigned long task_faults(struct task_struct
*p
, int nid
)
930 return p
->numa_faults
[task_faults_idx(nid
, 0)] +
931 p
->numa_faults
[task_faults_idx(nid
, 1)];
934 static inline unsigned long group_faults(struct task_struct
*p
, int nid
)
939 return atomic_long_read(&p
->numa_group
->faults
[2*nid
]) +
940 atomic_long_read(&p
->numa_group
->faults
[2*nid
+1]);
944 * These return the fraction of accesses done by a particular task, or
945 * task group, on a particular numa node. The group weight is given a
946 * larger multiplier, in order to group tasks together that are almost
947 * evenly spread out between numa nodes.
949 static inline unsigned long task_weight(struct task_struct
*p
, int nid
)
951 unsigned long total_faults
;
956 total_faults
= p
->total_numa_faults
;
961 return 1000 * task_faults(p
, nid
) / total_faults
;
964 static inline unsigned long group_weight(struct task_struct
*p
, int nid
)
966 unsigned long total_faults
;
971 total_faults
= atomic_long_read(&p
->numa_group
->total_faults
);
976 return 1000 * group_faults(p
, nid
) / total_faults
;
979 static unsigned long weighted_cpuload(const int cpu
);
980 static unsigned long source_load(int cpu
, int type
);
981 static unsigned long target_load(int cpu
, int type
);
982 static unsigned long power_of(int cpu
);
983 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
);
985 /* Cached statistics for all CPUs within a node */
987 unsigned long nr_running
;
990 /* Total compute capacity of CPUs on a node */
993 /* Approximate capacity in terms of runnable tasks on a node */
994 unsigned long capacity
;
999 * XXX borrowed from update_sg_lb_stats
1001 static void update_numa_stats(struct numa_stats
*ns
, int nid
)
1005 memset(ns
, 0, sizeof(*ns
));
1006 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
1007 struct rq
*rq
= cpu_rq(cpu
);
1009 ns
->nr_running
+= rq
->nr_running
;
1010 ns
->load
+= weighted_cpuload(cpu
);
1011 ns
->power
+= power_of(cpu
);
1014 ns
->load
= (ns
->load
* SCHED_POWER_SCALE
) / ns
->power
;
1015 ns
->capacity
= DIV_ROUND_CLOSEST(ns
->power
, SCHED_POWER_SCALE
);
1016 ns
->has_capacity
= (ns
->nr_running
< ns
->capacity
);
1019 struct task_numa_env
{
1020 struct task_struct
*p
;
1022 int src_cpu
, src_nid
;
1023 int dst_cpu
, dst_nid
;
1025 struct numa_stats src_stats
, dst_stats
;
1027 int imbalance_pct
, idx
;
1029 struct task_struct
*best_task
;
1034 static void task_numa_assign(struct task_numa_env
*env
,
1035 struct task_struct
*p
, long imp
)
1038 put_task_struct(env
->best_task
);
1043 env
->best_imp
= imp
;
1044 env
->best_cpu
= env
->dst_cpu
;
1048 * This checks if the overall compute and NUMA accesses of the system would
1049 * be improved if the source tasks was migrated to the target dst_cpu taking
1050 * into account that it might be best if task running on the dst_cpu should
1051 * be exchanged with the source task
1053 static void task_numa_compare(struct task_numa_env
*env
,
1054 long taskimp
, long groupimp
)
1056 struct rq
*src_rq
= cpu_rq(env
->src_cpu
);
1057 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1058 struct task_struct
*cur
;
1059 long dst_load
, src_load
;
1061 long imp
= (groupimp
> 0) ? groupimp
: taskimp
;
1064 cur
= ACCESS_ONCE(dst_rq
->curr
);
1065 if (cur
->pid
== 0) /* idle */
1069 * "imp" is the fault differential for the source task between the
1070 * source and destination node. Calculate the total differential for
1071 * the source task and potential destination task. The more negative
1072 * the value is, the more rmeote accesses that would be expected to
1073 * be incurred if the tasks were swapped.
1076 /* Skip this swap candidate if cannot move to the source cpu */
1077 if (!cpumask_test_cpu(env
->src_cpu
, tsk_cpus_allowed(cur
)))
1081 * If dst and source tasks are in the same NUMA group, or not
1082 * in any group then look only at task weights.
1084 if (cur
->numa_group
== env
->p
->numa_group
) {
1085 imp
= taskimp
+ task_weight(cur
, env
->src_nid
) -
1086 task_weight(cur
, env
->dst_nid
);
1088 * Add some hysteresis to prevent swapping the
1089 * tasks within a group over tiny differences.
1091 if (cur
->numa_group
)
1095 * Compare the group weights. If a task is all by
1096 * itself (not part of a group), use the task weight
1099 if (env
->p
->numa_group
)
1104 if (cur
->numa_group
)
1105 imp
+= group_weight(cur
, env
->src_nid
) -
1106 group_weight(cur
, env
->dst_nid
);
1108 imp
+= task_weight(cur
, env
->src_nid
) -
1109 task_weight(cur
, env
->dst_nid
);
1113 if (imp
< env
->best_imp
)
1117 /* Is there capacity at our destination? */
1118 if (env
->src_stats
.has_capacity
&&
1119 !env
->dst_stats
.has_capacity
)
1125 /* Balance doesn't matter much if we're running a task per cpu */
1126 if (src_rq
->nr_running
== 1 && dst_rq
->nr_running
== 1)
1130 * In the overloaded case, try and keep the load balanced.
1133 dst_load
= env
->dst_stats
.load
;
1134 src_load
= env
->src_stats
.load
;
1136 /* XXX missing power terms */
1137 load
= task_h_load(env
->p
);
1142 load
= task_h_load(cur
);
1147 /* make src_load the smaller */
1148 if (dst_load
< src_load
)
1149 swap(dst_load
, src_load
);
1151 if (src_load
* env
->imbalance_pct
< dst_load
* 100)
1155 task_numa_assign(env
, cur
, imp
);
1160 static void task_numa_find_cpu(struct task_numa_env
*env
,
1161 long taskimp
, long groupimp
)
1165 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1166 /* Skip this CPU if the source task cannot migrate */
1167 if (!cpumask_test_cpu(cpu
, tsk_cpus_allowed(env
->p
)))
1171 task_numa_compare(env
, taskimp
, groupimp
);
1175 static int task_numa_migrate(struct task_struct
*p
)
1177 struct task_numa_env env
= {
1180 .src_cpu
= task_cpu(p
),
1181 .src_nid
= task_node(p
),
1183 .imbalance_pct
= 112,
1189 struct sched_domain
*sd
;
1190 unsigned long taskweight
, groupweight
;
1192 long taskimp
, groupimp
;
1195 * Pick the lowest SD_NUMA domain, as that would have the smallest
1196 * imbalance and would be the first to start moving tasks about.
1198 * And we want to avoid any moving of tasks about, as that would create
1199 * random movement of tasks -- counter the numa conditions we're trying
1203 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1204 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1207 taskweight
= task_weight(p
, env
.src_nid
);
1208 groupweight
= group_weight(p
, env
.src_nid
);
1209 update_numa_stats(&env
.src_stats
, env
.src_nid
);
1210 env
.dst_nid
= p
->numa_preferred_nid
;
1211 taskimp
= task_weight(p
, env
.dst_nid
) - taskweight
;
1212 groupimp
= group_weight(p
, env
.dst_nid
) - groupweight
;
1213 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1215 /* If the preferred nid has capacity, try to use it. */
1216 if (env
.dst_stats
.has_capacity
)
1217 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1219 /* No space available on the preferred nid. Look elsewhere. */
1220 if (env
.best_cpu
== -1) {
1221 for_each_online_node(nid
) {
1222 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
1225 /* Only consider nodes where both task and groups benefit */
1226 taskimp
= task_weight(p
, nid
) - taskweight
;
1227 groupimp
= group_weight(p
, nid
) - groupweight
;
1228 if (taskimp
< 0 && groupimp
< 0)
1232 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1233 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1237 /* No better CPU than the current one was found. */
1238 if (env
.best_cpu
== -1)
1241 sched_setnuma(p
, env
.dst_nid
);
1244 * Reset the scan period if the task is being rescheduled on an
1245 * alternative node to recheck if the tasks is now properly placed.
1247 p
->numa_scan_period
= task_scan_min(p
);
1249 if (env
.best_task
== NULL
) {
1250 int ret
= migrate_task_to(p
, env
.best_cpu
);
1254 ret
= migrate_swap(p
, env
.best_task
);
1255 put_task_struct(env
.best_task
);
1259 /* Attempt to migrate a task to a CPU on the preferred node. */
1260 static void numa_migrate_preferred(struct task_struct
*p
)
1262 /* Success if task is already running on preferred CPU */
1263 p
->numa_migrate_retry
= 0;
1264 if (cpu_to_node(task_cpu(p
)) == p
->numa_preferred_nid
) {
1266 * If migration is temporarily disabled due to a task migration
1267 * then re-enable it now as the task is running on its
1268 * preferred node and memory should migrate locally
1270 if (!p
->numa_migrate_seq
)
1271 p
->numa_migrate_seq
++;
1275 /* This task has no NUMA fault statistics yet */
1276 if (unlikely(p
->numa_preferred_nid
== -1))
1279 /* Otherwise, try migrate to a CPU on the preferred node */
1280 if (task_numa_migrate(p
) != 0)
1281 p
->numa_migrate_retry
= jiffies
+ HZ
*5;
1285 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1286 * increments. The more local the fault statistics are, the higher the scan
1287 * period will be for the next scan window. If local/remote ratio is below
1288 * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the
1289 * scan period will decrease
1291 #define NUMA_PERIOD_SLOTS 10
1292 #define NUMA_PERIOD_THRESHOLD 3
1295 * Increase the scan period (slow down scanning) if the majority of
1296 * our memory is already on our local node, or if the majority of
1297 * the page accesses are shared with other processes.
1298 * Otherwise, decrease the scan period.
1300 static void update_task_scan_period(struct task_struct
*p
,
1301 unsigned long shared
, unsigned long private)
1303 unsigned int period_slot
;
1307 unsigned long remote
= p
->numa_faults_locality
[0];
1308 unsigned long local
= p
->numa_faults_locality
[1];
1311 * If there were no record hinting faults then either the task is
1312 * completely idle or all activity is areas that are not of interest
1313 * to automatic numa balancing. Scan slower
1315 if (local
+ shared
== 0) {
1316 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1317 p
->numa_scan_period
<< 1);
1319 p
->mm
->numa_next_scan
= jiffies
+
1320 msecs_to_jiffies(p
->numa_scan_period
);
1326 * Prepare to scale scan period relative to the current period.
1327 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1328 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1329 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1331 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
1332 ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
1333 if (ratio
>= NUMA_PERIOD_THRESHOLD
) {
1334 int slot
= ratio
- NUMA_PERIOD_THRESHOLD
;
1337 diff
= slot
* period_slot
;
1339 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
1342 * Scale scan rate increases based on sharing. There is an
1343 * inverse relationship between the degree of sharing and
1344 * the adjustment made to the scanning period. Broadly
1345 * speaking the intent is that there is little point
1346 * scanning faster if shared accesses dominate as it may
1347 * simply bounce migrations uselessly
1349 period_slot
= DIV_ROUND_UP(diff
, NUMA_PERIOD_SLOTS
);
1350 ratio
= DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS
, (private + shared
));
1351 diff
= (diff
* ratio
) / NUMA_PERIOD_SLOTS
;
1354 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
1355 task_scan_min(p
), task_scan_max(p
));
1356 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1359 static void task_numa_placement(struct task_struct
*p
)
1361 int seq
, nid
, max_nid
= -1, max_group_nid
= -1;
1362 unsigned long max_faults
= 0, max_group_faults
= 0;
1363 unsigned long fault_types
[2] = { 0, 0 };
1364 spinlock_t
*group_lock
= NULL
;
1366 seq
= ACCESS_ONCE(p
->mm
->numa_scan_seq
);
1367 if (p
->numa_scan_seq
== seq
)
1369 p
->numa_scan_seq
= seq
;
1370 p
->numa_migrate_seq
++;
1371 p
->numa_scan_period_max
= task_scan_max(p
);
1373 /* If the task is part of a group prevent parallel updates to group stats */
1374 if (p
->numa_group
) {
1375 group_lock
= &p
->numa_group
->lock
;
1376 spin_lock(group_lock
);
1379 /* Find the node with the highest number of faults */
1380 for_each_online_node(nid
) {
1381 unsigned long faults
= 0, group_faults
= 0;
1384 for (priv
= 0; priv
< 2; priv
++) {
1387 i
= task_faults_idx(nid
, priv
);
1388 diff
= -p
->numa_faults
[i
];
1390 /* Decay existing window, copy faults since last scan */
1391 p
->numa_faults
[i
] >>= 1;
1392 p
->numa_faults
[i
] += p
->numa_faults_buffer
[i
];
1393 fault_types
[priv
] += p
->numa_faults_buffer
[i
];
1394 p
->numa_faults_buffer
[i
] = 0;
1396 faults
+= p
->numa_faults
[i
];
1397 diff
+= p
->numa_faults
[i
];
1398 p
->total_numa_faults
+= diff
;
1399 if (p
->numa_group
) {
1400 /* safe because we can only change our own group */
1401 atomic_long_add(diff
, &p
->numa_group
->faults
[i
]);
1402 atomic_long_add(diff
, &p
->numa_group
->total_faults
);
1403 group_faults
+= atomic_long_read(&p
->numa_group
->faults
[i
]);
1407 if (faults
> max_faults
) {
1408 max_faults
= faults
;
1412 if (group_faults
> max_group_faults
) {
1413 max_group_faults
= group_faults
;
1414 max_group_nid
= nid
;
1418 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
1420 if (p
->numa_group
) {
1422 * If the preferred task and group nids are different,
1423 * iterate over the nodes again to find the best place.
1425 if (max_nid
!= max_group_nid
) {
1426 unsigned long weight
, max_weight
= 0;
1428 for_each_online_node(nid
) {
1429 weight
= task_weight(p
, nid
) + group_weight(p
, nid
);
1430 if (weight
> max_weight
) {
1431 max_weight
= weight
;
1437 spin_unlock(group_lock
);
1440 /* Preferred node as the node with the most faults */
1441 if (max_faults
&& max_nid
!= p
->numa_preferred_nid
) {
1442 /* Update the preferred nid and migrate task if possible */
1443 sched_setnuma(p
, max_nid
);
1444 numa_migrate_preferred(p
);
1448 static inline int get_numa_group(struct numa_group
*grp
)
1450 return atomic_inc_not_zero(&grp
->refcount
);
1453 static inline void put_numa_group(struct numa_group
*grp
)
1455 if (atomic_dec_and_test(&grp
->refcount
))
1456 kfree_rcu(grp
, rcu
);
1459 static void double_lock(spinlock_t
*l1
, spinlock_t
*l2
)
1465 spin_lock_nested(l2
, SINGLE_DEPTH_NESTING
);
1468 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
1471 struct numa_group
*grp
, *my_grp
;
1472 struct task_struct
*tsk
;
1474 int cpu
= cpupid_to_cpu(cpupid
);
1477 if (unlikely(!p
->numa_group
)) {
1478 unsigned int size
= sizeof(struct numa_group
) +
1479 2*nr_node_ids
*sizeof(atomic_long_t
);
1481 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
1485 atomic_set(&grp
->refcount
, 1);
1486 spin_lock_init(&grp
->lock
);
1487 INIT_LIST_HEAD(&grp
->task_list
);
1490 for (i
= 0; i
< 2*nr_node_ids
; i
++)
1491 atomic_long_set(&grp
->faults
[i
], p
->numa_faults
[i
]);
1493 atomic_long_set(&grp
->total_faults
, p
->total_numa_faults
);
1495 list_add(&p
->numa_entry
, &grp
->task_list
);
1497 rcu_assign_pointer(p
->numa_group
, grp
);
1501 tsk
= ACCESS_ONCE(cpu_rq(cpu
)->curr
);
1503 if (!cpupid_match_pid(tsk
, cpupid
))
1506 grp
= rcu_dereference(tsk
->numa_group
);
1510 my_grp
= p
->numa_group
;
1515 * Only join the other group if its bigger; if we're the bigger group,
1516 * the other task will join us.
1518 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
1522 * Tie-break on the grp address.
1524 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
1527 /* Always join threads in the same process. */
1528 if (tsk
->mm
== current
->mm
)
1531 /* Simple filter to avoid false positives due to PID collisions */
1532 if (flags
& TNF_SHARED
)
1535 /* Update priv based on whether false sharing was detected */
1538 if (join
&& !get_numa_group(grp
))
1547 for (i
= 0; i
< 2*nr_node_ids
; i
++) {
1548 atomic_long_sub(p
->numa_faults
[i
], &my_grp
->faults
[i
]);
1549 atomic_long_add(p
->numa_faults
[i
], &grp
->faults
[i
]);
1551 atomic_long_sub(p
->total_numa_faults
, &my_grp
->total_faults
);
1552 atomic_long_add(p
->total_numa_faults
, &grp
->total_faults
);
1554 double_lock(&my_grp
->lock
, &grp
->lock
);
1556 list_move(&p
->numa_entry
, &grp
->task_list
);
1560 spin_unlock(&my_grp
->lock
);
1561 spin_unlock(&grp
->lock
);
1563 rcu_assign_pointer(p
->numa_group
, grp
);
1565 put_numa_group(my_grp
);
1568 void task_numa_free(struct task_struct
*p
)
1570 struct numa_group
*grp
= p
->numa_group
;
1572 void *numa_faults
= p
->numa_faults
;
1575 for (i
= 0; i
< 2*nr_node_ids
; i
++)
1576 atomic_long_sub(p
->numa_faults
[i
], &grp
->faults
[i
]);
1578 atomic_long_sub(p
->total_numa_faults
, &grp
->total_faults
);
1580 spin_lock(&grp
->lock
);
1581 list_del(&p
->numa_entry
);
1583 spin_unlock(&grp
->lock
);
1584 rcu_assign_pointer(p
->numa_group
, NULL
);
1585 put_numa_group(grp
);
1588 p
->numa_faults
= NULL
;
1589 p
->numa_faults_buffer
= NULL
;
1594 * Got a PROT_NONE fault for a page on @node.
1596 void task_numa_fault(int last_cpupid
, int node
, int pages
, int flags
)
1598 struct task_struct
*p
= current
;
1599 bool migrated
= flags
& TNF_MIGRATED
;
1602 if (!numabalancing_enabled
)
1605 /* for example, ksmd faulting in a user's mm */
1609 /* Do not worry about placement if exiting */
1610 if (p
->state
== TASK_DEAD
)
1613 /* Allocate buffer to track faults on a per-node basis */
1614 if (unlikely(!p
->numa_faults
)) {
1615 int size
= sizeof(*p
->numa_faults
) * 2 * nr_node_ids
;
1617 /* numa_faults and numa_faults_buffer share the allocation */
1618 p
->numa_faults
= kzalloc(size
* 2, GFP_KERNEL
|__GFP_NOWARN
);
1619 if (!p
->numa_faults
)
1622 BUG_ON(p
->numa_faults_buffer
);
1623 p
->numa_faults_buffer
= p
->numa_faults
+ (2 * nr_node_ids
);
1624 p
->total_numa_faults
= 0;
1625 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1629 * First accesses are treated as private, otherwise consider accesses
1630 * to be private if the accessing pid has not changed
1632 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
1635 priv
= cpupid_match_pid(p
, last_cpupid
);
1636 if (!priv
&& !(flags
& TNF_NO_GROUP
))
1637 task_numa_group(p
, last_cpupid
, flags
, &priv
);
1640 task_numa_placement(p
);
1642 /* Retry task to preferred node migration if it previously failed */
1643 if (p
->numa_migrate_retry
&& time_after(jiffies
, p
->numa_migrate_retry
))
1644 numa_migrate_preferred(p
);
1647 p
->numa_pages_migrated
+= pages
;
1649 p
->numa_faults_buffer
[task_faults_idx(node
, priv
)] += pages
;
1650 p
->numa_faults_locality
[!!(flags
& TNF_FAULT_LOCAL
)] += pages
;
1653 static void reset_ptenuma_scan(struct task_struct
*p
)
1655 ACCESS_ONCE(p
->mm
->numa_scan_seq
)++;
1656 p
->mm
->numa_scan_offset
= 0;
1660 * The expensive part of numa migration is done from task_work context.
