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/cpuidle.h>
27 #include <linux/slab.h>
28 #include <linux/profile.h>
29 #include <linux/interrupt.h>
30 #include <linux/mempolicy.h>
31 #include <linux/migrate.h>
32 #include <linux/task_work.h>
34 #include <trace/events/sched.h>
39 * Targeted preemption latency for CPU-bound tasks:
40 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
42 * NOTE: this latency value is not the same as the concept of
43 * 'timeslice length' - timeslices in CFS are of variable length
44 * and have no persistent notion like in traditional, time-slice
45 * based scheduling concepts.
47 * (to see the precise effective timeslice length of your workload,
48 * run vmstat and monitor the context-switches (cs) field)
50 unsigned int sysctl_sched_latency
= 6000000ULL;
51 unsigned int normalized_sysctl_sched_latency
= 6000000ULL;
54 * The initial- and re-scaling of tunables is configurable
55 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
58 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
59 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
60 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
62 enum sched_tunable_scaling sysctl_sched_tunable_scaling
63 = SCHED_TUNABLESCALING_LOG
;
66 * Minimal preemption granularity for CPU-bound tasks:
67 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
69 unsigned int sysctl_sched_min_granularity
= 750000ULL;
70 unsigned int normalized_sysctl_sched_min_granularity
= 750000ULL;
73 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
75 static unsigned int sched_nr_latency
= 8;
78 * After fork, child runs first. If set to 0 (default) then
79 * parent will (try to) run first.
81 unsigned int sysctl_sched_child_runs_first __read_mostly
;
84 * SCHED_OTHER wake-up granularity.
85 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
87 * This option delays the preemption effects of decoupled workloads
88 * and reduces their over-scheduling. Synchronous workloads will still
89 * have immediate wakeup/sleep latencies.
91 unsigned int sysctl_sched_wakeup_granularity
= 1000000UL;
92 unsigned int normalized_sysctl_sched_wakeup_granularity
= 1000000UL;
94 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
97 * The exponential sliding window over which load is averaged for shares
101 unsigned int __read_mostly sysctl_sched_shares_window
= 10000000UL;
103 #ifdef CONFIG_CFS_BANDWIDTH
105 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
106 * each time a cfs_rq requests quota.
108 * Note: in the case that the slice exceeds the runtime remaining (either due
109 * to consumption or the quota being specified to be smaller than the slice)
110 * we will always only issue the remaining available time.
112 * default: 5 msec, units: microseconds
114 unsigned int sysctl_sched_cfs_bandwidth_slice
= 5000UL;
117 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
123 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
129 static inline void update_load_set(struct load_weight
*lw
, unsigned long w
)
136 * Increase the granularity value when there are more CPUs,
137 * because with more CPUs the 'effective latency' as visible
138 * to users decreases. But the relationship is not linear,
139 * so pick a second-best guess by going with the log2 of the
142 * This idea comes from the SD scheduler of Con Kolivas:
144 static int get_update_sysctl_factor(void)
146 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
149 switch (sysctl_sched_tunable_scaling
) {
150 case SCHED_TUNABLESCALING_NONE
:
153 case SCHED_TUNABLESCALING_LINEAR
:
156 case SCHED_TUNABLESCALING_LOG
:
158 factor
= 1 + ilog2(cpus
);
165 static void update_sysctl(void)
167 unsigned int factor
= get_update_sysctl_factor();
169 #define SET_SYSCTL(name) \
170 (sysctl_##name = (factor) * normalized_sysctl_##name)
171 SET_SYSCTL(sched_min_granularity
);
172 SET_SYSCTL(sched_latency
);
173 SET_SYSCTL(sched_wakeup_granularity
);
177 void sched_init_granularity(void)
182 #define WMULT_CONST (~0U)
183 #define WMULT_SHIFT 32
185 static void __update_inv_weight(struct load_weight
*lw
)
189 if (likely(lw
->inv_weight
))
192 w
= scale_load_down(lw
->weight
);
194 if (BITS_PER_LONG
> 32 && unlikely(w
>= WMULT_CONST
))
196 else if (unlikely(!w
))
197 lw
->inv_weight
= WMULT_CONST
;
199 lw
->inv_weight
= WMULT_CONST
/ w
;
203 * delta_exec * weight / lw.weight
205 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
207 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
208 * we're guaranteed shift stays positive because inv_weight is guaranteed to
209 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
211 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
212 * weight/lw.weight <= 1, and therefore our shift will also be positive.
