2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
38 * Targeted preemption latency for CPU-bound tasks:
39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41 * NOTE: this latency value is not the same as the concept of
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
49 unsigned int sysctl_sched_latency
= 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency
= 6000000ULL;
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG
;
65 * Minimal preemption granularity for CPU-bound tasks:
66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 unsigned int sysctl_sched_min_granularity
= 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity
= 750000ULL;
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74 static unsigned int sched_nr_latency
= 8;
77 * After fork, child runs first. If set to 0 (default) then
78 * parent will (try to) run first.
80 unsigned int sysctl_sched_child_runs_first __read_mostly
;
83 * SCHED_OTHER wake-up granularity.
84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
90 unsigned int sysctl_sched_wakeup_granularity
= 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity
= 1000000UL;
93 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
96 * The exponential sliding window over which load is averaged for shares
100 unsigned int __read_mostly sysctl_sched_shares_window
= 10000000UL;
102 #ifdef CONFIG_CFS_BANDWIDTH
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
111 * default: 5 msec, units: microseconds
113 unsigned int sysctl_sched_cfs_bandwidth_slice
= 5000UL;
116 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
122 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
128 static inline void update_load_set(struct load_weight
*lw
, unsigned long w
)
135 * Increase the granularity value when there are more CPUs,
136 * because with more CPUs the 'effective latency' as visible
137 * to users decreases. But the relationship is not linear,
138 * so pick a second-best guess by going with the log2 of the
141 * This idea comes from the SD scheduler of Con Kolivas:
143 static int get_update_sysctl_factor(void)
145 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
148 switch (sysctl_sched_tunable_scaling
) {
149 case SCHED_TUNABLESCALING_NONE
:
152 case SCHED_TUNABLESCALING_LINEAR
:
155 case SCHED_TUNABLESCALING_LOG
:
157 factor
= 1 + ilog2(cpus
);
164 static void update_sysctl(void)
166 unsigned int factor
= get_update_sysctl_factor();
168 #define SET_SYSCTL(name) \
169 (sysctl_##name = (factor) * normalized_sysctl_##name)
170 SET_SYSCTL(sched_min_granularity
);
171 SET_SYSCTL(sched_latency
);
172 SET_SYSCTL(sched_wakeup_granularity
);
176 void sched_init_granularity(void)
181 #if BITS_PER_LONG == 32
182 # define WMULT_CONST (~0UL)
184 # define WMULT_CONST (1UL << 32)
187 #define WMULT_SHIFT 32
190 * Shift right and round:
192 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
195 * delta *= weight / lw
198 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
199 struct load_weight
*lw
)
204 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
205 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
206 * 2^SCHED_LOAD_RESOLUTION.
208 if (likely(weight
> (1UL << SCHED_LOAD_RESOLUTION
)))
209 tmp
= (u64
)delta_exec
* scale_load_down(weight
);
211 tmp
= (u64
)delta_exec
;
213 if (!lw
->inv_weight
) {
214 unsigned long w
= scale_load_down(lw
->weight
);
216 if (BITS_PER_LONG
> 32 && unlikely(w
>= WMULT_CONST
))
218 else if (unlikely(!w
))
219 lw
->inv_weight
= WMULT_CONST
;
221 lw
->inv_weight
= WMULT_CONST
/ w
;
225 * Check whether we'd overflow the 64-bit multiplication:
227 if (unlikely(tmp
> WMULT_CONST
))
228 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
231 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
233 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
237 const struct sched_class fair_sched_class
;
239 /**************************************************************
240 * CFS operations on generic schedulable entities:
243 #ifdef CONFIG_FAIR_GROUP_SCHED
245 /* cpu runqueue to which this cfs_rq is attached */
246 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
251 /* An entity is a task if it doesn't "own" a runqueue */
252 #define entity_is_task(se) (!se->my_q)
254 static inline struct task_struct
*task_of(struct sched_entity
*se
)
256 #ifdef CONFIG_SCHED_DEBUG
257 WARN_ON_ONCE(!entity_is_task(se
));
259 return container_of(se
, struct task_struct
, se
);
262 /* Walk up scheduling entities hierarchy */
263 #define for_each_sched_entity(se) \
264 for (; se; se = se->parent)
266 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
271 /* runqueue on which this entity is (to be) queued */
272 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
277 /* runqueue "owned" by this group */
278 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
283 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
288 if (!cfs_rq
->on_list
) {
290 * Ensure we either appear before our parent (if already
291 * enqueued) or force our parent to appear after us when it is
292 * enqueued. The fact that we always enqueue bottom-up
293 * reduces this to two cases.
295 if (cfs_rq
->tg
->parent
&&
296 cfs_rq
->tg
->parent
->cfs_rq
[cpu_of(rq_of(cfs_rq
))]->on_list
) {
297 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
,
298 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
300 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
301 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
305 /* We should have no load, but we need to update last_decay. */
306 update_cfs_rq_blocked_load(cfs_rq
, 0);
310 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
312 if (cfs_rq
->on_list
) {
313 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
318 /* Iterate thr' all leaf cfs_rq's on a runqueue */
319 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
320 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
322 /* Do the two (enqueued) entities belong to the same group ? */
324 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
326 if (se
->cfs_rq
== pse
->cfs_rq
)
332 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
337 /* return depth at which a sched entity is present in the hierarchy */
338 static inline int depth_se(struct sched_entity
*se
)
342 for_each_sched_entity(se
)
349 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
351 int se_depth
, pse_depth
;
354 * preemption test can be made between sibling entities who are in the
355 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
356 * both tasks until we find their ancestors who are siblings of common
360 /* First walk up until both entities are at same depth */
361 se_depth
= depth_se(*se
);
362 pse_depth
= depth_se(*pse
);
364 while (se_depth
> pse_depth
) {
366 *se
= parent_entity(*se
);
369 while (pse_depth
> se_depth
) {
371 *pse
= parent_entity(*pse
);
374 while (!is_same_group(*se
, *pse
)) {
375 *se
= parent_entity(*se
);
376 *pse
= parent_entity(*pse
);
380 #else /* !CONFIG_FAIR_GROUP_SCHED */
382 static inline struct task_struct
*task_of(struct sched_entity
*se
)
384 return container_of(se
, struct task_struct
, se
);
387 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
389 return container_of(cfs_rq
, struct rq
, cfs
);
392 #define entity_is_task(se) 1
394 #define for_each_sched_entity(se) \
395 for (; se; se = NULL)
397 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
399 return &task_rq(p
)->cfs
;
402 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
404 struct task_struct
*p
= task_of(se
);
405 struct rq
*rq
= task_rq(p
);
410 /* runqueue "owned" by this group */
411 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
416 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
420 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
424 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
425 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
428 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
433 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
439 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
443 #endif /* CONFIG_FAIR_GROUP_SCHED */
445 static __always_inline
446 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, unsigned long delta_exec
);
448 /**************************************************************
449 * Scheduling class tree data structure manipulation methods:
452 static inline u64
max_vruntime(u64 max_vruntime
, u64 vruntime
)
454 s64 delta
= (s64
)(vruntime
- max_vruntime
);
456 max_vruntime
= vruntime
;
461 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
463 s64 delta
= (s64
)(vruntime
- min_vruntime
);
465 min_vruntime
= vruntime
;
470 static inline int entity_before(struct sched_entity
*a
,
471 struct sched_entity
*b
)
473 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
476 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
478 u64 vruntime
= cfs_rq
->min_vruntime
;
481 vruntime
= cfs_rq
->curr
->vruntime
;
483 if (cfs_rq
->rb_leftmost
) {
484 struct sched_entity
*se
= rb_entry(cfs_rq
->rb_leftmost
,
489 vruntime
= se
->vruntime
;
491 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
494 /* ensure we never gain time by being placed backwards. */
495 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
498 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
503 * Enqueue an entity into the rb-tree:
505 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
507 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
508 struct rb_node
*parent
= NULL
;
509 struct sched_entity
*entry
;
513 * Find the right place in the rbtree:
517 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
519 * We dont care about collisions. Nodes with
520 * the same key stay together.
522 if (entity_before(se
, entry
)) {
523 link
= &parent
->rb_left
;
525 link
= &parent
->rb_right
;
531 * Maintain a cache of leftmost tree entries (it is frequently
535 cfs_rq
->rb_leftmost
= &se
->run_node
;
537 rb_link_node(&se
->run_node
, parent
, link
);
538 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
541 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
543 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
544 struct rb_node
*next_node
;
546 next_node
= rb_next(&se
->run_node
);
547 cfs_rq
->rb_leftmost
= next_node
;
550 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
553 struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
)
555 struct rb_node
*left
= cfs_rq
->rb_leftmost
;
560 return rb_entry(left
, struct sched_entity
, run_node
);
563 static struct sched_entity
*__pick_next_entity(struct sched_entity
*se
)
565 struct rb_node
*next
= rb_next(&se
->run_node
);
570 return rb_entry(next
, struct sched_entity
, run_node
);
573 #ifdef CONFIG_SCHED_DEBUG
574 struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
576 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
581 return rb_entry(last
, struct sched_entity
, run_node
);
584 /**************************************************************
585 * Scheduling class statistics methods:
588 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
589 void __user
*buffer
, size_t *lenp
,
592 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
593 int factor
= get_update_sysctl_factor();
598 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
599 sysctl_sched_min_granularity
);
601 #define WRT_SYSCTL(name) \
602 (normalized_sysctl_##name = sysctl_##name / (factor))
603 WRT_SYSCTL(sched_min_granularity
);
604 WRT_SYSCTL(sched_latency
);
605 WRT_SYSCTL(sched_wakeup_granularity
);
615 static inline unsigned long
616 calc_delta_fair(unsigned long delta
, struct sched_entity
*se
)
618 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
619 delta
= calc_delta_mine(delta
, NICE_0_LOAD
, &se
->load
);
625 * The idea is to set a period in which each task runs once.
627 * When there are too many tasks (sched_nr_latency) we have to stretch
628 * this period because otherwise the slices get too small.
630 * p = (nr <= nl) ? l : l*nr/nl
632 static u64
__sched_period(unsigned long nr_running
)
634 u64 period
= sysctl_sched_latency
;
635 unsigned long nr_latency
= sched_nr_latency
;
637 if (unlikely(nr_running
> nr_latency
)) {
638 period
= sysctl_sched_min_granularity
;
639 period
*= nr_running
;
646 * We calculate the wall-time slice from the period by taking a part
647 * proportional to the weight.
651 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
653 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
655 for_each_sched_entity(se
) {
656 struct load_weight
*load
;
657 struct load_weight lw
;
659 cfs_rq
= cfs_rq_of(se
);
660 load
= &cfs_rq
->load
;
662 if (unlikely(!se
->on_rq
)) {
665 update_load_add(&lw
, se
->load
.weight
);
668 slice
= calc_delta_mine(slice
, se
->load
.weight
, load
);
674 * We calculate the vruntime slice of a to-be-inserted task.
678 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
680 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
684 static unsigned long task_h_load(struct task_struct
*p
);
686 static inline void __update_task_entity_contrib(struct sched_entity
*se
);
688 /* Give new task start runnable values to heavy its load in infant time */
689 void init_task_runnable_average(struct task_struct
*p
)
693 p
->se
.avg
.decay_count
= 0;
694 slice
= sched_slice(task_cfs_rq(p
), &p
->se
) >> 10;
695 p
->se
.avg
.runnable_avg_sum
= slice
;
696 p
->se
.avg
.runnable_avg_period
= slice
;
697 __update_task_entity_contrib(&p
->se
);
700 void init_task_runnable_average(struct task_struct
*p
)
706 * Update the current task's runtime statistics. Skip current tasks that
707 * are not in our scheduling class.
710 __update_curr(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
,
711 unsigned long delta_exec
)
713 unsigned long delta_exec_weighted
;
715 schedstat_set(curr
->statistics
.exec_max
,
716 max((u64
)delta_exec
, curr
->statistics
.exec_max
));
718 curr
->sum_exec_runtime
+= delta_exec
;
719 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
720 delta_exec_weighted
= calc_delta_fair(delta_exec
, curr
);
722 curr
->vruntime
+= delta_exec_weighted
;
723 update_min_vruntime(cfs_rq
);
726 static void update_curr(struct cfs_rq
*cfs_rq
)
728 struct sched_entity
*curr
= cfs_rq
->curr
;
729 u64 now
= rq_clock_task(rq_of(cfs_rq
));
730 unsigned long delta_exec
;
736 * Get the amount of time the current task was running
737 * since the last time we changed load (this cannot
738 * overflow on 32 bits):
740 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
744 __update_curr(cfs_rq
, curr
, delta_exec
);
745 curr
->exec_start
= now
;
747 if (entity_is_task(curr
)) {
748 struct task_struct
*curtask
= task_of(curr
);
750 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
751 cpuacct_charge(curtask
, delta_exec
);
752 account_group_exec_runtime(curtask
, delta_exec
);
755 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
759 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
761 schedstat_set(se
->statistics
.wait_start
, rq_clock(rq_of(cfs_rq
)));
765 * Task is being enqueued - update stats:
767 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
770 * Are we enqueueing a waiting task? (for current tasks
771 * a dequeue/enqueue event is a NOP)
773 if (se
!= cfs_rq
->curr
)
774 update_stats_wait_start(cfs_rq
, se
);
778 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
780 schedstat_set(se
->statistics
.wait_max
, max(se
->statistics
.wait_max
,
781 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
));
782 schedstat_set(se
->statistics
.wait_count
, se
->statistics
.wait_count
+ 1);
783 schedstat_set(se
->statistics
.wait_sum
, se
->statistics
.wait_sum
+
784 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
785 #ifdef CONFIG_SCHEDSTATS
786 if (entity_is_task(se
)) {
787 trace_sched_stat_wait(task_of(se
),
788 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
791 schedstat_set(se
->statistics
.wait_start
, 0);
795 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
798 * Mark the end of the wait period if dequeueing a
801 if (se
!= cfs_rq
->curr
)
802 update_stats_wait_end(cfs_rq
, se
);
806 * We are picking a new current task - update its stats:
809 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
812 * We are starting a new run period:
814 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
817 /**************************************************
818 * Scheduling class queueing methods:
821 #ifdef CONFIG_NUMA_BALANCING
823 * Approximate time to scan a full NUMA task in ms. The task scan period is
824 * calculated based on the tasks virtual memory size and
825 * numa_balancing_scan_size.
827 unsigned int sysctl_numa_balancing_scan_period_min
= 1000;
828 unsigned int sysctl_numa_balancing_scan_period_max
= 60000;
829 unsigned int sysctl_numa_balancing_scan_period_reset
= 60000;
831 /* Portion of address space to scan in MB */
832 unsigned int sysctl_numa_balancing_scan_size
= 256;
834 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
835 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
837 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
839 unsigned long rss
= 0;
840 unsigned long nr_scan_pages
;
843 * Calculations based on RSS as non-present and empty pages are skipped
844 * by the PTE scanner and NUMA hinting faults should be trapped based
847 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
848 rss
= get_mm_rss(p
->mm
);
852 rss
= round_up(rss
, nr_scan_pages
);
853 return rss
/ nr_scan_pages
;
856 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
857 #define MAX_SCAN_WINDOW 2560
859 static unsigned int task_scan_min(struct task_struct
*p
)
861 unsigned int scan
, floor
;
862 unsigned int windows
= 1;
864 if (sysctl_numa_balancing_scan_size
< MAX_SCAN_WINDOW
)
865 windows
= MAX_SCAN_WINDOW
/ sysctl_numa_balancing_scan_size
;
866 floor
= 1000 / windows
;
868 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
869 return max_t(unsigned int, floor
, scan
);
872 static unsigned int task_scan_max(struct task_struct
*p
)
874 unsigned int smin
= task_scan_min(p
);
877 /* Watch for min being lower than max due to floor calculations */
878 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
879 return max(smin
, smax
);
883 * Once a preferred node is selected the scheduler balancer will prefer moving
884 * a task to that node for sysctl_numa_balancing_settle_count number of PTE
885 * scans. This will give the process the chance to accumulate more faults on
886 * the preferred node but still allow the scheduler to move the task again if
887 * the nodes CPUs are overloaded.
889 unsigned int sysctl_numa_balancing_settle_count __read_mostly
= 4;
891 static void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
893 rq
->nr_numa_running
+= (p
->numa_preferred_nid
!= -1);
894 rq
->nr_preferred_running
+= (p
->numa_preferred_nid
== task_node(p
));
897 static void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
899 rq
->nr_numa_running
-= (p
->numa_preferred_nid
!= -1);
900 rq
->nr_preferred_running
-= (p
->numa_preferred_nid
== task_node(p
));
906 spinlock_t lock
; /* nr_tasks, tasks */
909 struct list_head task_list
;
912 atomic_long_t total_faults
;
913 atomic_long_t faults
[0];
916 pid_t
task_numa_group_id(struct task_struct
*p
)
918 return p
->numa_group
? p
->numa_group
->gid
: 0;
921 static inline int task_faults_idx(int nid
, int priv
)
923 return 2 * nid
+ priv
;
926 static inline unsigned long task_faults(struct task_struct
*p
, int nid
)
931 return p
->numa_faults
[task_faults_idx(nid
, 0)] +
932 p
->numa_faults
[task_faults_idx(nid
, 1)];
935 static inline unsigned long group_faults(struct task_struct
*p
, int nid
)
940 return atomic_long_read(&p
->numa_group
->faults
[2*nid
]) +
941 atomic_long_read(&p
->numa_group
->faults
[2*nid
+1]);
945 * These return the fraction of accesses done by a particular task, or
946 * task group, on a particular numa node. The group weight is given a
947 * larger multiplier, in order to group tasks together that are almost
948 * evenly spread out between numa nodes.
950 static inline unsigned long task_weight(struct task_struct
*p
, int nid
)
952 unsigned long total_faults
;
957 total_faults
= p
->total_numa_faults
;
962 return 1000 * task_faults(p
, nid
) / total_faults
;
965 static inline unsigned long group_weight(struct task_struct
*p
, int nid
)
967 unsigned long total_faults
;
972 total_faults
= atomic_long_read(&p
->numa_group
->total_faults
);
977 return 1000 * group_faults(p
, nid
) / total_faults
;
980 static unsigned long weighted_cpuload(const int cpu
);
981 static unsigned long source_load(int cpu
, int type
);
982 static unsigned long target_load(int cpu
, int type
);
983 static unsigned long power_of(int cpu
);
984 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
);
986 /* Cached statistics for all CPUs within a node */
988 unsigned long nr_running
;
991 /* Total compute capacity of CPUs on a node */
994 /* Approximate capacity in terms of runnable tasks on a node */
995 unsigned long capacity
;
1000 * XXX borrowed from update_sg_lb_stats
1002 static void update_numa_stats(struct numa_stats
*ns
, int nid
)
1006 memset(ns
, 0, sizeof(*ns
));
1007 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
1008 struct rq
*rq
= cpu_rq(cpu
);
1010 ns
->nr_running
+= rq
->nr_running
;
1011 ns
->load
+= weighted_cpuload(cpu
);
1012 ns
->power
+= power_of(cpu
);
1015 ns
->load
= (ns
->load
* SCHED_POWER_SCALE
) / ns
->power
;
1016 ns
->capacity
= DIV_ROUND_CLOSEST(ns
->power
, SCHED_POWER_SCALE
);
1017 ns
->has_capacity
= (ns
->nr_running
< ns
->capacity
);
1020 struct task_numa_env
{
1021 struct task_struct
*p
;
1023 int src_cpu
, src_nid
;
1024 int dst_cpu
, dst_nid
;
1026 struct numa_stats src_stats
, dst_stats
;
1028 int imbalance_pct
, idx
;
1030 struct task_struct
*best_task
;
1035 static void task_numa_assign(struct task_numa_env
*env
,
1036 struct task_struct
*p
, long imp
)
1039 put_task_struct(env
->best_task
);
1044 env
->best_imp
= imp
;
1045 env
->best_cpu
= env
->dst_cpu
;
1049 * This checks if the overall compute and NUMA accesses of the system would
1050 * be improved if the source tasks was migrated to the target dst_cpu taking
1051 * into account that it might be best if task running on the dst_cpu should
1052 * be exchanged with the source task
1054 static void task_numa_compare(struct task_numa_env
*env
,
1055 long taskimp
, long groupimp
)
1057 struct rq
*src_rq
= cpu_rq(env
->src_cpu
);
1058 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1059 struct task_struct
*cur
;
1060 long dst_load
, src_load
;
1062 long imp
= (groupimp
> 0) ? groupimp
: taskimp
;
1065 cur
= ACCESS_ONCE(dst_rq
->curr
);
1066 if (cur
->pid
== 0) /* idle */
1070 * "imp" is the fault differential for the source task between the
1071 * source and destination node. Calculate the total differential for
1072 * the source task and potential destination task. The more negative
1073 * the value is, the more rmeote accesses that would be expected to
1074 * be incurred if the tasks were swapped.
1077 /* Skip this swap candidate if cannot move to the source cpu */
1078 if (!cpumask_test_cpu(env
->src_cpu
, tsk_cpus_allowed(cur
)))
1082 * If dst and source tasks are in the same NUMA group, or not
1083 * in any group then look only at task weights.
1085 if (cur
->numa_group
== env
->p
->numa_group
) {
1086 imp
= taskimp
+ task_weight(cur
, env
->src_nid
) -
1087 task_weight(cur
, env
->dst_nid
);
1089 * Add some hysteresis to prevent swapping the
1090 * tasks within a group over tiny differences.
1092 if (cur
->numa_group
)
1096 * Compare the group weights. If a task is all by
1097 * itself (not part of a group), use the task weight
1100 if (env
->p
->numa_group
)
1105 if (cur
->numa_group
)
1106 imp
+= group_weight(cur
, env
->src_nid
) -
1107 group_weight(cur
, env
->dst_nid
);
1109 imp
+= task_weight(cur
, env
->src_nid
) -
1110 task_weight(cur
, env
->dst_nid
);
1114 if (imp
< env
->best_imp
)
1118 /* Is there capacity at our destination? */
1119 if (env
->src_stats
.has_capacity
&&
1120 !env
->dst_stats
.has_capacity
)
1126 /* Balance doesn't matter much if we're running a task per cpu */
1127 if (src_rq
->nr_running
== 1 && dst_rq
->nr_running
== 1)
1131 * In the overloaded case, try and keep the load balanced.