1661 * Triggered from task_tick_numa().
1663 void task_numa_work(struct callback_head
*work
)
1665 unsigned long migrate
, next_scan
, now
= jiffies
;
1666 struct task_struct
*p
= current
;
1667 struct mm_struct
*mm
= p
->mm
;
1668 struct vm_area_struct
*vma
;
1669 unsigned long start
, end
;
1670 unsigned long nr_pte_updates
= 0;
1673 WARN_ON_ONCE(p
!= container_of(work
, struct task_struct
, numa_work
));
1675 work
->next
= work
; /* protect against double add */
1677 * Who cares about NUMA placement when they're dying.
1679 * NOTE: make sure not to dereference p->mm before this check,
1680 * exit_task_work() happens _after_ exit_mm() so we could be called
1681 * without p->mm even though we still had it when we enqueued this
1684 if (p
->flags
& PF_EXITING
)
1687 if (!mm
->numa_next_scan
) {
1688 mm
->numa_next_scan
= now
+
1689 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
1693 * Enforce maximal scan/migration frequency..
1695 migrate
= mm
->numa_next_scan
;
1696 if (time_before(now
, migrate
))
1699 if (p
->numa_scan_period
== 0) {
1700 p
->numa_scan_period_max
= task_scan_max(p
);
1701 p
->numa_scan_period
= task_scan_min(p
);
1704 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
1705 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
1709 * Delay this task enough that another task of this mm will likely win
1710 * the next time around.
1712 p
->node_stamp
+= 2 * TICK_NSEC
;
1714 start
= mm
->numa_scan_offset
;
1715 pages
= sysctl_numa_balancing_scan_size
;
1716 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
1720 down_read(&mm
->mmap_sem
);
1721 vma
= find_vma(mm
, start
);
1723 reset_ptenuma_scan(p
);
1727 for (; vma
; vma
= vma
->vm_next
) {
1728 if (!vma_migratable(vma
) || !vma_policy_mof(p
, vma
))
1732 * Shared library pages mapped by multiple processes are not
1733 * migrated as it is expected they are cache replicated. Avoid
1734 * hinting faults in read-only file-backed mappings or the vdso
1735 * as migrating the pages will be of marginal benefit.
1738 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
1742 start
= max(start
, vma
->vm_start
);
1743 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
1744 end
= min(end
, vma
->vm_end
);
1745 nr_pte_updates
+= change_prot_numa(vma
, start
, end
);
1748 * Scan sysctl_numa_balancing_scan_size but ensure that
1749 * at least one PTE is updated so that unused virtual
1750 * address space is quickly skipped.
1753 pages
-= (end
- start
) >> PAGE_SHIFT
;
1758 } while (end
!= vma
->vm_end
);
1763 * It is possible to reach the end of the VMA list but the last few
1764 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1765 * would find the !migratable VMA on the next scan but not reset the
1766 * scanner to the start so check it now.
1769 mm
->numa_scan_offset
= start
;
1771 reset_ptenuma_scan(p
);
1772 up_read(&mm
->mmap_sem
);
1776 * Drive the periodic memory faults..
1778 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
1780 struct callback_head
*work
= &curr
->numa_work
;
1784 * We don't care about NUMA placement if we don't have memory.
1786 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
1790 * Using runtime rather than walltime has the dual advantage that
1791 * we (mostly) drive the selection from busy threads and that the
1792 * task needs to have done some actual work before we bother with
1795 now
= curr
->se
.sum_exec_runtime
;
1796 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
1798 if (now
- curr
->node_stamp
> period
) {
1799 if (!curr
->node_stamp
)
1800 curr
->numa_scan_period
= task_scan_min(curr
);
1801 curr
->node_stamp
+= period
;
1803 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
1804 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
1805 task_work_add(curr
, work
, true);
1810 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
1814 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
1818 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
1821 #endif /* CONFIG_NUMA_BALANCING */
1824 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1826 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
1827 if (!parent_entity(se
))
1828 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
1830 if (entity_is_task(se
)) {
1831 struct rq
*rq
= rq_of(cfs_rq
);
1833 account_numa_enqueue(rq
, task_of(se
));
1834 list_add(&se
->group_node
, &rq
->cfs_tasks
);
1837 cfs_rq
->nr_running
++;
1841 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1843 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
1844 if (!parent_entity(se
))
1845 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
1846 if (entity_is_task(se
)) {
1847 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
1848 list_del_init(&se
->group_node
);
1850 cfs_rq
->nr_running
--;
1853 #ifdef CONFIG_FAIR_GROUP_SCHED
1855 static inline long calc_tg_weight(struct task_group
*tg
, struct cfs_rq
*cfs_rq
)
1860 * Use this CPU's actual weight instead of the last load_contribution
1861 * to gain a more accurate current total weight. See
1862 * update_cfs_rq_load_contribution().
1864 tg_weight
= atomic_long_read(&tg
->load_avg
);
1865 tg_weight
-= cfs_rq
->tg_load_contrib
;
1866 tg_weight
+= cfs_rq
->load
.weight
;
1871 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
1873 long tg_weight
, load
, shares
;
1875 tg_weight
= calc_tg_weight(tg
, cfs_rq
);
1876 load
= cfs_rq
->load
.weight
;
1878 shares
= (tg
->shares
* load
);
1880 shares
/= tg_weight
;
1882 if (shares
< MIN_SHARES
)
1883 shares
= MIN_SHARES
;
1884 if (shares
> tg
->shares
)
1885 shares
= tg
->shares
;
1889 # else /* CONFIG_SMP */
1890 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
1894 # endif /* CONFIG_SMP */
1895 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
1896 unsigned long weight
)
1899 /* commit outstanding execution time */
1900 if (cfs_rq
->curr
== se
)
1901 update_curr(cfs_rq
);
1902 account_entity_dequeue(cfs_rq
, se
);
1905 update_load_set(&se
->load
, weight
);
1908 account_entity_enqueue(cfs_rq
, se
);
1911 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
1913 static void update_cfs_shares(struct cfs_rq
*cfs_rq
)
1915 struct task_group
*tg
;
1916 struct sched_entity
*se
;
1920 se
= tg
->se
[cpu_of(rq_of(cfs_rq
))];
1921 if (!se
|| throttled_hierarchy(cfs_rq
))
1924 if (likely(se
->load
.weight
== tg
->shares
))
1927 shares
= calc_cfs_shares(cfs_rq
, tg
);
1929 reweight_entity(cfs_rq_of(se
), se
, shares
);
1931 #else /* CONFIG_FAIR_GROUP_SCHED */
1932 static inline void update_cfs_shares(struct cfs_rq
*cfs_rq
)
1935 #endif /* CONFIG_FAIR_GROUP_SCHED */
1939 * We choose a half-life close to 1 scheduling period.
1940 * Note: The tables below are dependent on this value.
1942 #define LOAD_AVG_PERIOD 32
1943 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1944 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1946 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1947 static const u32 runnable_avg_yN_inv
[] = {
1948 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1949 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1950 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1951 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1952 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1953 0x85aac367, 0x82cd8698,
1957 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1958 * over-estimates when re-combining.
1960 static const u32 runnable_avg_yN_sum
[] = {
1961 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1962 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1963 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1968 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1970 static __always_inline u64
decay_load(u64 val
, u64 n
)
1972 unsigned int local_n
;
1976 else if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
1979 /* after bounds checking we can collapse to 32-bit */
1983 * As y^PERIOD = 1/2, we can combine
1984 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1985 * With a look-up table which covers k^n (n<PERIOD)
1987 * To achieve constant time decay_load.
1989 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
1990 val
>>= local_n
/ LOAD_AVG_PERIOD
;
1991 local_n
%= LOAD_AVG_PERIOD
;
1994 val
*= runnable_avg_yN_inv
[local_n
];
1995 /* We don't use SRR here since we always want to round down. */
2000 * For updates fully spanning n periods, the contribution to runnable
2001 * average will be: \Sum 1024*y^n
2003 * We can compute this reasonably efficiently by combining:
2004 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2006 static u32
__compute_runnable_contrib(u64 n
)
2010 if (likely(n
<= LOAD_AVG_PERIOD
))
2011 return runnable_avg_yN_sum
[n
];
2012 else if (unlikely(n
>= LOAD_AVG_MAX_N
))
2013 return LOAD_AVG_MAX
;
2015 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2017 contrib
/= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2018 contrib
+= runnable_avg_yN_sum
[LOAD_AVG_PERIOD
];
2020 n
-= LOAD_AVG_PERIOD
;
2021 } while (n
> LOAD_AVG_PERIOD
);
2023 contrib
= decay_load(contrib
, n
);
2024 return contrib
+ runnable_avg_yN_sum
[n
];
2028 * We can represent the historical contribution to runnable average as the
2029 * coefficients of a geometric series. To do this we sub-divide our runnable
2030 * history into segments of approximately 1ms (1024us); label the segment that
2031 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2033 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2035 * (now) (~1ms ago) (~2ms ago)
2037 * Let u_i denote the fraction of p_i that the entity was runnable.
2039 * We then designate the fractions u_i as our co-efficients, yielding the
2040 * following representation of historical load:
2041 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2043 * We choose y based on the with of a reasonably scheduling period, fixing:
2046 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2047 * approximately half as much as the contribution to load within the last ms
2050 * When a period "rolls over" and we have new u_0`, multiplying the previous
2051 * sum again by y is sufficient to update:
2052 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2053 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2055 static __always_inline
int __update_entity_runnable_avg(u64 now
,
2056 struct sched_avg
*sa
,
2060 u32 runnable_contrib
;
2061 int delta_w
, decayed
= 0;
2063 delta
= now
- sa
->last_runnable_update
;
2065 * This should only happen when time goes backwards, which it
2066 * unfortunately does during sched clock init when we swap over to TSC.
2068 if ((s64
)delta
< 0) {
2069 sa
->last_runnable_update
= now
;
2074 * Use 1024ns as the unit of measurement since it's a reasonable
2075 * approximation of 1us and fast to compute.
2080 sa
->last_runnable_update
= now
;
2082 /* delta_w is the amount already accumulated against our next period */
2083 delta_w
= sa
->runnable_avg_period
% 1024;
2084 if (delta
+ delta_w
>= 1024) {
2085 /* period roll-over */
2089 * Now that we know we're crossing a period boundary, figure
2090 * out how much from delta we need to complete the current
2091 * period and accrue it.
2093 delta_w
= 1024 - delta_w
;
2095 sa
->runnable_avg_sum
+= delta_w
;
2096 sa
->runnable_avg_period
+= delta_w
;
2100 /* Figure out how many additional periods this update spans */
2101 periods
= delta
/ 1024;
2104 sa
->runnable_avg_sum
= decay_load(sa
->runnable_avg_sum
,
2106 sa
->runnable_avg_period
= decay_load(sa
->runnable_avg_period
,
2109 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2110 runnable_contrib
= __compute_runnable_contrib(periods
);
2112 sa
->runnable_avg_sum
+= runnable_contrib
;
2113 sa
->runnable_avg_period
+= runnable_contrib
;
2116 /* Remainder of delta accrued against u_0` */
2118 sa
->runnable_avg_sum
+= delta
;
2119 sa
->runnable_avg_period
+= delta
;
2124 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2125 static inline u64
__synchronize_entity_decay(struct sched_entity
*se
)
2127 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2128 u64 decays
= atomic64_read(&cfs_rq
->decay_counter
);
2130 decays
-= se
->avg
.decay_count
;
2134 se
->avg
.load_avg_contrib
= decay_load(se
->avg
.load_avg_contrib
, decays
);
2135 se
->avg
.decay_count
= 0;
2140 #ifdef CONFIG_FAIR_GROUP_SCHED
2141 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2144 struct task_group
*tg
= cfs_rq
->tg
;
2147 tg_contrib
= cfs_rq
->runnable_load_avg
+ cfs_rq
->blocked_load_avg
;
2148 tg_contrib
-= cfs_rq
->tg_load_contrib
;
2150 if (force_update
|| abs(tg_contrib
) > cfs_rq
->tg_load_contrib
/ 8) {
2151 atomic_long_add(tg_contrib
, &tg
->load_avg
);
2152 cfs_rq
->tg_load_contrib
+= tg_contrib
;
2157 * Aggregate cfs_rq runnable averages into an equivalent task_group
2158 * representation for computing load contributions.
2160 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2161 struct cfs_rq
*cfs_rq
)
2163 struct task_group
*tg
= cfs_rq
->tg
;
2166 /* The fraction of a cpu used by this cfs_rq */
2167 contrib
= div_u64(sa
->runnable_avg_sum
<< NICE_0_SHIFT
,
2168 sa
->runnable_avg_period
+ 1);
2169 contrib
-= cfs_rq
->tg_runnable_contrib
;
2171 if (abs(contrib
) > cfs_rq
->tg_runnable_contrib
/ 64) {
2172 atomic_add(contrib
, &tg
->runnable_avg
);
2173 cfs_rq
->tg_runnable_contrib
+= contrib
;
2177 static inline void __update_group_entity_contrib(struct sched_entity
*se
)
2179 struct cfs_rq
*cfs_rq
= group_cfs_rq(se
);
2180 struct task_group
*tg
= cfs_rq
->tg
;
2185 contrib
= cfs_rq
->tg_load_contrib
* tg
->shares
;
2186 se
->avg
.load_avg_contrib
= div_u64(contrib
,
2187 atomic_long_read(&tg
->load_avg
) + 1);
2190 * For group entities we need to compute a correction term in the case
2191 * that they are consuming <1 cpu so that we would contribute the same
2192 * load as a task of equal weight.
2194 * Explicitly co-ordinating this measurement would be expensive, but
2195 * fortunately the sum of each cpus contribution forms a usable
2196 * lower-bound on the true value.
2198 * Consider the aggregate of 2 contributions. Either they are disjoint
2199 * (and the sum represents true value) or they are disjoint and we are
2200 * understating by the aggregate of their overlap.
2202 * Extending this to N cpus, for a given overlap, the maximum amount we
2203 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2204 * cpus that overlap for this interval and w_i is the interval width.
2206 * On a small machine; the first term is well-bounded which bounds the
2207 * total error since w_i is a subset of the period. Whereas on a
2208 * larger machine, while this first term can be larger, if w_i is the
2209 * of consequential size guaranteed to see n_i*w_i quickly converge to
2210 * our upper bound of 1-cpu.
2212 runnable_avg
= atomic_read(&tg
->runnable_avg
);
2213 if (runnable_avg
< NICE_0_LOAD
) {
2214 se
->avg
.load_avg_contrib
*= runnable_avg
;
2215 se
->avg
.load_avg_contrib
>>= NICE_0_SHIFT
;
2219 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2220 int force_update
) {}
2221 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2222 struct cfs_rq
*cfs_rq
) {}
2223 static inline void __update_group_entity_contrib(struct sched_entity
*se
) {}
2226 static inline void __update_task_entity_contrib(struct sched_entity
*se
)
2230 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2231 contrib
= se
->avg
.runnable_avg_sum
* scale_load_down(se
->load
.weight
);
2232 contrib
/= (se
->avg
.runnable_avg_period
+ 1);
2233 se
->avg
.load_avg_contrib
= scale_load(contrib
);
2236 /* Compute the current contribution to load_avg by se, return any delta */
2237 static long __update_entity_load_avg_contrib(struct sched_entity
*se
)
2239 long old_contrib
= se
->avg
.load_avg_contrib
;
2241 if (entity_is_task(se
)) {
2242 __update_task_entity_contrib(se
);
2244 __update_tg_runnable_avg(&se
->avg
, group_cfs_rq(se
));
2245 __update_group_entity_contrib(se
);
2248 return se
->avg
.load_avg_contrib
- old_contrib
;
2251 static inline void subtract_blocked_load_contrib(struct cfs_rq
*cfs_rq
,
2254 if (likely(load_contrib
< cfs_rq
->blocked_load_avg
))
2255 cfs_rq
->blocked_load_avg
-= load_contrib
;
2257 cfs_rq
->blocked_load_avg
= 0;
2260 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
2262 /* Update a sched_entity's runnable average */
2263 static inline void update_entity_load_avg(struct sched_entity
*se
,
2266 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2271 * For a group entity we need to use their owned cfs_rq_clock_task() in
2272 * case they are the parent of a throttled hierarchy.
2274 if (entity_is_task(se
))
2275 now
= cfs_rq_clock_task(cfs_rq
);
2277 now
= cfs_rq_clock_task(group_cfs_rq(se
));
2279 if (!__update_entity_runnable_avg(now
, &se
->avg
, se
->on_rq
))
2282 contrib_delta
= __update_entity_load_avg_contrib(se
);
2288 cfs_rq
->runnable_load_avg
+= contrib_delta
;
2290 subtract_blocked_load_contrib(cfs_rq
, -contrib_delta
);
2294 * Decay the load contributed by all blocked children and account this so that
2295 * their contribution may appropriately discounted when they wake up.
2297 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
, int force_update
)
2299 u64 now
= cfs_rq_clock_task(cfs_rq
) >> 20;
2302 decays
= now
- cfs_rq
->last_decay
;
2303 if (!decays
&& !force_update
)
2306 if (atomic_long_read(&cfs_rq
->removed_load
)) {
2307 unsigned long removed_load
;
2308 removed_load
= atomic_long_xchg(&cfs_rq
->removed_load
, 0);
2309 subtract_blocked_load_contrib(cfs_rq
, removed_load
);
2313 cfs_rq
->blocked_load_avg
= decay_load(cfs_rq
->blocked_load_avg
,
2315 atomic64_add(decays
, &cfs_rq
->decay_counter
);
2316 cfs_rq
->last_decay
= now
;
2319 __update_cfs_rq_tg_load_contrib(cfs_rq
, force_update
);
2322 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
)
2324 __update_entity_runnable_avg(rq_clock_task(rq
), &rq
->avg
, runnable
);
2325 __update_tg_runnable_avg(&rq
->avg
, &rq
->cfs
);
2328 /* Add the load generated by se into cfs_rq's child load-average */
2329 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2330 struct sched_entity
*se
,
2334 * We track migrations using entity decay_count <= 0, on a wake-up
2335 * migration we use a negative decay count to track the remote decays
2336 * accumulated while sleeping.
2338 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2339 * are seen by enqueue_entity_load_avg() as a migration with an already
2340 * constructed load_avg_contrib.
2342 if (unlikely(se
->avg
.decay_count
<= 0)) {
2343 se
->avg
.last_runnable_update
= rq_clock_task(rq_of(cfs_rq
));
2344 if (se
->avg
.decay_count
) {
2346 * In a wake-up migration we have to approximate the
2347 * time sleeping. This is because we can't synchronize
2348 * clock_task between the two cpus, and it is not
2349 * guaranteed to be read-safe. Instead, we can
2350 * approximate this using our carried decays, which are
2351 * explicitly atomically readable.
2353 se
->avg
.last_runnable_update
-= (-se
->avg
.decay_count
)
2355 update_entity_load_avg(se
, 0);
2356 /* Indicate that we're now synchronized and on-rq */
2357 se
->avg
.decay_count
= 0;
2362 * Task re-woke on same cpu (or else migrate_task_rq_fair()
2363 * would have made count negative); we must be careful to avoid
2364 * double-accounting blocked time after synchronizing decays.
2366 se
->avg
.last_runnable_update
+= __synchronize_entity_decay(se
)
2370 /* migrated tasks did not contribute to our blocked load */
2372 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
2373 update_entity_load_avg(se
, 0);
2376 cfs_rq
->runnable_load_avg
+= se
->avg
.load_avg_contrib
;
2377 /* we force update consideration on load-balancer moves */
2378 update_cfs_rq_blocked_load(cfs_rq
, !wakeup
);
2382 * Remove se's load from this cfs_rq child load-average, if the entity is
2383 * transitioning to a blocked state we track its projected decay using
2386 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2387 struct sched_entity
*se
,
2390 update_entity_load_avg(se
, 1);
2391 /* we force update consideration on load-balancer moves */
2392 update_cfs_rq_blocked_load(cfs_rq
, !sleep
);
2394 cfs_rq
->runnable_load_avg
-= se
->avg
.load_avg_contrib
;
2396 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
2397 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
2398 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2402 * Update the rq's load with the elapsed running time before entering
2403 * idle. if the last scheduled task is not a CFS task, idle_enter will
2404 * be the only way to update the runnable statistic.