214 static u64
__calc_delta(u64 delta_exec
, unsigned long weight
, struct load_weight
*lw
)
216 u64 fact
= scale_load_down(weight
);
217 int shift
= WMULT_SHIFT
;
219 __update_inv_weight(lw
);
221 if (unlikely(fact
>> 32)) {
228 /* hint to use a 32x32->64 mul */
229 fact
= (u64
)(u32
)fact
* lw
->inv_weight
;
236 return mul_u64_u32_shr(delta_exec
, fact
, shift
);
240 const struct sched_class fair_sched_class
;
242 /**************************************************************
243 * CFS operations on generic schedulable entities:
246 #ifdef CONFIG_FAIR_GROUP_SCHED
248 /* cpu runqueue to which this cfs_rq is attached */
249 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
254 /* An entity is a task if it doesn't "own" a runqueue */
255 #define entity_is_task(se) (!se->my_q)
257 static inline struct task_struct
*task_of(struct sched_entity
*se
)
259 #ifdef CONFIG_SCHED_DEBUG
260 WARN_ON_ONCE(!entity_is_task(se
));
262 return container_of(se
, struct task_struct
, se
);
265 /* Walk up scheduling entities hierarchy */
266 #define for_each_sched_entity(se) \
267 for (; se; se = se->parent)
269 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
274 /* runqueue on which this entity is (to be) queued */
275 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
280 /* runqueue "owned" by this group */
281 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
286 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
289 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
291 if (!cfs_rq
->on_list
) {
293 * Ensure we either appear before our parent (if already
294 * enqueued) or force our parent to appear after us when it is
295 * enqueued. The fact that we always enqueue bottom-up
296 * reduces this to two cases.
298 if (cfs_rq
->tg
->parent
&&
299 cfs_rq
->tg
->parent
->cfs_rq
[cpu_of(rq_of(cfs_rq
))]->on_list
) {
300 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
,
301 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
303 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
304 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
308 /* We should have no load, but we need to update last_decay. */
309 update_cfs_rq_blocked_load(cfs_rq
, 0);
313 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
315 if (cfs_rq
->on_list
) {
316 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
321 /* Iterate thr' all leaf cfs_rq's on a runqueue */
322 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
323 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
325 /* Do the two (enqueued) entities belong to the same group ? */
326 static inline struct cfs_rq
*
327 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
329 if (se
->cfs_rq
== pse
->cfs_rq
)
335 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
341 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
343 int se_depth
, pse_depth
;
346 * preemption test can be made between sibling entities who are in the
347 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
348 * both tasks until we find their ancestors who are siblings of common
352 /* First walk up until both entities are at same depth */
353 se_depth
= (*se
)->depth
;
354 pse_depth
= (*pse
)->depth
;
356 while (se_depth
> pse_depth
) {
358 *se
= parent_entity(*se
);
361 while (pse_depth
> se_depth
) {
363 *pse
= parent_entity(*pse
);
366 while (!is_same_group(*se
, *pse
)) {
367 *se
= parent_entity(*se
);
368 *pse
= parent_entity(*pse
);
372 #else /* !CONFIG_FAIR_GROUP_SCHED */
374 static inline struct task_struct
*task_of(struct sched_entity
*se
)
376 return container_of(se
, struct task_struct
, se
);
379 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
381 return container_of(cfs_rq
, struct rq
, cfs
);
384 #define entity_is_task(se) 1
386 #define for_each_sched_entity(se) \
387 for (; se; se = NULL)
389 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
391 return &task_rq(p
)->cfs
;
394 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
396 struct task_struct
*p
= task_of(se
);
397 struct rq
*rq
= task_rq(p
);
402 /* runqueue "owned" by this group */
403 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
408 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
412 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
416 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
417 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
419 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
425 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
429 #endif /* CONFIG_FAIR_GROUP_SCHED */
431 static __always_inline
432 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
);
434 /**************************************************************
435 * Scheduling class tree data structure manipulation methods:
438 static inline u64
max_vruntime(u64 max_vruntime
, u64 vruntime
)
440 s64 delta
= (s64
)(vruntime
- max_vruntime
);
442 max_vruntime
= vruntime
;
447 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
449 s64 delta
= (s64
)(vruntime
- min_vruntime
);
451 min_vruntime
= vruntime
;
456 static inline int entity_before(struct sched_entity
*a
,
457 struct sched_entity
*b
)
459 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
462 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
464 u64 vruntime
= cfs_rq
->min_vruntime
;
467 vruntime
= cfs_rq
->curr
->vruntime
;
469 if (cfs_rq
->rb_leftmost
) {
470 struct sched_entity
*se
= rb_entry(cfs_rq
->rb_leftmost
,
475 vruntime
= se
->vruntime
;
477 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
480 /* ensure we never gain time by being placed backwards. */
481 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
484 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
489 * Enqueue an entity into the rb-tree:
491 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
493 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
494 struct rb_node
*parent
= NULL
;
495 struct sched_entity
*entry
;
499 * Find the right place in the rbtree:
503 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
505 * We dont care about collisions. Nodes with
506 * the same key stay together.