1134 dst_load
= env
->dst_stats
.load
;
1135 src_load
= env
->src_stats
.load
;
1137 /* XXX missing power terms */
1138 load
= task_h_load(env
->p
);
1143 load
= task_h_load(cur
);
1148 /* make src_load the smaller */
1149 if (dst_load
< src_load
)
1150 swap(dst_load
, src_load
);
1152 if (src_load
* env
->imbalance_pct
< dst_load
* 100)
1156 task_numa_assign(env
, cur
, imp
);
1161 static void task_numa_find_cpu(struct task_numa_env
*env
,
1162 long taskimp
, long groupimp
)
1166 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1167 /* Skip this CPU if the source task cannot migrate */
1168 if (!cpumask_test_cpu(cpu
, tsk_cpus_allowed(env
->p
)))
1172 task_numa_compare(env
, taskimp
, groupimp
);
1176 static int task_numa_migrate(struct task_struct
*p
)
1178 struct task_numa_env env
= {
1181 .src_cpu
= task_cpu(p
),
1182 .src_nid
= task_node(p
),
1184 .imbalance_pct
= 112,
1190 struct sched_domain
*sd
;
1191 unsigned long taskweight
, groupweight
;
1193 long taskimp
, groupimp
;
1196 * Pick the lowest SD_NUMA domain, as that would have the smallest
1197 * imbalance and would be the first to start moving tasks about.
1199 * And we want to avoid any moving of tasks about, as that would create
1200 * random movement of tasks -- counter the numa conditions we're trying
1204 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1205 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1208 taskweight
= task_weight(p
, env
.src_nid
);
1209 groupweight
= group_weight(p
, env
.src_nid
);
1210 update_numa_stats(&env
.src_stats
, env
.src_nid
);
1211 env
.dst_nid
= p
->numa_preferred_nid
;
1212 taskimp
= task_weight(p
, env
.dst_nid
) - taskweight
;
1213 groupimp
= group_weight(p
, env
.dst_nid
) - groupweight
;
1214 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1216 /* If the preferred nid has capacity, try to use it. */
1217 if (env
.dst_stats
.has_capacity
)
1218 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1220 /* No space available on the preferred nid. Look elsewhere. */
1221 if (env
.best_cpu
== -1) {
1222 for_each_online_node(nid
) {
1223 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
1226 /* Only consider nodes where both task and groups benefit */
1227 taskimp
= task_weight(p
, nid
) - taskweight
;
1228 groupimp
= group_weight(p
, nid
) - groupweight
;
1229 if (taskimp
< 0 && groupimp
< 0)
1233 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1234 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1238 /* No better CPU than the current one was found. */
1239 if (env
.best_cpu
== -1)
1242 sched_setnuma(p
, env
.dst_nid
);
1245 * Reset the scan period if the task is being rescheduled on an
1246 * alternative node to recheck if the tasks is now properly placed.
1248 p
->numa_scan_period
= task_scan_min(p
);
1250 if (env
.best_task
== NULL
) {
1251 int ret
= migrate_task_to(p
, env
.best_cpu
);
1255 ret
= migrate_swap(p
, env
.best_task
);
1256 put_task_struct(env
.best_task
);
1260 /* Attempt to migrate a task to a CPU on the preferred node. */
1261 static void numa_migrate_preferred(struct task_struct
*p
)
1263 /* Success if task is already running on preferred CPU */
1264 p
->numa_migrate_retry
= 0;
1265 if (cpu_to_node(task_cpu(p
)) == p
->numa_preferred_nid
) {
1267 * If migration is temporarily disabled due to a task migration
1268 * then re-enable it now as the task is running on its
1269 * preferred node and memory should migrate locally
1271 if (!p
->numa_migrate_seq
)
1272 p
->numa_migrate_seq
++;
1276 /* This task has no NUMA fault statistics yet */
1277 if (unlikely(p
->numa_preferred_nid
== -1))
1280 /* Otherwise, try migrate to a CPU on the preferred node */
1281 if (task_numa_migrate(p
) != 0)
1282 p
->numa_migrate_retry
= jiffies
+ HZ
*5;
1286 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1287 * increments. The more local the fault statistics are, the higher the scan
1288 * period will be for the next scan window. If local/remote ratio is below
1289 * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the
1290 * scan period will decrease
1292 #define NUMA_PERIOD_SLOTS 10
1293 #define NUMA_PERIOD_THRESHOLD 3
1296 * Increase the scan period (slow down scanning) if the majority of
1297 * our memory is already on our local node, or if the majority of
1298 * the page accesses are shared with other processes.
1299 * Otherwise, decrease the scan period.
1301 static void update_task_scan_period(struct task_struct
*p
,
1302 unsigned long shared
, unsigned long private)
1304 unsigned int period_slot
;
1308 unsigned long remote
= p
->numa_faults_locality
[0];
1309 unsigned long local
= p
->numa_faults_locality
[1];
1312 * If there were no record hinting faults then either the task is
1313 * completely idle or all activity is areas that are not of interest
1314 * to automatic numa balancing. Scan slower
1316 if (local
+ shared
== 0) {
1317 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1318 p
->numa_scan_period
<< 1);
1320 p
->mm
->numa_next_scan
= jiffies
+
1321 msecs_to_jiffies(p
->numa_scan_period
);
1327 * Prepare to scale scan period relative to the current period.
1328 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1329 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1330 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1332 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
1333 ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
1334 if (ratio
>= NUMA_PERIOD_THRESHOLD
) {
1335 int slot
= ratio
- NUMA_PERIOD_THRESHOLD
;
1338 diff
= slot
* period_slot
;
1340 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
1343 * Scale scan rate increases based on sharing. There is an
1344 * inverse relationship between the degree of sharing and
1345 * the adjustment made to the scanning period. Broadly
1346 * speaking the intent is that there is little point
1347 * scanning faster if shared accesses dominate as it may
1348 * simply bounce migrations uselessly
1350 period_slot
= DIV_ROUND_UP(diff
, NUMA_PERIOD_SLOTS
);
1351 ratio
= DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS
, (private + shared
));
1352 diff
= (diff
* ratio
) / NUMA_PERIOD_SLOTS
;
1355 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
1356 task_scan_min(p
), task_scan_max(p
));
1357 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1360 static void task_numa_placement(struct task_struct
*p
)
1362 int seq
, nid
, max_nid
= -1, max_group_nid
= -1;
1363 unsigned long max_faults
= 0, max_group_faults
= 0;
1364 unsigned long fault_types
[2] = { 0, 0 };
1365 spinlock_t
*group_lock
= NULL
;
1367 seq
= ACCESS_ONCE(p
->mm
->numa_scan_seq
);
1368 if (p
->numa_scan_seq
== seq
)
1370 p
->numa_scan_seq
= seq
;
1371 p
->numa_migrate_seq
++;
1372 p
->numa_scan_period_max
= task_scan_max(p
);
1374 /* If the task is part of a group prevent parallel updates to group stats */
1375 if (p
->numa_group
) {
1376 group_lock
= &p
->numa_group
->lock
;
1377 spin_lock(group_lock
);
1380 /* Find the node with the highest number of faults */
1381 for_each_online_node(nid
) {
1382 unsigned long faults
= 0, group_faults
= 0;
1385 for (priv
= 0; priv
< 2; priv
++) {
1388 i
= task_faults_idx(nid
, priv
);
1389 diff
= -p
->numa_faults
[i
];
1391 /* Decay existing window, copy faults since last scan */
1392 p
->numa_faults
[i
] >>= 1;
1393 p
->numa_faults
[i
] += p
->numa_faults_buffer
[i
];
1394 fault_types
[priv
] += p
->numa_faults_buffer
[i
];
1395 p
->numa_faults_buffer
[i
] = 0;
1397 faults
+= p
->numa_faults
[i
];
1398 diff
+= p
->numa_faults
[i
];
1399 p
->total_numa_faults
+= diff
;
1400 if (p
->numa_group
) {
1401 /* safe because we can only change our own group */
1402 atomic_long_add(diff
, &p
->numa_group
->faults
[i
]);
1403 atomic_long_add(diff
, &p
->numa_group
->total_faults
);
1404 group_faults
+= atomic_long_read(&p
->numa_group
->faults
[i
]);
1408 if (faults
> max_faults
) {
1409 max_faults
= faults
;
1413 if (group_faults
> max_group_faults
) {
1414 max_group_faults
= group_faults
;
1415 max_group_nid
= nid
;
1419 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
1421 if (p
->numa_group
) {
1423 * If the preferred task and group nids are different,
1424 * iterate over the nodes again to find the best place.
1426 if (max_nid
!= max_group_nid
) {
1427 unsigned long weight
, max_weight
= 0;
1429 for_each_online_node(nid
) {
1430 weight
= task_weight(p
, nid
) + group_weight(p
, nid
);
1431 if (weight
> max_weight
) {
1432 max_weight
= weight
;
1438 spin_unlock(group_lock
);
1441 /* Preferred node as the node with the most faults */
1442 if (max_faults
&& max_nid
!= p
->numa_preferred_nid
) {
1443 /* Update the preferred nid and migrate task if possible */
1444 sched_setnuma(p
, max_nid
);
1445 numa_migrate_preferred(p
);
1449 static inline int get_numa_group(struct numa_group
*grp
)
1451 return atomic_inc_not_zero(&grp
->refcount
);
1454 static inline void put_numa_group(struct numa_group
*grp
)
1456 if (atomic_dec_and_test(&grp
->refcount
))
1457 kfree_rcu(grp
, rcu
);
1460 static void double_lock(spinlock_t
*l1
, spinlock_t
*l2
)
1466 spin_lock_nested(l2
, SINGLE_DEPTH_NESTING
);
1469 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
1472 struct numa_group
*grp
, *my_grp
;
1473 struct task_struct
*tsk
;
1475 int cpu
= cpupid_to_cpu(cpupid
);
1478 if (unlikely(!p
->numa_group
)) {
1479 unsigned int size
= sizeof(struct numa_group
) +
1480 2*nr_node_ids
*sizeof(atomic_long_t
);
1482 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
1486 atomic_set(&grp
->refcount
, 1);
1487 spin_lock_init(&grp
->lock
);
1488 INIT_LIST_HEAD(&grp
->task_list
);
1491 for (i
= 0; i
< 2*nr_node_ids
; i
++)
1492 atomic_long_set(&grp
->faults
[i
], p
->numa_faults
[i
]);
1494 atomic_long_set(&grp
->total_faults
, p
->total_numa_faults
);
1496 list_add(&p
->numa_entry
, &grp
->task_list
);
1498 rcu_assign_pointer(p
->numa_group
, grp
);
1502 tsk
= ACCESS_ONCE(cpu_rq(cpu
)->curr
);
1504 if (!cpupid_match_pid(tsk
, cpupid
))
1507 grp
= rcu_dereference(tsk
->numa_group
);
1511 my_grp
= p
->numa_group
;
1516 * Only join the other group if its bigger; if we're the bigger group,
1517 * the other task will join us.
1519 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
1523 * Tie-break on the grp address.
1525 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
1528 /* Always join threads in the same process. */
1529 if (tsk
->mm
== current
->mm
)
1532 /* Simple filter to avoid false positives due to PID collisions */
1533 if (flags
& TNF_SHARED
)
1536 /* Update priv based on whether false sharing was detected */
1539 if (join
&& !get_numa_group(grp
))
1548 for (i
= 0; i
< 2*nr_node_ids
; i
++) {
1549 atomic_long_sub(p
->numa_faults
[i
], &my_grp
->faults
[i
]);
1550 atomic_long_add(p
->numa_faults
[i
], &grp
->faults
[i
]);
1552 atomic_long_sub(p
->total_numa_faults
, &my_grp
->total_faults
);
1553 atomic_long_add(p
->total_numa_faults
, &grp
->total_faults
);
1555 double_lock(&my_grp
->lock
, &grp
->lock
);
1557 list_move(&p
->numa_entry
, &grp
->task_list
);
1561 spin_unlock(&my_grp
->lock
);
1562 spin_unlock(&grp
->lock
);
1564 rcu_assign_pointer(p
->numa_group
, grp
);
1566 put_numa_group(my_grp
);
1569 void task_numa_free(struct task_struct
*p
)
1571 struct numa_group
*grp
= p
->numa_group
;
1573 void *numa_faults
= p
->numa_faults
;
1576 for (i
= 0; i
< 2*nr_node_ids
; i
++)
1577 atomic_long_sub(p
->numa_faults
[i
], &grp
->faults
[i
]);
1579 atomic_long_sub(p
->total_numa_faults
, &grp
->total_faults
);
1581 spin_lock(&grp
->lock
);
1582 list_del(&p
->numa_entry
);
1584 spin_unlock(&grp
->lock
);
1585 rcu_assign_pointer(p
->numa_group
, NULL
);
1586 put_numa_group(grp
);
1589 p
->numa_faults
= NULL
;
1590 p
->numa_faults_buffer
= NULL
;
1595 * Got a PROT_NONE fault for a page on @node.
1597 void task_numa_fault(int last_cpupid
, int node
, int pages
, int flags
)
1599 struct task_struct
*p
= current
;
1600 bool migrated
= flags
& TNF_MIGRATED
;
1603 if (!numabalancing_enabled
)
1606 /* for example, ksmd faulting in a user's mm */
1610 /* Do not worry about placement if exiting */
1611 if (p
->state
== TASK_DEAD
)
1614 /* Allocate buffer to track faults on a per-node basis */
1615 if (unlikely(!p
->numa_faults
)) {
1616 int size
= sizeof(*p
->numa_faults
) * 2 * nr_node_ids
;
1618 /* numa_faults and numa_faults_buffer share the allocation */
1619 p
->numa_faults
= kzalloc(size
* 2, GFP_KERNEL
|__GFP_NOWARN
);
1620 if (!p
->numa_faults
)
1623 BUG_ON(p
->numa_faults_buffer
);
1624 p
->numa_faults_buffer
= p
->numa_faults
+ (2 * nr_node_ids
);
1625 p
->total_numa_faults
= 0;
1626 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1630 * First accesses are treated as private, otherwise consider accesses
1631 * to be private if the accessing pid has not changed
1633 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
1636 priv
= cpupid_match_pid(p
, last_cpupid
);
1637 if (!priv
&& !(flags
& TNF_NO_GROUP
))
1638 task_numa_group(p
, last_cpupid
, flags
, &priv
);
1641 task_numa_placement(p
);
1643 /* Retry task to preferred node migration if it previously failed */
1644 if (p
->numa_migrate_retry
&& time_after(jiffies
, p
->numa_migrate_retry
))
1645 numa_migrate_preferred(p
);
1648 p
->numa_pages_migrated
+= pages
;
1650 p
->numa_faults_buffer
[task_faults_idx(node
, priv
)] += pages
;
1651 p
->numa_faults_locality
[!!(flags
& TNF_FAULT_LOCAL
)] += pages
;
1654 static void reset_ptenuma_scan(struct task_struct
*p
)
1656 ACCESS_ONCE(p
->mm
->numa_scan_seq
)++;
1657 p
->mm
->numa_scan_offset
= 0;
1661 * The expensive part of numa migration is done from task_work context.
1662 * Triggered from task_tick_numa().
1664 void task_numa_work(struct callback_head
*work
)
1666 unsigned long migrate
, next_scan
, now
= jiffies
;
1667 struct task_struct
*p
= current
;
1668 struct mm_struct
*mm
= p
->mm
;
1669 struct vm_area_struct
*vma
;
1670 unsigned long start
, end
;
1671 unsigned long nr_pte_updates
= 0;
1674 WARN_ON_ONCE(p
!= container_of(work
, struct task_struct
, numa_work
));
1676 work
->next
= work
; /* protect against double add */
1678 * Who cares about NUMA placement when they're dying.
1680 * NOTE: make sure not to dereference p->mm before this check,
1681 * exit_task_work() happens _after_ exit_mm() so we could be called
1682 * without p->mm even though we still had it when we enqueued this
1685 if (p
->flags
& PF_EXITING
)
1688 if (!mm
->numa_next_reset
|| !mm
->numa_next_scan
) {
1689 mm
->numa_next_scan
= now
+
1690 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
1691 mm
->numa_next_reset
= now
+
1692 msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset
);
1696 * Reset the scan period if enough time has gone by. Objective is that
1697 * scanning will be reduced if pages are properly placed. As tasks
1698 * can enter different phases this needs to be re-examined. Lacking
1699 * proper tracking of reference behaviour, this blunt hammer is used.
1701 migrate
= mm
->numa_next_reset
;
1702 if (time_after(now
, migrate
)) {
1703 p
->numa_scan_period
= task_scan_min(p
);
1704 next_scan
= now
+ msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset
);
1705 xchg(&mm
->numa_next_reset
, next_scan
);
1709 * Enforce maximal scan/migration frequency..
1711 migrate
= mm
->numa_next_scan
;
1712 if (time_before(now
, migrate
))
1715 if (p
->numa_scan_period
== 0) {
1716 p
->numa_scan_period_max
= task_scan_max(p
);
1717 p
->numa_scan_period
= task_scan_min(p
);
1720 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
1721 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
1725 * Delay this task enough that another task of this mm will likely win
1726 * the next time around.
1728 p
->node_stamp
+= 2 * TICK_NSEC
;
1730 start
= mm
->numa_scan_offset
;
1731 pages
= sysctl_numa_balancing_scan_size
;
1732 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
1736 down_read(&mm
->mmap_sem
);
1737 vma
= find_vma(mm
, start
);
1739 reset_ptenuma_scan(p
);
1743 for (; vma
; vma
= vma
->vm_next
) {
1744 if (!vma_migratable(vma
) || !vma_policy_mof(p
, vma
))
1748 * Shared library pages mapped by multiple processes are not
1749 * migrated as it is expected they are cache replicated. Avoid
1750 * hinting faults in read-only file-backed mappings or the vdso
1751 * as migrating the pages will be of marginal benefit.
1754 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
1758 start
= max(start
, vma
->vm_start
);
1759 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
1760 end
= min(end
, vma
->vm_end
);
1761 nr_pte_updates
+= change_prot_numa(vma
, start
, end
);
1764 * Scan sysctl_numa_balancing_scan_size but ensure that
1765 * at least one PTE is updated so that unused virtual
1766 * address space is quickly skipped.
1769 pages
-= (end
- start
) >> PAGE_SHIFT
;
1774 } while (end
!= vma
->vm_end
);
1779 * It is possible to reach the end of the VMA list but the last few
1780 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1781 * would find the !migratable VMA on the next scan but not reset the
1782 * scanner to the start so check it now.
1785 mm
->numa_scan_offset
= start
;
1787 reset_ptenuma_scan(p
);
1788 up_read(&mm
->mmap_sem
);
1792 * Drive the periodic memory faults..
1794 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
1796 struct callback_head
*work
= &curr
->numa_work
;
1800 * We don't care about NUMA placement if we don't have memory.
1802 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
1806 * Using runtime rather than walltime has the dual advantage that
1807 * we (mostly) drive the selection from busy threads and that the
1808 * task needs to have done some actual work before we bother with
1811 now
= curr
->se
.sum_exec_runtime
;
1812 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
1814 if (now
- curr
->node_stamp
> period
) {
1815 if (!curr
->node_stamp
)
1816 curr
->numa_scan_period
= task_scan_min(curr
);
1817 curr
->node_stamp
+= period
;
1819 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
1820 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
1821 task_work_add(curr
, work
, true);
1826 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
1830 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
1834 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
1837 #endif /* CONFIG_NUMA_BALANCING */
1840 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1842 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
1843 if (!parent_entity(se
))
1844 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
1846 if (entity_is_task(se
)) {
1847 struct rq
*rq
= rq_of(cfs_rq
);
1849 account_numa_enqueue(rq
, task_of(se
));
1850 list_add(&se
->group_node
, &rq
->cfs_tasks
);
1853 cfs_rq
->nr_running
++;
1857 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1859 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
1860 if (!parent_entity(se
))
1861 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
1862 if (entity_is_task(se
)) {
1863 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
1864 list_del_init(&se
->group_node
);
1866 cfs_rq
->nr_running
--;
1869 #ifdef CONFIG_FAIR_GROUP_SCHED
1871 static inline long calc_tg_weight(struct task_group
*tg
, struct cfs_rq
*cfs_rq
)
1876 * Use this CPU's actual weight instead of the last load_contribution
1877 * to gain a more accurate current total weight. See
1878 * update_cfs_rq_load_contribution().
1880 tg_weight
= atomic_long_read(&tg
->load_avg
);
1881 tg_weight
-= cfs_rq
->tg_load_contrib
;
1882 tg_weight
+= cfs_rq
->load
.weight
;
1887 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
1889 long tg_weight
, load
, shares
;
1891 tg_weight
= calc_tg_weight(tg
, cfs_rq
);
1892 load
= cfs_rq
->load
.weight
;
1894 shares
= (tg
->shares
* load
);
1896 shares
/= tg_weight
;
1898 if (shares
< MIN_SHARES
)
1899 shares
= MIN_SHARES
;
1900 if (shares
> tg
->shares
)
1901 shares
= tg
->shares
;
1905 # else /* CONFIG_SMP */
1906 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
1910 # endif /* CONFIG_SMP */
1911 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
1912 unsigned long weight
)
1915 /* commit outstanding execution time */
1916 if (cfs_rq
->curr
== se
)
1917 update_curr(cfs_rq
);
1918 account_entity_dequeue(cfs_rq
, se
);
1921 update_load_set(&se
->load
, weight
);
1924 account_entity_enqueue(cfs_rq
, se
);
1927 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
1929 static void update_cfs_shares(struct cfs_rq
*cfs_rq
)
1931 struct task_group
*tg
;
1932 struct sched_entity
*se
;
1936 se
= tg
->se
[cpu_of(rq_of(cfs_rq
))];
1937 if (!se
|| throttled_hierarchy(cfs_rq
))
1940 if (likely(se
->load
.weight
== tg
->shares
))
1943 shares
= calc_cfs_shares(cfs_rq
, tg
);
1945 reweight_entity(cfs_rq_of(se
), se
, shares
);
1947 #else /* CONFIG_FAIR_GROUP_SCHED */
1948 static inline void update_cfs_shares(struct cfs_rq
*cfs_rq
)
1951 #endif /* CONFIG_FAIR_GROUP_SCHED */
1955 * We choose a half-life close to 1 scheduling period.
1956 * Note: The tables below are dependent on this value.