2406 void idle_enter_fair(struct rq
*this_rq
)
2408 update_rq_runnable_avg(this_rq
, 1);
2412 * Update the rq's load with the elapsed idle time before a task is
2413 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2414 * be the only way to update the runnable statistic.
2416 void idle_exit_fair(struct rq
*this_rq
)
2418 update_rq_runnable_avg(this_rq
, 0);
2422 static inline void update_entity_load_avg(struct sched_entity
*se
,
2423 int update_cfs_rq
) {}
2424 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2425 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2426 struct sched_entity
*se
,
2428 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2429 struct sched_entity
*se
,
2431 static inline void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
2432 int force_update
) {}
2435 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2437 #ifdef CONFIG_SCHEDSTATS
2438 struct task_struct
*tsk
= NULL
;
2440 if (entity_is_task(se
))
2443 if (se
->statistics
.sleep_start
) {
2444 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.sleep_start
;
2449 if (unlikely(delta
> se
->statistics
.sleep_max
))
2450 se
->statistics
.sleep_max
= delta
;
2452 se
->statistics
.sleep_start
= 0;
2453 se
->statistics
.sum_sleep_runtime
+= delta
;
2456 account_scheduler_latency(tsk
, delta
>> 10, 1);
2457 trace_sched_stat_sleep(tsk
, delta
);
2460 if (se
->statistics
.block_start
) {
2461 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.block_start
;
2466 if (unlikely(delta
> se
->statistics
.block_max
))
2467 se
->statistics
.block_max
= delta
;
2469 se
->statistics
.block_start
= 0;
2470 se
->statistics
.sum_sleep_runtime
+= delta
;
2473 if (tsk
->in_iowait
) {
2474 se
->statistics
.iowait_sum
+= delta
;
2475 se
->statistics
.iowait_count
++;
2476 trace_sched_stat_iowait(tsk
, delta
);
2479 trace_sched_stat_blocked(tsk
, delta
);
2482 * Blocking time is in units of nanosecs, so shift by
2483 * 20 to get a milliseconds-range estimation of the
2484 * amount of time that the task spent sleeping:
2486 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
2487 profile_hits(SLEEP_PROFILING
,
2488 (void *)get_wchan(tsk
),
2491 account_scheduler_latency(tsk
, delta
>> 10, 0);
2497 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2499 #ifdef CONFIG_SCHED_DEBUG
2500 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
2505 if (d
> 3*sysctl_sched_latency
)
2506 schedstat_inc(cfs_rq
, nr_spread_over
);
2511 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
2513 u64 vruntime
= cfs_rq
->min_vruntime
;
2516 * The 'current' period is already promised to the current tasks,
2517 * however the extra weight of the new task will slow them down a
2518 * little, place the new task so that it fits in the slot that
2519 * stays open at the end.
2521 if (initial
&& sched_feat(START_DEBIT
))
2522 vruntime
+= sched_vslice(cfs_rq
, se
);
2524 /* sleeps up to a single latency don't count. */
2526 unsigned long thresh
= sysctl_sched_latency
;
2529 * Halve their sleep time's effect, to allow
2530 * for a gentler effect of sleepers:
2532 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
2538 /* ensure we never gain time by being placed backwards. */
2539 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
2542 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
2545 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2548 * Update the normalized vruntime before updating min_vruntime
2549 * through calling update_curr().
2551 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
2552 se
->vruntime
+= cfs_rq
->min_vruntime
;
2555 * Update run-time statistics of the 'current'.
2557 update_curr(cfs_rq
);
2558 enqueue_entity_load_avg(cfs_rq
, se
, flags
& ENQUEUE_WAKEUP
);
2559 account_entity_enqueue(cfs_rq
, se
);
2560 update_cfs_shares(cfs_rq
);
2562 if (flags
& ENQUEUE_WAKEUP
) {
2563 place_entity(cfs_rq
, se
, 0);
2564 enqueue_sleeper(cfs_rq
, se
);
2567 update_stats_enqueue(cfs_rq
, se
);
2568 check_spread(cfs_rq
, se
);
2569 if (se
!= cfs_rq
->curr
)
2570 __enqueue_entity(cfs_rq
, se
);
2573 if (cfs_rq
->nr_running
== 1) {
2574 list_add_leaf_cfs_rq(cfs_rq
);
2575 check_enqueue_throttle(cfs_rq
);
2579 static void __clear_buddies_last(struct sched_entity
*se
)
2581 for_each_sched_entity(se
) {
2582 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2583 if (cfs_rq
->last
== se
)
2584 cfs_rq
->last
= NULL
;
2590 static void __clear_buddies_next(struct sched_entity
*se
)
2592 for_each_sched_entity(se
) {
2593 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2594 if (cfs_rq
->next
== se
)
2595 cfs_rq
->next
= NULL
;
2601 static void __clear_buddies_skip(struct sched_entity
*se
)
2603 for_each_sched_entity(se
) {
2604 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2605 if (cfs_rq
->skip
== se
)
2606 cfs_rq
->skip
= NULL
;
2612 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2614 if (cfs_rq
->last
== se
)
2615 __clear_buddies_last(se
);
2617 if (cfs_rq
->next
== se
)
2618 __clear_buddies_next(se
);
2620 if (cfs_rq
->skip
== se
)
2621 __clear_buddies_skip(se
);
2624 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
2627 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2630 * Update run-time statistics of the 'current'.
2632 update_curr(cfs_rq
);
2633 dequeue_entity_load_avg(cfs_rq
, se
, flags
& DEQUEUE_SLEEP
);
2635 update_stats_dequeue(cfs_rq
, se
);
2636 if (flags
& DEQUEUE_SLEEP
) {
2637 #ifdef CONFIG_SCHEDSTATS
2638 if (entity_is_task(se
)) {
2639 struct task_struct
*tsk
= task_of(se
);
2641 if (tsk
->state
& TASK_INTERRUPTIBLE
)
2642 se
->statistics
.sleep_start
= rq_clock(rq_of(cfs_rq
));
2643 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
2644 se
->statistics
.block_start
= rq_clock(rq_of(cfs_rq
));
2649 clear_buddies(cfs_rq
, se
);
2651 if (se
!= cfs_rq
->curr
)
2652 __dequeue_entity(cfs_rq
, se
);
2654 account_entity_dequeue(cfs_rq
, se
);
2657 * Normalize the entity after updating the min_vruntime because the
2658 * update can refer to the ->curr item and we need to reflect this
2659 * movement in our normalized position.
2661 if (!(flags
& DEQUEUE_SLEEP
))
2662 se
->vruntime
-= cfs_rq
->min_vruntime
;
2664 /* return excess runtime on last dequeue */
2665 return_cfs_rq_runtime(cfs_rq
);
2667 update_min_vruntime(cfs_rq
);
2668 update_cfs_shares(cfs_rq
);
2672 * Preempt the current task with a newly woken task if needed:
2675 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
2677 unsigned long ideal_runtime
, delta_exec
;
2678 struct sched_entity
*se
;
2681 ideal_runtime
= sched_slice(cfs_rq
, curr
);
2682 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
2683 if (delta_exec
> ideal_runtime
) {
2684 resched_task(rq_of(cfs_rq
)->curr
);
2686 * The current task ran long enough, ensure it doesn't get
2687 * re-elected due to buddy favours.
2689 clear_buddies(cfs_rq
, curr
);
2694 * Ensure that a task that missed wakeup preemption by a
2695 * narrow margin doesn't have to wait for a full slice.
2696 * This also mitigates buddy induced latencies under load.
2698 if (delta_exec
< sysctl_sched_min_granularity
)
2701 se
= __pick_first_entity(cfs_rq
);
2702 delta
= curr
->vruntime
- se
->vruntime
;
2707 if (delta
> ideal_runtime
)
2708 resched_task(rq_of(cfs_rq
)->curr
);
2712 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2714 /* 'current' is not kept within the tree. */
2717 * Any task has to be enqueued before it get to execute on
2718 * a CPU. So account for the time it spent waiting on the
2721 update_stats_wait_end(cfs_rq
, se
);
2722 __dequeue_entity(cfs_rq
, se
);
2725 update_stats_curr_start(cfs_rq
, se
);
2727 #ifdef CONFIG_SCHEDSTATS
2729 * Track our maximum slice length, if the CPU's load is at
2730 * least twice that of our own weight (i.e. dont track it
2731 * when there are only lesser-weight tasks around):
2733 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
2734 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
2735 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
2738 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
2742 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
2745 * Pick the next process, keeping these things in mind, in this order:
2746 * 1) keep things fair between processes/task groups
2747 * 2) pick the "next" process, since someone really wants that to run
2748 * 3) pick the "last" process, for cache locality
2749 * 4) do not run the "skip" process, if something else is available
2751 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
2753 struct sched_entity
*se
= __pick_first_entity(cfs_rq
);
2754 struct sched_entity
*left
= se
;
2757 * Avoid running the skip buddy, if running something else can
2758 * be done without getting too unfair.
2760 if (cfs_rq
->skip
== se
) {
2761 struct sched_entity
*second
= __pick_next_entity(se
);
2762 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
2767 * Prefer last buddy, try to return the CPU to a preempted task.
2769 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
2773 * Someone really wants this to run. If it's not unfair, run it.
2775 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
2778 clear_buddies(cfs_rq
, se
);
2783 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
2785 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
2788 * If still on the runqueue then deactivate_task()
2789 * was not called and update_curr() has to be done:
2792 update_curr(cfs_rq
);
2794 /* throttle cfs_rqs exceeding runtime */
2795 check_cfs_rq_runtime(cfs_rq
);
2797 check_spread(cfs_rq
, prev
);
2799 update_stats_wait_start(cfs_rq
, prev
);
2800 /* Put 'current' back into the tree. */
2801 __enqueue_entity(cfs_rq
, prev
);
2802 /* in !on_rq case, update occurred at dequeue */
2803 update_entity_load_avg(prev
, 1);
2805 cfs_rq
->curr
= NULL
;
2809 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
2812 * Update run-time statistics of the 'current'.
2814 update_curr(cfs_rq
);
2817 * Ensure that runnable average is periodically updated.
2819 update_entity_load_avg(curr
, 1);
2820 update_cfs_rq_blocked_load(cfs_rq
, 1);
2821 update_cfs_shares(cfs_rq
);
2823 #ifdef CONFIG_SCHED_HRTICK
2825 * queued ticks are scheduled to match the slice, so don't bother
2826 * validating it and just reschedule.
2829 resched_task(rq_of(cfs_rq
)->curr
);
2833 * don't let the period tick interfere with the hrtick preemption
2835 if (!sched_feat(DOUBLE_TICK
) &&
2836 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
2840 if (cfs_rq
->nr_running
> 1)
2841 check_preempt_tick(cfs_rq
, curr
);
2845 /**************************************************
2846 * CFS bandwidth control machinery
2849 #ifdef CONFIG_CFS_BANDWIDTH
2851 #ifdef HAVE_JUMP_LABEL
2852 static struct static_key __cfs_bandwidth_used
;
2854 static inline bool cfs_bandwidth_used(void)
2856 return static_key_false(&__cfs_bandwidth_used
);
2859 void account_cfs_bandwidth_used(int enabled
, int was_enabled
)
2861 /* only need to count groups transitioning between enabled/!enabled */
2862 if (enabled
&& !was_enabled
)
2863 static_key_slow_inc(&__cfs_bandwidth_used
);
2864 else if (!enabled
&& was_enabled
)
2865 static_key_slow_dec(&__cfs_bandwidth_used
);
2867 #else /* HAVE_JUMP_LABEL */
2868 static bool cfs_bandwidth_used(void)
2873 void account_cfs_bandwidth_used(int enabled
, int was_enabled
) {}
2874 #endif /* HAVE_JUMP_LABEL */
2877 * default period for cfs group bandwidth.
2878 * default: 0.1s, units: nanoseconds
2880 static inline u64
default_cfs_period(void)
2882 return 100000000ULL;
2885 static inline u64
sched_cfs_bandwidth_slice(void)
2887 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
2891 * Replenish runtime according to assigned quota and update expiration time.
2892 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2893 * additional synchronization around rq->lock.
2895 * requires cfs_b->lock
2897 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
2901 if (cfs_b
->quota
== RUNTIME_INF
)
2904 now
= sched_clock_cpu(smp_processor_id());
2905 cfs_b
->runtime
= cfs_b
->quota
;
2906 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
2909 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
2911 return &tg
->cfs_bandwidth
;
2914 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2915 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
2917 if (unlikely(cfs_rq
->throttle_count
))
2918 return cfs_rq
->throttled_clock_task
;
2920 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
2923 /* returns 0 on failure to allocate runtime */
2924 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2926 struct task_group
*tg
= cfs_rq
->tg
;
2927 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
2928 u64 amount
= 0, min_amount
, expires
;
2930 /* note: this is a positive sum as runtime_remaining <= 0 */
2931 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
2933 raw_spin_lock(&cfs_b
->lock
);
2934 if (cfs_b
->quota
== RUNTIME_INF
)
2935 amount
= min_amount
;
2938 * If the bandwidth pool has become inactive, then at least one
2939 * period must have elapsed since the last consumption.
2940 * Refresh the global state and ensure bandwidth timer becomes
2943 if (!cfs_b
->timer_active
) {
2944 __refill_cfs_bandwidth_runtime(cfs_b
);
2945 __start_cfs_bandwidth(cfs_b
);
2948 if (cfs_b
->runtime
> 0) {
2949 amount
= min(cfs_b
->runtime
, min_amount
);
2950 cfs_b
->runtime
-= amount
;
2954 expires
= cfs_b
->runtime_expires
;
2955 raw_spin_unlock(&cfs_b
->lock
);
2957 cfs_rq
->runtime_remaining
+= amount
;
2959 * we may have advanced our local expiration to account for allowed
2960 * spread between our sched_clock and the one on which runtime was
2963 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
2964 cfs_rq
->runtime_expires
= expires
;
2966 return cfs_rq
->runtime_remaining
> 0;
2970 * Note: This depends on the synchronization provided by sched_clock and the
2971 * fact that rq->clock snapshots this value.
2973 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2975 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
2977 /* if the deadline is ahead of our clock, nothing to do */
2978 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
2981 if (cfs_rq
->runtime_remaining
< 0)
2985 * If the local deadline has passed we have to consider the
2986 * possibility that our sched_clock is 'fast' and the global deadline
2987 * has not truly expired.
2989 * Fortunately we can check determine whether this the case by checking
2990 * whether the global deadline has advanced.
2993 if ((s64
)(cfs_rq
->runtime_expires
- cfs_b
->runtime_expires
) >= 0) {
2994 /* extend local deadline, drift is bounded above by 2 ticks */
2995 cfs_rq
->runtime_expires
+= TICK_NSEC
;
2997 /* global deadline is ahead, expiration has passed */
2998 cfs_rq
->runtime_remaining
= 0;
3002 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
,
3003 unsigned long delta_exec
)
3005 /* dock delta_exec before expiring quota (as it could span periods) */
3006 cfs_rq
->runtime_remaining
-= delta_exec
;
3007 expire_cfs_rq_runtime(cfs_rq
);
3009 if (likely(cfs_rq
->runtime_remaining
> 0))
3013 * if we're unable to extend our runtime we resched so that the active
3014 * hierarchy can be throttled
3016 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
3017 resched_task(rq_of(cfs_rq
)->curr
);
3020 static __always_inline
3021 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, unsigned long delta_exec
)
3023 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
3026 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
3029 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3031 return cfs_bandwidth_used() && cfs_rq
->throttled
;
3034 /* check whether cfs_rq, or any parent, is throttled */
3035 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3037 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
3041 * Ensure that neither of the group entities corresponding to src_cpu or
3042 * dest_cpu are members of a throttled hierarchy when performing group
3043 * load-balance operations.
3045 static inline int throttled_lb_pair(struct task_group
*tg
,
3046 int src_cpu
, int dest_cpu
)
3048 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
3050 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
3051 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
3053 return throttled_hierarchy(src_cfs_rq
) ||
3054 throttled_hierarchy(dest_cfs_rq
);
3057 /* updated child weight may affect parent so we have to do this bottom up */
3058 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
3060 struct rq
*rq
= data
;
3061 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3063 cfs_rq
->throttle_count
--;
3065 if (!cfs_rq
->throttle_count
) {
3066 /* adjust cfs_rq_clock_task() */
3067 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
3068 cfs_rq
->throttled_clock_task
;
3075 static int tg_throttle_down(struct task_group
*tg
, void *data
)
3077 struct rq
*rq
= data
;
3078 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3080 /* group is entering throttled state, stop time */
3081 if (!cfs_rq
->throttle_count
)
3082 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
3083 cfs_rq
->throttle_count
++;
3088 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
3090 struct rq
*rq
= rq_of(cfs_rq
);
3091 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3092 struct sched_entity
*se
;
3093 long task_delta
, dequeue
= 1;
3095 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
3097 /* freeze hierarchy runnable averages while throttled */
3099 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
3102 task_delta
= cfs_rq
->h_nr_running
;
3103 for_each_sched_entity(se
) {
3104 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
3105 /* throttled entity or throttle-on-deactivate */
3110 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
3111 qcfs_rq
->h_nr_running
-= task_delta
;
3113 if (qcfs_rq
->load
.weight
)
3118 rq
->nr_running
-= task_delta
;
3120 cfs_rq
->throttled
= 1;
3121 cfs_rq
->throttled_clock
= rq_clock(rq
);
3122 raw_spin_lock(&cfs_b
->lock
);
3123 list_add_tail_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
3124 raw_spin_unlock(&cfs_b
->lock
);
3127 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
3129 struct rq
*rq
= rq_of(cfs_rq
);
3130 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3131 struct sched_entity
*se
;
3135 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
3137 cfs_rq
->throttled
= 0;
3139 update_rq_clock(rq
);
3141 raw_spin_lock(&cfs_b
->lock
);
3142 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
3143 list_del_rcu(&cfs_rq
->throttled_list
);
3144 raw_spin_unlock(&cfs_b
->lock
);
3146 /* update hierarchical throttle state */
3147 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
3149 if (!cfs_rq
->load
.weight
)
3152 task_delta
= cfs_rq
->h_nr_running
;
3153 for_each_sched_entity(se
) {
3157 cfs_rq
= cfs_rq_of(se
);
3159 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
3160 cfs_rq
->h_nr_running
+= task_delta
;
3162 if (cfs_rq_throttled(cfs_rq
))
3167 rq
->nr_running
+= task_delta
;
3169 /* determine whether we need to wake up potentially idle cpu */
3170 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
3171 resched_task(rq
->curr
);
3174 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
3175 u64 remaining
, u64 expires
)
3177 struct cfs_rq
*cfs_rq
;
3178 u64 runtime
= remaining
;
3181 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
3183 struct rq
*rq
= rq_of(cfs_rq
);
3185 raw_spin_lock(&rq
->lock
);
3186 if (!cfs_rq_throttled(cfs_rq
))
3189 runtime
= -cfs_rq
->runtime_remaining
+ 1;
3190 if (runtime
> remaining
)
3191 runtime
= remaining
;
3192 remaining
-= runtime
;
3194 cfs_rq
->runtime_remaining
+= runtime
;
3195 cfs_rq
->runtime_expires
= expires
;
3197 /* we check whether we're throttled above */
3198 if (cfs_rq
->runtime_remaining
> 0)
3199 unthrottle_cfs_rq(cfs_rq
);
3202 raw_spin_unlock(&rq
->lock
);
3213 * Responsible for refilling a task_group's bandwidth and unthrottling its
3214 * cfs_rqs as appropriate. If there has been no activity within the last
3215 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3216 * used to track this state.
3218 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
3220 u64 runtime
, runtime_expires
;
3221 int idle
= 1, throttled
;
3223 raw_spin_lock(&cfs_b
->lock
);
3224 /* no need to continue the timer with no bandwidth constraint */
3225 if (cfs_b
->quota
== RUNTIME_INF
)
3228 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3229 /* idle depends on !throttled (for the case of a large deficit) */
3230 idle
= cfs_b
->idle
&& !throttled
;
3231 cfs_b
->nr_periods
+= overrun
;
3233 /* if we're going inactive then everything else can be deferred */
3237 __refill_cfs_bandwidth_runtime(cfs_b
);
3240 /* mark as potentially idle for the upcoming period */
3245 /* account preceding periods in which throttling occurred */
3246 cfs_b
->nr_throttled
+= overrun
;
3249 * There are throttled entities so we must first use the new bandwidth
3250 * to unthrottle them before making it generally available. This
3251 * ensures that all existing debts will be paid before a new cfs_rq is
3254 runtime
= cfs_b
->runtime
;
3255 runtime_expires
= cfs_b
->runtime_expires
;
3259 * This check is repeated as we are holding onto the new bandwidth
3260 * while we unthrottle. This can potentially race with an unthrottled
3261 * group trying to acquire new bandwidth from the global pool.
3263 while (throttled
&& runtime
> 0) {
3264 raw_spin_unlock(&cfs_b
->lock
);
3265 /* we can't nest cfs_b->lock while distributing bandwidth */
3266 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
3268 raw_spin_lock(&cfs_b
->lock
);
3270 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3273 /* return (any) remaining runtime */
3274 cfs_b
->runtime
= runtime
;
3276 * While we are ensured activity in the period following an
3277 * unthrottle, this also covers the case in which the new bandwidth is
3278 * insufficient to cover the existing bandwidth deficit. (Forcing the
3279 * timer to remain active while there are any throttled entities.)
3284 cfs_b
->timer_active
= 0;
3285 raw_spin_unlock(&cfs_b
->lock
);
3290 /* a cfs_rq won't donate quota below this amount */
3291 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
3292 /* minimum remaining period time to redistribute slack quota */
3293 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
3294 /* how long we wait to gather additional slack before distributing */
3295 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
3297 /* are we near the end of the current quota period? */
3298 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
3300 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
3303 /* if the call-back is running a quota refresh is already occurring */
3304 if (hrtimer_callback_running(refresh_timer
))
3307 /* is a quota refresh about to occur? */
3308 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
3309 if (remaining
< min_expire
)
3315 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
3317 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
3319 /* if there's a quota refresh soon don't bother with slack */
3320 if (runtime_refresh_within(cfs_b
, min_left
))
3323 start_bandwidth_timer(&cfs_b
->slack_timer
,
3324 ns_to_ktime(cfs_bandwidth_slack_period
));
3327 /* we know any runtime found here is valid as update_curr() precedes return */
3328 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3330 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3331 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
3333 if (slack_runtime
<= 0)
3336 raw_spin_lock(&cfs_b
->lock
);
3337 if (cfs_b
->quota
!= RUNTIME_INF
&&
3338 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
3339 cfs_b
->runtime
+= slack_runtime
;
3341 /* we are under rq->lock, defer unthrottling using a timer */
3342 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
3343 !list_empty(&cfs_b
->throttled_cfs_rq
))
3344 start_cfs_slack_bandwidth(cfs_b
);
3346 raw_spin_unlock(&cfs_b
->lock
);
3348 /* even if it's not valid for return we don't want to try again */
3349 cfs_rq
->runtime_remaining
-= slack_runtime
;
3352 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3354 if (!cfs_bandwidth_used())
3357 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
3360 __return_cfs_rq_runtime(cfs_rq
);
3364 * This is done with a timer (instead of inline with bandwidth return) since
3365 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3367 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
3369 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
3372 /* confirm we're still not at a refresh boundary */
3373 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
))
3376 raw_spin_lock(&cfs_b
->lock
);
3377 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
) {
3378 runtime
= cfs_b
->runtime
;
3381 expires
= cfs_b
->runtime_expires
;
3382 raw_spin_unlock(&cfs_b
->lock
);
3387 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
3389 raw_spin_lock(&cfs_b
->lock
);
3390 if (expires
== cfs_b
->runtime_expires
)
3391 cfs_b
->runtime
= runtime
;
3392 raw_spin_unlock(&cfs_b
->lock
);
3396 * When a group wakes up we want to make sure that its quota is not already
3397 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3398 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3400 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
3402 if (!cfs_bandwidth_used())
3405 /* an active group must be handled by the update_curr()->put() path */
3406 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
3409 /* ensure the group is not already throttled */
3410 if (cfs_rq_throttled(cfs_rq
))
3413 /* update runtime allocation */
3414 account_cfs_rq_runtime(cfs_rq
, 0);
3415 if (cfs_rq
->runtime_remaining
<= 0)
3416 throttle_cfs_rq(cfs_rq
);
3419 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3420 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3422 if (!cfs_bandwidth_used())
3425 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
3429 * it's possible for a throttled entity to be forced into a running
3430 * state (e.g. set_curr_task), in this case we're finished.
3432 if (cfs_rq_throttled(cfs_rq
))
3435 throttle_cfs_rq(cfs_rq
);
3438 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
3440 struct cfs_bandwidth
*cfs_b
=
3441 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
3442 do_sched_cfs_slack_timer(cfs_b
);
3444 return HRTIMER_NORESTART
;
3447 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
3449 struct cfs_bandwidth
*cfs_b
=
3450 container_of(timer
, struct cfs_bandwidth
, period_timer
);
3456 now
= hrtimer_cb_get_time(timer
);
3457 overrun
= hrtimer_forward(timer
, now
, cfs_b
->period
);
3462 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
3465 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
3468 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3470 raw_spin_lock_init(&cfs_b
->lock
);
3472 cfs_b
->quota
= RUNTIME_INF
;
3473 cfs_b
->period
= ns_to_ktime(default_cfs_period());
3475 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
3476 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3477 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
3478 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3479 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
3482 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3484 cfs_rq
->runtime_enabled
= 0;
3485 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
3488 /* requires cfs_b->lock, may release to reprogram timer */
3489 void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3492 * The timer may be active because we're trying to set a new bandwidth
3493 * period or because we're racing with the tear-down path
3494 * (timer_active==0 becomes visible before the hrtimer call-back
3495 * terminates). In either case we ensure that it's re-programmed
3497 while (unlikely(hrtimer_active(&cfs_b
->period_timer
))) {
3498 raw_spin_unlock(&cfs_b
->lock
);
3499 /* ensure cfs_b->lock is available while we wait */
3500 hrtimer_cancel(&cfs_b
->period_timer
);
3502 raw_spin_lock(&cfs_b
->lock
);
3503 /* if someone else restarted the timer then we're done */
3504 if (cfs_b
->timer_active
)
3508 cfs_b
->timer_active
= 1;
3509 start_bandwidth_timer(&cfs_b
->period_timer
, cfs_b
->period
);
3512 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3514 hrtimer_cancel(&cfs_b
->period_timer
);
3515 hrtimer_cancel(&cfs_b
->slack_timer
);
3518 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
3520 struct cfs_rq
*cfs_rq
;
3522 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
3523 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3525 if (!cfs_rq
->runtime_enabled
)
3529 * clock_task is not advancing so we just need to make sure
3530 * there's some valid quota amount
3532 cfs_rq
->runtime_remaining
= cfs_b
->quota
;
3533 if (cfs_rq_throttled(cfs_rq
))
3534 unthrottle_cfs_rq(cfs_rq
);
3538 #else /* CONFIG_CFS_BANDWIDTH */
3539 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3541 return rq_clock_task(rq_of(cfs_rq
));
3544 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
,
3545 unsigned long delta_exec
) {}
3546 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3547 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
3548 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3550 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3555 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3560 static inline int throttled_lb_pair(struct task_group
*tg
,
3561 int src_cpu
, int dest_cpu
)
3566 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3568 #ifdef CONFIG_FAIR_GROUP_SCHED
3569 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3572 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3576 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3577 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
3579 #endif /* CONFIG_CFS_BANDWIDTH */
3581 /**************************************************
3582 * CFS operations on tasks:
3585 #ifdef CONFIG_SCHED_HRTICK
3586 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3588 struct sched_entity
*se
= &p
->se
;
3589 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3591 WARN_ON(task_rq(p
) != rq
);
3593 if (cfs_rq
->nr_running
> 1) {
3594 u64 slice
= sched_slice(cfs_rq
, se
);
3595 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
3596 s64 delta
= slice
- ran
;
3605 * Don't schedule slices shorter than 10000ns, that just
3606 * doesn't make sense. Rely on vruntime for fairness.
3609 delta
= max_t(s64
, 10000LL, delta
);
3611 hrtick_start(rq
, delta
);
3616 * called from enqueue/dequeue and updates the hrtick when the
3617 * current task is from our class and nr_running is low enough
3620 static void hrtick_update(struct rq
*rq
)
3622 struct task_struct
*curr
= rq
->curr
;
3624 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
3627 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
3628 hrtick_start_fair(rq
, curr
);
3630 #else /* !CONFIG_SCHED_HRTICK */
3632 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3636 static inline void hrtick_update(struct rq
*rq
)
3642 * The enqueue_task method is called before nr_running is
3643 * increased. Here we update the fair scheduling stats and
3644 * then put the task into the rbtree:
3647 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
3649 struct cfs_rq
*cfs_rq
;
3650 struct sched_entity
*se
= &p
->se
;
3652 for_each_sched_entity(se
) {
3655 cfs_rq
= cfs_rq_of(se
);
3656 enqueue_entity(cfs_rq
, se
, flags
);
3659 * end evaluation on encountering a throttled cfs_rq
3661 * note: in the case of encountering a throttled cfs_rq we will
3662 * post the final h_nr_running increment below.
3664 if (cfs_rq_throttled(cfs_rq
))
3666 cfs_rq
->h_nr_running
++;
3668 flags
= ENQUEUE_WAKEUP
;
3671 for_each_sched_entity(se
) {
3672 cfs_rq
= cfs_rq_of(se
);
3673 cfs_rq
->h_nr_running
++;
3675 if (cfs_rq_throttled(cfs_rq
))
3678 update_cfs_shares(cfs_rq
);
3679 update_entity_load_avg(se
, 1);
3683 update_rq_runnable_avg(rq
, rq
->nr_running
);
3689 static void set_next_buddy(struct sched_entity
*se
);
3692 * The dequeue_task method is called before nr_running is
3693 * decreased. We remove the task from the rbtree and
3694 * update the fair scheduling stats:
3696 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
3698 struct cfs_rq
*cfs_rq
;
3699 struct sched_entity
*se
= &p
->se
;
3700 int task_sleep
= flags
& DEQUEUE_SLEEP
;
3702 for_each_sched_entity(se
) {
3703 cfs_rq
= cfs_rq_of(se
);
3704 dequeue_entity(cfs_rq
, se
, flags
);
3707 * end evaluation on encountering a throttled cfs_rq
3709 * note: in the case of encountering a throttled cfs_rq we will
3710 * post the final h_nr_running decrement below.
3712 if (cfs_rq_throttled(cfs_rq
))
3714 cfs_rq
->h_nr_running
--;
3716 /* Don't dequeue parent if it has other entities besides us */
3717 if (cfs_rq
->load
.weight
) {
3719 * Bias pick_next to pick a task from this cfs_rq, as
3720 * p is sleeping when it is within its sched_slice.
3722 if (task_sleep
&& parent_entity(se
))
3723 set_next_buddy(parent_entity(se
));
3725 /* avoid re-evaluating load for this entity */
3726 se
= parent_entity(se
);
3729 flags
|= DEQUEUE_SLEEP
;
3732 for_each_sched_entity(se
) {
3733 cfs_rq
= cfs_rq_of(se
);
3734 cfs_rq
->h_nr_running
--;
3736 if (cfs_rq_throttled(cfs_rq
))
3739 update_cfs_shares(cfs_rq
);
3740 update_entity_load_avg(se
, 1);
3745 update_rq_runnable_avg(rq
, 1);
3751 /* Used instead of source_load when we know the type == 0 */
3752 static unsigned long weighted_cpuload(const int cpu
)
3754 return cpu_rq(cpu
)->cfs
.runnable_load_avg
;
3758 * Return a low guess at the load of a migration-source cpu weighted
3759 * according to the scheduling class and "nice" value.
3761 * We want to under-estimate the load of migration sources, to
3762 * balance conservatively.
3764 static unsigned long source_load(int cpu
, int type
)
3766 struct rq
*rq
= cpu_rq(cpu
);
3767 unsigned long total
= weighted_cpuload(cpu
);
3769 if (type
== 0 || !sched_feat(LB_BIAS
))
3772 return min(rq
->cpu_load
[type
-1], total
);
3776 * Return a high guess at the load of a migration-target cpu weighted
3777 * according to the scheduling class and "nice" value.
3779 static unsigned long target_load(int cpu
, int type
)
3781 struct rq
*rq
= cpu_rq(cpu
);
3782 unsigned long total
= weighted_cpuload(cpu
);
3784 if (type
== 0 || !sched_feat(LB_BIAS
))
3787 return max(rq
->cpu_load
[type
-1], total
);
3790 static unsigned long power_of(int cpu
)
3792 return cpu_rq(cpu
)->cpu_power
;
3795 static unsigned long cpu_avg_load_per_task(int cpu
)
3797 struct rq
*rq
= cpu_rq(cpu
);
3798 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
3799 unsigned long load_avg
= rq
->cfs
.runnable_load_avg
;
3802 return load_avg
/ nr_running
;
3807 static void record_wakee(struct task_struct
*p
)
3810 * Rough decay (wiping) for cost saving, don't worry
3811 * about the boundary, really active task won't care
3814 if (jiffies
> current
->wakee_flip_decay_ts
+ HZ
) {
3815 current
->wakee_flips
= 0;
3816 current
->wakee_flip_decay_ts
= jiffies
;
3819 if (current
->last_wakee
!= p
) {
3820 current
->last_wakee
= p
;
3821 current
->wakee_flips
++;
3825 static void task_waking_fair(struct task_struct
*p
)
3827 struct sched_entity
*se
= &p
->se
;
3828 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3831 #ifndef CONFIG_64BIT
3832 u64 min_vruntime_copy
;
3835 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
3837 min_vruntime
= cfs_rq
->min_vruntime
;
3838 } while (min_vruntime
!= min_vruntime_copy
);
3840 min_vruntime
= cfs_rq
->min_vruntime
;
3843 se
->vruntime
-= min_vruntime
;
3847 #ifdef CONFIG_FAIR_GROUP_SCHED
3849 * effective_load() calculates the load change as seen from the root_task_group
3851 * Adding load to a group doesn't make a group heavier, but can cause movement
3852 * of group shares between cpus. Assuming the shares were perfectly aligned one
3853 * can calculate the shift in shares.
3855 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3856 * on this @cpu and results in a total addition (subtraction) of @wg to the
3857 * total group weight.
3859 * Given a runqueue weight distribution (rw_i) we can compute a shares
3860 * distribution (s_i) using:
3862 * s_i = rw_i / \Sum rw_j (1)
3864 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3865 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3866 * shares distribution (s_i):
3868 * rw_i = { 2, 4, 1, 0 }
3869 * s_i = { 2/7, 4/7, 1/7, 0 }
3871 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3872 * task used to run on and the CPU the waker is running on), we need to
3873 * compute the effect of waking a task on either CPU and, in case of a sync
3874 * wakeup, compute the effect of the current task going to sleep.
3876 * So for a change of @wl to the local @cpu with an overall group weight change
3877 * of @wl we can compute the new shares distribution (s'_i) using:
3879 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3881 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3882 * differences in waking a task to CPU 0. The additional task changes the
3883 * weight and shares distributions like:
3885 * rw'_i = { 3, 4, 1, 0 }
3886 * s'_i = { 3/8, 4/8, 1/8, 0 }
3888 * We can then compute the difference in effective weight by using:
3890 * dw_i = S * (s'_i - s_i) (3)
3892 * Where 'S' is the group weight as seen by its parent.
3894 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3895 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3896 * 4/7) times the weight of the group.
3898 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
3900 struct sched_entity
*se
= tg
->se
[cpu
];
3902 if (!tg
->parent
|| !wl
) /* the trivial, non-cgroup case */
3905 for_each_sched_entity(se
) {
3911 * W = @wg + \Sum rw_j
3913 W
= wg
+ calc_tg_weight(tg
, se
->my_q
);
3918 w
= se
->my_q
->load
.weight
+ wl
;
3921 * wl = S * s'_i; see (2)
3924 wl
= (w
* tg
->shares
) / W
;
3929 * Per the above, wl is the new se->load.weight value; since
3930 * those are clipped to [MIN_SHARES, ...) do so now. See
3931 * calc_cfs_shares().
3933 if (wl
< MIN_SHARES
)
3937 * wl = dw_i = S * (s'_i - s_i); see (3)
3939 wl
-= se
->load
.weight
;
3942 * Recursively apply this logic to all parent groups to compute
3943 * the final effective load change on the root group. Since
3944 * only the @tg group gets extra weight, all parent groups can
3945 * only redistribute existing shares. @wl is the shift in shares
3946 * resulting from this level per the above.
3955 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
3962 static int wake_wide(struct task_struct
*p
)
3964 int factor
= this_cpu_read(sd_llc_size
);
3967 * Yeah, it's the switching-frequency, could means many wakee or
3968 * rapidly switch, use factor here will just help to automatically
3969 * adjust the loose-degree, so bigger node will lead to more pull.
3971 if (p
->wakee_flips
> factor
) {
3973 * wakee is somewhat hot, it needs certain amount of cpu
3974 * resource, so if waker is far more hot, prefer to leave
3977 if (current
->wakee_flips
> (factor
* p
->wakee_flips
))
3984 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
3986 s64 this_load
, load
;
3987 int idx
, this_cpu
, prev_cpu
;
3988 unsigned long tl_per_task
;
3989 struct task_group
*tg
;
3990 unsigned long weight
;
3994 * If we wake multiple tasks be careful to not bounce
3995 * ourselves around too much.
4001 this_cpu
= smp_processor_id();
4002 prev_cpu
= task_cpu(p
);
4003 load
= source_load(prev_cpu
, idx
);
4004 this_load
= target_load(this_cpu
, idx
);
4007 * If sync wakeup then subtract the (maximum possible)
4008 * effect of the currently running task from the load
4009 * of the current CPU:
4012 tg
= task_group(current
);
4013 weight
= current
->se
.load
.weight
;
4015 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
4016 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
4020 weight
= p
->se
.load
.weight
;
4023 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4024 * due to the sync cause above having dropped this_load to 0, we'll
4025 * always have an imbalance, but there's really nothing you can do
4026 * about that, so that's good too.
4028 * Otherwise check if either cpus are near enough in load to allow this
4029 * task to be woken on this_cpu.
4031 if (this_load
> 0) {
4032 s64 this_eff_load
, prev_eff_load
;
4034 this_eff_load
= 100;
4035 this_eff_load
*= power_of(prev_cpu
);
4036 this_eff_load
*= this_load
+
4037 effective_load(tg
, this_cpu
, weight
, weight
);
4039 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
4040 prev_eff_load
*= power_of(this_cpu
);
4041 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
4043 balanced
= this_eff_load
<= prev_eff_load
;
4048 * If the currently running task will sleep within
4049 * a reasonable amount of time then attract this newly
4052 if (sync
&& balanced
)
4055 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
4056 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
4059 (this_load
<= load
&&
4060 this_load
+ target_load(prev_cpu
, idx
) <= tl_per_task
)) {
4062 * This domain has SD_WAKE_AFFINE and
4063 * p is cache cold in this domain, and
4064 * there is no bad imbalance.