1958 #define LOAD_AVG_PERIOD 32
1959 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1960 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1962 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1963 static const u32 runnable_avg_yN_inv
[] = {
1964 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1965 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1966 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1967 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1968 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1969 0x85aac367, 0x82cd8698,
1973 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1974 * over-estimates when re-combining.
1976 static const u32 runnable_avg_yN_sum
[] = {
1977 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1978 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1979 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1984 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1986 static __always_inline u64
decay_load(u64 val
, u64 n
)
1988 unsigned int local_n
;
1992 else if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
1995 /* after bounds checking we can collapse to 32-bit */
1999 * As y^PERIOD = 1/2, we can combine
2000 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
2001 * With a look-up table which covers k^n (n<PERIOD)
2003 * To achieve constant time decay_load.
2005 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
2006 val
>>= local_n
/ LOAD_AVG_PERIOD
;
2007 local_n
%= LOAD_AVG_PERIOD
;
2010 val
*= runnable_avg_yN_inv
[local_n
];
2011 /* We don't use SRR here since we always want to round down. */
2016 * For updates fully spanning n periods, the contribution to runnable
2017 * average will be: \Sum 1024*y^n
2019 * We can compute this reasonably efficiently by combining:
2020 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2022 static u32
__compute_runnable_contrib(u64 n
)
2026 if (likely(n
<= LOAD_AVG_PERIOD
))
2027 return runnable_avg_yN_sum
[n
];
2028 else if (unlikely(n
>= LOAD_AVG_MAX_N
))
2029 return LOAD_AVG_MAX
;
2031 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2033 contrib
/= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2034 contrib
+= runnable_avg_yN_sum
[LOAD_AVG_PERIOD
];
2036 n
-= LOAD_AVG_PERIOD
;
2037 } while (n
> LOAD_AVG_PERIOD
);
2039 contrib
= decay_load(contrib
, n
);
2040 return contrib
+ runnable_avg_yN_sum
[n
];
2044 * We can represent the historical contribution to runnable average as the
2045 * coefficients of a geometric series. To do this we sub-divide our runnable
2046 * history into segments of approximately 1ms (1024us); label the segment that
2047 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2049 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2051 * (now) (~1ms ago) (~2ms ago)
2053 * Let u_i denote the fraction of p_i that the entity was runnable.
2055 * We then designate the fractions u_i as our co-efficients, yielding the
2056 * following representation of historical load:
2057 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2059 * We choose y based on the with of a reasonably scheduling period, fixing:
2062 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2063 * approximately half as much as the contribution to load within the last ms
2066 * When a period "rolls over" and we have new u_0`, multiplying the previous
2067 * sum again by y is sufficient to update:
2068 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2069 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2071 static __always_inline
int __update_entity_runnable_avg(u64 now
,
2072 struct sched_avg
*sa
,
2076 u32 runnable_contrib
;
2077 int delta_w
, decayed
= 0;
2079 delta
= now
- sa
->last_runnable_update
;
2081 * This should only happen when time goes backwards, which it
2082 * unfortunately does during sched clock init when we swap over to TSC.
2084 if ((s64
)delta
< 0) {
2085 sa
->last_runnable_update
= now
;
2090 * Use 1024ns as the unit of measurement since it's a reasonable
2091 * approximation of 1us and fast to compute.
2096 sa
->last_runnable_update
= now
;
2098 /* delta_w is the amount already accumulated against our next period */
2099 delta_w
= sa
->runnable_avg_period
% 1024;
2100 if (delta
+ delta_w
>= 1024) {
2101 /* period roll-over */
2105 * Now that we know we're crossing a period boundary, figure
2106 * out how much from delta we need to complete the current
2107 * period and accrue it.
2109 delta_w
= 1024 - delta_w
;
2111 sa
->runnable_avg_sum
+= delta_w
;
2112 sa
->runnable_avg_period
+= delta_w
;
2116 /* Figure out how many additional periods this update spans */
2117 periods
= delta
/ 1024;
2120 sa
->runnable_avg_sum
= decay_load(sa
->runnable_avg_sum
,
2122 sa
->runnable_avg_period
= decay_load(sa
->runnable_avg_period
,
2125 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2126 runnable_contrib
= __compute_runnable_contrib(periods
);
2128 sa
->runnable_avg_sum
+= runnable_contrib
;
2129 sa
->runnable_avg_period
+= runnable_contrib
;
2132 /* Remainder of delta accrued against u_0` */
2134 sa
->runnable_avg_sum
+= delta
;
2135 sa
->runnable_avg_period
+= delta
;
2140 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2141 static inline u64
__synchronize_entity_decay(struct sched_entity
*se
)
2143 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2144 u64 decays
= atomic64_read(&cfs_rq
->decay_counter
);
2146 decays
-= se
->avg
.decay_count
;
2150 se
->avg
.load_avg_contrib
= decay_load(se
->avg
.load_avg_contrib
, decays
);
2151 se
->avg
.decay_count
= 0;
2156 #ifdef CONFIG_FAIR_GROUP_SCHED
2157 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2160 struct task_group
*tg
= cfs_rq
->tg
;
2163 tg_contrib
= cfs_rq
->runnable_load_avg
+ cfs_rq
->blocked_load_avg
;
2164 tg_contrib
-= cfs_rq
->tg_load_contrib
;
2166 if (force_update
|| abs(tg_contrib
) > cfs_rq
->tg_load_contrib
/ 8) {
2167 atomic_long_add(tg_contrib
, &tg
->load_avg
);
2168 cfs_rq
->tg_load_contrib
+= tg_contrib
;
2173 * Aggregate cfs_rq runnable averages into an equivalent task_group
2174 * representation for computing load contributions.
2176 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2177 struct cfs_rq
*cfs_rq
)
2179 struct task_group
*tg
= cfs_rq
->tg
;
2182 /* The fraction of a cpu used by this cfs_rq */
2183 contrib
= div_u64(sa
->runnable_avg_sum
<< NICE_0_SHIFT
,
2184 sa
->runnable_avg_period
+ 1);
2185 contrib
-= cfs_rq
->tg_runnable_contrib
;
2187 if (abs(contrib
) > cfs_rq
->tg_runnable_contrib
/ 64) {
2188 atomic_add(contrib
, &tg
->runnable_avg
);
2189 cfs_rq
->tg_runnable_contrib
+= contrib
;
2193 static inline void __update_group_entity_contrib(struct sched_entity
*se
)
2195 struct cfs_rq
*cfs_rq
= group_cfs_rq(se
);
2196 struct task_group
*tg
= cfs_rq
->tg
;
2201 contrib
= cfs_rq
->tg_load_contrib
* tg
->shares
;
2202 se
->avg
.load_avg_contrib
= div_u64(contrib
,
2203 atomic_long_read(&tg
->load_avg
) + 1);
2206 * For group entities we need to compute a correction term in the case
2207 * that they are consuming <1 cpu so that we would contribute the same
2208 * load as a task of equal weight.
2210 * Explicitly co-ordinating this measurement would be expensive, but
2211 * fortunately the sum of each cpus contribution forms a usable
2212 * lower-bound on the true value.
2214 * Consider the aggregate of 2 contributions. Either they are disjoint
2215 * (and the sum represents true value) or they are disjoint and we are
2216 * understating by the aggregate of their overlap.
2218 * Extending this to N cpus, for a given overlap, the maximum amount we
2219 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2220 * cpus that overlap for this interval and w_i is the interval width.
2222 * On a small machine; the first term is well-bounded which bounds the
2223 * total error since w_i is a subset of the period. Whereas on a
2224 * larger machine, while this first term can be larger, if w_i is the
2225 * of consequential size guaranteed to see n_i*w_i quickly converge to
2226 * our upper bound of 1-cpu.
2228 runnable_avg
= atomic_read(&tg
->runnable_avg
);
2229 if (runnable_avg
< NICE_0_LOAD
) {
2230 se
->avg
.load_avg_contrib
*= runnable_avg
;
2231 se
->avg
.load_avg_contrib
>>= NICE_0_SHIFT
;
2235 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2236 int force_update
) {}
2237 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2238 struct cfs_rq
*cfs_rq
) {}
2239 static inline void __update_group_entity_contrib(struct sched_entity
*se
) {}
2242 static inline void __update_task_entity_contrib(struct sched_entity
*se
)
2246 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2247 contrib
= se
->avg
.runnable_avg_sum
* scale_load_down(se
->load
.weight
);
2248 contrib
/= (se
->avg
.runnable_avg_period
+ 1);
2249 se
->avg
.load_avg_contrib
= scale_load(contrib
);
2252 /* Compute the current contribution to load_avg by se, return any delta */
2253 static long __update_entity_load_avg_contrib(struct sched_entity
*se
)
2255 long old_contrib
= se
->avg
.load_avg_contrib
;
2257 if (entity_is_task(se
)) {
2258 __update_task_entity_contrib(se
);
2260 __update_tg_runnable_avg(&se
->avg
, group_cfs_rq(se
));
2261 __update_group_entity_contrib(se
);
2264 return se
->avg
.load_avg_contrib
- old_contrib
;
2267 static inline void subtract_blocked_load_contrib(struct cfs_rq
*cfs_rq
,
2270 if (likely(load_contrib
< cfs_rq
->blocked_load_avg
))
2271 cfs_rq
->blocked_load_avg
-= load_contrib
;
2273 cfs_rq
->blocked_load_avg
= 0;
2276 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
2278 /* Update a sched_entity's runnable average */
2279 static inline void update_entity_load_avg(struct sched_entity
*se
,
2282 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2287 * For a group entity we need to use their owned cfs_rq_clock_task() in
2288 * case they are the parent of a throttled hierarchy.
2290 if (entity_is_task(se
))
2291 now
= cfs_rq_clock_task(cfs_rq
);
2293 now
= cfs_rq_clock_task(group_cfs_rq(se
));
2295 if (!__update_entity_runnable_avg(now
, &se
->avg
, se
->on_rq
))
2298 contrib_delta
= __update_entity_load_avg_contrib(se
);
2304 cfs_rq
->runnable_load_avg
+= contrib_delta
;
2306 subtract_blocked_load_contrib(cfs_rq
, -contrib_delta
);
2310 * Decay the load contributed by all blocked children and account this so that
2311 * their contribution may appropriately discounted when they wake up.
2313 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
, int force_update
)
2315 u64 now
= cfs_rq_clock_task(cfs_rq
) >> 20;
2318 decays
= now
- cfs_rq
->last_decay
;
2319 if (!decays
&& !force_update
)
2322 if (atomic_long_read(&cfs_rq
->removed_load
)) {
2323 unsigned long removed_load
;
2324 removed_load
= atomic_long_xchg(&cfs_rq
->removed_load
, 0);
2325 subtract_blocked_load_contrib(cfs_rq
, removed_load
);
2329 cfs_rq
->blocked_load_avg
= decay_load(cfs_rq
->blocked_load_avg
,
2331 atomic64_add(decays
, &cfs_rq
->decay_counter
);
2332 cfs_rq
->last_decay
= now
;
2335 __update_cfs_rq_tg_load_contrib(cfs_rq
, force_update
);
2338 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
)
2340 __update_entity_runnable_avg(rq_clock_task(rq
), &rq
->avg
, runnable
);
2341 __update_tg_runnable_avg(&rq
->avg
, &rq
->cfs
);
2344 /* Add the load generated by se into cfs_rq's child load-average */
2345 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2346 struct sched_entity
*se
,
2350 * We track migrations using entity decay_count <= 0, on a wake-up
2351 * migration we use a negative decay count to track the remote decays
2352 * accumulated while sleeping.
2354 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2355 * are seen by enqueue_entity_load_avg() as a migration with an already
2356 * constructed load_avg_contrib.
2358 if (unlikely(se
->avg
.decay_count
<= 0)) {
2359 se
->avg
.last_runnable_update
= rq_clock_task(rq_of(cfs_rq
));
2360 if (se
->avg
.decay_count
) {
2362 * In a wake-up migration we have to approximate the
2363 * time sleeping. This is because we can't synchronize
2364 * clock_task between the two cpus, and it is not
2365 * guaranteed to be read-safe. Instead, we can
2366 * approximate this using our carried decays, which are
2367 * explicitly atomically readable.
2369 se
->avg
.last_runnable_update
-= (-se
->avg
.decay_count
)
2371 update_entity_load_avg(se
, 0);
2372 /* Indicate that we're now synchronized and on-rq */
2373 se
->avg
.decay_count
= 0;
2378 * Task re-woke on same cpu (or else migrate_task_rq_fair()
2379 * would have made count negative); we must be careful to avoid
2380 * double-accounting blocked time after synchronizing decays.
2382 se
->avg
.last_runnable_update
+= __synchronize_entity_decay(se
)
2386 /* migrated tasks did not contribute to our blocked load */
2388 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
2389 update_entity_load_avg(se
, 0);
2392 cfs_rq
->runnable_load_avg
+= se
->avg
.load_avg_contrib
;
2393 /* we force update consideration on load-balancer moves */
2394 update_cfs_rq_blocked_load(cfs_rq
, !wakeup
);
2398 * Remove se's load from this cfs_rq child load-average, if the entity is
2399 * transitioning to a blocked state we track its projected decay using
2402 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2403 struct sched_entity
*se
,
2406 update_entity_load_avg(se
, 1);
2407 /* we force update consideration on load-balancer moves */
2408 update_cfs_rq_blocked_load(cfs_rq
, !sleep
);
2410 cfs_rq
->runnable_load_avg
-= se
->avg
.load_avg_contrib
;
2412 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
2413 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
2414 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2418 * Update the rq's load with the elapsed running time before entering
2419 * idle. if the last scheduled task is not a CFS task, idle_enter will
2420 * be the only way to update the runnable statistic.
2422 void idle_enter_fair(struct rq
*this_rq
)
2424 update_rq_runnable_avg(this_rq
, 1);
2428 * Update the rq's load with the elapsed idle time before a task is
2429 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2430 * be the only way to update the runnable statistic.
2432 void idle_exit_fair(struct rq
*this_rq
)
2434 update_rq_runnable_avg(this_rq
, 0);
2438 static inline void update_entity_load_avg(struct sched_entity
*se
,
2439 int update_cfs_rq
) {}
2440 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2441 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2442 struct sched_entity
*se
,
2444 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2445 struct sched_entity
*se
,
2447 static inline void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
2448 int force_update
) {}
2451 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2453 #ifdef CONFIG_SCHEDSTATS
2454 struct task_struct
*tsk
= NULL
;
2456 if (entity_is_task(se
))
2459 if (se
->statistics
.sleep_start
) {
2460 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.sleep_start
;
2465 if (unlikely(delta
> se
->statistics
.sleep_max
))
2466 se
->statistics
.sleep_max
= delta
;
2468 se
->statistics
.sleep_start
= 0;
2469 se
->statistics
.sum_sleep_runtime
+= delta
;
2472 account_scheduler_latency(tsk
, delta
>> 10, 1);
2473 trace_sched_stat_sleep(tsk
, delta
);
2476 if (se
->statistics
.block_start
) {
2477 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.block_start
;
2482 if (unlikely(delta
> se
->statistics
.block_max
))
2483 se
->statistics
.block_max
= delta
;
2485 se
->statistics
.block_start
= 0;
2486 se
->statistics
.sum_sleep_runtime
+= delta
;
2489 if (tsk
->in_iowait
) {
2490 se
->statistics
.iowait_sum
+= delta
;
2491 se
->statistics
.iowait_count
++;
2492 trace_sched_stat_iowait(tsk
, delta
);
2495 trace_sched_stat_blocked(tsk
, delta
);
2498 * Blocking time is in units of nanosecs, so shift by
2499 * 20 to get a milliseconds-range estimation of the
2500 * amount of time that the task spent sleeping:
2502 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
2503 profile_hits(SLEEP_PROFILING
,
2504 (void *)get_wchan(tsk
),
2507 account_scheduler_latency(tsk
, delta
>> 10, 0);
2513 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2515 #ifdef CONFIG_SCHED_DEBUG
2516 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
2521 if (d
> 3*sysctl_sched_latency
)
2522 schedstat_inc(cfs_rq
, nr_spread_over
);
2527 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
2529 u64 vruntime
= cfs_rq
->min_vruntime
;
2532 * The 'current' period is already promised to the current tasks,
2533 * however the extra weight of the new task will slow them down a
2534 * little, place the new task so that it fits in the slot that
2535 * stays open at the end.
2537 if (initial
&& sched_feat(START_DEBIT
))
2538 vruntime
+= sched_vslice(cfs_rq
, se
);
2540 /* sleeps up to a single latency don't count. */
2542 unsigned long thresh
= sysctl_sched_latency
;
2545 * Halve their sleep time's effect, to allow
2546 * for a gentler effect of sleepers:
2548 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
2554 /* ensure we never gain time by being placed backwards. */
2555 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
2558 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
2561 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2564 * Update the normalized vruntime before updating min_vruntime
2565 * through calling update_curr().
2567 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
2568 se
->vruntime
+= cfs_rq
->min_vruntime
;
2571 * Update run-time statistics of the 'current'.
2573 update_curr(cfs_rq
);
2574 enqueue_entity_load_avg(cfs_rq
, se
, flags
& ENQUEUE_WAKEUP
);
2575 account_entity_enqueue(cfs_rq
, se
);
2576 update_cfs_shares(cfs_rq
);
2578 if (flags
& ENQUEUE_WAKEUP
) {
2579 place_entity(cfs_rq
, se
, 0);
2580 enqueue_sleeper(cfs_rq
, se
);
2583 update_stats_enqueue(cfs_rq
, se
);
2584 check_spread(cfs_rq
, se
);
2585 if (se
!= cfs_rq
->curr
)
2586 __enqueue_entity(cfs_rq
, se
);
2589 if (cfs_rq
->nr_running
== 1) {
2590 list_add_leaf_cfs_rq(cfs_rq
);
2591 check_enqueue_throttle(cfs_rq
);
2595 static void __clear_buddies_last(struct sched_entity
*se
)
2597 for_each_sched_entity(se
) {
2598 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2599 if (cfs_rq
->last
== se
)
2600 cfs_rq
->last
= NULL
;
2606 static void __clear_buddies_next(struct sched_entity
*se
)
2608 for_each_sched_entity(se
) {
2609 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2610 if (cfs_rq
->next
== se
)
2611 cfs_rq
->next
= NULL
;
2617 static void __clear_buddies_skip(struct sched_entity
*se
)
2619 for_each_sched_entity(se
) {
2620 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2621 if (cfs_rq
->skip
== se
)
2622 cfs_rq
->skip
= NULL
;
2628 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2630 if (cfs_rq
->last
== se
)
2631 __clear_buddies_last(se
);
2633 if (cfs_rq
->next
== se
)
2634 __clear_buddies_next(se
);
2636 if (cfs_rq
->skip
== se
)
2637 __clear_buddies_skip(se
);
2640 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
2643 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2646 * Update run-time statistics of the 'current'.
2648 update_curr(cfs_rq
);
2649 dequeue_entity_load_avg(cfs_rq
, se
, flags
& DEQUEUE_SLEEP
);
2651 update_stats_dequeue(cfs_rq
, se
);
2652 if (flags
& DEQUEUE_SLEEP
) {
2653 #ifdef CONFIG_SCHEDSTATS
2654 if (entity_is_task(se
)) {
2655 struct task_struct
*tsk
= task_of(se
);
2657 if (tsk
->state
& TASK_INTERRUPTIBLE
)
2658 se
->statistics
.sleep_start
= rq_clock(rq_of(cfs_rq
));
2659 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
2660 se
->statistics
.block_start
= rq_clock(rq_of(cfs_rq
));
2665 clear_buddies(cfs_rq
, se
);
2667 if (se
!= cfs_rq
->curr
)
2668 __dequeue_entity(cfs_rq
, se
);
2670 account_entity_dequeue(cfs_rq
, se
);
2673 * Normalize the entity after updating the min_vruntime because the
2674 * update can refer to the ->curr item and we need to reflect this
2675 * movement in our normalized position.
2677 if (!(flags
& DEQUEUE_SLEEP
))
2678 se
->vruntime
-= cfs_rq
->min_vruntime
;
2680 /* return excess runtime on last dequeue */
2681 return_cfs_rq_runtime(cfs_rq
);
2683 update_min_vruntime(cfs_rq
);
2684 update_cfs_shares(cfs_rq
);
2688 * Preempt the current task with a newly woken task if needed:
2691 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
2693 unsigned long ideal_runtime
, delta_exec
;
2694 struct sched_entity
*se
;
2697 ideal_runtime
= sched_slice(cfs_rq
, curr
);
2698 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
2699 if (delta_exec
> ideal_runtime
) {
2700 resched_task(rq_of(cfs_rq
)->curr
);
2702 * The current task ran long enough, ensure it doesn't get
2703 * re-elected due to buddy favours.
2705 clear_buddies(cfs_rq
, curr
);
2710 * Ensure that a task that missed wakeup preemption by a
2711 * narrow margin doesn't have to wait for a full slice.
2712 * This also mitigates buddy induced latencies under load.
2714 if (delta_exec
< sysctl_sched_min_granularity
)
2717 se
= __pick_first_entity(cfs_rq
);
2718 delta
= curr
->vruntime
- se
->vruntime
;
2723 if (delta
> ideal_runtime
)
2724 resched_task(rq_of(cfs_rq
)->curr
);
2728 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2730 /* 'current' is not kept within the tree. */
2733 * Any task has to be enqueued before it get to execute on
2734 * a CPU. So account for the time it spent waiting on the
2737 update_stats_wait_end(cfs_rq
, se
);
2738 __dequeue_entity(cfs_rq
, se
);
2741 update_stats_curr_start(cfs_rq
, se
);
2743 #ifdef CONFIG_SCHEDSTATS
2745 * Track our maximum slice length, if the CPU's load is at
2746 * least twice that of our own weight (i.e. dont track it
2747 * when there are only lesser-weight tasks around):
2749 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
2750 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
2751 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
2754 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
2758 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
2761 * Pick the next process, keeping these things in mind, in this order:
2762 * 1) keep things fair between processes/task groups
2763 * 2) pick the "next" process, since someone really wants that to run
2764 * 3) pick the "last" process, for cache locality
2765 * 4) do not run the "skip" process, if something else is available
2767 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
2769 struct sched_entity
*se
= __pick_first_entity(cfs_rq
);
2770 struct sched_entity
*left
= se
;
2773 * Avoid running the skip buddy, if running something else can
2774 * be done without getting too unfair.