4066 schedstat_inc(sd
, ttwu_move_affine
);
4067 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
4075 * find_idlest_group finds and returns the least busy CPU group within the
4078 static struct sched_group
*
4079 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
4080 int this_cpu
, int load_idx
)
4082 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
4083 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
4084 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
4087 unsigned long load
, avg_load
;
4091 /* Skip over this group if it has no CPUs allowed */
4092 if (!cpumask_intersects(sched_group_cpus(group
),
4093 tsk_cpus_allowed(p
)))
4096 local_group
= cpumask_test_cpu(this_cpu
,
4097 sched_group_cpus(group
));
4099 /* Tally up the load of all CPUs in the group */
4102 for_each_cpu(i
, sched_group_cpus(group
)) {
4103 /* Bias balancing toward cpus of our domain */
4105 load
= source_load(i
, load_idx
);
4107 load
= target_load(i
, load_idx
);
4112 /* Adjust by relative CPU power of the group */
4113 avg_load
= (avg_load
* SCHED_POWER_SCALE
) / group
->sgp
->power
;
4116 this_load
= avg_load
;
4117 } else if (avg_load
< min_load
) {
4118 min_load
= avg_load
;
4121 } while (group
= group
->next
, group
!= sd
->groups
);
4123 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
4129 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4132 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
4134 unsigned long load
, min_load
= ULONG_MAX
;
4138 /* Traverse only the allowed CPUs */
4139 for_each_cpu_and(i
, sched_group_cpus(group
), tsk_cpus_allowed(p
)) {
4140 load
= weighted_cpuload(i
);
4142 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
4152 * Try and locate an idle CPU in the sched_domain.
4154 static int select_idle_sibling(struct task_struct
*p
, int target
)
4156 struct sched_domain
*sd
;
4157 struct sched_group
*sg
;
4158 int i
= task_cpu(p
);
4160 if (idle_cpu(target
))
4164 * If the prevous cpu is cache affine and idle, don't be stupid.
4166 if (i
!= target
&& cpus_share_cache(i
, target
) && idle_cpu(i
))
4170 * Otherwise, iterate the domains and find an elegible idle cpu.
4172 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
4173 for_each_lower_domain(sd
) {
4176 if (!cpumask_intersects(sched_group_cpus(sg
),
4177 tsk_cpus_allowed(p
)))
4180 for_each_cpu(i
, sched_group_cpus(sg
)) {
4181 if (i
== target
|| !idle_cpu(i
))
4185 target
= cpumask_first_and(sched_group_cpus(sg
),
4186 tsk_cpus_allowed(p
));
4190 } while (sg
!= sd
->groups
);
4197 * sched_balance_self: balance the current task (running on cpu) in domains
4198 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
4201 * Balance, ie. select the least loaded group.
4203 * Returns the target CPU number, or the same CPU if no balancing is needed.
4205 * preempt must be disabled.
4208 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
4210 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
4211 int cpu
= smp_processor_id();
4213 int want_affine
= 0;
4214 int sync
= wake_flags
& WF_SYNC
;
4216 if (p
->nr_cpus_allowed
== 1)
4219 if (sd_flag
& SD_BALANCE_WAKE
) {
4220 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)))
4226 for_each_domain(cpu
, tmp
) {
4227 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
4231 * If both cpu and prev_cpu are part of this domain,
4232 * cpu is a valid SD_WAKE_AFFINE target.
4234 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
4235 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
4240 if (tmp
->flags
& sd_flag
)
4245 if (cpu
!= prev_cpu
&& wake_affine(affine_sd
, p
, sync
))
4248 new_cpu
= select_idle_sibling(p
, prev_cpu
);
4253 int load_idx
= sd
->forkexec_idx
;
4254 struct sched_group
*group
;
4257 if (!(sd
->flags
& sd_flag
)) {
4262 if (sd_flag
& SD_BALANCE_WAKE
)
4263 load_idx
= sd
->wake_idx
;
4265 group
= find_idlest_group(sd
, p
, cpu
, load_idx
);
4271 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
4272 if (new_cpu
== -1 || new_cpu
== cpu
) {
4273 /* Now try balancing at a lower domain level of cpu */
4278 /* Now try balancing at a lower domain level of new_cpu */
4280 weight
= sd
->span_weight
;
4282 for_each_domain(cpu
, tmp
) {
4283 if (weight
<= tmp
->span_weight
)
4285 if (tmp
->flags
& sd_flag
)
4288 /* while loop will break here if sd == NULL */
4297 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4298 * cfs_rq_of(p) references at time of call are still valid and identify the
4299 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4300 * other assumptions, including the state of rq->lock, should be made.
4303 migrate_task_rq_fair(struct task_struct
*p
, int next_cpu
)
4305 struct sched_entity
*se
= &p
->se
;
4306 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4309 * Load tracking: accumulate removed load so that it can be processed
4310 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4311 * to blocked load iff they have a positive decay-count. It can never
4312 * be negative here since on-rq tasks have decay-count == 0.
4314 if (se
->avg
.decay_count
) {
4315 se
->avg
.decay_count
= -__synchronize_entity_decay(se
);
4316 atomic_long_add(se
->avg
.load_avg_contrib
,
4317 &cfs_rq
->removed_load
);
4320 #endif /* CONFIG_SMP */
4322 static unsigned long
4323 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
4325 unsigned long gran
= sysctl_sched_wakeup_granularity
;
4328 * Since its curr running now, convert the gran from real-time
4329 * to virtual-time in his units.
4331 * By using 'se' instead of 'curr' we penalize light tasks, so
4332 * they get preempted easier. That is, if 'se' < 'curr' then
4333 * the resulting gran will be larger, therefore penalizing the
4334 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4335 * be smaller, again penalizing the lighter task.
4337 * This is especially important for buddies when the leftmost
4338 * task is higher priority than the buddy.
4340 return calc_delta_fair(gran
, se
);
4344 * Should 'se' preempt 'curr'.
4358 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
4360 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
4365 gran
= wakeup_gran(curr
, se
);
4372 static void set_last_buddy(struct sched_entity
*se
)
4374 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4377 for_each_sched_entity(se
)
4378 cfs_rq_of(se
)->last
= se
;
4381 static void set_next_buddy(struct sched_entity
*se
)
4383 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4386 for_each_sched_entity(se
)
4387 cfs_rq_of(se
)->next
= se
;
4390 static void set_skip_buddy(struct sched_entity
*se
)
4392 for_each_sched_entity(se
)
4393 cfs_rq_of(se
)->skip
= se
;
4397 * Preempt the current task with a newly woken task if needed:
4399 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
4401 struct task_struct
*curr
= rq
->curr
;
4402 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
4403 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4404 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
4405 int next_buddy_marked
= 0;
4407 if (unlikely(se
== pse
))
4411 * This is possible from callers such as move_task(), in which we
4412 * unconditionally check_prempt_curr() after an enqueue (which may have
4413 * lead to a throttle). This both saves work and prevents false
4414 * next-buddy nomination below.
4416 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
4419 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
4420 set_next_buddy(pse
);
4421 next_buddy_marked
= 1;
4425 * We can come here with TIF_NEED_RESCHED already set from new task
4428 * Note: this also catches the edge-case of curr being in a throttled
4429 * group (e.g. via set_curr_task), since update_curr() (in the
4430 * enqueue of curr) will have resulted in resched being set. This
4431 * prevents us from potentially nominating it as a false LAST_BUDDY
4434 if (test_tsk_need_resched(curr
))
4437 /* Idle tasks are by definition preempted by non-idle tasks. */
4438 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
4439 likely(p
->policy
!= SCHED_IDLE
))
4443 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4444 * is driven by the tick):
4446 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
4449 find_matching_se(&se
, &pse
);
4450 update_curr(cfs_rq_of(se
));
4452 if (wakeup_preempt_entity(se
, pse
) == 1) {
4454 * Bias pick_next to pick the sched entity that is
4455 * triggering this preemption.
4457 if (!next_buddy_marked
)
4458 set_next_buddy(pse
);
4467 * Only set the backward buddy when the current task is still
4468 * on the rq. This can happen when a wakeup gets interleaved
4469 * with schedule on the ->pre_schedule() or idle_balance()
4470 * point, either of which can * drop the rq lock.
4472 * Also, during early boot the idle thread is in the fair class,
4473 * for obvious reasons its a bad idea to schedule back to it.
4475 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
4478 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
4482 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
4484 struct task_struct
*p
;
4485 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
4486 struct sched_entity
*se
;
4488 if (!cfs_rq
->nr_running
)
4492 se
= pick_next_entity(cfs_rq
);
4493 set_next_entity(cfs_rq
, se
);
4494 cfs_rq
= group_cfs_rq(se
);
4498 if (hrtick_enabled(rq
))
4499 hrtick_start_fair(rq
, p
);
4505 * Account for a descheduled task:
4507 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
4509 struct sched_entity
*se
= &prev
->se
;
4510 struct cfs_rq
*cfs_rq
;
4512 for_each_sched_entity(se
) {
4513 cfs_rq
= cfs_rq_of(se
);
4514 put_prev_entity(cfs_rq
, se
);
4519 * sched_yield() is very simple
4521 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4523 static void yield_task_fair(struct rq
*rq
)
4525 struct task_struct
*curr
= rq
->curr
;
4526 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4527 struct sched_entity
*se
= &curr
->se
;
4530 * Are we the only task in the tree?
4532 if (unlikely(rq
->nr_running
== 1))
4535 clear_buddies(cfs_rq
, se
);
4537 if (curr
->policy
!= SCHED_BATCH
) {
4538 update_rq_clock(rq
);
4540 * Update run-time statistics of the 'current'.
4542 update_curr(cfs_rq
);
4544 * Tell update_rq_clock() that we've just updated,
4545 * so we don't do microscopic update in schedule()
4546 * and double the fastpath cost.
4548 rq
->skip_clock_update
= 1;
4554 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
4556 struct sched_entity
*se
= &p
->se
;
4558 /* throttled hierarchies are not runnable */
4559 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
4562 /* Tell the scheduler that we'd really like pse to run next. */
4565 yield_task_fair(rq
);
4571 /**************************************************
4572 * Fair scheduling class load-balancing methods.
4576 * The purpose of load-balancing is to achieve the same basic fairness the
4577 * per-cpu scheduler provides, namely provide a proportional amount of compute
4578 * time to each task. This is expressed in the following equation:
4580 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4582 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4583 * W_i,0 is defined as:
4585 * W_i,0 = \Sum_j w_i,j (2)
4587 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4588 * is derived from the nice value as per prio_to_weight[].
4590 * The weight average is an exponential decay average of the instantaneous
4593 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4595 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4596 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4597 * can also include other factors [XXX].
4599 * To achieve this balance we define a measure of imbalance which follows
4600 * directly from (1):
4602 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4604 * We them move tasks around to minimize the imbalance. In the continuous
4605 * function space it is obvious this converges, in the discrete case we get
4606 * a few fun cases generally called infeasible weight scenarios.
4609 * - infeasible weights;
4610 * - local vs global optima in the discrete case. ]
4615 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4616 * for all i,j solution, we create a tree of cpus that follows the hardware
4617 * topology where each level pairs two lower groups (or better). This results
4618 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4619 * tree to only the first of the previous level and we decrease the frequency
4620 * of load-balance at each level inv. proportional to the number of cpus in
4626 * \Sum { --- * --- * 2^i } = O(n) (5)
4628 * `- size of each group
4629 * | | `- number of cpus doing load-balance
4631 * `- sum over all levels
4633 * Coupled with a limit on how many tasks we can migrate every balance pass,
4634 * this makes (5) the runtime complexity of the balancer.
4636 * An important property here is that each CPU is still (indirectly) connected
4637 * to every other cpu in at most O(log n) steps:
4639 * The adjacency matrix of the resulting graph is given by:
4642 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4645 * And you'll find that:
4647 * A^(log_2 n)_i,j != 0 for all i,j (7)
4649 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4650 * The task movement gives a factor of O(m), giving a convergence complexity
4653 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4658 * In order to avoid CPUs going idle while there's still work to do, new idle
4659 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4660 * tree itself instead of relying on other CPUs to bring it work.
4662 * This adds some complexity to both (5) and (8) but it reduces the total idle
4670 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4673 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4678 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4680 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4682 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4685 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4686 * rewrite all of this once again.]
4689 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
4691 enum fbq_type
{ regular
, remote
, all
};
4693 #define LBF_ALL_PINNED 0x01
4694 #define LBF_NEED_BREAK 0x02
4695 #define LBF_DST_PINNED 0x04
4696 #define LBF_SOME_PINNED 0x08
4699 struct sched_domain
*sd
;
4707 struct cpumask
*dst_grpmask
;
4709 enum cpu_idle_type idle
;
4711 /* The set of CPUs under consideration for load-balancing */
4712 struct cpumask
*cpus
;
4717 unsigned int loop_break
;
4718 unsigned int loop_max
;
4720 enum fbq_type fbq_type
;
4724 * move_task - move a task from one runqueue to another runqueue.
4725 * Both runqueues must be locked.
4727 static void move_task(struct task_struct
*p
, struct lb_env
*env
)
4729 deactivate_task(env
->src_rq
, p
, 0);
4730 set_task_cpu(p
, env
->dst_cpu
);
4731 activate_task(env
->dst_rq
, p
, 0);
4732 check_preempt_curr(env
->dst_rq
, p
, 0);
4733 #ifdef CONFIG_NUMA_BALANCING
4734 if (p
->numa_preferred_nid
!= -1) {
4735 int src_nid
= cpu_to_node(env
->src_cpu
);
4736 int dst_nid
= cpu_to_node(env
->dst_cpu
);
4739 * If the load balancer has moved the task then limit
4740 * migrations from taking place in the short term in
4741 * case this is a short-lived migration.
4743 if (src_nid
!= dst_nid
&& dst_nid
!= p
->numa_preferred_nid
)
4744 p
->numa_migrate_seq
= 0;
4750 * Is this task likely cache-hot:
4753 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
4757 if (p
->sched_class
!= &fair_sched_class
)
4760 if (unlikely(p
->policy
== SCHED_IDLE
))
4764 * Buddy candidates are cache hot:
4766 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
4767 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
4768 &p
->se
== cfs_rq_of(&p
->se
)->last
))
4771 if (sysctl_sched_migration_cost
== -1)
4773 if (sysctl_sched_migration_cost
== 0)
4776 delta
= now
- p
->se
.exec_start
;
4778 return delta
< (s64
)sysctl_sched_migration_cost
;
4781 #ifdef CONFIG_NUMA_BALANCING
4782 /* Returns true if the destination node has incurred more faults */
4783 static bool migrate_improves_locality(struct task_struct
*p
, struct lb_env
*env
)
4785 int src_nid
, dst_nid
;
4787 if (!sched_feat(NUMA_FAVOUR_HIGHER
) || !p
->numa_faults
||
4788 !(env
->sd
->flags
& SD_NUMA
)) {
4792 src_nid
= cpu_to_node(env
->src_cpu
);
4793 dst_nid
= cpu_to_node(env
->dst_cpu
);
4795 if (src_nid
== dst_nid
)
4798 /* Always encourage migration to the preferred node. */
4799 if (dst_nid
== p
->numa_preferred_nid
)
4802 /* If both task and group weight improve, this move is a winner. */
4803 if (task_weight(p
, dst_nid
) > task_weight(p
, src_nid
) &&
4804 group_weight(p
, dst_nid
) > group_weight(p
, src_nid
))
4811 static bool migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
4813 int src_nid
, dst_nid
;
4815 if (!sched_feat(NUMA
) || !sched_feat(NUMA_RESIST_LOWER
))
4818 if (!p
->numa_faults
|| !(env
->sd
->flags
& SD_NUMA
))
4821 src_nid
= cpu_to_node(env
->src_cpu
);
4822 dst_nid
= cpu_to_node(env
->dst_cpu
);
4824 if (src_nid
== dst_nid
)
4827 /* Migrating away from the preferred node is always bad. */
4828 if (src_nid
== p
->numa_preferred_nid
)
4831 /* If either task or group weight get worse, don't do it. */
4832 if (task_weight(p
, dst_nid
) < task_weight(p
, src_nid
) ||
4833 group_weight(p
, dst_nid
) < group_weight(p
, src_nid
))
4840 static inline bool migrate_improves_locality(struct task_struct
*p
,
4846 static inline bool migrate_degrades_locality(struct task_struct
*p
,
4854 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4857 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
4859 int tsk_cache_hot
= 0;
4861 * We do not migrate tasks that are:
4862 * 1) throttled_lb_pair, or
4863 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4864 * 3) running (obviously), or
4865 * 4) are cache-hot on their current CPU.
4867 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
4870 if (!cpumask_test_cpu(env
->dst_cpu
, tsk_cpus_allowed(p
))) {
4873 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
4875 env
->flags
|= LBF_SOME_PINNED
;
4878 * Remember if this task can be migrated to any other cpu in
4879 * our sched_group. We may want to revisit it if we couldn't
4880 * meet load balance goals by pulling other tasks on src_cpu.
4882 * Also avoid computing new_dst_cpu if we have already computed
4883 * one in current iteration.
4885 if (!env
->dst_grpmask
|| (env
->flags
& LBF_DST_PINNED
))
4888 /* Prevent to re-select dst_cpu via env's cpus */
4889 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
4890 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
))) {
4891 env
->flags
|= LBF_DST_PINNED
;
4892 env
->new_dst_cpu
= cpu
;
4900 /* Record that we found atleast one task that could run on dst_cpu */
4901 env
->flags
&= ~LBF_ALL_PINNED
;
4903 if (task_running(env
->src_rq
, p
)) {
4904 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
4909 * Aggressive migration if:
4910 * 1) destination numa is preferred
4911 * 2) task is cache cold, or
4912 * 3) too many balance attempts have failed.
4914 tsk_cache_hot
= task_hot(p
, rq_clock_task(env
->src_rq
), env
->sd
);
4916 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
4918 if (migrate_improves_locality(p
, env
)) {
4919 #ifdef CONFIG_SCHEDSTATS
4920 if (tsk_cache_hot
) {
4921 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
4922 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
4928 if (!tsk_cache_hot
||
4929 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
4931 if (tsk_cache_hot
) {
4932 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
4933 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
4939 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
4944 * move_one_task tries to move exactly one task from busiest to this_rq, as
4945 * part of active balancing operations within "domain".
4946 * Returns 1 if successful and 0 otherwise.
4948 * Called with both runqueues locked.
4950 static int move_one_task(struct lb_env
*env
)
4952 struct task_struct
*p
, *n
;
4954 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
4955 if (!can_migrate_task(p
, env
))
4960 * Right now, this is only the second place move_task()
4961 * is called, so we can safely collect move_task()
4962 * stats here rather than inside move_task().
4964 schedstat_inc(env
->sd
, lb_gained
[env
->idle
]);
4970 static const unsigned int sched_nr_migrate_break
= 32;
4973 * move_tasks tries to move up to imbalance weighted load from busiest to
4974 * this_rq, as part of a balancing operation within domain "sd".
4975 * Returns 1 if successful and 0 otherwise.
4977 * Called with both runqueues locked.
4979 static int move_tasks(struct lb_env
*env
)
4981 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
4982 struct task_struct
*p
;
4986 if (env
->imbalance
<= 0)
4989 while (!list_empty(tasks
)) {
4990 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
4993 /* We've more or less seen every task there is, call it quits */
4994 if (env
->loop
> env
->loop_max
)
4997 /* take a breather every nr_migrate tasks */
4998 if (env
->loop
> env
->loop_break
) {
4999 env
->loop_break
+= sched_nr_migrate_break
;
5000 env
->flags
|= LBF_NEED_BREAK
;
5004 if (!can_migrate_task(p
, env
))
5007 load
= task_h_load(p
);
5009 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
5012 if ((load
/ 2) > env
->imbalance
)
5017 env
->imbalance
-= load
;
5019 #ifdef CONFIG_PREEMPT
5021 * NEWIDLE balancing is a source of latency, so preemptible
5022 * kernels will stop after the first task is pulled to minimize
5023 * the critical section.
5025 if (env
->idle
== CPU_NEWLY_IDLE
)
5030 * We only want to steal up to the prescribed amount of
5033 if (env
->imbalance
<= 0)
5038 list_move_tail(&p
->se
.group_node
, tasks
);
5042 * Right now, this is one of only two places move_task() is called,
5043 * so we can safely collect move_task() stats here rather than
5044 * inside move_task().
5046 schedstat_add(env
->sd
, lb_gained
[env
->idle
], pulled
);
5051 #ifdef CONFIG_FAIR_GROUP_SCHED
5053 * update tg->load_weight by folding this cpu's load_avg
5055 static void __update_blocked_averages_cpu(struct task_group
*tg
, int cpu
)
5057 struct sched_entity
*se
= tg
->se
[cpu
];
5058 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu
];
5060 /* throttled entities do not contribute to load */
5061 if (throttled_hierarchy(cfs_rq
))
5064 update_cfs_rq_blocked_load(cfs_rq
, 1);
5067 update_entity_load_avg(se
, 1);
5069 * We pivot on our runnable average having decayed to zero for
5070 * list removal. This generally implies that all our children
5071 * have also been removed (modulo rounding error or bandwidth
5072 * control); however, such cases are rare and we can fix these
5075 * TODO: fix up out-of-order children on enqueue.