2776 if (cfs_rq
->skip
== se
) {
2777 struct sched_entity
*second
= __pick_next_entity(se
);
2778 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
2783 * Prefer last buddy, try to return the CPU to a preempted task.
2785 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
2789 * Someone really wants this to run. If it's not unfair, run it.
2791 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
2794 clear_buddies(cfs_rq
, se
);
2799 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
2801 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
2804 * If still on the runqueue then deactivate_task()
2805 * was not called and update_curr() has to be done:
2808 update_curr(cfs_rq
);
2810 /* throttle cfs_rqs exceeding runtime */
2811 check_cfs_rq_runtime(cfs_rq
);
2813 check_spread(cfs_rq
, prev
);
2815 update_stats_wait_start(cfs_rq
, prev
);
2816 /* Put 'current' back into the tree. */
2817 __enqueue_entity(cfs_rq
, prev
);
2818 /* in !on_rq case, update occurred at dequeue */
2819 update_entity_load_avg(prev
, 1);
2821 cfs_rq
->curr
= NULL
;
2825 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
2828 * Update run-time statistics of the 'current'.
2830 update_curr(cfs_rq
);
2833 * Ensure that runnable average is periodically updated.
2835 update_entity_load_avg(curr
, 1);
2836 update_cfs_rq_blocked_load(cfs_rq
, 1);
2837 update_cfs_shares(cfs_rq
);
2839 #ifdef CONFIG_SCHED_HRTICK
2841 * queued ticks are scheduled to match the slice, so don't bother
2842 * validating it and just reschedule.
2845 resched_task(rq_of(cfs_rq
)->curr
);
2849 * don't let the period tick interfere with the hrtick preemption
2851 if (!sched_feat(DOUBLE_TICK
) &&
2852 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
2856 if (cfs_rq
->nr_running
> 1)
2857 check_preempt_tick(cfs_rq
, curr
);
2861 /**************************************************
2862 * CFS bandwidth control machinery
2865 #ifdef CONFIG_CFS_BANDWIDTH
2867 #ifdef HAVE_JUMP_LABEL
2868 static struct static_key __cfs_bandwidth_used
;
2870 static inline bool cfs_bandwidth_used(void)
2872 return static_key_false(&__cfs_bandwidth_used
);
2875 void account_cfs_bandwidth_used(int enabled
, int was_enabled
)
2877 /* only need to count groups transitioning between enabled/!enabled */
2878 if (enabled
&& !was_enabled
)
2879 static_key_slow_inc(&__cfs_bandwidth_used
);
2880 else if (!enabled
&& was_enabled
)
2881 static_key_slow_dec(&__cfs_bandwidth_used
);
2883 #else /* HAVE_JUMP_LABEL */
2884 static bool cfs_bandwidth_used(void)
2889 void account_cfs_bandwidth_used(int enabled
, int was_enabled
) {}
2890 #endif /* HAVE_JUMP_LABEL */
2893 * default period for cfs group bandwidth.
2894 * default: 0.1s, units: nanoseconds
2896 static inline u64
default_cfs_period(void)
2898 return 100000000ULL;
2901 static inline u64
sched_cfs_bandwidth_slice(void)
2903 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
2907 * Replenish runtime according to assigned quota and update expiration time.
2908 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2909 * additional synchronization around rq->lock.
2911 * requires cfs_b->lock
2913 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
2917 if (cfs_b
->quota
== RUNTIME_INF
)
2920 now
= sched_clock_cpu(smp_processor_id());
2921 cfs_b
->runtime
= cfs_b
->quota
;
2922 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
2925 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
2927 return &tg
->cfs_bandwidth
;
2930 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2931 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
2933 if (unlikely(cfs_rq
->throttle_count
))
2934 return cfs_rq
->throttled_clock_task
;
2936 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
2939 /* returns 0 on failure to allocate runtime */
2940 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2942 struct task_group
*tg
= cfs_rq
->tg
;
2943 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
2944 u64 amount
= 0, min_amount
, expires
;
2946 /* note: this is a positive sum as runtime_remaining <= 0 */
2947 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
2949 raw_spin_lock(&cfs_b
->lock
);
2950 if (cfs_b
->quota
== RUNTIME_INF
)
2951 amount
= min_amount
;
2954 * If the bandwidth pool has become inactive, then at least one
2955 * period must have elapsed since the last consumption.
2956 * Refresh the global state and ensure bandwidth timer becomes
2959 if (!cfs_b
->timer_active
) {
2960 __refill_cfs_bandwidth_runtime(cfs_b
);
2961 __start_cfs_bandwidth(cfs_b
);
2964 if (cfs_b
->runtime
> 0) {
2965 amount
= min(cfs_b
->runtime
, min_amount
);
2966 cfs_b
->runtime
-= amount
;
2970 expires
= cfs_b
->runtime_expires
;
2971 raw_spin_unlock(&cfs_b
->lock
);
2973 cfs_rq
->runtime_remaining
+= amount
;
2975 * we may have advanced our local expiration to account for allowed
2976 * spread between our sched_clock and the one on which runtime was
2979 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
2980 cfs_rq
->runtime_expires
= expires
;
2982 return cfs_rq
->runtime_remaining
> 0;
2986 * Note: This depends on the synchronization provided by sched_clock and the
2987 * fact that rq->clock snapshots this value.
2989 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2991 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
2993 /* if the deadline is ahead of our clock, nothing to do */
2994 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
2997 if (cfs_rq
->runtime_remaining
< 0)
3001 * If the local deadline has passed we have to consider the
3002 * possibility that our sched_clock is 'fast' and the global deadline
3003 * has not truly expired.
3005 * Fortunately we can check determine whether this the case by checking
3006 * whether the global deadline has advanced.
3009 if ((s64
)(cfs_rq
->runtime_expires
- cfs_b
->runtime_expires
) >= 0) {
3010 /* extend local deadline, drift is bounded above by 2 ticks */
3011 cfs_rq
->runtime_expires
+= TICK_NSEC
;
3013 /* global deadline is ahead, expiration has passed */
3014 cfs_rq
->runtime_remaining
= 0;
3018 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
,
3019 unsigned long delta_exec
)
3021 /* dock delta_exec before expiring quota (as it could span periods) */
3022 cfs_rq
->runtime_remaining
-= delta_exec
;
3023 expire_cfs_rq_runtime(cfs_rq
);
3025 if (likely(cfs_rq
->runtime_remaining
> 0))
3029 * if we're unable to extend our runtime we resched so that the active
3030 * hierarchy can be throttled
3032 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
3033 resched_task(rq_of(cfs_rq
)->curr
);
3036 static __always_inline
3037 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, unsigned long delta_exec
)
3039 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
3042 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
3045 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3047 return cfs_bandwidth_used() && cfs_rq
->throttled
;
3050 /* check whether cfs_rq, or any parent, is throttled */
3051 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3053 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
3057 * Ensure that neither of the group entities corresponding to src_cpu or
3058 * dest_cpu are members of a throttled hierarchy when performing group
3059 * load-balance operations.
3061 static inline int throttled_lb_pair(struct task_group
*tg
,
3062 int src_cpu
, int dest_cpu
)
3064 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
3066 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
3067 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
3069 return throttled_hierarchy(src_cfs_rq
) ||
3070 throttled_hierarchy(dest_cfs_rq
);
3073 /* updated child weight may affect parent so we have to do this bottom up */
3074 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
3076 struct rq
*rq
= data
;
3077 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3079 cfs_rq
->throttle_count
--;
3081 if (!cfs_rq
->throttle_count
) {
3082 /* adjust cfs_rq_clock_task() */
3083 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
3084 cfs_rq
->throttled_clock_task
;
3091 static int tg_throttle_down(struct task_group
*tg
, void *data
)
3093 struct rq
*rq
= data
;
3094 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3096 /* group is entering throttled state, stop time */
3097 if (!cfs_rq
->throttle_count
)
3098 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
3099 cfs_rq
->throttle_count
++;
3104 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
3106 struct rq
*rq
= rq_of(cfs_rq
);
3107 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3108 struct sched_entity
*se
;
3109 long task_delta
, dequeue
= 1;
3111 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
3113 /* freeze hierarchy runnable averages while throttled */
3115 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
3118 task_delta
= cfs_rq
->h_nr_running
;
3119 for_each_sched_entity(se
) {
3120 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
3121 /* throttled entity or throttle-on-deactivate */
3126 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
3127 qcfs_rq
->h_nr_running
-= task_delta
;
3129 if (qcfs_rq
->load
.weight
)
3134 rq
->nr_running
-= task_delta
;
3136 cfs_rq
->throttled
= 1;
3137 cfs_rq
->throttled_clock
= rq_clock(rq
);
3138 raw_spin_lock(&cfs_b
->lock
);
3139 list_add_tail_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
3140 raw_spin_unlock(&cfs_b
->lock
);
3143 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
3145 struct rq
*rq
= rq_of(cfs_rq
);
3146 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3147 struct sched_entity
*se
;
3151 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
3153 cfs_rq
->throttled
= 0;
3155 update_rq_clock(rq
);
3157 raw_spin_lock(&cfs_b
->lock
);
3158 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
3159 list_del_rcu(&cfs_rq
->throttled_list
);
3160 raw_spin_unlock(&cfs_b
->lock
);
3162 /* update hierarchical throttle state */
3163 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
3165 if (!cfs_rq
->load
.weight
)
3168 task_delta
= cfs_rq
->h_nr_running
;
3169 for_each_sched_entity(se
) {
3173 cfs_rq
= cfs_rq_of(se
);
3175 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
3176 cfs_rq
->h_nr_running
+= task_delta
;
3178 if (cfs_rq_throttled(cfs_rq
))
3183 rq
->nr_running
+= task_delta
;
3185 /* determine whether we need to wake up potentially idle cpu */
3186 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
3187 resched_task(rq
->curr
);
3190 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
3191 u64 remaining
, u64 expires
)
3193 struct cfs_rq
*cfs_rq
;
3194 u64 runtime
= remaining
;
3197 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
3199 struct rq
*rq
= rq_of(cfs_rq
);
3201 raw_spin_lock(&rq
->lock
);
3202 if (!cfs_rq_throttled(cfs_rq
))
3205 runtime
= -cfs_rq
->runtime_remaining
+ 1;
3206 if (runtime
> remaining
)
3207 runtime
= remaining
;
3208 remaining
-= runtime
;
3210 cfs_rq
->runtime_remaining
+= runtime
;
3211 cfs_rq
->runtime_expires
= expires
;
3213 /* we check whether we're throttled above */
3214 if (cfs_rq
->runtime_remaining
> 0)
3215 unthrottle_cfs_rq(cfs_rq
);
3218 raw_spin_unlock(&rq
->lock
);
3229 * Responsible for refilling a task_group's bandwidth and unthrottling its
3230 * cfs_rqs as appropriate. If there has been no activity within the last
3231 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3232 * used to track this state.
3234 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
3236 u64 runtime
, runtime_expires
;
3237 int idle
= 1, throttled
;
3239 raw_spin_lock(&cfs_b
->lock
);
3240 /* no need to continue the timer with no bandwidth constraint */
3241 if (cfs_b
->quota
== RUNTIME_INF
)
3244 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3245 /* idle depends on !throttled (for the case of a large deficit) */
3246 idle
= cfs_b
->idle
&& !throttled
;
3247 cfs_b
->nr_periods
+= overrun
;
3249 /* if we're going inactive then everything else can be deferred */
3253 __refill_cfs_bandwidth_runtime(cfs_b
);
3256 /* mark as potentially idle for the upcoming period */
3261 /* account preceding periods in which throttling occurred */
3262 cfs_b
->nr_throttled
+= overrun
;
3265 * There are throttled entities so we must first use the new bandwidth
3266 * to unthrottle them before making it generally available. This
3267 * ensures that all existing debts will be paid before a new cfs_rq is
3270 runtime
= cfs_b
->runtime
;
3271 runtime_expires
= cfs_b
->runtime_expires
;
3275 * This check is repeated as we are holding onto the new bandwidth
3276 * while we unthrottle. This can potentially race with an unthrottled
3277 * group trying to acquire new bandwidth from the global pool.
3279 while (throttled
&& runtime
> 0) {
3280 raw_spin_unlock(&cfs_b
->lock
);
3281 /* we can't nest cfs_b->lock while distributing bandwidth */
3282 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
3284 raw_spin_lock(&cfs_b
->lock
);
3286 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3289 /* return (any) remaining runtime */
3290 cfs_b
->runtime
= runtime
;
3292 * While we are ensured activity in the period following an
3293 * unthrottle, this also covers the case in which the new bandwidth is
3294 * insufficient to cover the existing bandwidth deficit. (Forcing the
3295 * timer to remain active while there are any throttled entities.)
3300 cfs_b
->timer_active
= 0;
3301 raw_spin_unlock(&cfs_b
->lock
);
3306 /* a cfs_rq won't donate quota below this amount */
3307 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
3308 /* minimum remaining period time to redistribute slack quota */
3309 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
3310 /* how long we wait to gather additional slack before distributing */
3311 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
3313 /* are we near the end of the current quota period? */
3314 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
3316 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
3319 /* if the call-back is running a quota refresh is already occurring */
3320 if (hrtimer_callback_running(refresh_timer
))
3323 /* is a quota refresh about to occur? */
3324 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
3325 if (remaining
< min_expire
)
3331 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
3333 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
3335 /* if there's a quota refresh soon don't bother with slack */
3336 if (runtime_refresh_within(cfs_b
, min_left
))
3339 start_bandwidth_timer(&cfs_b
->slack_timer
,
3340 ns_to_ktime(cfs_bandwidth_slack_period
));
3343 /* we know any runtime found here is valid as update_curr() precedes return */
3344 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3346 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3347 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
3349 if (slack_runtime
<= 0)
3352 raw_spin_lock(&cfs_b
->lock
);
3353 if (cfs_b
->quota
!= RUNTIME_INF
&&
3354 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
3355 cfs_b
->runtime
+= slack_runtime
;
3357 /* we are under rq->lock, defer unthrottling using a timer */
3358 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
3359 !list_empty(&cfs_b
->throttled_cfs_rq
))
3360 start_cfs_slack_bandwidth(cfs_b
);
3362 raw_spin_unlock(&cfs_b
->lock
);
3364 /* even if it's not valid for return we don't want to try again */
3365 cfs_rq
->runtime_remaining
-= slack_runtime
;
3368 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3370 if (!cfs_bandwidth_used())
3373 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
3376 __return_cfs_rq_runtime(cfs_rq
);
3380 * This is done with a timer (instead of inline with bandwidth return) since
3381 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3383 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
3385 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
3388 /* confirm we're still not at a refresh boundary */
3389 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
))
3392 raw_spin_lock(&cfs_b
->lock
);
3393 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
) {
3394 runtime
= cfs_b
->runtime
;
3397 expires
= cfs_b
->runtime_expires
;
3398 raw_spin_unlock(&cfs_b
->lock
);
3403 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
3405 raw_spin_lock(&cfs_b
->lock
);
3406 if (expires
== cfs_b
->runtime_expires
)
3407 cfs_b
->runtime
= runtime
;
3408 raw_spin_unlock(&cfs_b
->lock
);
3412 * When a group wakes up we want to make sure that its quota is not already
3413 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3414 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3416 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
3418 if (!cfs_bandwidth_used())
3421 /* an active group must be handled by the update_curr()->put() path */
3422 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
3425 /* ensure the group is not already throttled */
3426 if (cfs_rq_throttled(cfs_rq
))
3429 /* update runtime allocation */
3430 account_cfs_rq_runtime(cfs_rq
, 0);
3431 if (cfs_rq
->runtime_remaining
<= 0)
3432 throttle_cfs_rq(cfs_rq
);
3435 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3436 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3438 if (!cfs_bandwidth_used())
3441 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
3445 * it's possible for a throttled entity to be forced into a running
3446 * state (e.g. set_curr_task), in this case we're finished.
3448 if (cfs_rq_throttled(cfs_rq
))
3451 throttle_cfs_rq(cfs_rq
);
3454 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
3456 struct cfs_bandwidth
*cfs_b
=
3457 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
3458 do_sched_cfs_slack_timer(cfs_b
);
3460 return HRTIMER_NORESTART
;
3463 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
3465 struct cfs_bandwidth
*cfs_b
=
3466 container_of(timer
, struct cfs_bandwidth
, period_timer
);
3472 now
= hrtimer_cb_get_time(timer
);
3473 overrun
= hrtimer_forward(timer
, now
, cfs_b
->period
);
3478 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
3481 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
3484 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3486 raw_spin_lock_init(&cfs_b
->lock
);
3488 cfs_b
->quota
= RUNTIME_INF
;
3489 cfs_b
->period
= ns_to_ktime(default_cfs_period());
3491 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
3492 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3493 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
3494 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3495 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
3498 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3500 cfs_rq
->runtime_enabled
= 0;
3501 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
3504 /* requires cfs_b->lock, may release to reprogram timer */
3505 void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3508 * The timer may be active because we're trying to set a new bandwidth
3509 * period or because we're racing with the tear-down path
3510 * (timer_active==0 becomes visible before the hrtimer call-back
3511 * terminates). In either case we ensure that it's re-programmed
3513 while (unlikely(hrtimer_active(&cfs_b
->period_timer
))) {
3514 raw_spin_unlock(&cfs_b
->lock
);
3515 /* ensure cfs_b->lock is available while we wait */
3516 hrtimer_cancel(&cfs_b
->period_timer
);
3518 raw_spin_lock(&cfs_b
->lock
);
3519 /* if someone else restarted the timer then we're done */
3520 if (cfs_b
->timer_active
)
3524 cfs_b
->timer_active
= 1;
3525 start_bandwidth_timer(&cfs_b
->period_timer
, cfs_b
->period
);
3528 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3530 hrtimer_cancel(&cfs_b
->period_timer
);
3531 hrtimer_cancel(&cfs_b
->slack_timer
);
3534 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
3536 struct cfs_rq
*cfs_rq
;
3538 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
3539 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3541 if (!cfs_rq
->runtime_enabled
)
3545 * clock_task is not advancing so we just need to make sure
3546 * there's some valid quota amount
3548 cfs_rq
->runtime_remaining
= cfs_b
->quota
;
3549 if (cfs_rq_throttled(cfs_rq
))
3550 unthrottle_cfs_rq(cfs_rq
);
3554 #else /* CONFIG_CFS_BANDWIDTH */
3555 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3557 return rq_clock_task(rq_of(cfs_rq
));
3560 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
,
3561 unsigned long delta_exec
) {}
3562 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3563 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
3564 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3566 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3571 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3576 static inline int throttled_lb_pair(struct task_group
*tg
,
3577 int src_cpu
, int dest_cpu
)
3582 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3584 #ifdef CONFIG_FAIR_GROUP_SCHED
3585 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3588 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3592 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3593 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
3595 #endif /* CONFIG_CFS_BANDWIDTH */
3597 /**************************************************
3598 * CFS operations on tasks:
3601 #ifdef CONFIG_SCHED_HRTICK
3602 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3604 struct sched_entity
*se
= &p
->se
;
3605 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3607 WARN_ON(task_rq(p
) != rq
);
3609 if (cfs_rq
->nr_running
> 1) {
3610 u64 slice
= sched_slice(cfs_rq
, se
);
3611 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
3612 s64 delta
= slice
- ran
;
3621 * Don't schedule slices shorter than 10000ns, that just
3622 * doesn't make sense. Rely on vruntime for fairness.
3625 delta
= max_t(s64
, 10000LL, delta
);
3627 hrtick_start(rq
, delta
);
3632 * called from enqueue/dequeue and updates the hrtick when the
3633 * current task is from our class and nr_running is low enough
3636 static void hrtick_update(struct rq
*rq
)
3638 struct task_struct
*curr
= rq
->curr
;
3640 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
3643 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
3644 hrtick_start_fair(rq
, curr
);
3646 #else /* !CONFIG_SCHED_HRTICK */
3648 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3652 static inline void hrtick_update(struct rq
*rq
)
3658 * The enqueue_task method is called before nr_running is
3659 * increased. Here we update the fair scheduling stats and
3660 * then put the task into the rbtree:
3663 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
3665 struct cfs_rq
*cfs_rq
;
3666 struct sched_entity
*se
= &p
->se
;
3668 for_each_sched_entity(se
) {
3671 cfs_rq
= cfs_rq_of(se
);
3672 enqueue_entity(cfs_rq
, se
, flags
);
3675 * end evaluation on encountering a throttled cfs_rq
3677 * note: in the case of encountering a throttled cfs_rq we will
3678 * post the final h_nr_running increment below.
3680 if (cfs_rq_throttled(cfs_rq
))
3682 cfs_rq
->h_nr_running
++;
3684 flags
= ENQUEUE_WAKEUP
;
3687 for_each_sched_entity(se
) {
3688 cfs_rq
= cfs_rq_of(se
);
3689 cfs_rq
->h_nr_running
++;
3691 if (cfs_rq_throttled(cfs_rq
))
3694 update_cfs_shares(cfs_rq
);
3695 update_entity_load_avg(se
, 1);
3699 update_rq_runnable_avg(rq
, rq
->nr_running
);
3705 static void set_next_buddy(struct sched_entity
*se
);
3708 * The dequeue_task method is called before nr_running is
3709 * decreased. We remove the task from the rbtree and
3710 * update the fair scheduling stats:
3712 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
3714 struct cfs_rq
*cfs_rq
;
3715 struct sched_entity
*se
= &p
->se
;
3716 int task_sleep
= flags
& DEQUEUE_SLEEP
;
3718 for_each_sched_entity(se
) {
3719 cfs_rq
= cfs_rq_of(se
);
3720 dequeue_entity(cfs_rq
, se
, flags
);
3723 * end evaluation on encountering a throttled cfs_rq
3725 * note: in the case of encountering a throttled cfs_rq we will
3726 * post the final h_nr_running decrement below.
3728 if (cfs_rq_throttled(cfs_rq
))
3730 cfs_rq
->h_nr_running
--;
3732 /* Don't dequeue parent if it has other entities besides us */
3733 if (cfs_rq
->load
.weight
) {
3735 * Bias pick_next to pick a task from this cfs_rq, as
3736 * p is sleeping when it is within its sched_slice.
3738 if (task_sleep
&& parent_entity(se
))
3739 set_next_buddy(parent_entity(se
));
3741 /* avoid re-evaluating load for this entity */
3742 se
= parent_entity(se
);
3745 flags
|= DEQUEUE_SLEEP
;
3748 for_each_sched_entity(se
) {
3749 cfs_rq
= cfs_rq_of(se
);
3750 cfs_rq
->h_nr_running
--;
3752 if (cfs_rq_throttled(cfs_rq
))
3755 update_cfs_shares(cfs_rq
);
3756 update_entity_load_avg(se
, 1);
3761 update_rq_runnable_avg(rq
, 1);
3767 /* Used instead of source_load when we know the type == 0 */
3768 static unsigned long weighted_cpuload(const int cpu
)
3770 return cpu_rq(cpu
)->cfs
.runnable_load_avg
;
3774 * Return a low guess at the load of a migration-source cpu weighted
3775 * according to the scheduling class and "nice" value.
3777 * We want to under-estimate the load of migration sources, to
3778 * balance conservatively.
3780 static unsigned long source_load(int cpu
, int type
)
3782 struct rq
*rq
= cpu_rq(cpu
);
3783 unsigned long total
= weighted_cpuload(cpu
);
3785 if (type
== 0 || !sched_feat(LB_BIAS
))
3788 return min(rq
->cpu_load
[type
-1], total
);
3792 * Return a high guess at the load of a migration-target cpu weighted
3793 * according to the scheduling class and "nice" value.
3795 static unsigned long target_load(int cpu
, int type
)
3797 struct rq
*rq
= cpu_rq(cpu
);
3798 unsigned long total
= weighted_cpuload(cpu
);
3800 if (type
== 0 || !sched_feat(LB_BIAS
))
3803 return max(rq
->cpu_load
[type
-1], total
);
3806 static unsigned long power_of(int cpu
)
3808 return cpu_rq(cpu
)->cpu_power
;
3811 static unsigned long cpu_avg_load_per_task(int cpu
)
3813 struct rq
*rq
= cpu_rq(cpu
);
3814 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
3815 unsigned long load_avg
= rq
->cfs
.runnable_load_avg
;
3818 return load_avg
/ nr_running
;
3823 static void record_wakee(struct task_struct
*p
)
3826 * Rough decay (wiping) for cost saving, don't worry
3827 * about the boundary, really active task won't care
3830 if (jiffies
> current
->wakee_flip_decay_ts
+ HZ
) {
3831 current
->wakee_flips
= 0;
3832 current
->wakee_flip_decay_ts
= jiffies
;
3835 if (current
->last_wakee
!= p
) {
3836 current
->last_wakee
= p
;
3837 current
->wakee_flips
++;
3841 static void task_waking_fair(struct task_struct
*p
)
3843 struct sched_entity
*se
= &p
->se
;
3844 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3847 #ifndef CONFIG_64BIT
3848 u64 min_vruntime_copy
;
3851 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
3853 min_vruntime
= cfs_rq
->min_vruntime
;
3854 } while (min_vruntime
!= min_vruntime_copy
);
3856 min_vruntime
= cfs_rq
->min_vruntime
;
3859 se
->vruntime
-= min_vruntime
;
3863 #ifdef CONFIG_FAIR_GROUP_SCHED
3865 * effective_load() calculates the load change as seen from the root_task_group
3867 * Adding load to a group doesn't make a group heavier, but can cause movement
3868 * of group shares between cpus. Assuming the shares were perfectly aligned one
3869 * can calculate the shift in shares.
3871 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3872 * on this @cpu and results in a total addition (subtraction) of @wg to the
3873 * total group weight.
3875 * Given a runqueue weight distribution (rw_i) we can compute a shares
3876 * distribution (s_i) using:
3878 * s_i = rw_i / \Sum rw_j (1)
3880 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3881 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3882 * shares distribution (s_i):
3884 * rw_i = { 2, 4, 1, 0 }
3885 * s_i = { 2/7, 4/7, 1/7, 0 }
3887 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3888 * task used to run on and the CPU the waker is running on), we need to
3889 * compute the effect of waking a task on either CPU and, in case of a sync
3890 * wakeup, compute the effect of the current task going to sleep.
3892 * So for a change of @wl to the local @cpu with an overall group weight change
3893 * of @wl we can compute the new shares distribution (s'_i) using:
3895 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3897 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3898 * differences in waking a task to CPU 0. The additional task changes the
3899 * weight and shares distributions like:
3901 * rw'_i = { 3, 4, 1, 0 }
3902 * s'_i = { 3/8, 4/8, 1/8, 0 }
3904 * We can then compute the difference in effective weight by using:
3906 * dw_i = S * (s'_i - s_i) (3)
3908 * Where 'S' is the group weight as seen by its parent.
3910 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3911 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3912 * 4/7) times the weight of the group.
3914 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
3916 struct sched_entity
*se
= tg
->se
[cpu
];
3918 if (!tg
->parent
|| !wl
) /* the trivial, non-cgroup case */
3921 for_each_sched_entity(se
) {
3927 * W = @wg + \Sum rw_j
3929 W
= wg
+ calc_tg_weight(tg
, se
->my_q
);
3934 w
= se
->my_q
->load
.weight
+ wl
;
3937 * wl = S * s'_i; see (2)
3940 wl
= (w
* tg
->shares
) / W
;
3945 * Per the above, wl is the new se->load.weight value; since
3946 * those are clipped to [MIN_SHARES, ...) do so now. See
3947 * calc_cfs_shares().
3949 if (wl
< MIN_SHARES
)
3953 * wl = dw_i = S * (s'_i - s_i); see (3)
3955 wl
-= se
->load
.weight
;
3958 * Recursively apply this logic to all parent groups to compute
3959 * the final effective load change on the root group. Since
3960 * only the @tg group gets extra weight, all parent groups can
3961 * only redistribute existing shares. @wl is the shift in shares
3962 * resulting from this level per the above.
3971 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
3978 static int wake_wide(struct task_struct
*p
)
3980 int factor
= this_cpu_read(sd_llc_size
);
3983 * Yeah, it's the switching-frequency, could means many wakee or
3984 * rapidly switch, use factor here will just help to automatically
3985 * adjust the loose-degree, so bigger node will lead to more pull.
3987 if (p
->wakee_flips
> factor
) {
3989 * wakee is somewhat hot, it needs certain amount of cpu
3990 * resource, so if waker is far more hot, prefer to leave
3993 if (current
->wakee_flips
> (factor
* p
->wakee_flips
))
4000 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
4002 s64 this_load
, load
;
4003 int idx
, this_cpu
, prev_cpu
;
4004 unsigned long tl_per_task
;
4005 struct task_group
*tg
;
4006 unsigned long weight
;
4010 * If we wake multiple tasks be careful to not bounce
4011 * ourselves around too much.
4017 this_cpu
= smp_processor_id();
4018 prev_cpu
= task_cpu(p
);
4019 load
= source_load(prev_cpu
, idx
);
4020 this_load
= target_load(this_cpu
, idx
);
4023 * If sync wakeup then subtract the (maximum possible)
4024 * effect of the currently running task from the load
4025 * of the current CPU:
4028 tg
= task_group(current
);
4029 weight
= current
->se
.load
.weight
;
4031 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
4032 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
4036 weight
= p
->se
.load
.weight
;
4039 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4040 * due to the sync cause above having dropped this_load to 0, we'll
4041 * always have an imbalance, but there's really nothing you can do
4042 * about that, so that's good too.
4044 * Otherwise check if either cpus are near enough in load to allow this
4045 * task to be woken on this_cpu.
4047 if (this_load
> 0) {
4048 s64 this_eff_load
, prev_eff_load
;
4050 this_eff_load
= 100;
4051 this_eff_load
*= power_of(prev_cpu
);
4052 this_eff_load
*= this_load
+
4053 effective_load(tg
, this_cpu
, weight
, weight
);
4055 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
4056 prev_eff_load
*= power_of(this_cpu
);
4057 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
4059 balanced
= this_eff_load
<= prev_eff_load
;
4064 * If the currently running task will sleep within
4065 * a reasonable amount of time then attract this newly
4068 if (sync
&& balanced
)
4071 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
4072 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
4075 (this_load
<= load
&&
4076 this_load
+ target_load(prev_cpu
, idx
) <= tl_per_task
)) {
4078 * This domain has SD_WAKE_AFFINE and
4079 * p is cache cold in this domain, and
4080 * there is no bad imbalance.
4082 schedstat_inc(sd
, ttwu_move_affine
);
4083 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
4091 * find_idlest_group finds and returns the least busy CPU group within the
4094 static struct sched_group
*
4095 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
4096 int this_cpu
, int load_idx
)
4098 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
4099 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
4100 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
4103 unsigned long load
, avg_load
;
4107 /* Skip over this group if it has no CPUs allowed */
4108 if (!cpumask_intersects(sched_group_cpus(group
),
4109 tsk_cpus_allowed(p
)))
4112 local_group
= cpumask_test_cpu(this_cpu
,
4113 sched_group_cpus(group
));
4115 /* Tally up the load of all CPUs in the group */
4118 for_each_cpu(i
, sched_group_cpus(group
)) {
4119 /* Bias balancing toward cpus of our domain */
4121 load
= source_load(i
, load_idx
);
4123 load
= target_load(i
, load_idx
);
4128 /* Adjust by relative CPU power of the group */
4129 avg_load
= (avg_load
* SCHED_POWER_SCALE
) / group
->sgp
->power
;
4132 this_load
= avg_load
;
4133 } else if (avg_load
< min_load
) {
4134 min_load
= avg_load
;
4137 } while (group
= group
->next
, group
!= sd
->groups
);
4139 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
4145 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4148 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
4150 unsigned long load
, min_load
= ULONG_MAX
;
4154 /* Traverse only the allowed CPUs */
4155 for_each_cpu_and(i
, sched_group_cpus(group
), tsk_cpus_allowed(p
)) {
4156 load
= weighted_cpuload(i
);
4158 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
4168 * Try and locate an idle CPU in the sched_domain.
4170 static int select_idle_sibling(struct task_struct
*p
, int target
)
4172 struct sched_domain
*sd
;
4173 struct sched_group
*sg
;
4174 int i
= task_cpu(p
);
4176 if (idle_cpu(target
))
4180 * If the prevous cpu is cache affine and idle, don't be stupid.
4182 if (i
!= target
&& cpus_share_cache(i
, target
) && idle_cpu(i
))
4186 * Otherwise, iterate the domains and find an elegible idle cpu.
4188 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
4189 for_each_lower_domain(sd
) {
4192 if (!cpumask_intersects(sched_group_cpus(sg
),
4193 tsk_cpus_allowed(p
)))
4196 for_each_cpu(i
, sched_group_cpus(sg
)) {
4197 if (i
== target
|| !idle_cpu(i
))
4201 target
= cpumask_first_and(sched_group_cpus(sg
),
4202 tsk_cpus_allowed(p
));
4206 } while (sg
!= sd
->groups
);
4213 * sched_balance_self: balance the current task (running on cpu) in domains
4214 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
4217 * Balance, ie. select the least loaded group.
4219 * Returns the target CPU number, or the same CPU if no balancing is needed.
4221 * preempt must be disabled.
4224 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
4226 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
4227 int cpu
= smp_processor_id();
4229 int want_affine
= 0;
4230 int sync
= wake_flags
& WF_SYNC
;
4232 if (p
->nr_cpus_allowed
== 1)
4235 if (sd_flag
& SD_BALANCE_WAKE
) {
4236 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)))
4242 for_each_domain(cpu
, tmp
) {
4243 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
4247 * If both cpu and prev_cpu are part of this domain,
4248 * cpu is a valid SD_WAKE_AFFINE target.
4250 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
4251 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
4256 if (tmp
->flags
& sd_flag
)
4261 if (cpu
!= prev_cpu
&& wake_affine(affine_sd
, p
, sync
))
4264 new_cpu
= select_idle_sibling(p
, prev_cpu
);
4269 int load_idx
= sd
->forkexec_idx
;
4270 struct sched_group
*group
;
4273 if (!(sd
->flags
& sd_flag
)) {
4278 if (sd_flag
& SD_BALANCE_WAKE
)
4279 load_idx
= sd
->wake_idx
;
4281 group
= find_idlest_group(sd
, p
, cpu
, load_idx
);
4287 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
4288 if (new_cpu
== -1 || new_cpu
== cpu
) {
4289 /* Now try balancing at a lower domain level of cpu */
4294 /* Now try balancing at a lower domain level of new_cpu */
4296 weight
= sd
->span_weight
;
4298 for_each_domain(cpu
, tmp
) {
4299 if (weight
<= tmp
->span_weight
)
4301 if (tmp
->flags
& sd_flag
)
4304 /* while loop will break here if sd == NULL */
4313 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4314 * cfs_rq_of(p) references at time of call are still valid and identify the
4315 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4316 * other assumptions, including the state of rq->lock, should be made.
4319 migrate_task_rq_fair(struct task_struct
*p
, int next_cpu
)
4321 struct sched_entity
*se
= &p
->se
;
4322 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4325 * Load tracking: accumulate removed load so that it can be processed
4326 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4327 * to blocked load iff they have a positive decay-count. It can never
4328 * be negative here since on-rq tasks have decay-count == 0.
4330 if (se
->avg
.decay_count
) {
4331 se
->avg
.decay_count
= -__synchronize_entity_decay(se
);
4332 atomic_long_add(se
->avg
.load_avg_contrib
,
4333 &cfs_rq
->removed_load
);
4336 #endif /* CONFIG_SMP */
4338 static unsigned long
4339 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
4341 unsigned long gran
= sysctl_sched_wakeup_granularity
;
4344 * Since its curr running now, convert the gran from real-time
4345 * to virtual-time in his units.
4347 * By using 'se' instead of 'curr' we penalize light tasks, so
4348 * they get preempted easier. That is, if 'se' < 'curr' then
4349 * the resulting gran will be larger, therefore penalizing the
4350 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4351 * be smaller, again penalizing the lighter task.
4353 * This is especially important for buddies when the leftmost
4354 * task is higher priority than the buddy.
4356 return calc_delta_fair(gran
, se
);
4360 * Should 'se' preempt 'curr'.
4374 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
4376 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
4381 gran
= wakeup_gran(curr
, se
);
4388 static void set_last_buddy(struct sched_entity
*se
)
4390 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4393 for_each_sched_entity(se
)
4394 cfs_rq_of(se
)->last
= se
;
4397 static void set_next_buddy(struct sched_entity
*se
)
4399 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4402 for_each_sched_entity(se
)
4403 cfs_rq_of(se
)->next
= se
;
4406 static void set_skip_buddy(struct sched_entity
*se
)
4408 for_each_sched_entity(se
)
4409 cfs_rq_of(se
)->skip
= se
;
4413 * Preempt the current task with a newly woken task if needed:
4415 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
4417 struct task_struct
*curr
= rq
->curr
;
4418 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
4419 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4420 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
4421 int next_buddy_marked
= 0;
4423 if (unlikely(se
== pse
))
4427 * This is possible from callers such as move_task(), in which we
4428 * unconditionally check_prempt_curr() after an enqueue (which may have
4429 * lead to a throttle). This both saves work and prevents false
4430 * next-buddy nomination below.
4432 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
4435 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
4436 set_next_buddy(pse
);
4437 next_buddy_marked
= 1;
4441 * We can come here with TIF_NEED_RESCHED already set from new task
4444 * Note: this also catches the edge-case of curr being in a throttled
4445 * group (e.g. via set_curr_task), since update_curr() (in the
4446 * enqueue of curr) will have resulted in resched being set. This
4447 * prevents us from potentially nominating it as a false LAST_BUDDY
4450 if (test_tsk_need_resched(curr
))
4453 /* Idle tasks are by definition preempted by non-idle tasks. */
4454 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
4455 likely(p
->policy
!= SCHED_IDLE
))
4459 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4460 * is driven by the tick):
4462 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
4465 find_matching_se(&se
, &pse
);
4466 update_curr(cfs_rq_of(se
));
4468 if (wakeup_preempt_entity(se
, pse
) == 1) {
4470 * Bias pick_next to pick the sched entity that is
4471 * triggering this preemption.
4473 if (!next_buddy_marked
)
4474 set_next_buddy(pse
);
4483 * Only set the backward buddy when the current task is still
4484 * on the rq. This can happen when a wakeup gets interleaved
4485 * with schedule on the ->pre_schedule() or idle_balance()
4486 * point, either of which can * drop the rq lock.
4488 * Also, during early boot the idle thread is in the fair class,
4489 * for obvious reasons its a bad idea to schedule back to it.
4491 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
4494 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
4498 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
4500 struct task_struct
*p
;
4501 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
4502 struct sched_entity
*se
;
4504 if (!cfs_rq
->nr_running
)
4508 se
= pick_next_entity(cfs_rq
);
4509 set_next_entity(cfs_rq
, se
);
4510 cfs_rq
= group_cfs_rq(se
);
4514 if (hrtick_enabled(rq
))
4515 hrtick_start_fair(rq
, p
);
4521 * Account for a descheduled task:
4523 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
4525 struct sched_entity
*se
= &prev
->se
;
4526 struct cfs_rq
*cfs_rq
;
4528 for_each_sched_entity(se
) {
4529 cfs_rq
= cfs_rq_of(se
);
4530 put_prev_entity(cfs_rq
, se
);
4535 * sched_yield() is very simple
4537 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4539 static void yield_task_fair(struct rq
*rq
)
4541 struct task_struct
*curr
= rq
->curr
;
4542 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4543 struct sched_entity
*se
= &curr
->se
;
4546 * Are we the only task in the tree?
4548 if (unlikely(rq
->nr_running
== 1))
4551 clear_buddies(cfs_rq
, se
);
4553 if (curr
->policy
!= SCHED_BATCH
) {
4554 update_rq_clock(rq
);
4556 * Update run-time statistics of the 'current'.
4558 update_curr(cfs_rq
);
4560 * Tell update_rq_clock() that we've just updated,
4561 * so we don't do microscopic update in schedule()
4562 * and double the fastpath cost.
4564 rq
->skip_clock_update
= 1;
4570 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
4572 struct sched_entity
*se
= &p
->se
;
4574 /* throttled hierarchies are not runnable */
4575 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
4578 /* Tell the scheduler that we'd really like pse to run next. */
4581 yield_task_fair(rq
);
4587 /**************************************************
4588 * Fair scheduling class load-balancing methods.
4592 * The purpose of load-balancing is to achieve the same basic fairness the
4593 * per-cpu scheduler provides, namely provide a proportional amount of compute
4594 * time to each task. This is expressed in the following equation:
4596 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4598 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4599 * W_i,0 is defined as:
4601 * W_i,0 = \Sum_j w_i,j (2)
4603 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4604 * is derived from the nice value as per prio_to_weight[].
4606 * The weight average is an exponential decay average of the instantaneous
4609 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4611 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4612 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4613 * can also include other factors [XXX].
4615 * To achieve this balance we define a measure of imbalance which follows
4616 * directly from (1):
4618 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4620 * We them move tasks around to minimize the imbalance. In the continuous
4621 * function space it is obvious this converges, in the discrete case we get
4622 * a few fun cases generally called infeasible weight scenarios.
4625 * - infeasible weights;
4626 * - local vs global optima in the discrete case. ]
4631 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4632 * for all i,j solution, we create a tree of cpus that follows the hardware
4633 * topology where each level pairs two lower groups (or better). This results
4634 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4635 * tree to only the first of the previous level and we decrease the frequency
4636 * of load-balance at each level inv. proportional to the number of cpus in
4642 * \Sum { --- * --- * 2^i } = O(n) (5)
4644 * `- size of each group
4645 * | | `- number of cpus doing load-balance
4647 * `- sum over all levels
4649 * Coupled with a limit on how many tasks we can migrate every balance pass,
4650 * this makes (5) the runtime complexity of the balancer.
4652 * An important property here is that each CPU is still (indirectly) connected
4653 * to every other cpu in at most O(log n) steps:
4655 * The adjacency matrix of the resulting graph is given by:
4658 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4661 * And you'll find that:
4663 * A^(log_2 n)_i,j != 0 for all i,j (7)
4665 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4666 * The task movement gives a factor of O(m), giving a convergence complexity
4669 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4674 * In order to avoid CPUs going idle while there's still work to do, new idle
4675 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4676 * tree itself instead of relying on other CPUs to bring it work.
4678 * This adds some complexity to both (5) and (8) but it reduces the total idle
4686 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4689 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4694 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4696 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4698 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4701 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4702 * rewrite all of this once again.]
4705 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
4707 enum fbq_type
{ regular
, remote
, all
};
4709 #define LBF_ALL_PINNED 0x01
4710 #define LBF_NEED_BREAK 0x02
4711 #define LBF_DST_PINNED 0x04
4712 #define LBF_SOME_PINNED 0x08
4715 struct sched_domain
*sd
;
4723 struct cpumask
*dst_grpmask
;
4725 enum cpu_idle_type idle
;
4727 /* The set of CPUs under consideration for load-balancing */
4728 struct cpumask
*cpus
;
4733 unsigned int loop_break
;
4734 unsigned int loop_max
;
4736 enum fbq_type fbq_type
;
4740 * move_task - move a task from one runqueue to another runqueue.
4741 * Both runqueues must be locked.
4743 static void move_task(struct task_struct
*p
, struct lb_env
*env
)
4745 deactivate_task(env
->src_rq
, p
, 0);
4746 set_task_cpu(p
, env
->dst_cpu
);
4747 activate_task(env
->dst_rq
, p
, 0);
4748 check_preempt_curr(env
->dst_rq
, p
, 0);
4749 #ifdef CONFIG_NUMA_BALANCING
4750 if (p
->numa_preferred_nid
!= -1) {
4751 int src_nid
= cpu_to_node(env
->src_cpu
);
4752 int dst_nid
= cpu_to_node(env
->dst_cpu
);
4755 * If the load balancer has moved the task then limit
4756 * migrations from taking place in the short term in
4757 * case this is a short-lived migration.