5077 if (!se
->avg
.runnable_avg_sum
&& !cfs_rq
->nr_running
)
5078 list_del_leaf_cfs_rq(cfs_rq
);
5080 struct rq
*rq
= rq_of(cfs_rq
);
5081 update_rq_runnable_avg(rq
, rq
->nr_running
);
5085 static void update_blocked_averages(int cpu
)
5087 struct rq
*rq
= cpu_rq(cpu
);
5088 struct cfs_rq
*cfs_rq
;
5089 unsigned long flags
;
5091 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5092 update_rq_clock(rq
);
5094 * Iterates the task_group tree in a bottom up fashion, see
5095 * list_add_leaf_cfs_rq() for details.
5097 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
5099 * Note: We may want to consider periodically releasing
5100 * rq->lock about these updates so that creating many task
5101 * groups does not result in continually extending hold time.
5103 __update_blocked_averages_cpu(cfs_rq
->tg
, rq
->cpu
);
5106 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5110 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5111 * This needs to be done in a top-down fashion because the load of a child
5112 * group is a fraction of its parents load.
5114 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
5116 struct rq
*rq
= rq_of(cfs_rq
);
5117 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
5118 unsigned long now
= jiffies
;
5121 if (cfs_rq
->last_h_load_update
== now
)
5124 cfs_rq
->h_load_next
= NULL
;
5125 for_each_sched_entity(se
) {
5126 cfs_rq
= cfs_rq_of(se
);
5127 cfs_rq
->h_load_next
= se
;
5128 if (cfs_rq
->last_h_load_update
== now
)
5133 cfs_rq
->h_load
= cfs_rq
->runnable_load_avg
;
5134 cfs_rq
->last_h_load_update
= now
;
5137 while ((se
= cfs_rq
->h_load_next
) != NULL
) {
5138 load
= cfs_rq
->h_load
;
5139 load
= div64_ul(load
* se
->avg
.load_avg_contrib
,
5140 cfs_rq
->runnable_load_avg
+ 1);
5141 cfs_rq
= group_cfs_rq(se
);
5142 cfs_rq
->h_load
= load
;
5143 cfs_rq
->last_h_load_update
= now
;
5147 static unsigned long task_h_load(struct task_struct
*p
)
5149 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
5151 update_cfs_rq_h_load(cfs_rq
);
5152 return div64_ul(p
->se
.avg
.load_avg_contrib
* cfs_rq
->h_load
,
5153 cfs_rq
->runnable_load_avg
+ 1);
5156 static inline void update_blocked_averages(int cpu
)
5160 static unsigned long task_h_load(struct task_struct
*p
)
5162 return p
->se
.avg
.load_avg_contrib
;
5166 /********** Helpers for find_busiest_group ************************/
5168 * sg_lb_stats - stats of a sched_group required for load_balancing
5170 struct sg_lb_stats
{
5171 unsigned long avg_load
; /*Avg load across the CPUs of the group */
5172 unsigned long group_load
; /* Total load over the CPUs of the group */
5173 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
5174 unsigned long load_per_task
;
5175 unsigned long group_power
;
5176 unsigned int sum_nr_running
; /* Nr tasks running in the group */
5177 unsigned int group_capacity
;
5178 unsigned int idle_cpus
;
5179 unsigned int group_weight
;
5180 int group_imb
; /* Is there an imbalance in the group ? */
5181 int group_has_capacity
; /* Is there extra capacity in the group? */
5182 #ifdef CONFIG_NUMA_BALANCING
5183 unsigned int nr_numa_running
;
5184 unsigned int nr_preferred_running
;
5189 * sd_lb_stats - Structure to store the statistics of a sched_domain
5190 * during load balancing.
5192 struct sd_lb_stats
{
5193 struct sched_group
*busiest
; /* Busiest group in this sd */
5194 struct sched_group
*local
; /* Local group in this sd */
5195 unsigned long total_load
; /* Total load of all groups in sd */
5196 unsigned long total_pwr
; /* Total power of all groups in sd */
5197 unsigned long avg_load
; /* Average load across all groups in sd */
5199 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
5200 struct sg_lb_stats local_stat
; /* Statistics of the local group */
5203 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
5206 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5207 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5208 * We must however clear busiest_stat::avg_load because
5209 * update_sd_pick_busiest() reads this before assignment.
5211 *sds
= (struct sd_lb_stats
){
5223 * get_sd_load_idx - Obtain the load index for a given sched domain.
5224 * @sd: The sched_domain whose load_idx is to be obtained.
5225 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
5227 * Return: The load index.
5229 static inline int get_sd_load_idx(struct sched_domain
*sd
,
5230 enum cpu_idle_type idle
)
5236 load_idx
= sd
->busy_idx
;
5239 case CPU_NEWLY_IDLE
:
5240 load_idx
= sd
->newidle_idx
;
5243 load_idx
= sd
->idle_idx
;
5250 static unsigned long default_scale_freq_power(struct sched_domain
*sd
, int cpu
)
5252 return SCHED_POWER_SCALE
;
5255 unsigned long __weak
arch_scale_freq_power(struct sched_domain
*sd
, int cpu
)
5257 return default_scale_freq_power(sd
, cpu
);
5260 static unsigned long default_scale_smt_power(struct sched_domain
*sd
, int cpu
)
5262 unsigned long weight
= sd
->span_weight
;
5263 unsigned long smt_gain
= sd
->smt_gain
;
5270 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
5272 return default_scale_smt_power(sd
, cpu
);
5275 static unsigned long scale_rt_power(int cpu
)
5277 struct rq
*rq
= cpu_rq(cpu
);
5278 u64 total
, available
, age_stamp
, avg
;
5281 * Since we're reading these variables without serialization make sure
5282 * we read them once before doing sanity checks on them.
5284 age_stamp
= ACCESS_ONCE(rq
->age_stamp
);
5285 avg
= ACCESS_ONCE(rq
->rt_avg
);
5287 total
= sched_avg_period() + (rq_clock(rq
) - age_stamp
);
5289 if (unlikely(total
< avg
)) {
5290 /* Ensures that power won't end up being negative */
5293 available
= total
- avg
;
5296 if (unlikely((s64
)total
< SCHED_POWER_SCALE
))
5297 total
= SCHED_POWER_SCALE
;
5299 total
>>= SCHED_POWER_SHIFT
;
5301 return div_u64(available
, total
);
5304 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
5306 unsigned long weight
= sd
->span_weight
;
5307 unsigned long power
= SCHED_POWER_SCALE
;
5308 struct sched_group
*sdg
= sd
->groups
;
5310 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
5311 if (sched_feat(ARCH_POWER
))
5312 power
*= arch_scale_smt_power(sd
, cpu
);
5314 power
*= default_scale_smt_power(sd
, cpu
);
5316 power
>>= SCHED_POWER_SHIFT
;
5319 sdg
->sgp
->power_orig
= power
;
5321 if (sched_feat(ARCH_POWER
))
5322 power
*= arch_scale_freq_power(sd
, cpu
);
5324 power
*= default_scale_freq_power(sd
, cpu
);
5326 power
>>= SCHED_POWER_SHIFT
;
5328 power
*= scale_rt_power(cpu
);
5329 power
>>= SCHED_POWER_SHIFT
;
5334 cpu_rq(cpu
)->cpu_power
= power
;
5335 sdg
->sgp
->power
= power
;
5338 void update_group_power(struct sched_domain
*sd
, int cpu
)
5340 struct sched_domain
*child
= sd
->child
;
5341 struct sched_group
*group
, *sdg
= sd
->groups
;
5342 unsigned long power
, power_orig
;
5343 unsigned long interval
;
5345 interval
= msecs_to_jiffies(sd
->balance_interval
);
5346 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
5347 sdg
->sgp
->next_update
= jiffies
+ interval
;
5350 update_cpu_power(sd
, cpu
);
5354 power_orig
= power
= 0;
5356 if (child
->flags
& SD_OVERLAP
) {
5358 * SD_OVERLAP domains cannot assume that child groups
5359 * span the current group.
5362 for_each_cpu(cpu
, sched_group_cpus(sdg
)) {
5363 struct sched_group
*sg
= cpu_rq(cpu
)->sd
->groups
;
5365 power_orig
+= sg
->sgp
->power_orig
;
5366 power
+= sg
->sgp
->power
;
5370 * !SD_OVERLAP domains can assume that child groups
5371 * span the current group.
5374 group
= child
->groups
;
5376 power_orig
+= group
->sgp
->power_orig
;
5377 power
+= group
->sgp
->power
;
5378 group
= group
->next
;
5379 } while (group
!= child
->groups
);
5382 sdg
->sgp
->power_orig
= power_orig
;
5383 sdg
->sgp
->power
= power
;
5387 * Try and fix up capacity for tiny siblings, this is needed when
5388 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5389 * which on its own isn't powerful enough.
5391 * See update_sd_pick_busiest() and check_asym_packing().
5394 fix_small_capacity(struct sched_domain
*sd
, struct sched_group
*group
)
5397 * Only siblings can have significantly less than SCHED_POWER_SCALE
5399 if (!(sd
->flags
& SD_SHARE_CPUPOWER
))
5403 * If ~90% of the cpu_power is still there, we're good.
5405 if (group
->sgp
->power
* 32 > group
->sgp
->power_orig
* 29)
5412 * Group imbalance indicates (and tries to solve) the problem where balancing
5413 * groups is inadequate due to tsk_cpus_allowed() constraints.
5415 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5416 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5419 * { 0 1 2 3 } { 4 5 6 7 }
5422 * If we were to balance group-wise we'd place two tasks in the first group and
5423 * two tasks in the second group. Clearly this is undesired as it will overload
5424 * cpu 3 and leave one of the cpus in the second group unused.
5426 * The current solution to this issue is detecting the skew in the first group
5427 * by noticing the lower domain failed to reach balance and had difficulty
5428 * moving tasks due to affinity constraints.
5430 * When this is so detected; this group becomes a candidate for busiest; see
5431 * update_sd_pick_busiest(). And calculcate_imbalance() and
5432 * find_busiest_group() avoid some of the usual balance conditions to allow it
5433 * to create an effective group imbalance.
5435 * This is a somewhat tricky proposition since the next run might not find the
5436 * group imbalance and decide the groups need to be balanced again. A most
5437 * subtle and fragile situation.
5440 static inline int sg_imbalanced(struct sched_group
*group
)
5442 return group
->sgp
->imbalance
;
5446 * Compute the group capacity.
5448 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5449 * first dividing out the smt factor and computing the actual number of cores
5450 * and limit power unit capacity with that.
5452 static inline int sg_capacity(struct lb_env
*env
, struct sched_group
*group
)
5454 unsigned int capacity
, smt
, cpus
;
5455 unsigned int power
, power_orig
;
5457 power
= group
->sgp
->power
;
5458 power_orig
= group
->sgp
->power_orig
;
5459 cpus
= group
->group_weight
;
5461 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5462 smt
= DIV_ROUND_UP(SCHED_POWER_SCALE
* cpus
, power_orig
);
5463 capacity
= cpus
/ smt
; /* cores */
5465 capacity
= min_t(unsigned, capacity
, DIV_ROUND_CLOSEST(power
, SCHED_POWER_SCALE
));
5467 capacity
= fix_small_capacity(env
->sd
, group
);
5473 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5474 * @env: The load balancing environment.
5475 * @group: sched_group whose statistics are to be updated.
5476 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5477 * @local_group: Does group contain this_cpu.
5478 * @sgs: variable to hold the statistics for this group.
5480 static inline void update_sg_lb_stats(struct lb_env
*env
,
5481 struct sched_group
*group
, int load_idx
,
5482 int local_group
, struct sg_lb_stats
*sgs
)
5484 unsigned long nr_running
;
5488 memset(sgs
, 0, sizeof(*sgs
));
5490 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
5491 struct rq
*rq
= cpu_rq(i
);
5493 nr_running
= rq
->nr_running
;
5495 /* Bias balancing toward cpus of our domain */
5497 load
= target_load(i
, load_idx
);
5499 load
= source_load(i
, load_idx
);
5501 sgs
->group_load
+= load
;
5502 sgs
->sum_nr_running
+= nr_running
;
5503 #ifdef CONFIG_NUMA_BALANCING
5504 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
5505 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
5507 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
5512 /* Adjust by relative CPU power of the group */
5513 sgs
->group_power
= group
->sgp
->power
;
5514 sgs
->avg_load
= (sgs
->group_load
*SCHED_POWER_SCALE
) / sgs
->group_power
;
5516 if (sgs
->sum_nr_running
)
5517 sgs
->load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
5519 sgs
->group_weight
= group
->group_weight
;
5521 sgs
->group_imb
= sg_imbalanced(group
);
5522 sgs
->group_capacity
= sg_capacity(env
, group
);
5524 if (sgs
->group_capacity
> sgs
->sum_nr_running
)
5525 sgs
->group_has_capacity
= 1;
5529 * update_sd_pick_busiest - return 1 on busiest group
5530 * @env: The load balancing environment.
5531 * @sds: sched_domain statistics
5532 * @sg: sched_group candidate to be checked for being the busiest
5533 * @sgs: sched_group statistics
5535 * Determine if @sg is a busier group than the previously selected
5538 * Return: %true if @sg is a busier group than the previously selected
5539 * busiest group. %false otherwise.
5541 static bool update_sd_pick_busiest(struct lb_env
*env
,
5542 struct sd_lb_stats
*sds
,
5543 struct sched_group
*sg
,
5544 struct sg_lb_stats
*sgs
)
5546 if (sgs
->avg_load
<= sds
->busiest_stat
.avg_load
)
5549 if (sgs
->sum_nr_running
> sgs
->group_capacity
)
5556 * ASYM_PACKING needs to move all the work to the lowest
5557 * numbered CPUs in the group, therefore mark all groups
5558 * higher than ourself as busy.
5560 if ((env
->sd
->flags
& SD_ASYM_PACKING
) && sgs
->sum_nr_running
&&
5561 env
->dst_cpu
< group_first_cpu(sg
)) {
5565 if (group_first_cpu(sds
->busiest
) > group_first_cpu(sg
))
5572 #ifdef CONFIG_NUMA_BALANCING
5573 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
5575 if (sgs
->sum_nr_running
> sgs
->nr_numa_running
)
5577 if (sgs
->sum_nr_running
> sgs
->nr_preferred_running
)
5582 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
5584 if (rq
->nr_running
> rq
->nr_numa_running
)
5586 if (rq
->nr_running
> rq
->nr_preferred_running
)
5591 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
5596 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
5600 #endif /* CONFIG_NUMA_BALANCING */
5603 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5604 * @env: The load balancing environment.
5605 * @balance: Should we balance.
5606 * @sds: variable to hold the statistics for this sched_domain.
5608 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
5610 struct sched_domain
*child
= env
->sd
->child
;
5611 struct sched_group
*sg
= env
->sd
->groups
;
5612 struct sg_lb_stats tmp_sgs
;
5613 int load_idx
, prefer_sibling
= 0;
5615 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
5618 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
5621 struct sg_lb_stats
*sgs
= &tmp_sgs
;
5624 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_cpus(sg
));
5627 sgs
= &sds
->local_stat
;
5629 if (env
->idle
!= CPU_NEWLY_IDLE
||
5630 time_after_eq(jiffies
, sg
->sgp
->next_update
))
5631 update_group_power(env
->sd
, env
->dst_cpu
);
5634 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, sgs
);
5640 * In case the child domain prefers tasks go to siblings
5641 * first, lower the sg capacity to one so that we'll try
5642 * and move all the excess tasks away. We lower the capacity
5643 * of a group only if the local group has the capacity to fit
5644 * these excess tasks, i.e. nr_running < group_capacity. The
5645 * extra check prevents the case where you always pull from the
5646 * heaviest group when it is already under-utilized (possible
5647 * with a large weight task outweighs the tasks on the system).
5649 if (prefer_sibling
&& sds
->local
&&
5650 sds
->local_stat
.group_has_capacity
)
5651 sgs
->group_capacity
= min(sgs
->group_capacity
, 1U);
5653 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
5655 sds
->busiest_stat
= *sgs
;
5659 /* Now, start updating sd_lb_stats */
5660 sds
->total_load
+= sgs
->group_load
;
5661 sds
->total_pwr
+= sgs
->group_power
;
5664 } while (sg
!= env
->sd
->groups
);
5666 if (env
->sd
->flags
& SD_NUMA
)
5667 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
5671 * check_asym_packing - Check to see if the group is packed into the
5674 * This is primarily intended to used at the sibling level. Some
5675 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5676 * case of POWER7, it can move to lower SMT modes only when higher
5677 * threads are idle. When in lower SMT modes, the threads will
5678 * perform better since they share less core resources. Hence when we
5679 * have idle threads, we want them to be the higher ones.
5681 * This packing function is run on idle threads. It checks to see if
5682 * the busiest CPU in this domain (core in the P7 case) has a higher
5683 * CPU number than the packing function is being run on. Here we are
5684 * assuming lower CPU number will be equivalent to lower a SMT thread
5687 * Return: 1 when packing is required and a task should be moved to
5688 * this CPU. The amount of the imbalance is returned in *imbalance.
5690 * @env: The load balancing environment.
5691 * @sds: Statistics of the sched_domain which is to be packed
5693 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
5697 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
5703 busiest_cpu
= group_first_cpu(sds
->busiest
);
5704 if (env
->dst_cpu
> busiest_cpu
)
5707 env
->imbalance
= DIV_ROUND_CLOSEST(
5708 sds
->busiest_stat
.avg_load
* sds
->busiest_stat
.group_power
,
5715 * fix_small_imbalance - Calculate the minor imbalance that exists
5716 * amongst the groups of a sched_domain, during
5718 * @env: The load balancing environment.
5719 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5722 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
5724 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
5725 unsigned int imbn
= 2;
5726 unsigned long scaled_busy_load_per_task
;
5727 struct sg_lb_stats
*local
, *busiest
;
5729 local
= &sds
->local_stat
;
5730 busiest
= &sds
->busiest_stat
;
5732 if (!local
->sum_nr_running
)
5733 local
->load_per_task
= cpu_avg_load_per_task(env
->dst_cpu
);
5734 else if (busiest
->load_per_task
> local
->load_per_task
)
5737 scaled_busy_load_per_task
=
5738 (busiest
->load_per_task
* SCHED_POWER_SCALE
) /
5739 busiest
->group_power
;
5741 if (busiest
->avg_load
+ scaled_busy_load_per_task
>=
5742 local
->avg_load
+ (scaled_busy_load_per_task
* imbn
)) {
5743 env
->imbalance
= busiest
->load_per_task
;
5748 * OK, we don't have enough imbalance to justify moving tasks,
5749 * however we may be able to increase total CPU power used by
5753 pwr_now
+= busiest
->group_power
*
5754 min(busiest
->load_per_task
, busiest
->avg_load
);
5755 pwr_now
+= local
->group_power
*
5756 min(local
->load_per_task
, local
->avg_load
);
5757 pwr_now
/= SCHED_POWER_SCALE
;
5759 /* Amount of load we'd subtract */
5760 tmp
= (busiest
->load_per_task
* SCHED_POWER_SCALE
) /
5761 busiest
->group_power
;
5762 if (busiest
->avg_load
> tmp
) {
5763 pwr_move
+= busiest
->group_power
*
5764 min(busiest
->load_per_task
,
5765 busiest
->avg_load
- tmp
);
5768 /* Amount of load we'd add */
5769 if (busiest
->avg_load
* busiest
->group_power
<
5770 busiest
->load_per_task
* SCHED_POWER_SCALE
) {
5771 tmp
= (busiest
->avg_load
* busiest
->group_power
) /
5774 tmp
= (busiest
->load_per_task
* SCHED_POWER_SCALE
) /
5777 pwr_move
+= local
->group_power
*
5778 min(local
->load_per_task
, local
->avg_load
+ tmp
);
5779 pwr_move
/= SCHED_POWER_SCALE
;
5781 /* Move if we gain throughput */
5782 if (pwr_move
> pwr_now
)
5783 env
->imbalance
= busiest
->load_per_task
;
5787 * calculate_imbalance - Calculate the amount of imbalance present within the
5788 * groups of a given sched_domain during load balance.