4759 if (src_nid
!= dst_nid
&& dst_nid
!= p
->numa_preferred_nid
)
4760 p
->numa_migrate_seq
= 0;
4766 * Is this task likely cache-hot:
4769 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
4773 if (p
->sched_class
!= &fair_sched_class
)
4776 if (unlikely(p
->policy
== SCHED_IDLE
))
4780 * Buddy candidates are cache hot:
4782 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
4783 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
4784 &p
->se
== cfs_rq_of(&p
->se
)->last
))
4787 if (sysctl_sched_migration_cost
== -1)
4789 if (sysctl_sched_migration_cost
== 0)
4792 delta
= now
- p
->se
.exec_start
;
4794 return delta
< (s64
)sysctl_sched_migration_cost
;
4797 #ifdef CONFIG_NUMA_BALANCING
4798 /* Returns true if the destination node has incurred more faults */
4799 static bool migrate_improves_locality(struct task_struct
*p
, struct lb_env
*env
)
4801 int src_nid
, dst_nid
;
4803 if (!sched_feat(NUMA_FAVOUR_HIGHER
) || !p
->numa_faults
||
4804 !(env
->sd
->flags
& SD_NUMA
)) {
4808 src_nid
= cpu_to_node(env
->src_cpu
);
4809 dst_nid
= cpu_to_node(env
->dst_cpu
);
4811 if (src_nid
== dst_nid
)
4814 /* Always encourage migration to the preferred node. */
4815 if (dst_nid
== p
->numa_preferred_nid
)
4818 /* If both task and group weight improve, this move is a winner. */
4819 if (task_weight(p
, dst_nid
) > task_weight(p
, src_nid
) &&
4820 group_weight(p
, dst_nid
) > group_weight(p
, src_nid
))
4827 static bool migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
4829 int src_nid
, dst_nid
;
4831 if (!sched_feat(NUMA
) || !sched_feat(NUMA_RESIST_LOWER
))
4834 if (!p
->numa_faults
|| !(env
->sd
->flags
& SD_NUMA
))
4837 src_nid
= cpu_to_node(env
->src_cpu
);
4838 dst_nid
= cpu_to_node(env
->dst_cpu
);
4840 if (src_nid
== dst_nid
)
4843 /* Migrating away from the preferred node is always bad. */
4844 if (src_nid
== p
->numa_preferred_nid
)
4847 /* If either task or group weight get worse, don't do it. */
4848 if (task_weight(p
, dst_nid
) < task_weight(p
, src_nid
) ||
4849 group_weight(p
, dst_nid
) < group_weight(p
, src_nid
))
4856 static inline bool migrate_improves_locality(struct task_struct
*p
,
4862 static inline bool migrate_degrades_locality(struct task_struct
*p
,
4870 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4873 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
4875 int tsk_cache_hot
= 0;
4877 * We do not migrate tasks that are:
4878 * 1) throttled_lb_pair, or
4879 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4880 * 3) running (obviously), or
4881 * 4) are cache-hot on their current CPU.
4883 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
4886 if (!cpumask_test_cpu(env
->dst_cpu
, tsk_cpus_allowed(p
))) {
4889 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
4891 env
->flags
|= LBF_SOME_PINNED
;
4894 * Remember if this task can be migrated to any other cpu in
4895 * our sched_group. We may want to revisit it if we couldn't
4896 * meet load balance goals by pulling other tasks on src_cpu.
4898 * Also avoid computing new_dst_cpu if we have already computed
4899 * one in current iteration.
4901 if (!env
->dst_grpmask
|| (env
->flags
& LBF_DST_PINNED
))
4904 /* Prevent to re-select dst_cpu via env's cpus */
4905 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
4906 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
))) {
4907 env
->flags
|= LBF_DST_PINNED
;
4908 env
->new_dst_cpu
= cpu
;
4916 /* Record that we found atleast one task that could run on dst_cpu */
4917 env
->flags
&= ~LBF_ALL_PINNED
;
4919 if (task_running(env
->src_rq
, p
)) {
4920 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
4925 * Aggressive migration if:
4926 * 1) destination numa is preferred
4927 * 2) task is cache cold, or
4928 * 3) too many balance attempts have failed.
4930 tsk_cache_hot
= task_hot(p
, rq_clock_task(env
->src_rq
), env
->sd
);
4932 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
4934 if (migrate_improves_locality(p
, env
)) {
4935 #ifdef CONFIG_SCHEDSTATS
4936 if (tsk_cache_hot
) {
4937 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
4938 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
4944 if (!tsk_cache_hot
||
4945 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
4947 if (tsk_cache_hot
) {
4948 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
4949 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
4955 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
4960 * move_one_task tries to move exactly one task from busiest to this_rq, as
4961 * part of active balancing operations within "domain".
4962 * Returns 1 if successful and 0 otherwise.
4964 * Called with both runqueues locked.
4966 static int move_one_task(struct lb_env
*env
)
4968 struct task_struct
*p
, *n
;
4970 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
4971 if (!can_migrate_task(p
, env
))
4976 * Right now, this is only the second place move_task()
4977 * is called, so we can safely collect move_task()
4978 * stats here rather than inside move_task().
4980 schedstat_inc(env
->sd
, lb_gained
[env
->idle
]);
4986 static const unsigned int sched_nr_migrate_break
= 32;
4989 * move_tasks tries to move up to imbalance weighted load from busiest to
4990 * this_rq, as part of a balancing operation within domain "sd".
4991 * Returns 1 if successful and 0 otherwise.
4993 * Called with both runqueues locked.
4995 static int move_tasks(struct lb_env
*env
)
4997 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
4998 struct task_struct
*p
;
5002 if (env
->imbalance
<= 0)
5005 while (!list_empty(tasks
)) {
5006 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
5009 /* We've more or less seen every task there is, call it quits */
5010 if (env
->loop
> env
->loop_max
)
5013 /* take a breather every nr_migrate tasks */
5014 if (env
->loop
> env
->loop_break
) {
5015 env
->loop_break
+= sched_nr_migrate_break
;
5016 env
->flags
|= LBF_NEED_BREAK
;
5020 if (!can_migrate_task(p
, env
))
5023 load
= task_h_load(p
);
5025 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
5028 if ((load
/ 2) > env
->imbalance
)
5033 env
->imbalance
-= load
;
5035 #ifdef CONFIG_PREEMPT
5037 * NEWIDLE balancing is a source of latency, so preemptible
5038 * kernels will stop after the first task is pulled to minimize
5039 * the critical section.
5041 if (env
->idle
== CPU_NEWLY_IDLE
)
5046 * We only want to steal up to the prescribed amount of
5049 if (env
->imbalance
<= 0)
5054 list_move_tail(&p
->se
.group_node
, tasks
);
5058 * Right now, this is one of only two places move_task() is called,
5059 * so we can safely collect move_task() stats here rather than
5060 * inside move_task().
5062 schedstat_add(env
->sd
, lb_gained
[env
->idle
], pulled
);
5067 #ifdef CONFIG_FAIR_GROUP_SCHED
5069 * update tg->load_weight by folding this cpu's load_avg
5071 static void __update_blocked_averages_cpu(struct task_group
*tg
, int cpu
)
5073 struct sched_entity
*se
= tg
->se
[cpu
];
5074 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu
];
5076 /* throttled entities do not contribute to load */
5077 if (throttled_hierarchy(cfs_rq
))
5080 update_cfs_rq_blocked_load(cfs_rq
, 1);
5083 update_entity_load_avg(se
, 1);
5085 * We pivot on our runnable average having decayed to zero for
5086 * list removal. This generally implies that all our children
5087 * have also been removed (modulo rounding error or bandwidth
5088 * control); however, such cases are rare and we can fix these
5091 * TODO: fix up out-of-order children on enqueue.
5093 if (!se
->avg
.runnable_avg_sum
&& !cfs_rq
->nr_running
)
5094 list_del_leaf_cfs_rq(cfs_rq
);
5096 struct rq
*rq
= rq_of(cfs_rq
);
5097 update_rq_runnable_avg(rq
, rq
->nr_running
);
5101 static void update_blocked_averages(int cpu
)
5103 struct rq
*rq
= cpu_rq(cpu
);
5104 struct cfs_rq
*cfs_rq
;
5105 unsigned long flags
;
5107 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5108 update_rq_clock(rq
);
5110 * Iterates the task_group tree in a bottom up fashion, see
5111 * list_add_leaf_cfs_rq() for details.
5113 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
5115 * Note: We may want to consider periodically releasing
5116 * rq->lock about these updates so that creating many task
5117 * groups does not result in continually extending hold time.
5119 __update_blocked_averages_cpu(cfs_rq
->tg
, rq
->cpu
);
5122 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5126 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5127 * This needs to be done in a top-down fashion because the load of a child
5128 * group is a fraction of its parents load.
5130 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
5132 struct rq
*rq
= rq_of(cfs_rq
);
5133 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
5134 unsigned long now
= jiffies
;
5137 if (cfs_rq
->last_h_load_update
== now
)
5140 cfs_rq
->h_load_next
= NULL
;
5141 for_each_sched_entity(se
) {
5142 cfs_rq
= cfs_rq_of(se
);
5143 cfs_rq
->h_load_next
= se
;
5144 if (cfs_rq
->last_h_load_update
== now
)
5149 cfs_rq
->h_load
= cfs_rq
->runnable_load_avg
;
5150 cfs_rq
->last_h_load_update
= now
;
5153 while ((se
= cfs_rq
->h_load_next
) != NULL
) {
5154 load
= cfs_rq
->h_load
;
5155 load
= div64_ul(load
* se
->avg
.load_avg_contrib
,
5156 cfs_rq
->runnable_load_avg
+ 1);
5157 cfs_rq
= group_cfs_rq(se
);
5158 cfs_rq
->h_load
= load
;
5159 cfs_rq
->last_h_load_update
= now
;
5163 static unsigned long task_h_load(struct task_struct
*p
)
5165 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
5167 update_cfs_rq_h_load(cfs_rq
);
5168 return div64_ul(p
->se
.avg
.load_avg_contrib
* cfs_rq
->h_load
,
5169 cfs_rq
->runnable_load_avg
+ 1);
5172 static inline void update_blocked_averages(int cpu
)
5176 static unsigned long task_h_load(struct task_struct
*p
)
5178 return p
->se
.avg
.load_avg_contrib
;
5182 /********** Helpers for find_busiest_group ************************/
5184 * sg_lb_stats - stats of a sched_group required for load_balancing
5186 struct sg_lb_stats
{
5187 unsigned long avg_load
; /*Avg load across the CPUs of the group */
5188 unsigned long group_load
; /* Total load over the CPUs of the group */
5189 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
5190 unsigned long load_per_task
;
5191 unsigned long group_power
;
5192 unsigned int sum_nr_running
; /* Nr tasks running in the group */
5193 unsigned int group_capacity
;
5194 unsigned int idle_cpus
;
5195 unsigned int group_weight
;
5196 int group_imb
; /* Is there an imbalance in the group ? */
5197 int group_has_capacity
; /* Is there extra capacity in the group? */
5198 #ifdef CONFIG_NUMA_BALANCING
5199 unsigned int nr_numa_running
;
5200 unsigned int nr_preferred_running
;
5205 * sd_lb_stats - Structure to store the statistics of a sched_domain
5206 * during load balancing.
5208 struct sd_lb_stats
{
5209 struct sched_group
*busiest
; /* Busiest group in this sd */
5210 struct sched_group
*local
; /* Local group in this sd */
5211 unsigned long total_load
; /* Total load of all groups in sd */
5212 unsigned long total_pwr
; /* Total power of all groups in sd */
5213 unsigned long avg_load
; /* Average load across all groups in sd */
5215 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
5216 struct sg_lb_stats local_stat
; /* Statistics of the local group */
5219 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
5222 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5223 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5224 * We must however clear busiest_stat::avg_load because
5225 * update_sd_pick_busiest() reads this before assignment.
5227 *sds
= (struct sd_lb_stats
){
5239 * get_sd_load_idx - Obtain the load index for a given sched domain.
5240 * @sd: The sched_domain whose load_idx is to be obtained.
5241 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
5243 * Return: The load index.
5245 static inline int get_sd_load_idx(struct sched_domain
*sd
,
5246 enum cpu_idle_type idle
)
5252 load_idx
= sd
->busy_idx
;
5255 case CPU_NEWLY_IDLE
:
5256 load_idx
= sd
->newidle_idx
;
5259 load_idx
= sd
->idle_idx
;
5266 static unsigned long default_scale_freq_power(struct sched_domain
*sd
, int cpu
)
5268 return SCHED_POWER_SCALE
;
5271 unsigned long __weak
arch_scale_freq_power(struct sched_domain
*sd
, int cpu
)
5273 return default_scale_freq_power(sd
, cpu
);
5276 static unsigned long default_scale_smt_power(struct sched_domain
*sd
, int cpu
)
5278 unsigned long weight
= sd
->span_weight
;
5279 unsigned long smt_gain
= sd
->smt_gain
;
5286 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
5288 return default_scale_smt_power(sd
, cpu
);
5291 static unsigned long scale_rt_power(int cpu
)
5293 struct rq
*rq
= cpu_rq(cpu
);
5294 u64 total
, available
, age_stamp
, avg
;
5297 * Since we're reading these variables without serialization make sure
5298 * we read them once before doing sanity checks on them.
5300 age_stamp
= ACCESS_ONCE(rq
->age_stamp
);
5301 avg
= ACCESS_ONCE(rq
->rt_avg
);
5303 total
= sched_avg_period() + (rq_clock(rq
) - age_stamp
);
5305 if (unlikely(total
< avg
)) {
5306 /* Ensures that power won't end up being negative */
5309 available
= total
- avg
;
5312 if (unlikely((s64
)total
< SCHED_POWER_SCALE
))
5313 total
= SCHED_POWER_SCALE
;
5315 total
>>= SCHED_POWER_SHIFT
;
5317 return div_u64(available
, total
);
5320 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
5322 unsigned long weight
= sd
->span_weight
;
5323 unsigned long power
= SCHED_POWER_SCALE
;
5324 struct sched_group
*sdg
= sd
->groups
;
5326 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
5327 if (sched_feat(ARCH_POWER
))
5328 power
*= arch_scale_smt_power(sd
, cpu
);
5330 power
*= default_scale_smt_power(sd
, cpu
);
5332 power
>>= SCHED_POWER_SHIFT
;
5335 sdg
->sgp
->power_orig
= power
;
5337 if (sched_feat(ARCH_POWER
))
5338 power
*= arch_scale_freq_power(sd
, cpu
);
5340 power
*= default_scale_freq_power(sd
, cpu
);
5342 power
>>= SCHED_POWER_SHIFT
;
5344 power
*= scale_rt_power(cpu
);
5345 power
>>= SCHED_POWER_SHIFT
;
5350 cpu_rq(cpu
)->cpu_power
= power
;
5351 sdg
->sgp
->power
= power
;
5354 void update_group_power(struct sched_domain
*sd
, int cpu
)
5356 struct sched_domain
*child
= sd
->child
;
5357 struct sched_group
*group
, *sdg
= sd
->groups
;
5358 unsigned long power
, power_orig
;
5359 unsigned long interval
;
5361 interval
= msecs_to_jiffies(sd
->balance_interval
);
5362 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
5363 sdg
->sgp
->next_update
= jiffies
+ interval
;
5366 update_cpu_power(sd
, cpu
);
5370 power_orig
= power
= 0;
5372 if (child
->flags
& SD_OVERLAP
) {
5374 * SD_OVERLAP domains cannot assume that child groups
5375 * span the current group.
5378 for_each_cpu(cpu
, sched_group_cpus(sdg
)) {
5379 struct sched_group
*sg
= cpu_rq(cpu
)->sd
->groups
;
5381 power_orig
+= sg
->sgp
->power_orig
;
5382 power
+= sg
->sgp
->power
;
5386 * !SD_OVERLAP domains can assume that child groups
5387 * span the current group.
5390 group
= child
->groups
;
5392 power_orig
+= group
->sgp
->power_orig
;
5393 power
+= group
->sgp
->power
;
5394 group
= group
->next
;
5395 } while (group
!= child
->groups
);
5398 sdg
->sgp
->power_orig
= power_orig
;
5399 sdg
->sgp
->power
= power
;
5403 * Try and fix up capacity for tiny siblings, this is needed when
5404 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5405 * which on its own isn't powerful enough.
5407 * See update_sd_pick_busiest() and check_asym_packing().
5410 fix_small_capacity(struct sched_domain
*sd
, struct sched_group
*group
)
5413 * Only siblings can have significantly less than SCHED_POWER_SCALE
5415 if (!(sd
->flags
& SD_SHARE_CPUPOWER
))
5419 * If ~90% of the cpu_power is still there, we're good.
5421 if (group
->sgp
->power
* 32 > group
->sgp
->power_orig
* 29)
5428 * Group imbalance indicates (and tries to solve) the problem where balancing
5429 * groups is inadequate due to tsk_cpus_allowed() constraints.
5431 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5432 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5435 * { 0 1 2 3 } { 4 5 6 7 }
5438 * If we were to balance group-wise we'd place two tasks in the first group and
5439 * two tasks in the second group. Clearly this is undesired as it will overload
5440 * cpu 3 and leave one of the cpus in the second group unused.
5442 * The current solution to this issue is detecting the skew in the first group
5443 * by noticing the lower domain failed to reach balance and had difficulty
5444 * moving tasks due to affinity constraints.
5446 * When this is so detected; this group becomes a candidate for busiest; see
5447 * update_sd_pick_busiest(). And calculcate_imbalance() and
5448 * find_busiest_group() avoid some of the usual balance conditions to allow it
5449 * to create an effective group imbalance.
5451 * This is a somewhat tricky proposition since the next run might not find the
5452 * group imbalance and decide the groups need to be balanced again. A most
5453 * subtle and fragile situation.
5456 static inline int sg_imbalanced(struct sched_group
*group
)
5458 return group
->sgp
->imbalance
;
5462 * Compute the group capacity.
5464 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5465 * first dividing out the smt factor and computing the actual number of cores
5466 * and limit power unit capacity with that.
5468 static inline int sg_capacity(struct lb_env
*env
, struct sched_group
*group
)
5470 unsigned int capacity
, smt
, cpus
;
5471 unsigned int power
, power_orig
;
5473 power
= group
->sgp
->power
;
5474 power_orig
= group
->sgp
->power_orig
;
5475 cpus
= group
->group_weight
;
5477 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5478 smt
= DIV_ROUND_UP(SCHED_POWER_SCALE
* cpus
, power_orig
);
5479 capacity
= cpus
/ smt
; /* cores */
5481 capacity
= min_t(unsigned, capacity
, DIV_ROUND_CLOSEST(power
, SCHED_POWER_SCALE
));
5483 capacity
= fix_small_capacity(env
->sd
, group
);
5489 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5490 * @env: The load balancing environment.
5491 * @group: sched_group whose statistics are to be updated.
5492 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5493 * @local_group: Does group contain this_cpu.
5494 * @sgs: variable to hold the statistics for this group.
5496 static inline void update_sg_lb_stats(struct lb_env
*env
,
5497 struct sched_group
*group
, int load_idx
,
5498 int local_group
, struct sg_lb_stats
*sgs
)
5500 unsigned long nr_running
;
5504 memset(sgs
, 0, sizeof(*sgs
));
5506 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
5507 struct rq
*rq
= cpu_rq(i
);
5509 nr_running
= rq
->nr_running
;
5511 /* Bias balancing toward cpus of our domain */
5513 load
= target_load(i
, load_idx
);
5515 load
= source_load(i
, load_idx
);
5517 sgs
->group_load
+= load
;
5518 sgs
->sum_nr_running
+= nr_running
;
5519 #ifdef CONFIG_NUMA_BALANCING
5520 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
5521 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
5523 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
5528 /* Adjust by relative CPU power of the group */
5529 sgs
->group_power
= group
->sgp
->power
;
5530 sgs
->avg_load
= (sgs
->group_load
*SCHED_POWER_SCALE
) / sgs
->group_power
;
5532 if (sgs
->sum_nr_running
)
5533 sgs
->load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
5535 sgs
->group_weight
= group
->group_weight
;
5537 sgs
->group_imb
= sg_imbalanced(group
);
5538 sgs
->group_capacity
= sg_capacity(env
, group
);
5540 if (sgs
->group_capacity
> sgs
->sum_nr_running
)
5541 sgs
->group_has_capacity
= 1;
5545 * update_sd_pick_busiest - return 1 on busiest group
5546 * @env: The load balancing environment.
5547 * @sds: sched_domain statistics
5548 * @sg: sched_group candidate to be checked for being the busiest
5549 * @sgs: sched_group statistics
5551 * Determine if @sg is a busier group than the previously selected
5554 * Return: %true if @sg is a busier group than the previously selected
5555 * busiest group. %false otherwise.
5557 static bool update_sd_pick_busiest(struct lb_env
*env
,
5558 struct sd_lb_stats
*sds
,
5559 struct sched_group
*sg
,
5560 struct sg_lb_stats
*sgs
)
5562 if (sgs
->avg_load
<= sds
->busiest_stat
.avg_load
)
5565 if (sgs
->sum_nr_running
> sgs
->group_capacity
)
5572 * ASYM_PACKING needs to move all the work to the lowest
5573 * numbered CPUs in the group, therefore mark all groups
5574 * higher than ourself as busy.
5576 if ((env
->sd
->flags
& SD_ASYM_PACKING
) && sgs
->sum_nr_running
&&
5577 env
->dst_cpu
< group_first_cpu(sg
)) {
5581 if (group_first_cpu(sds
->busiest
) > group_first_cpu(sg
))
5588 #ifdef CONFIG_NUMA_BALANCING
5589 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
5591 if (sgs
->sum_nr_running
> sgs
->nr_numa_running
)
5593 if (sgs
->sum_nr_running
> sgs
->nr_preferred_running
)
5598 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
5600 if (rq
->nr_running
> rq
->nr_numa_running
)
5602 if (rq
->nr_running
> rq
->nr_preferred_running
)
5607 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
5612 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
5616 #endif /* CONFIG_NUMA_BALANCING */
5619 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5620 * @env: The load balancing environment.
5621 * @balance: Should we balance.
5622 * @sds: variable to hold the statistics for this sched_domain.
5624 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
5626 struct sched_domain
*child
= env
->sd
->child
;
5627 struct sched_group
*sg
= env
->sd
->groups
;
5628 struct sg_lb_stats tmp_sgs
;
5629 int load_idx
, prefer_sibling
= 0;
5631 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
5634 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
5637 struct sg_lb_stats
*sgs
= &tmp_sgs
;
5640 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_cpus(sg
));
5643 sgs
= &sds
->local_stat
;
5645 if (env
->idle
!= CPU_NEWLY_IDLE
||
5646 time_after_eq(jiffies
, sg
->sgp
->next_update
))
5647 update_group_power(env
->sd
, env
->dst_cpu
);
5650 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, sgs
);
5656 * In case the child domain prefers tasks go to siblings
5657 * first, lower the sg capacity to one so that we'll try
5658 * and move all the excess tasks away. We lower the capacity
5659 * of a group only if the local group has the capacity to fit
5660 * these excess tasks, i.e. nr_running < group_capacity. The
5661 * extra check prevents the case where you always pull from the
5662 * heaviest group when it is already under-utilized (possible
5663 * with a large weight task outweighs the tasks on the system).
5665 if (prefer_sibling
&& sds
->local
&&
5666 sds
->local_stat
.group_has_capacity
)
5667 sgs
->group_capacity
= min(sgs
->group_capacity
, 1U);
5669 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
5671 sds
->busiest_stat
= *sgs
;
5675 /* Now, start updating sd_lb_stats */
5676 sds
->total_load
+= sgs
->group_load
;
5677 sds
->total_pwr
+= sgs
->group_power
;
5680 } while (sg
!= env
->sd
->groups
);
5682 if (env
->sd
->flags
& SD_NUMA
)
5683 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
5687 * check_asym_packing - Check to see if the group is packed into the
5690 * This is primarily intended to used at the sibling level. Some
5691 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5692 * case of POWER7, it can move to lower SMT modes only when higher
5693 * threads are idle. When in lower SMT modes, the threads will
5694 * perform better since they share less core resources. Hence when we
5695 * have idle threads, we want them to be the higher ones.
5697 * This packing function is run on idle threads. It checks to see if
5698 * the busiest CPU in this domain (core in the P7 case) has a higher
5699 * CPU number than the packing function is being run on. Here we are
5700 * assuming lower CPU number will be equivalent to lower a SMT thread
5703 * Return: 1 when packing is required and a task should be moved to
5704 * this CPU. The amount of the imbalance is returned in *imbalance.
5706 * @env: The load balancing environment.
5707 * @sds: Statistics of the sched_domain which is to be packed
5709 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
5713 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
5719 busiest_cpu
= group_first_cpu(sds
->busiest
);
5720 if (env
->dst_cpu
> busiest_cpu
)
5723 env
->imbalance
= DIV_ROUND_CLOSEST(
5724 sds
->busiest_stat
.avg_load
* sds
->busiest_stat
.group_power
,
5731 * fix_small_imbalance - Calculate the minor imbalance that exists
5732 * amongst the groups of a sched_domain, during
5734 * @env: The load balancing environment.
5735 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5738 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
5740 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
5741 unsigned int imbn
= 2;
5742 unsigned long scaled_busy_load_per_task
;
5743 struct sg_lb_stats
*local
, *busiest
;
5745 local
= &sds
->local_stat
;
5746 busiest
= &sds
->busiest_stat
;
5748 if (!local
->sum_nr_running
)
5749 local
->load_per_task
= cpu_avg_load_per_task(env
->dst_cpu
);
5750 else if (busiest
->load_per_task
> local
->load_per_task
)
5753 scaled_busy_load_per_task
=
5754 (busiest
->load_per_task
* SCHED_POWER_SCALE
) /
5755 busiest
->group_power
;
5757 if (busiest
->avg_load
+ scaled_busy_load_per_task
>=
5758 local
->avg_load
+ (scaled_busy_load_per_task
* imbn
)) {
5759 env
->imbalance
= busiest
->load_per_task
;
5764 * OK, we don't have enough imbalance to justify moving tasks,
5765 * however we may be able to increase total CPU power used by
5769 pwr_now
+= busiest
->group_power
*
5770 min(busiest
->load_per_task
, busiest
->avg_load
);
5771 pwr_now
+= local
->group_power
*
5772 min(local
->load_per_task
, local
->avg_load
);
5773 pwr_now
/= SCHED_POWER_SCALE
;
5775 /* Amount of load we'd subtract */
5776 tmp
= (busiest
->load_per_task
* SCHED_POWER_SCALE
) /
5777 busiest
->group_power
;
5778 if (busiest
->avg_load
> tmp
) {
5779 pwr_move
+= busiest
->group_power
*
5780 min(busiest
->load_per_task
,
5781 busiest
->avg_load
- tmp
);
5784 /* Amount of load we'd add */
5785 if (busiest
->avg_load
* busiest
->group_power
<
5786 busiest
->load_per_task
* SCHED_POWER_SCALE
) {
5787 tmp
= (busiest
->avg_load
* busiest
->group_power
) /
5790 tmp
= (busiest
->load_per_task
* SCHED_POWER_SCALE
) /
5793 pwr_move
+= local
->group_power
*
5794 min(local
->load_per_task
, local
->avg_load
+ tmp
);
5795 pwr_move
/= SCHED_POWER_SCALE
;
5797 /* Move if we gain throughput */
5798 if (pwr_move
> pwr_now
)
5799 env
->imbalance
= busiest
->load_per_task
;
5803 * calculate_imbalance - Calculate the amount of imbalance present within the
5804 * groups of a given sched_domain during load balance.
5805 * @env: load balance environment
5806 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5808 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
5810 unsigned long max_pull
, load_above_capacity
= ~0UL;
5811 struct sg_lb_stats
*local
, *busiest
;
5813 local
= &sds
->local_stat
;
5814 busiest
= &sds
->busiest_stat
;
5816 if (busiest
->group_imb
) {
5818 * In the group_imb case we cannot rely on group-wide averages
5819 * to ensure cpu-load equilibrium, look at wider averages. XXX
5821 busiest
->load_per_task
=
5822 min(busiest
->load_per_task
, sds
->avg_load
);
5826 * In the presence of smp nice balancing, certain scenarios can have
5827 * max load less than avg load(as we skip the groups at or below
5828 * its cpu_power, while calculating max_load..)
5830 if (busiest
->avg_load
<= sds
->avg_load
||
5831 local
->avg_load
>= sds
->avg_load
) {
5833 return fix_small_imbalance(env
, sds
);
5836 if (!busiest
->group_imb
) {
5838 * Don't want to pull so many tasks that a group would go idle.
5839 * Except of course for the group_imb case, since then we might
5840 * have to drop below capacity to reach cpu-load equilibrium.
5842 load_above_capacity
=
5843 (busiest
->sum_nr_running
- busiest
->group_capacity
);
5845 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_POWER_SCALE
);
5846 load_above_capacity
/= busiest
->group_power
;
5850 * We're trying to get all the cpus to the average_load, so we don't
5851 * want to push ourselves above the average load, nor do we wish to
5852 * reduce the max loaded cpu below the average load. At the same time,
5853 * we also don't want to reduce the group load below the group capacity
5854 * (so that we can implement power-savings policies etc). Thus we look
5855 * for the minimum possible imbalance.
5857 max_pull
= min(busiest
->avg_load
- sds
->avg_load
, load_above_capacity
);
5859 /* How much load to actually move to equalise the imbalance */
5860 env
->imbalance
= min(
5861 max_pull
* busiest
->group_power
,
5862 (sds
->avg_load
- local
->avg_load
) * local
->group_power
5863 ) / SCHED_POWER_SCALE
;
5866 * if *imbalance is less than the average load per runnable task
5867 * there is no guarantee that any tasks will be moved so we'll have
5868 * a think about bumping its value to force at least one task to be
5871 if (env
->imbalance
< busiest
->load_per_task
)
5872 return fix_small_imbalance(env
, sds
);
5875 /******* find_busiest_group() helpers end here *********************/
5878 * find_busiest_group - Returns the busiest group within the sched_domain
5879 * if there is an imbalance. If there isn't an imbalance, and
5880 * the user has opted for power-savings, it returns a group whose
5881 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5882 * such a group exists.
5884 * Also calculates the amount of weighted load which should be moved
5885 * to restore balance.
5887 * @env: The load balancing environment.
5889 * Return: - The busiest group if imbalance exists.
5890 * - If no imbalance and user has opted for power-savings balance,
5891 * return the least loaded group whose CPUs can be
5892 * put to idle by rebalancing its tasks onto our group.
5894 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
5896 struct sg_lb_stats
*local
, *busiest
;
5897 struct sd_lb_stats sds
;
5899 init_sd_lb_stats(&sds
);
5902 * Compute the various statistics relavent for load balancing at
5905 update_sd_lb_stats(env
, &sds
);
5906 local
= &sds
.local_stat
;
5907 busiest
= &sds
.busiest_stat
;
5909 if ((env
->idle
== CPU_IDLE
|| env
->idle
== CPU_NEWLY_IDLE
) &&
5910 check_asym_packing(env
, &sds
))
5913 /* There is no busy sibling group to pull tasks from */
5914 if (!sds
.busiest
|| busiest
->sum_nr_running
== 0)
5917 sds
.avg_load
= (SCHED_POWER_SCALE
* sds
.total_load
) / sds
.total_pwr
;
5920 * If the busiest group is imbalanced the below checks don't
5921 * work because they assume all things are equal, which typically
5922 * isn't true due to cpus_allowed constraints and the like.
5924 if (busiest
->group_imb
)
5927 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5928 if (env
->idle
== CPU_NEWLY_IDLE
&& local
->group_has_capacity
&&
5929 !busiest
->group_has_capacity
)
5933 * If the local group is more busy than the selected busiest group
5934 * don't try and pull any tasks.
5936 if (local
->avg_load
>= busiest
->avg_load
)
5940 * Don't pull any tasks if this group is already above the domain
5943 if (local
->avg_load
>= sds
.avg_load
)
5946 if (env
->idle
== CPU_IDLE
) {
5948 * This cpu is idle. If the busiest group load doesn't
5949 * have more tasks than the number of available cpu's and
5950 * there is no imbalance between this and busiest group
5951 * wrt to idle cpu's, it is balanced.
5953 if ((local
->idle_cpus
< busiest
->idle_cpus
) &&
5954 busiest
->sum_nr_running
<= busiest
->group_weight
)
5958 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5959 * imbalance_pct to be conservative.
5961 if (100 * busiest
->avg_load
<=
5962 env
->sd
->imbalance_pct
* local
->avg_load
)
5967 /* Looks like there is an imbalance. Compute it */
5968 calculate_imbalance(env
, &sds
);
5977 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5979 static struct rq
*find_busiest_queue(struct lb_env
*env
,
5980 struct sched_group
*group
)
5982 struct rq
*busiest
= NULL
, *rq
;
5983 unsigned long busiest_load
= 0, busiest_power
= 1;
5986 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
5987 unsigned long power
, capacity
, wl
;
5991 rt
= fbq_classify_rq(rq
);
5994 * We classify groups/runqueues into three groups:
5995 * - regular: there are !numa tasks
5996 * - remote: there are numa tasks that run on the 'wrong' node
5997 * - all: there is no distinction
5999 * In order to avoid migrating ideally placed numa tasks,
6000 * ignore those when there's better options.
6002 * If we ignore the actual busiest queue to migrate another
6003 * task, the next balance pass can still reduce the busiest
6004 * queue by moving tasks around inside the node.
6006 * If we cannot move enough load due to this classification
6007 * the next pass will adjust the group classification and
6008 * allow migration of more tasks.
6010 * Both cases only affect the total convergence complexity.
6012 if (rt
> env
->fbq_type
)
6015 power
= power_of(i
);
6016 capacity
= DIV_ROUND_CLOSEST(power
, SCHED_POWER_SCALE
);
6018 capacity
= fix_small_capacity(env
->sd
, group
);
6020 wl
= weighted_cpuload(i
);
6023 * When comparing with imbalance, use weighted_cpuload()
6024 * which is not scaled with the cpu power.
6026 if (capacity
&& rq
->nr_running
== 1 && wl
> env
->imbalance
)
6030 * For the load comparisons with the other cpu's, consider
6031 * the weighted_cpuload() scaled with the cpu power, so that
6032 * the load can be moved away from the cpu that is potentially
6033 * running at a lower capacity.
6035 * Thus we're looking for max(wl_i / power_i), crosswise
6036 * multiplication to rid ourselves of the division works out
6037 * to: wl_i * power_j > wl_j * power_i; where j is our
6040 if (wl
* busiest_power
> busiest_load
* power
) {
6042 busiest_power
= power
;
6051 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6052 * so long as it is large enough.
6054 #define MAX_PINNED_INTERVAL 512
6056 /* Working cpumask for load_balance and load_balance_newidle. */
6057 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6059 static int need_active_balance(struct lb_env
*env
)
6061 struct sched_domain
*sd
= env
->sd
;
6063 if (env
->idle
== CPU_NEWLY_IDLE
) {
6066 * ASYM_PACKING needs to force migrate tasks from busy but
6067 * higher numbered CPUs in order to pack all tasks in the
6068 * lowest numbered CPUs.
6070 if ((sd
->flags
& SD_ASYM_PACKING
) && env
->src_cpu
> env
->dst_cpu
)
6074 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
6077 static int active_load_balance_cpu_stop(void *data
);
6079 static int should_we_balance(struct lb_env
*env
)
6081 struct sched_group
*sg
= env
->sd
->groups
;
6082 struct cpumask
*sg_cpus
, *sg_mask
;
6083 int cpu
, balance_cpu
= -1;
6086 * In the newly idle case, we will allow all the cpu's
6087 * to do the newly idle load balance.
6089 if (env
->idle
== CPU_NEWLY_IDLE
)
6092 sg_cpus
= sched_group_cpus(sg
);
6093 sg_mask
= sched_group_mask(sg
);
6094 /* Try to find first idle cpu */
6095 for_each_cpu_and(cpu
, sg_cpus
, env
->cpus
) {
6096 if (!cpumask_test_cpu(cpu
, sg_mask
) || !idle_cpu(cpu
))
6103 if (balance_cpu
== -1)
6104 balance_cpu
= group_balance_cpu(sg
);
6107 * First idle cpu or the first cpu(busiest) in this sched group
6108 * is eligible for doing load balancing at this and above domains.
6110 return balance_cpu
== env
->dst_cpu
;
6114 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6115 * tasks if there is an imbalance.
6117 static int load_balance(int this_cpu
, struct rq
*this_rq
,
6118 struct sched_domain
*sd
, enum cpu_idle_type idle
,
6119 int *continue_balancing
)
6121 int ld_moved
, cur_ld_moved
, active_balance
= 0;
6122 struct sched_domain
*sd_parent
= sd
->parent
;
6123 struct sched_group
*group
;
6125 unsigned long flags
;
6126 struct cpumask
*cpus
= __get_cpu_var(load_balance_mask
);
6128 struct lb_env env
= {
6130 .dst_cpu
= this_cpu
,
6132 .dst_grpmask
= sched_group_cpus(sd
->groups
),
6134 .loop_break
= sched_nr_migrate_break
,
6140 * For NEWLY_IDLE load_balancing, we don't need to consider
6141 * other cpus in our group
6143 if (idle
== CPU_NEWLY_IDLE
)
6144 env
.dst_grpmask
= NULL
;
6146 cpumask_copy(cpus
, cpu_active_mask
);
6148 schedstat_inc(sd
, lb_count
[idle
]);
6151 if (!should_we_balance(&env
)) {
6152 *continue_balancing
= 0;
6156 group
= find_busiest_group(&env
);
6158 schedstat_inc(sd
, lb_nobusyg
[idle
]);
6162 busiest
= find_busiest_queue(&env
, group
);
6164 schedstat_inc(sd
, lb_nobusyq
[idle
]);
6168 BUG_ON(busiest
== env
.dst_rq
);
6170 schedstat_add(sd
, lb_imbalance
[idle
], env
.imbalance
);
6173 if (busiest
->nr_running
> 1) {
6175 * Attempt to move tasks. If find_busiest_group has found
6176 * an imbalance but busiest->nr_running <= 1, the group is
6177 * still unbalanced. ld_moved simply stays zero, so it is
6178 * correctly treated as an imbalance.
6180 env
.flags
|= LBF_ALL_PINNED
;
6181 env
.src_cpu
= busiest
->cpu
;
6182 env
.src_rq
= busiest
;
6183 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
6186 local_irq_save(flags
);
6187 double_rq_lock(env
.dst_rq
, busiest
);
6190 * cur_ld_moved - load moved in current iteration
6191 * ld_moved - cumulative load moved across iterations
6193 cur_ld_moved
= move_tasks(&env
);
6194 ld_moved
+= cur_ld_moved
;
6195 double_rq_unlock(env
.dst_rq
, busiest
);
6196 local_irq_restore(flags
);
6199 * some other cpu did the load balance for us.
6201 if (cur_ld_moved
&& env
.dst_cpu
!= smp_processor_id())
6202 resched_cpu(env
.dst_cpu
);
6204 if (env
.flags
& LBF_NEED_BREAK
) {
6205 env
.flags
&= ~LBF_NEED_BREAK
;
6210 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6211 * us and move them to an alternate dst_cpu in our sched_group
6212 * where they can run. The upper limit on how many times we
6213 * iterate on same src_cpu is dependent on number of cpus in our
6216 * This changes load balance semantics a bit on who can move
6217 * load to a given_cpu. In addition to the given_cpu itself
6218 * (or a ilb_cpu acting on its behalf where given_cpu is
6219 * nohz-idle), we now have balance_cpu in a position to move
6220 * load to given_cpu. In rare situations, this may cause
6221 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6222 * _independently_ and at _same_ time to move some load to
6223 * given_cpu) causing exceess load to be moved to given_cpu.
6224 * This however should not happen so much in practice and
6225 * moreover subsequent load balance cycles should correct the
6226 * excess load moved.
6228 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
6230 /* Prevent to re-select dst_cpu via env's cpus */
6231 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
6233 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
6234 env
.dst_cpu
= env
.new_dst_cpu
;
6235 env
.flags
&= ~LBF_DST_PINNED
;
6237 env
.loop_break
= sched_nr_migrate_break
;
6240 * Go back to "more_balance" rather than "redo" since we
6241 * need to continue with same src_cpu.
6247 * We failed to reach balance because of affinity.
6250 int *group_imbalance
= &sd_parent
->groups
->sgp
->imbalance
;
6252 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0) {
6253 *group_imbalance
= 1;
6254 } else if (*group_imbalance
)
6255 *group_imbalance
= 0;
6258 /* All tasks on this runqueue were pinned by CPU affinity */
6259 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
6260 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
6261 if (!cpumask_empty(cpus
)) {
6263 env
.loop_break
= sched_nr_migrate_break
;
6271 schedstat_inc(sd
, lb_failed
[idle
]);
6273 * Increment the failure counter only on periodic balance.
6274 * We do not want newidle balance, which can be very
6275 * frequent, pollute the failure counter causing
6276 * excessive cache_hot migrations and active balances.
6278 if (idle
!= CPU_NEWLY_IDLE
)
6279 sd
->nr_balance_failed
++;
6281 if (need_active_balance(&env
)) {
6282 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
6284 /* don't kick the active_load_balance_cpu_stop,
6285 * if the curr task on busiest cpu can't be
6288 if (!cpumask_test_cpu(this_cpu
,
6289 tsk_cpus_allowed(busiest
->curr
))) {
6290 raw_spin_unlock_irqrestore(&busiest
->lock
,
6292 env
.flags
|= LBF_ALL_PINNED
;
6293 goto out_one_pinned
;
6297 * ->active_balance synchronizes accesses to
6298 * ->active_balance_work. Once set, it's cleared
6299 * only after active load balance is finished.
6301 if (!busiest
->active_balance
) {
6302 busiest
->active_balance
= 1;
6303 busiest
->push_cpu
= this_cpu
;
6306 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
6308 if (active_balance
) {
6309 stop_one_cpu_nowait(cpu_of(busiest
),
6310 active_load_balance_cpu_stop
, busiest
,
6311 &busiest
->active_balance_work
);
6315 * We've kicked active balancing, reset the failure
6318 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
6321 sd
->nr_balance_failed
= 0;
6323 if (likely(!active_balance
)) {
6324 /* We were unbalanced, so reset the balancing interval */
6325 sd
->balance_interval
= sd
->min_interval
;
6328 * If we've begun active balancing, start to back off. This
6329 * case may not be covered by the all_pinned logic if there
6330 * is only 1 task on the busy runqueue (because we don't call
6333 if (sd
->balance_interval
< sd
->max_interval
)
6334 sd
->balance_interval
*= 2;
6340 schedstat_inc(sd
, lb_balanced
[idle
]);
6342 sd
->nr_balance_failed
= 0;
6345 /* tune up the balancing interval */
6346 if (((env
.flags
& LBF_ALL_PINNED
) &&
6347 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
6348 (sd
->balance_interval
< sd
->max_interval
))
6349 sd
->balance_interval
*= 2;
6357 * idle_balance is called by schedule() if this_cpu is about to become
6358 * idle. Attempts to pull tasks from other CPUs.
6360 void idle_balance(int this_cpu
, struct rq
*this_rq
)
6362 struct sched_domain
*sd
;
6363 int pulled_task
= 0;
6364 unsigned long next_balance
= jiffies
+ HZ
;
6367 this_rq
->idle_stamp
= rq_clock(this_rq
);
6369 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
6373 * Drop the rq->lock, but keep IRQ/preempt disabled.
6375 raw_spin_unlock(&this_rq
->lock
);
6377 update_blocked_averages(this_cpu
);
6379 for_each_domain(this_cpu
, sd
) {
6380 unsigned long interval
;
6381 int continue_balancing
= 1;
6382 u64 t0
, domain_cost
;
6384 if (!(sd
->flags
& SD_LOAD_BALANCE
))
6387 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
)
6390 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
6391 t0
= sched_clock_cpu(this_cpu
);
6393 /* If we've pulled tasks over stop searching: */
6394 pulled_task
= load_balance(this_cpu
, this_rq
,
6396 &continue_balancing
);
6398 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
6399 if (domain_cost
> sd
->max_newidle_lb_cost
)
6400 sd
->max_newidle_lb_cost
= domain_cost
;
6402 curr_cost
+= domain_cost
;
6405 interval
= msecs_to_jiffies(sd
->balance_interval
);
6406 if (time_after(next_balance
, sd
->last_balance
+ interval
))
6407 next_balance
= sd
->last_balance
+ interval
;
6409 this_rq
->idle_stamp
= 0;
6415 raw_spin_lock(&this_rq
->lock
);
6417 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
6419 * We are going idle. next_balance may be set based on
6420 * a busy processor. So reset next_balance.
6422 this_rq
->next_balance
= next_balance
;
6425 if (curr_cost
> this_rq
->max_idle_balance_cost
)
6426 this_rq
->max_idle_balance_cost
= curr_cost
;
6430 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6431 * running tasks off the busiest CPU onto idle CPUs. It requires at
6432 * least 1 task to be running on each physical CPU where possible, and
6433 * avoids physical / logical imbalances.