5789 * @env: load balance environment
5790 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5792 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
5794 unsigned long max_pull
, load_above_capacity
= ~0UL;
5795 struct sg_lb_stats
*local
, *busiest
;
5797 local
= &sds
->local_stat
;
5798 busiest
= &sds
->busiest_stat
;
5800 if (busiest
->group_imb
) {
5802 * In the group_imb case we cannot rely on group-wide averages
5803 * to ensure cpu-load equilibrium, look at wider averages. XXX
5805 busiest
->load_per_task
=
5806 min(busiest
->load_per_task
, sds
->avg_load
);
5810 * In the presence of smp nice balancing, certain scenarios can have
5811 * max load less than avg load(as we skip the groups at or below
5812 * its cpu_power, while calculating max_load..)
5814 if (busiest
->avg_load
<= sds
->avg_load
||
5815 local
->avg_load
>= sds
->avg_load
) {
5817 return fix_small_imbalance(env
, sds
);
5820 if (!busiest
->group_imb
) {
5822 * Don't want to pull so many tasks that a group would go idle.
5823 * Except of course for the group_imb case, since then we might
5824 * have to drop below capacity to reach cpu-load equilibrium.
5826 load_above_capacity
=
5827 (busiest
->sum_nr_running
- busiest
->group_capacity
);
5829 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_POWER_SCALE
);
5830 load_above_capacity
/= busiest
->group_power
;
5834 * We're trying to get all the cpus to the average_load, so we don't
5835 * want to push ourselves above the average load, nor do we wish to
5836 * reduce the max loaded cpu below the average load. At the same time,
5837 * we also don't want to reduce the group load below the group capacity
5838 * (so that we can implement power-savings policies etc). Thus we look
5839 * for the minimum possible imbalance.
5841 max_pull
= min(busiest
->avg_load
- sds
->avg_load
, load_above_capacity
);
5843 /* How much load to actually move to equalise the imbalance */
5844 env
->imbalance
= min(
5845 max_pull
* busiest
->group_power
,
5846 (sds
->avg_load
- local
->avg_load
) * local
->group_power
5847 ) / SCHED_POWER_SCALE
;
5850 * if *imbalance is less than the average load per runnable task
5851 * there is no guarantee that any tasks will be moved so we'll have
5852 * a think about bumping its value to force at least one task to be
5855 if (env
->imbalance
< busiest
->load_per_task
)
5856 return fix_small_imbalance(env
, sds
);
5859 /******* find_busiest_group() helpers end here *********************/
5862 * find_busiest_group - Returns the busiest group within the sched_domain
5863 * if there is an imbalance. If there isn't an imbalance, and
5864 * the user has opted for power-savings, it returns a group whose
5865 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5866 * such a group exists.
5868 * Also calculates the amount of weighted load which should be moved
5869 * to restore balance.
5871 * @env: The load balancing environment.
5873 * Return: - The busiest group if imbalance exists.
5874 * - If no imbalance and user has opted for power-savings balance,
5875 * return the least loaded group whose CPUs can be
5876 * put to idle by rebalancing its tasks onto our group.
5878 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
5880 struct sg_lb_stats
*local
, *busiest
;
5881 struct sd_lb_stats sds
;
5883 init_sd_lb_stats(&sds
);
5886 * Compute the various statistics relavent for load balancing at
5889 update_sd_lb_stats(env
, &sds
);
5890 local
= &sds
.local_stat
;
5891 busiest
= &sds
.busiest_stat
;
5893 if ((env
->idle
== CPU_IDLE
|| env
->idle
== CPU_NEWLY_IDLE
) &&
5894 check_asym_packing(env
, &sds
))
5897 /* There is no busy sibling group to pull tasks from */
5898 if (!sds
.busiest
|| busiest
->sum_nr_running
== 0)
5901 sds
.avg_load
= (SCHED_POWER_SCALE
* sds
.total_load
) / sds
.total_pwr
;
5904 * If the busiest group is imbalanced the below checks don't
5905 * work because they assume all things are equal, which typically
5906 * isn't true due to cpus_allowed constraints and the like.
5908 if (busiest
->group_imb
)
5911 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5912 if (env
->idle
== CPU_NEWLY_IDLE
&& local
->group_has_capacity
&&
5913 !busiest
->group_has_capacity
)
5917 * If the local group is more busy than the selected busiest group
5918 * don't try and pull any tasks.
5920 if (local
->avg_load
>= busiest
->avg_load
)
5924 * Don't pull any tasks if this group is already above the domain
5927 if (local
->avg_load
>= sds
.avg_load
)
5930 if (env
->idle
== CPU_IDLE
) {
5932 * This cpu is idle. If the busiest group load doesn't
5933 * have more tasks than the number of available cpu's and
5934 * there is no imbalance between this and busiest group
5935 * wrt to idle cpu's, it is balanced.
5937 if ((local
->idle_cpus
< busiest
->idle_cpus
) &&
5938 busiest
->sum_nr_running
<= busiest
->group_weight
)
5942 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5943 * imbalance_pct to be conservative.
5945 if (100 * busiest
->avg_load
<=
5946 env
->sd
->imbalance_pct
* local
->avg_load
)
5951 /* Looks like there is an imbalance. Compute it */
5952 calculate_imbalance(env
, &sds
);
5961 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5963 static struct rq
*find_busiest_queue(struct lb_env
*env
,
5964 struct sched_group
*group
)
5966 struct rq
*busiest
= NULL
, *rq
;
5967 unsigned long busiest_load
= 0, busiest_power
= 1;
5970 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
5971 unsigned long power
, capacity
, wl
;
5975 rt
= fbq_classify_rq(rq
);
5978 * We classify groups/runqueues into three groups:
5979 * - regular: there are !numa tasks
5980 * - remote: there are numa tasks that run on the 'wrong' node
5981 * - all: there is no distinction
5983 * In order to avoid migrating ideally placed numa tasks,
5984 * ignore those when there's better options.
5986 * If we ignore the actual busiest queue to migrate another
5987 * task, the next balance pass can still reduce the busiest
5988 * queue by moving tasks around inside the node.
5990 * If we cannot move enough load due to this classification
5991 * the next pass will adjust the group classification and
5992 * allow migration of more tasks.
5994 * Both cases only affect the total convergence complexity.
5996 if (rt
> env
->fbq_type
)
5999 power
= power_of(i
);
6000 capacity
= DIV_ROUND_CLOSEST(power
, SCHED_POWER_SCALE
);
6002 capacity
= fix_small_capacity(env
->sd
, group
);
6004 wl
= weighted_cpuload(i
);
6007 * When comparing with imbalance, use weighted_cpuload()
6008 * which is not scaled with the cpu power.
6010 if (capacity
&& rq
->nr_running
== 1 && wl
> env
->imbalance
)
6014 * For the load comparisons with the other cpu's, consider
6015 * the weighted_cpuload() scaled with the cpu power, so that
6016 * the load can be moved away from the cpu that is potentially
6017 * running at a lower capacity.
6019 * Thus we're looking for max(wl_i / power_i), crosswise
6020 * multiplication to rid ourselves of the division works out
6021 * to: wl_i * power_j > wl_j * power_i; where j is our
6024 if (wl
* busiest_power
> busiest_load
* power
) {
6026 busiest_power
= power
;
6035 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6036 * so long as it is large enough.
6038 #define MAX_PINNED_INTERVAL 512
6040 /* Working cpumask for load_balance and load_balance_newidle. */
6041 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6043 static int need_active_balance(struct lb_env
*env
)
6045 struct sched_domain
*sd
= env
->sd
;
6047 if (env
->idle
== CPU_NEWLY_IDLE
) {
6050 * ASYM_PACKING needs to force migrate tasks from busy but
6051 * higher numbered CPUs in order to pack all tasks in the
6052 * lowest numbered CPUs.
6054 if ((sd
->flags
& SD_ASYM_PACKING
) && env
->src_cpu
> env
->dst_cpu
)
6058 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
6061 static int active_load_balance_cpu_stop(void *data
);
6063 static int should_we_balance(struct lb_env
*env
)
6065 struct sched_group
*sg
= env
->sd
->groups
;
6066 struct cpumask
*sg_cpus
, *sg_mask
;
6067 int cpu
, balance_cpu
= -1;
6070 * In the newly idle case, we will allow all the cpu's
6071 * to do the newly idle load balance.
6073 if (env
->idle
== CPU_NEWLY_IDLE
)
6076 sg_cpus
= sched_group_cpus(sg
);
6077 sg_mask
= sched_group_mask(sg
);
6078 /* Try to find first idle cpu */
6079 for_each_cpu_and(cpu
, sg_cpus
, env
->cpus
) {
6080 if (!cpumask_test_cpu(cpu
, sg_mask
) || !idle_cpu(cpu
))
6087 if (balance_cpu
== -1)
6088 balance_cpu
= group_balance_cpu(sg
);
6091 * First idle cpu or the first cpu(busiest) in this sched group
6092 * is eligible for doing load balancing at this and above domains.
6094 return balance_cpu
== env
->dst_cpu
;
6098 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6099 * tasks if there is an imbalance.
6101 static int load_balance(int this_cpu
, struct rq
*this_rq
,
6102 struct sched_domain
*sd
, enum cpu_idle_type idle
,
6103 int *continue_balancing
)
6105 int ld_moved
, cur_ld_moved
, active_balance
= 0;
6106 struct sched_domain
*sd_parent
= sd
->parent
;
6107 struct sched_group
*group
;
6109 unsigned long flags
;
6110 struct cpumask
*cpus
= __get_cpu_var(load_balance_mask
);
6112 struct lb_env env
= {
6114 .dst_cpu
= this_cpu
,
6116 .dst_grpmask
= sched_group_cpus(sd
->groups
),
6118 .loop_break
= sched_nr_migrate_break
,
6124 * For NEWLY_IDLE load_balancing, we don't need to consider
6125 * other cpus in our group
6127 if (idle
== CPU_NEWLY_IDLE
)
6128 env
.dst_grpmask
= NULL
;
6130 cpumask_copy(cpus
, cpu_active_mask
);
6132 schedstat_inc(sd
, lb_count
[idle
]);
6135 if (!should_we_balance(&env
)) {
6136 *continue_balancing
= 0;
6140 group
= find_busiest_group(&env
);
6142 schedstat_inc(sd
, lb_nobusyg
[idle
]);
6146 busiest
= find_busiest_queue(&env
, group
);
6148 schedstat_inc(sd
, lb_nobusyq
[idle
]);
6152 BUG_ON(busiest
== env
.dst_rq
);
6154 schedstat_add(sd
, lb_imbalance
[idle
], env
.imbalance
);
6157 if (busiest
->nr_running
> 1) {
6159 * Attempt to move tasks. If find_busiest_group has found
6160 * an imbalance but busiest->nr_running <= 1, the group is
6161 * still unbalanced. ld_moved simply stays zero, so it is
6162 * correctly treated as an imbalance.
6164 env
.flags
|= LBF_ALL_PINNED
;
6165 env
.src_cpu
= busiest
->cpu
;
6166 env
.src_rq
= busiest
;
6167 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
6170 local_irq_save(flags
);
6171 double_rq_lock(env
.dst_rq
, busiest
);
6174 * cur_ld_moved - load moved in current iteration
6175 * ld_moved - cumulative load moved across iterations
6177 cur_ld_moved
= move_tasks(&env
);
6178 ld_moved
+= cur_ld_moved
;
6179 double_rq_unlock(env
.dst_rq
, busiest
);
6180 local_irq_restore(flags
);
6183 * some other cpu did the load balance for us.
6185 if (cur_ld_moved
&& env
.dst_cpu
!= smp_processor_id())
6186 resched_cpu(env
.dst_cpu
);
6188 if (env
.flags
& LBF_NEED_BREAK
) {
6189 env
.flags
&= ~LBF_NEED_BREAK
;
6194 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6195 * us and move them to an alternate dst_cpu in our sched_group
6196 * where they can run. The upper limit on how many times we
6197 * iterate on same src_cpu is dependent on number of cpus in our
6200 * This changes load balance semantics a bit on who can move
6201 * load to a given_cpu. In addition to the given_cpu itself
6202 * (or a ilb_cpu acting on its behalf where given_cpu is
6203 * nohz-idle), we now have balance_cpu in a position to move
6204 * load to given_cpu. In rare situations, this may cause
6205 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6206 * _independently_ and at _same_ time to move some load to
6207 * given_cpu) causing exceess load to be moved to given_cpu.
6208 * This however should not happen so much in practice and
6209 * moreover subsequent load balance cycles should correct the
6210 * excess load moved.
6212 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
6214 /* Prevent to re-select dst_cpu via env's cpus */
6215 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
6217 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
6218 env
.dst_cpu
= env
.new_dst_cpu
;
6219 env
.flags
&= ~LBF_DST_PINNED
;
6221 env
.loop_break
= sched_nr_migrate_break
;
6224 * Go back to "more_balance" rather than "redo" since we
6225 * need to continue with same src_cpu.
6231 * We failed to reach balance because of affinity.
6234 int *group_imbalance
= &sd_parent
->groups
->sgp
->imbalance
;
6236 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0) {
6237 *group_imbalance
= 1;
6238 } else if (*group_imbalance
)
6239 *group_imbalance
= 0;
6242 /* All tasks on this runqueue were pinned by CPU affinity */
6243 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
6244 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
6245 if (!cpumask_empty(cpus
)) {
6247 env
.loop_break
= sched_nr_migrate_break
;
6255 schedstat_inc(sd
, lb_failed
[idle
]);
6257 * Increment the failure counter only on periodic balance.
6258 * We do not want newidle balance, which can be very
6259 * frequent, pollute the failure counter causing
6260 * excessive cache_hot migrations and active balances.
6262 if (idle
!= CPU_NEWLY_IDLE
)
6263 sd
->nr_balance_failed
++;
6265 if (need_active_balance(&env
)) {
6266 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
6268 /* don't kick the active_load_balance_cpu_stop,
6269 * if the curr task on busiest cpu can't be
6272 if (!cpumask_test_cpu(this_cpu
,
6273 tsk_cpus_allowed(busiest
->curr
))) {
6274 raw_spin_unlock_irqrestore(&busiest
->lock
,
6276 env
.flags
|= LBF_ALL_PINNED
;
6277 goto out_one_pinned
;
6281 * ->active_balance synchronizes accesses to
6282 * ->active_balance_work. Once set, it's cleared
6283 * only after active load balance is finished.
6285 if (!busiest
->active_balance
) {
6286 busiest
->active_balance
= 1;
6287 busiest
->push_cpu
= this_cpu
;
6290 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
6292 if (active_balance
) {
6293 stop_one_cpu_nowait(cpu_of(busiest
),
6294 active_load_balance_cpu_stop
, busiest
,
6295 &busiest
->active_balance_work
);
6299 * We've kicked active balancing, reset the failure
6302 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
6305 sd
->nr_balance_failed
= 0;
6307 if (likely(!active_balance
)) {
6308 /* We were unbalanced, so reset the balancing interval */
6309 sd
->balance_interval
= sd
->min_interval
;
6312 * If we've begun active balancing, start to back off. This
6313 * case may not be covered by the all_pinned logic if there
6314 * is only 1 task on the busy runqueue (because we don't call
6317 if (sd
->balance_interval
< sd
->max_interval
)
6318 sd
->balance_interval
*= 2;
6324 schedstat_inc(sd
, lb_balanced
[idle
]);
6326 sd
->nr_balance_failed
= 0;
6329 /* tune up the balancing interval */
6330 if (((env
.flags
& LBF_ALL_PINNED
) &&
6331 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
6332 (sd
->balance_interval
< sd
->max_interval
))
6333 sd
->balance_interval
*= 2;
6341 * idle_balance is called by schedule() if this_cpu is about to become
6342 * idle. Attempts to pull tasks from other CPUs.
6344 void idle_balance(int this_cpu
, struct rq
*this_rq
)
6346 struct sched_domain
*sd
;
6347 int pulled_task
= 0;
6348 unsigned long next_balance
= jiffies
+ HZ
;
6351 this_rq
->idle_stamp
= rq_clock(this_rq
);
6353 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
6357 * Drop the rq->lock, but keep IRQ/preempt disabled.
6359 raw_spin_unlock(&this_rq
->lock
);
6361 update_blocked_averages(this_cpu
);
6363 for_each_domain(this_cpu
, sd
) {
6364 unsigned long interval
;
6365 int continue_balancing
= 1;
6366 u64 t0
, domain_cost
;
6368 if (!(sd
->flags
& SD_LOAD_BALANCE
))
6371 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
)
6374 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
6375 t0
= sched_clock_cpu(this_cpu
);
6377 /* If we've pulled tasks over stop searching: */
6378 pulled_task
= load_balance(this_cpu
, this_rq
,
6380 &continue_balancing
);
6382 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
6383 if (domain_cost
> sd
->max_newidle_lb_cost
)
6384 sd
->max_newidle_lb_cost
= domain_cost
;
6386 curr_cost
+= domain_cost
;
6389 interval
= msecs_to_jiffies(sd
->balance_interval
);
6390 if (time_after(next_balance
, sd
->last_balance
+ interval
))
6391 next_balance
= sd
->last_balance
+ interval
;
6393 this_rq
->idle_stamp
= 0;
6399 raw_spin_lock(&this_rq
->lock
);
6401 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
6403 * We are going idle. next_balance may be set based on
6404 * a busy processor. So reset next_balance.
6406 this_rq
->next_balance
= next_balance
;
6409 if (curr_cost
> this_rq
->max_idle_balance_cost
)
6410 this_rq
->max_idle_balance_cost
= curr_cost
;
6414 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6415 * running tasks off the busiest CPU onto idle CPUs. It requires at
6416 * least 1 task to be running on each physical CPU where possible, and
6417 * avoids physical / logical imbalances.
6419 static int active_load_balance_cpu_stop(void *data
)
6421 struct rq
*busiest_rq
= data
;
6422 int busiest_cpu
= cpu_of(busiest_rq
);
6423 int target_cpu
= busiest_rq
->push_cpu
;
6424 struct rq
*target_rq
= cpu_rq(target_cpu
);
6425 struct sched_domain
*sd
;
6427 raw_spin_lock_irq(&busiest_rq
->lock
);
6429 /* make sure the requested cpu hasn't gone down in the meantime */
6430 if (unlikely(busiest_cpu
!= smp_processor_id() ||
6431 !busiest_rq
->active_balance
))
6434 /* Is there any task to move? */
6435 if (busiest_rq
->nr_running
<= 1)
6439 * This condition is "impossible", if it occurs
6440 * we need to fix it. Originally reported by
6441 * Bjorn Helgaas on a 128-cpu setup.
6443 BUG_ON(busiest_rq
== target_rq
);
6445 /* move a task from busiest_rq to target_rq */
6446 double_lock_balance(busiest_rq
, target_rq
);
6448 /* Search for an sd spanning us and the target CPU. */
6450 for_each_domain(target_cpu
, sd
) {
6451 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
6452 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
6457 struct lb_env env
= {
6459 .dst_cpu
= target_cpu
,
6460 .dst_rq
= target_rq
,
6461 .src_cpu
= busiest_rq
->cpu
,
6462 .src_rq
= busiest_rq
,
6466 schedstat_inc(sd
, alb_count
);
6468 if (move_one_task(&env
))
6469 schedstat_inc(sd
, alb_pushed
);
6471 schedstat_inc(sd
, alb_failed
);
6474 double_unlock_balance(busiest_rq
, target_rq
);
6476 busiest_rq
->active_balance
= 0;
6477 raw_spin_unlock_irq(&busiest_rq
->lock
);
6481 #ifdef CONFIG_NO_HZ_COMMON
6483 * idle load balancing details
6484 * - When one of the busy CPUs notice that there may be an idle rebalancing
6485 * needed, they will kick the idle load balancer, which then does idle
6486 * load balancing for all the idle CPUs.