6435 static int active_load_balance_cpu_stop(void *data
)
6437 struct rq
*busiest_rq
= data
;
6438 int busiest_cpu
= cpu_of(busiest_rq
);
6439 int target_cpu
= busiest_rq
->push_cpu
;
6440 struct rq
*target_rq
= cpu_rq(target_cpu
);
6441 struct sched_domain
*sd
;
6443 raw_spin_lock_irq(&busiest_rq
->lock
);
6445 /* make sure the requested cpu hasn't gone down in the meantime */
6446 if (unlikely(busiest_cpu
!= smp_processor_id() ||
6447 !busiest_rq
->active_balance
))
6450 /* Is there any task to move? */
6451 if (busiest_rq
->nr_running
<= 1)
6455 * This condition is "impossible", if it occurs
6456 * we need to fix it. Originally reported by
6457 * Bjorn Helgaas on a 128-cpu setup.
6459 BUG_ON(busiest_rq
== target_rq
);
6461 /* move a task from busiest_rq to target_rq */
6462 double_lock_balance(busiest_rq
, target_rq
);
6464 /* Search for an sd spanning us and the target CPU. */
6466 for_each_domain(target_cpu
, sd
) {
6467 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
6468 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
6473 struct lb_env env
= {
6475 .dst_cpu
= target_cpu
,
6476 .dst_rq
= target_rq
,
6477 .src_cpu
= busiest_rq
->cpu
,
6478 .src_rq
= busiest_rq
,
6482 schedstat_inc(sd
, alb_count
);
6484 if (move_one_task(&env
))
6485 schedstat_inc(sd
, alb_pushed
);
6487 schedstat_inc(sd
, alb_failed
);
6490 double_unlock_balance(busiest_rq
, target_rq
);
6492 busiest_rq
->active_balance
= 0;
6493 raw_spin_unlock_irq(&busiest_rq
->lock
);
6497 #ifdef CONFIG_NO_HZ_COMMON
6499 * idle load balancing details
6500 * - When one of the busy CPUs notice that there may be an idle rebalancing
6501 * needed, they will kick the idle load balancer, which then does idle
6502 * load balancing for all the idle CPUs.
6505 cpumask_var_t idle_cpus_mask
;
6507 unsigned long next_balance
; /* in jiffy units */
6508 } nohz ____cacheline_aligned
;
6510 static inline int find_new_ilb(int call_cpu
)
6512 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
6514 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
6521 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6522 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6523 * CPU (if there is one).
6525 static void nohz_balancer_kick(int cpu
)
6529 nohz
.next_balance
++;
6531 ilb_cpu
= find_new_ilb(cpu
);
6533 if (ilb_cpu
>= nr_cpu_ids
)
6536 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
6539 * Use smp_send_reschedule() instead of resched_cpu().
6540 * This way we generate a sched IPI on the target cpu which
6541 * is idle. And the softirq performing nohz idle load balance
6542 * will be run before returning from the IPI.
6544 smp_send_reschedule(ilb_cpu
);
6548 static inline void nohz_balance_exit_idle(int cpu
)
6550 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
6551 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
6552 atomic_dec(&nohz
.nr_cpus
);
6553 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
6557 static inline void set_cpu_sd_state_busy(void)
6559 struct sched_domain
*sd
;
6562 sd
= rcu_dereference_check_sched_domain(this_rq()->sd
);
6564 if (!sd
|| !sd
->nohz_idle
)
6568 for (; sd
; sd
= sd
->parent
)
6569 atomic_inc(&sd
->groups
->sgp
->nr_busy_cpus
);
6574 void set_cpu_sd_state_idle(void)
6576 struct sched_domain
*sd
;
6579 sd
= rcu_dereference_check_sched_domain(this_rq()->sd
);
6581 if (!sd
|| sd
->nohz_idle
)
6585 for (; sd
; sd
= sd
->parent
)
6586 atomic_dec(&sd
->groups
->sgp
->nr_busy_cpus
);
6592 * This routine will record that the cpu is going idle with tick stopped.
6593 * This info will be used in performing idle load balancing in the future.
6595 void nohz_balance_enter_idle(int cpu
)
6598 * If this cpu is going down, then nothing needs to be done.
6600 if (!cpu_active(cpu
))
6603 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
6606 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
6607 atomic_inc(&nohz
.nr_cpus
);
6608 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
6611 static int sched_ilb_notifier(struct notifier_block
*nfb
,
6612 unsigned long action
, void *hcpu
)
6614 switch (action
& ~CPU_TASKS_FROZEN
) {
6616 nohz_balance_exit_idle(smp_processor_id());
6624 static DEFINE_SPINLOCK(balancing
);
6627 * Scale the max load_balance interval with the number of CPUs in the system.
6628 * This trades load-balance latency on larger machines for less cross talk.
6630 void update_max_interval(void)
6632 max_load_balance_interval
= HZ
*num_online_cpus()/10;
6636 * It checks each scheduling domain to see if it is due to be balanced,
6637 * and initiates a balancing operation if so.
6639 * Balancing parameters are set up in init_sched_domains.
6641 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
6643 int continue_balancing
= 1;
6644 struct rq
*rq
= cpu_rq(cpu
);
6645 unsigned long interval
;
6646 struct sched_domain
*sd
;
6647 /* Earliest time when we have to do rebalance again */
6648 unsigned long next_balance
= jiffies
+ 60*HZ
;
6649 int update_next_balance
= 0;
6650 int need_serialize
, need_decay
= 0;
6653 update_blocked_averages(cpu
);
6656 for_each_domain(cpu
, sd
) {
6658 * Decay the newidle max times here because this is a regular
6659 * visit to all the domains. Decay ~1% per second.
6661 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
6662 sd
->max_newidle_lb_cost
=
6663 (sd
->max_newidle_lb_cost
* 253) / 256;
6664 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
6667 max_cost
+= sd
->max_newidle_lb_cost
;
6669 if (!(sd
->flags
& SD_LOAD_BALANCE
))
6673 * Stop the load balance at this level. There is another
6674 * CPU in our sched group which is doing load balancing more
6677 if (!continue_balancing
) {
6683 interval
= sd
->balance_interval
;
6684 if (idle
!= CPU_IDLE
)
6685 interval
*= sd
->busy_factor
;
6687 /* scale ms to jiffies */
6688 interval
= msecs_to_jiffies(interval
);
6689 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
6691 need_serialize
= sd
->flags
& SD_SERIALIZE
;
6693 if (need_serialize
) {
6694 if (!spin_trylock(&balancing
))
6698 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
6699 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
6701 * The LBF_DST_PINNED logic could have changed
6702 * env->dst_cpu, so we can't know our idle
6703 * state even if we migrated tasks. Update it.
6705 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
6707 sd
->last_balance
= jiffies
;
6710 spin_unlock(&balancing
);
6712 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
6713 next_balance
= sd
->last_balance
+ interval
;
6714 update_next_balance
= 1;
6719 * Ensure the rq-wide value also decays but keep it at a
6720 * reasonable floor to avoid funnies with rq->avg_idle.
6722 rq
->max_idle_balance_cost
=
6723 max((u64
)sysctl_sched_migration_cost
, max_cost
);
6728 * next_balance will be updated only when there is a need.
6729 * When the cpu is attached to null domain for ex, it will not be
6732 if (likely(update_next_balance
))
6733 rq
->next_balance
= next_balance
;
6736 #ifdef CONFIG_NO_HZ_COMMON
6738 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6739 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6741 static void nohz_idle_balance(int this_cpu
, enum cpu_idle_type idle
)
6743 struct rq
*this_rq
= cpu_rq(this_cpu
);
6747 if (idle
!= CPU_IDLE
||
6748 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
6751 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
6752 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
6756 * If this cpu gets work to do, stop the load balancing
6757 * work being done for other cpus. Next load
6758 * balancing owner will pick it up.
6763 rq
= cpu_rq(balance_cpu
);
6765 raw_spin_lock_irq(&rq
->lock
);
6766 update_rq_clock(rq
);
6767 update_idle_cpu_load(rq
);
6768 raw_spin_unlock_irq(&rq
->lock
);
6770 rebalance_domains(balance_cpu
, CPU_IDLE
);
6772 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
6773 this_rq
->next_balance
= rq
->next_balance
;
6775 nohz
.next_balance
= this_rq
->next_balance
;
6777 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
6781 * Current heuristic for kicking the idle load balancer in the presence
6782 * of an idle cpu is the system.
6783 * - This rq has more than one task.
6784 * - At any scheduler domain level, this cpu's scheduler group has multiple
6785 * busy cpu's exceeding the group's power.
6786 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6787 * domain span are idle.
6789 static inline int nohz_kick_needed(struct rq
*rq
, int cpu
)
6791 unsigned long now
= jiffies
;
6792 struct sched_domain
*sd
;
6794 if (unlikely(idle_cpu(cpu
)))
6798 * We may be recently in ticked or tickless idle mode. At the first
6799 * busy tick after returning from idle, we will update the busy stats.
6801 set_cpu_sd_state_busy();
6802 nohz_balance_exit_idle(cpu
);
6805 * None are in tickless mode and hence no need for NOHZ idle load
6808 if (likely(!atomic_read(&nohz
.nr_cpus
)))
6811 if (time_before(now
, nohz
.next_balance
))
6814 if (rq
->nr_running
>= 2)
6818 for_each_domain(cpu
, sd
) {
6819 struct sched_group
*sg
= sd
->groups
;
6820 struct sched_group_power
*sgp
= sg
->sgp
;
6821 int nr_busy
= atomic_read(&sgp
->nr_busy_cpus
);
6823 if (sd
->flags
& SD_SHARE_PKG_RESOURCES
&& nr_busy
> 1)
6824 goto need_kick_unlock
;
6826 if (sd
->flags
& SD_ASYM_PACKING
&& nr_busy
!= sg
->group_weight
6827 && (cpumask_first_and(nohz
.idle_cpus_mask
,
6828 sched_domain_span(sd
)) < cpu
))
6829 goto need_kick_unlock
;
6831 if (!(sd
->flags
& (SD_SHARE_PKG_RESOURCES
| SD_ASYM_PACKING
)))
6843 static void nohz_idle_balance(int this_cpu
, enum cpu_idle_type idle
) { }
6847 * run_rebalance_domains is triggered when needed from the scheduler tick.
6848 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6850 static void run_rebalance_domains(struct softirq_action
*h
)
6852 int this_cpu
= smp_processor_id();
6853 struct rq
*this_rq
= cpu_rq(this_cpu
);
6854 enum cpu_idle_type idle
= this_rq
->idle_balance
?
6855 CPU_IDLE
: CPU_NOT_IDLE
;
6857 rebalance_domains(this_cpu
, idle
);
6860 * If this cpu has a pending nohz_balance_kick, then do the
6861 * balancing on behalf of the other idle cpus whose ticks are
6864 nohz_idle_balance(this_cpu
, idle
);
6867 static inline int on_null_domain(int cpu
)
6869 return !rcu_dereference_sched(cpu_rq(cpu
)->sd
);
6873 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6875 void trigger_load_balance(struct rq
*rq
, int cpu
)
6877 /* Don't need to rebalance while attached to NULL domain */
6878 if (time_after_eq(jiffies
, rq
->next_balance
) &&
6879 likely(!on_null_domain(cpu
)))
6880 raise_softirq(SCHED_SOFTIRQ
);
6881 #ifdef CONFIG_NO_HZ_COMMON
6882 if (nohz_kick_needed(rq
, cpu
) && likely(!on_null_domain(cpu
)))
6883 nohz_balancer_kick(cpu
);
6887 static void rq_online_fair(struct rq
*rq
)
6892 static void rq_offline_fair(struct rq
*rq
)
6896 /* Ensure any throttled groups are reachable by pick_next_task */
6897 unthrottle_offline_cfs_rqs(rq
);
6900 #endif /* CONFIG_SMP */
6903 * scheduler tick hitting a task of our scheduling class:
6905 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
6907 struct cfs_rq
*cfs_rq
;
6908 struct sched_entity
*se
= &curr
->se
;
6910 for_each_sched_entity(se
) {
6911 cfs_rq
= cfs_rq_of(se
);
6912 entity_tick(cfs_rq
, se
, queued
);
6915 if (numabalancing_enabled
)
6916 task_tick_numa(rq
, curr
);
6918 update_rq_runnable_avg(rq
, 1);
6922 * called on fork with the child task as argument from the parent's context
6923 * - child not yet on the tasklist
6924 * - preemption disabled
6926 static void task_fork_fair(struct task_struct
*p
)
6928 struct cfs_rq
*cfs_rq
;
6929 struct sched_entity
*se
= &p
->se
, *curr
;
6930 int this_cpu
= smp_processor_id();
6931 struct rq
*rq
= this_rq();
6932 unsigned long flags
;
6934 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6936 update_rq_clock(rq
);
6938 cfs_rq
= task_cfs_rq(current
);
6939 curr
= cfs_rq
->curr
;
6942 * Not only the cpu but also the task_group of the parent might have
6943 * been changed after parent->se.parent,cfs_rq were copied to
6944 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6945 * of child point to valid ones.
6948 __set_task_cpu(p
, this_cpu
);
6951 update_curr(cfs_rq
);
6954 se
->vruntime
= curr
->vruntime
;
6955 place_entity(cfs_rq
, se
, 1);
6957 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
6959 * Upon rescheduling, sched_class::put_prev_task() will place
6960 * 'current' within the tree based on its new key value.
6962 swap(curr
->vruntime
, se
->vruntime
);
6963 resched_task(rq
->curr
);
6966 se
->vruntime
-= cfs_rq
->min_vruntime
;
6968 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6972 * Priority of the task has changed. Check to see if we preempt
6976 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
6982 * Reschedule if we are currently running on this runqueue and
6983 * our priority decreased, or if we are not currently running on
6984 * this runqueue and our priority is higher than the current's
6986 if (rq
->curr
== p
) {
6987 if (p
->prio
> oldprio
)
6988 resched_task(rq
->curr
);
6990 check_preempt_curr(rq
, p
, 0);
6993 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
6995 struct sched_entity
*se
= &p
->se
;
6996 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
6999 * Ensure the task's vruntime is normalized, so that when its
7000 * switched back to the fair class the enqueue_entity(.flags=0) will
7001 * do the right thing.
7003 * If it was on_rq, then the dequeue_entity(.flags=0) will already
7004 * have normalized the vruntime, if it was !on_rq, then only when
7005 * the task is sleeping will it still have non-normalized vruntime.
7007 if (!se
->on_rq
&& p
->state
!= TASK_RUNNING
) {
7009 * Fix up our vruntime so that the current sleep doesn't
7010 * cause 'unlimited' sleep bonus.
7012 place_entity(cfs_rq
, se
, 0);
7013 se
->vruntime
-= cfs_rq
->min_vruntime
;
7018 * Remove our load from contribution when we leave sched_fair
7019 * and ensure we don't carry in an old decay_count if we
7022 if (se
->avg
.decay_count
) {
7023 __synchronize_entity_decay(se
);
7024 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
7030 * We switched to the sched_fair class.
7032 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
7038 * We were most likely switched from sched_rt, so
7039 * kick off the schedule if running, otherwise just see
7040 * if we can still preempt the current task.
7043 resched_task(rq
->curr
);
7045 check_preempt_curr(rq
, p
, 0);
7048 /* Account for a task changing its policy or group.
7050 * This routine is mostly called to set cfs_rq->curr field when a task
7051 * migrates between groups/classes.
7053 static void set_curr_task_fair(struct rq
*rq
)
7055 struct sched_entity
*se
= &rq
->curr
->se
;
7057 for_each_sched_entity(se
) {
7058 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7060 set_next_entity(cfs_rq
, se
);
7061 /* ensure bandwidth has been allocated on our new cfs_rq */
7062 account_cfs_rq_runtime(cfs_rq
, 0);
7066 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
7068 cfs_rq
->tasks_timeline
= RB_ROOT
;
7069 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7070 #ifndef CONFIG_64BIT
7071 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
7074 atomic64_set(&cfs_rq
->decay_counter
, 1);
7075 atomic_long_set(&cfs_rq
->removed_load
, 0);
7079 #ifdef CONFIG_FAIR_GROUP_SCHED
7080 static void task_move_group_fair(struct task_struct
*p
, int on_rq
)
7082 struct cfs_rq
*cfs_rq
;
7084 * If the task was not on the rq at the time of this cgroup movement
7085 * it must have been asleep, sleeping tasks keep their ->vruntime
7086 * absolute on their old rq until wakeup (needed for the fair sleeper
7087 * bonus in place_entity()).
7089 * If it was on the rq, we've just 'preempted' it, which does convert
7090 * ->vruntime to a relative base.
7092 * Make sure both cases convert their relative position when migrating
7093 * to another cgroup's rq. This does somewhat interfere with the
7094 * fair sleeper stuff for the first placement, but who cares.
7097 * When !on_rq, vruntime of the task has usually NOT been normalized.
7098 * But there are some cases where it has already been normalized:
7100 * - Moving a forked child which is waiting for being woken up by
7101 * wake_up_new_task().
7102 * - Moving a task which has been woken up by try_to_wake_up() and
7103 * waiting for actually being woken up by sched_ttwu_pending().
7105 * To prevent boost or penalty in the new cfs_rq caused by delta
7106 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7108 if (!on_rq
&& (!p
->se
.sum_exec_runtime
|| p
->state
== TASK_WAKING
))
7112 p
->se
.vruntime
-= cfs_rq_of(&p
->se
)->min_vruntime
;
7113 set_task_rq(p
, task_cpu(p
));
7115 cfs_rq
= cfs_rq_of(&p
->se
);
7116 p
->se
.vruntime
+= cfs_rq
->min_vruntime
;
7119 * migrate_task_rq_fair() will have removed our previous
7120 * contribution, but we must synchronize for ongoing future
7123 p
->se
.avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
7124 cfs_rq
->blocked_load_avg
+= p
->se
.avg
.load_avg_contrib
;
7129 void free_fair_sched_group(struct task_group
*tg
)
7133 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7135 for_each_possible_cpu(i
) {
7137 kfree(tg
->cfs_rq
[i
]);
7146 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7148 struct cfs_rq
*cfs_rq
;
7149 struct sched_entity
*se
;
7152 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
7155 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
7159 tg
->shares
= NICE_0_LOAD
;
7161 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7163 for_each_possible_cpu(i
) {
7164 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
7165 GFP_KERNEL
, cpu_to_node(i
));
7169 se
= kzalloc_node(sizeof(struct sched_entity
),
7170 GFP_KERNEL
, cpu_to_node(i
));
7174 init_cfs_rq(cfs_rq
);
7175 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
7186 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
7188 struct rq
*rq
= cpu_rq(cpu
);
7189 unsigned long flags
;
7192 * Only empty task groups can be destroyed; so we can speculatively
7193 * check on_list without danger of it being re-added.
7195 if (!tg
->cfs_rq
[cpu
]->on_list
)
7198 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7199 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
7200 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7203 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7204 struct sched_entity
*se
, int cpu
,
7205 struct sched_entity
*parent
)
7207 struct rq
*rq
= cpu_rq(cpu
);
7211 init_cfs_rq_runtime(cfs_rq
);
7213 tg
->cfs_rq
[cpu
] = cfs_rq
;
7216 /* se could be NULL for root_task_group */
7221 se
->cfs_rq
= &rq
->cfs
;
7223 se
->cfs_rq
= parent
->my_q
;
7226 update_load_set(&se
->load
, 0);
7227 se
->parent
= parent
;
7230 static DEFINE_MUTEX(shares_mutex
);
7232 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
7235 unsigned long flags
;
7238 * We can't change the weight of the root cgroup.
7243 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
7245 mutex_lock(&shares_mutex
);
7246 if (tg
->shares
== shares
)
7249 tg
->shares
= shares
;
7250 for_each_possible_cpu(i
) {
7251 struct rq
*rq
= cpu_rq(i
);
7252 struct sched_entity
*se
;
7255 /* Propagate contribution to hierarchy */
7256 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7258 /* Possible calls to update_curr() need rq clock */
7259 update_rq_clock(rq
);
7260 for_each_sched_entity(se
)
7261 update_cfs_shares(group_cfs_rq(se
));
7262 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7266 mutex_unlock(&shares_mutex
);
7269 #else /* CONFIG_FAIR_GROUP_SCHED */
7271 void free_fair_sched_group(struct task_group
*tg
) { }
7273 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7278 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
) { }
7280 #endif /* CONFIG_FAIR_GROUP_SCHED */
7283 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
7285 struct sched_entity
*se
= &task
->se
;
7286 unsigned int rr_interval
= 0;
7289 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7292 if (rq
->cfs
.load
.weight
)
7293 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
7299 * All the scheduling class methods:
7301 const struct sched_class fair_sched_class
= {
7302 .next
= &idle_sched_class
,
7303 .enqueue_task
= enqueue_task_fair
,
7304 .dequeue_task
= dequeue_task_fair
,
7305 .yield_task
= yield_task_fair
,
7306 .yield_to_task
= yield_to_task_fair
,
7308 .check_preempt_curr
= check_preempt_wakeup
,
7310 .pick_next_task
= pick_next_task_fair
,
7311 .put_prev_task
= put_prev_task_fair
,
7314 .select_task_rq
= select_task_rq_fair
,
7315 .migrate_task_rq
= migrate_task_rq_fair
,
7317 .rq_online
= rq_online_fair
,
7318 .rq_offline
= rq_offline_fair
,
7320 .task_waking
= task_waking_fair
,
7323 .set_curr_task
= set_curr_task_fair
,
7324 .task_tick
= task_tick_fair
,
7325 .task_fork
= task_fork_fair
,
7327 .prio_changed
= prio_changed_fair
,
7328 .switched_from
= switched_from_fair
,
7329 .switched_to
= switched_to_fair
,
7331 .get_rr_interval
= get_rr_interval_fair
,
7333 #ifdef CONFIG_FAIR_GROUP_SCHED
7334 .task_move_group
= task_move_group_fair
,
7338 #ifdef CONFIG_SCHED_DEBUG
7339 void print_cfs_stats(struct seq_file
*m
, int cpu
)
7341 struct cfs_rq
*cfs_rq
;
7344 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
7345 print_cfs_rq(m
, cpu
, cfs_rq
);
7350 __init
void init_sched_fair_class(void)
7353 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7355 #ifdef CONFIG_NO_HZ_COMMON
7356 nohz
.next_balance
= jiffies
;
7357 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
7358 cpu_notifier(sched_ilb_notifier
, 0);