6489 cpumask_var_t idle_cpus_mask
;
6491 unsigned long next_balance
; /* in jiffy units */
6492 } nohz ____cacheline_aligned
;
6494 static inline int find_new_ilb(int call_cpu
)
6496 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
6498 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
6505 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6506 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6507 * CPU (if there is one).
6509 static void nohz_balancer_kick(int cpu
)
6513 nohz
.next_balance
++;
6515 ilb_cpu
= find_new_ilb(cpu
);
6517 if (ilb_cpu
>= nr_cpu_ids
)
6520 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
6523 * Use smp_send_reschedule() instead of resched_cpu().
6524 * This way we generate a sched IPI on the target cpu which
6525 * is idle. And the softirq performing nohz idle load balance
6526 * will be run before returning from the IPI.
6528 smp_send_reschedule(ilb_cpu
);
6532 static inline void nohz_balance_exit_idle(int cpu
)
6534 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
6535 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
6536 atomic_dec(&nohz
.nr_cpus
);
6537 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
6541 static inline void set_cpu_sd_state_busy(void)
6543 struct sched_domain
*sd
;
6546 sd
= rcu_dereference_check_sched_domain(this_rq()->sd
);
6548 if (!sd
|| !sd
->nohz_idle
)
6552 for (; sd
; sd
= sd
->parent
)
6553 atomic_inc(&sd
->groups
->sgp
->nr_busy_cpus
);
6558 void set_cpu_sd_state_idle(void)
6560 struct sched_domain
*sd
;
6563 sd
= rcu_dereference_check_sched_domain(this_rq()->sd
);
6565 if (!sd
|| sd
->nohz_idle
)
6569 for (; sd
; sd
= sd
->parent
)
6570 atomic_dec(&sd
->groups
->sgp
->nr_busy_cpus
);
6576 * This routine will record that the cpu is going idle with tick stopped.
6577 * This info will be used in performing idle load balancing in the future.
6579 void nohz_balance_enter_idle(int cpu
)
6582 * If this cpu is going down, then nothing needs to be done.
6584 if (!cpu_active(cpu
))
6587 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
6590 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
6591 atomic_inc(&nohz
.nr_cpus
);
6592 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
6595 static int sched_ilb_notifier(struct notifier_block
*nfb
,
6596 unsigned long action
, void *hcpu
)
6598 switch (action
& ~CPU_TASKS_FROZEN
) {
6600 nohz_balance_exit_idle(smp_processor_id());
6608 static DEFINE_SPINLOCK(balancing
);
6611 * Scale the max load_balance interval with the number of CPUs in the system.
6612 * This trades load-balance latency on larger machines for less cross talk.
6614 void update_max_interval(void)
6616 max_load_balance_interval
= HZ
*num_online_cpus()/10;
6620 * It checks each scheduling domain to see if it is due to be balanced,
6621 * and initiates a balancing operation if so.
6623 * Balancing parameters are set up in init_sched_domains.
6625 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
6627 int continue_balancing
= 1;
6628 struct rq
*rq
= cpu_rq(cpu
);
6629 unsigned long interval
;
6630 struct sched_domain
*sd
;
6631 /* Earliest time when we have to do rebalance again */
6632 unsigned long next_balance
= jiffies
+ 60*HZ
;
6633 int update_next_balance
= 0;
6634 int need_serialize
, need_decay
= 0;
6637 update_blocked_averages(cpu
);
6640 for_each_domain(cpu
, sd
) {
6642 * Decay the newidle max times here because this is a regular
6643 * visit to all the domains. Decay ~1% per second.
6645 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
6646 sd
->max_newidle_lb_cost
=
6647 (sd
->max_newidle_lb_cost
* 253) / 256;
6648 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
6651 max_cost
+= sd
->max_newidle_lb_cost
;
6653 if (!(sd
->flags
& SD_LOAD_BALANCE
))
6657 * Stop the load balance at this level. There is another
6658 * CPU in our sched group which is doing load balancing more
6661 if (!continue_balancing
) {
6667 interval
= sd
->balance_interval
;
6668 if (idle
!= CPU_IDLE
)
6669 interval
*= sd
->busy_factor
;
6671 /* scale ms to jiffies */
6672 interval
= msecs_to_jiffies(interval
);
6673 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
6675 need_serialize
= sd
->flags
& SD_SERIALIZE
;
6677 if (need_serialize
) {
6678 if (!spin_trylock(&balancing
))
6682 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
6683 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
6685 * The LBF_DST_PINNED logic could have changed
6686 * env->dst_cpu, so we can't know our idle
6687 * state even if we migrated tasks. Update it.
6689 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
6691 sd
->last_balance
= jiffies
;
6694 spin_unlock(&balancing
);
6696 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
6697 next_balance
= sd
->last_balance
+ interval
;
6698 update_next_balance
= 1;
6703 * Ensure the rq-wide value also decays but keep it at a
6704 * reasonable floor to avoid funnies with rq->avg_idle.
6706 rq
->max_idle_balance_cost
=
6707 max((u64
)sysctl_sched_migration_cost
, max_cost
);
6712 * next_balance will be updated only when there is a need.
6713 * When the cpu is attached to null domain for ex, it will not be
6716 if (likely(update_next_balance
))
6717 rq
->next_balance
= next_balance
;
6720 #ifdef CONFIG_NO_HZ_COMMON
6722 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6723 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6725 static void nohz_idle_balance(int this_cpu
, enum cpu_idle_type idle
)
6727 struct rq
*this_rq
= cpu_rq(this_cpu
);
6731 if (idle
!= CPU_IDLE
||
6732 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
6735 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
6736 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
6740 * If this cpu gets work to do, stop the load balancing
6741 * work being done for other cpus. Next load
6742 * balancing owner will pick it up.
6747 rq
= cpu_rq(balance_cpu
);
6749 raw_spin_lock_irq(&rq
->lock
);
6750 update_rq_clock(rq
);
6751 update_idle_cpu_load(rq
);
6752 raw_spin_unlock_irq(&rq
->lock
);
6754 rebalance_domains(balance_cpu
, CPU_IDLE
);
6756 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
6757 this_rq
->next_balance
= rq
->next_balance
;
6759 nohz
.next_balance
= this_rq
->next_balance
;
6761 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
6765 * Current heuristic for kicking the idle load balancer in the presence
6766 * of an idle cpu is the system.
6767 * - This rq has more than one task.
6768 * - At any scheduler domain level, this cpu's scheduler group has multiple
6769 * busy cpu's exceeding the group's power.
6770 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6771 * domain span are idle.
6773 static inline int nohz_kick_needed(struct rq
*rq
, int cpu
)
6775 unsigned long now
= jiffies
;
6776 struct sched_domain
*sd
;
6778 if (unlikely(idle_cpu(cpu
)))
6782 * We may be recently in ticked or tickless idle mode. At the first
6783 * busy tick after returning from idle, we will update the busy stats.
6785 set_cpu_sd_state_busy();
6786 nohz_balance_exit_idle(cpu
);
6789 * None are in tickless mode and hence no need for NOHZ idle load
6792 if (likely(!atomic_read(&nohz
.nr_cpus
)))
6795 if (time_before(now
, nohz
.next_balance
))
6798 if (rq
->nr_running
>= 2)
6802 for_each_domain(cpu
, sd
) {
6803 struct sched_group
*sg
= sd
->groups
;
6804 struct sched_group_power
*sgp
= sg
->sgp
;
6805 int nr_busy
= atomic_read(&sgp
->nr_busy_cpus
);
6807 if (sd
->flags
& SD_SHARE_PKG_RESOURCES
&& nr_busy
> 1)
6808 goto need_kick_unlock
;
6810 if (sd
->flags
& SD_ASYM_PACKING
&& nr_busy
!= sg
->group_weight
6811 && (cpumask_first_and(nohz
.idle_cpus_mask
,
6812 sched_domain_span(sd
)) < cpu
))
6813 goto need_kick_unlock
;
6815 if (!(sd
->flags
& (SD_SHARE_PKG_RESOURCES
| SD_ASYM_PACKING
)))
6827 static void nohz_idle_balance(int this_cpu
, enum cpu_idle_type idle
) { }
6831 * run_rebalance_domains is triggered when needed from the scheduler tick.
6832 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6834 static void run_rebalance_domains(struct softirq_action
*h
)
6836 int this_cpu
= smp_processor_id();
6837 struct rq
*this_rq
= cpu_rq(this_cpu
);
6838 enum cpu_idle_type idle
= this_rq
->idle_balance
?
6839 CPU_IDLE
: CPU_NOT_IDLE
;
6841 rebalance_domains(this_cpu
, idle
);
6844 * If this cpu has a pending nohz_balance_kick, then do the
6845 * balancing on behalf of the other idle cpus whose ticks are
6848 nohz_idle_balance(this_cpu
, idle
);
6851 static inline int on_null_domain(int cpu
)
6853 return !rcu_dereference_sched(cpu_rq(cpu
)->sd
);
6857 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6859 void trigger_load_balance(struct rq
*rq
, int cpu
)
6861 /* Don't need to rebalance while attached to NULL domain */
6862 if (time_after_eq(jiffies
, rq
->next_balance
) &&
6863 likely(!on_null_domain(cpu
)))
6864 raise_softirq(SCHED_SOFTIRQ
);
6865 #ifdef CONFIG_NO_HZ_COMMON
6866 if (nohz_kick_needed(rq
, cpu
) && likely(!on_null_domain(cpu
)))
6867 nohz_balancer_kick(cpu
);
6871 static void rq_online_fair(struct rq
*rq
)
6876 static void rq_offline_fair(struct rq
*rq
)
6880 /* Ensure any throttled groups are reachable by pick_next_task */
6881 unthrottle_offline_cfs_rqs(rq
);
6884 #endif /* CONFIG_SMP */
6887 * scheduler tick hitting a task of our scheduling class:
6889 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
6891 struct cfs_rq
*cfs_rq
;
6892 struct sched_entity
*se
= &curr
->se
;
6894 for_each_sched_entity(se
) {
6895 cfs_rq
= cfs_rq_of(se
);
6896 entity_tick(cfs_rq
, se
, queued
);
6899 if (numabalancing_enabled
)
6900 task_tick_numa(rq
, curr
);
6902 update_rq_runnable_avg(rq
, 1);
6906 * called on fork with the child task as argument from the parent's context
6907 * - child not yet on the tasklist
6908 * - preemption disabled
6910 static void task_fork_fair(struct task_struct
*p
)
6912 struct cfs_rq
*cfs_rq
;
6913 struct sched_entity
*se
= &p
->se
, *curr
;
6914 int this_cpu
= smp_processor_id();
6915 struct rq
*rq
= this_rq();
6916 unsigned long flags
;
6918 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6920 update_rq_clock(rq
);
6922 cfs_rq
= task_cfs_rq(current
);
6923 curr
= cfs_rq
->curr
;
6926 * Not only the cpu but also the task_group of the parent might have
6927 * been changed after parent->se.parent,cfs_rq were copied to
6928 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6929 * of child point to valid ones.
6932 __set_task_cpu(p
, this_cpu
);
6935 update_curr(cfs_rq
);
6938 se
->vruntime
= curr
->vruntime
;
6939 place_entity(cfs_rq
, se
, 1);
6941 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
6943 * Upon rescheduling, sched_class::put_prev_task() will place
6944 * 'current' within the tree based on its new key value.
6946 swap(curr
->vruntime
, se
->vruntime
);
6947 resched_task(rq
->curr
);
6950 se
->vruntime
-= cfs_rq
->min_vruntime
;
6952 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6956 * Priority of the task has changed. Check to see if we preempt
6960 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
6966 * Reschedule if we are currently running on this runqueue and
6967 * our priority decreased, or if we are not currently running on
6968 * this runqueue and our priority is higher than the current's
6970 if (rq
->curr
== p
) {
6971 if (p
->prio
> oldprio
)
6972 resched_task(rq
->curr
);
6974 check_preempt_curr(rq
, p
, 0);
6977 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
6979 struct sched_entity
*se
= &p
->se
;
6980 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
6983 * Ensure the task's vruntime is normalized, so that when its
6984 * switched back to the fair class the enqueue_entity(.flags=0) will
6985 * do the right thing.
6987 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6988 * have normalized the vruntime, if it was !on_rq, then only when
6989 * the task is sleeping will it still have non-normalized vruntime.
6991 if (!se
->on_rq
&& p
->state
!= TASK_RUNNING
) {
6993 * Fix up our vruntime so that the current sleep doesn't
6994 * cause 'unlimited' sleep bonus.
6996 place_entity(cfs_rq
, se
, 0);
6997 se
->vruntime
-= cfs_rq
->min_vruntime
;
7002 * Remove our load from contribution when we leave sched_fair
7003 * and ensure we don't carry in an old decay_count if we
7006 if (se
->avg
.decay_count
) {
7007 __synchronize_entity_decay(se
);
7008 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
7014 * We switched to the sched_fair class.
7016 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
7022 * We were most likely switched from sched_rt, so
7023 * kick off the schedule if running, otherwise just see
7024 * if we can still preempt the current task.
7027 resched_task(rq
->curr
);
7029 check_preempt_curr(rq
, p
, 0);
7032 /* Account for a task changing its policy or group.
7034 * This routine is mostly called to set cfs_rq->curr field when a task
7035 * migrates between groups/classes.
7037 static void set_curr_task_fair(struct rq
*rq
)
7039 struct sched_entity
*se
= &rq
->curr
->se
;
7041 for_each_sched_entity(se
) {
7042 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7044 set_next_entity(cfs_rq
, se
);
7045 /* ensure bandwidth has been allocated on our new cfs_rq */
7046 account_cfs_rq_runtime(cfs_rq
, 0);
7050 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
7052 cfs_rq
->tasks_timeline
= RB_ROOT
;
7053 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7054 #ifndef CONFIG_64BIT
7055 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
7058 atomic64_set(&cfs_rq
->decay_counter
, 1);
7059 atomic_long_set(&cfs_rq
->removed_load
, 0);
7063 #ifdef CONFIG_FAIR_GROUP_SCHED
7064 static void task_move_group_fair(struct task_struct
*p
, int on_rq
)
7066 struct cfs_rq
*cfs_rq
;
7068 * If the task was not on the rq at the time of this cgroup movement
7069 * it must have been asleep, sleeping tasks keep their ->vruntime
7070 * absolute on their old rq until wakeup (needed for the fair sleeper
7071 * bonus in place_entity()).
7073 * If it was on the rq, we've just 'preempted' it, which does convert
7074 * ->vruntime to a relative base.
7076 * Make sure both cases convert their relative position when migrating
7077 * to another cgroup's rq. This does somewhat interfere with the
7078 * fair sleeper stuff for the first placement, but who cares.
7081 * When !on_rq, vruntime of the task has usually NOT been normalized.
7082 * But there are some cases where it has already been normalized:
7084 * - Moving a forked child which is waiting for being woken up by
7085 * wake_up_new_task().
7086 * - Moving a task which has been woken up by try_to_wake_up() and
7087 * waiting for actually being woken up by sched_ttwu_pending().
7089 * To prevent boost or penalty in the new cfs_rq caused by delta
7090 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7092 if (!on_rq
&& (!p
->se
.sum_exec_runtime
|| p
->state
== TASK_WAKING
))
7096 p
->se
.vruntime
-= cfs_rq_of(&p
->se
)->min_vruntime
;
7097 set_task_rq(p
, task_cpu(p
));
7099 cfs_rq
= cfs_rq_of(&p
->se
);
7100 p
->se
.vruntime
+= cfs_rq
->min_vruntime
;
7103 * migrate_task_rq_fair() will have removed our previous
7104 * contribution, but we must synchronize for ongoing future
7107 p
->se
.avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
7108 cfs_rq
->blocked_load_avg
+= p
->se
.avg
.load_avg_contrib
;
7113 void free_fair_sched_group(struct task_group
*tg
)
7117 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7119 for_each_possible_cpu(i
) {
7121 kfree(tg
->cfs_rq
[i
]);
7130 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7132 struct cfs_rq
*cfs_rq
;
7133 struct sched_entity
*se
;
7136 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
7139 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
7143 tg
->shares
= NICE_0_LOAD
;
7145 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7147 for_each_possible_cpu(i
) {
7148 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
7149 GFP_KERNEL
, cpu_to_node(i
));
7153 se
= kzalloc_node(sizeof(struct sched_entity
),
7154 GFP_KERNEL
, cpu_to_node(i
));
7158 init_cfs_rq(cfs_rq
);
7159 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
7170 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
7172 struct rq
*rq
= cpu_rq(cpu
);
7173 unsigned long flags
;
7176 * Only empty task groups can be destroyed; so we can speculatively
7177 * check on_list without danger of it being re-added.
7179 if (!tg
->cfs_rq
[cpu
]->on_list
)
7182 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7183 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
7184 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7187 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7188 struct sched_entity
*se
, int cpu
,
7189 struct sched_entity
*parent
)
7191 struct rq
*rq
= cpu_rq(cpu
);
7195 init_cfs_rq_runtime(cfs_rq
);
7197 tg
->cfs_rq
[cpu
] = cfs_rq
;
7200 /* se could be NULL for root_task_group */
7205 se
->cfs_rq
= &rq
->cfs
;
7207 se
->cfs_rq
= parent
->my_q
;
7210 update_load_set(&se
->load
, 0);
7211 se
->parent
= parent
;
7214 static DEFINE_MUTEX(shares_mutex
);
7216 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
7219 unsigned long flags
;
7222 * We can't change the weight of the root cgroup.
7227 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
7229 mutex_lock(&shares_mutex
);
7230 if (tg
->shares
== shares
)
7233 tg
->shares
= shares
;
7234 for_each_possible_cpu(i
) {
7235 struct rq
*rq
= cpu_rq(i
);
7236 struct sched_entity
*se
;
7239 /* Propagate contribution to hierarchy */
7240 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7242 /* Possible calls to update_curr() need rq clock */
7243 update_rq_clock(rq
);
7244 for_each_sched_entity(se
)
7245 update_cfs_shares(group_cfs_rq(se
));
7246 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7250 mutex_unlock(&shares_mutex
);
7253 #else /* CONFIG_FAIR_GROUP_SCHED */
7255 void free_fair_sched_group(struct task_group
*tg
) { }
7257 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7262 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
) { }
7264 #endif /* CONFIG_FAIR_GROUP_SCHED */
7267 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
7269 struct sched_entity
*se
= &task
->se
;
7270 unsigned int rr_interval
= 0;
7273 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7276 if (rq
->cfs
.load
.weight
)
7277 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
7283 * All the scheduling class methods:
7285 const struct sched_class fair_sched_class
= {
7286 .next
= &idle_sched_class
,
7287 .enqueue_task
= enqueue_task_fair
,
7288 .dequeue_task
= dequeue_task_fair
,
7289 .yield_task
= yield_task_fair
,
7290 .yield_to_task
= yield_to_task_fair
,
7292 .check_preempt_curr
= check_preempt_wakeup
,
7294 .pick_next_task
= pick_next_task_fair
,
7295 .put_prev_task
= put_prev_task_fair
,
7298 .select_task_rq
= select_task_rq_fair
,
7299 .migrate_task_rq
= migrate_task_rq_fair
,
7301 .rq_online
= rq_online_fair
,
7302 .rq_offline
= rq_offline_fair
,
7304 .task_waking
= task_waking_fair
,
7307 .set_curr_task
= set_curr_task_fair
,
7308 .task_tick
= task_tick_fair
,
7309 .task_fork
= task_fork_fair
,
7311 .prio_changed
= prio_changed_fair
,
7312 .switched_from
= switched_from_fair
,
7313 .switched_to
= switched_to_fair
,
7315 .get_rr_interval
= get_rr_interval_fair
,
7317 #ifdef CONFIG_FAIR_GROUP_SCHED
7318 .task_move_group
= task_move_group_fair
,
7322 #ifdef CONFIG_SCHED_DEBUG
7323 void print_cfs_stats(struct seq_file
*m
, int cpu
)
7325 struct cfs_rq
*cfs_rq
;
7328 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
7329 print_cfs_rq(m
, cpu
, cfs_rq
);
7334 __init
void init_sched_fair_class(void)
7337 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7339 #ifdef CONFIG_NO_HZ_COMMON
7340 nohz
.next_balance
= jiffies
;
7341 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
7342 cpu_notifier(sched_ilb_notifier
, 0);