508 if (entity_before(se
, entry
)) {
509 link
= &parent
->rb_left
;
511 link
= &parent
->rb_right
;
517 * Maintain a cache of leftmost tree entries (it is frequently
521 cfs_rq
->rb_leftmost
= &se
->run_node
;
523 rb_link_node(&se
->run_node
, parent
, link
);
524 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
527 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
529 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
530 struct rb_node
*next_node
;
532 next_node
= rb_next(&se
->run_node
);
533 cfs_rq
->rb_leftmost
= next_node
;
536 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
539 struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
)
541 struct rb_node
*left
= cfs_rq
->rb_leftmost
;
546 return rb_entry(left
, struct sched_entity
, run_node
);
549 static struct sched_entity
*__pick_next_entity(struct sched_entity
*se
)
551 struct rb_node
*next
= rb_next(&se
->run_node
);
556 return rb_entry(next
, struct sched_entity
, run_node
);
559 #ifdef CONFIG_SCHED_DEBUG
560 struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
562 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
567 return rb_entry(last
, struct sched_entity
, run_node
);
570 /**************************************************************
571 * Scheduling class statistics methods:
574 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
575 void __user
*buffer
, size_t *lenp
,
578 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
579 int factor
= get_update_sysctl_factor();
584 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
585 sysctl_sched_min_granularity
);
587 #define WRT_SYSCTL(name) \
588 (normalized_sysctl_##name = sysctl_##name / (factor))
589 WRT_SYSCTL(sched_min_granularity
);
590 WRT_SYSCTL(sched_latency
);
591 WRT_SYSCTL(sched_wakeup_granularity
);
601 static inline u64
calc_delta_fair(u64 delta
, struct sched_entity
*se
)
603 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
604 delta
= __calc_delta(delta
, NICE_0_LOAD
, &se
->load
);
610 * The idea is to set a period in which each task runs once.
612 * When there are too many tasks (sched_nr_latency) we have to stretch
613 * this period because otherwise the slices get too small.
615 * p = (nr <= nl) ? l : l*nr/nl
617 static u64
__sched_period(unsigned long nr_running
)
619 u64 period
= sysctl_sched_latency
;
620 unsigned long nr_latency
= sched_nr_latency
;
622 if (unlikely(nr_running
> nr_latency
)) {
623 period
= sysctl_sched_min_granularity
;
624 period
*= nr_running
;
631 * We calculate the wall-time slice from the period by taking a part
632 * proportional to the weight.
636 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
638 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
640 for_each_sched_entity(se
) {
641 struct load_weight
*load
;
642 struct load_weight lw
;
644 cfs_rq
= cfs_rq_of(se
);
645 load
= &cfs_rq
->load
;
647 if (unlikely(!se
->on_rq
)) {
650 update_load_add(&lw
, se
->load
.weight
);
653 slice
= __calc_delta(slice
, se
->load
.weight
, load
);
659 * We calculate the vruntime slice of a to-be-inserted task.
663 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
665 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
669 static int select_idle_sibling(struct task_struct
*p
, int cpu
);
670 static unsigned long task_h_load(struct task_struct
*p
);
672 static inline void __update_task_entity_contrib(struct sched_entity
*se
);
674 /* Give new task start runnable values to heavy its load in infant time */
675 void init_task_runnable_average(struct task_struct
*p
)
679 p
->se
.avg
.decay_count
= 0;
680 slice
= sched_slice(task_cfs_rq(p
), &p
->se
) >> 10;
681 p
->se
.avg
.runnable_avg_sum
= slice
;
682 p
->se
.avg
.runnable_avg_period
= slice
;
683 __update_task_entity_contrib(&p
->se
);
686 void init_task_runnable_average(struct task_struct
*p
)
692 * Update the current task's runtime statistics.
694 static void update_curr(struct cfs_rq
*cfs_rq
)
696 struct sched_entity
*curr
= cfs_rq
->curr
;
697 u64 now
= rq_clock_task(rq_of(cfs_rq
));
703 delta_exec
= now
- curr
->exec_start
;
704 if (unlikely((s64
)delta_exec
<= 0))
707 curr
->exec_start
= now
;
709 schedstat_set(curr
->statistics
.exec_max
,
710 max(delta_exec
, curr
->statistics
.exec_max
));
712 curr
->sum_exec_runtime
+= delta_exec
;
713 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
715 curr
->vruntime
+= calc_delta_fair(delta_exec
, curr
);
716 update_min_vruntime(cfs_rq
);
718 if (entity_is_task(curr
)) {
719 struct task_struct
*curtask
= task_of(curr
);
721 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
722 cpuacct_charge(curtask
, delta_exec
);
723 account_group_exec_runtime(curtask
, delta_exec
);
726 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
729 static void update_curr_fair(struct rq
*rq
)
731 update_curr(cfs_rq_of(&rq
->curr
->se
));
735 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
737 schedstat_set(se
->statistics
.wait_start
, rq_clock(rq_of(cfs_rq
)));
741 * Task is being enqueued - update stats:
743 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
746 * Are we enqueueing a waiting task? (for current tasks
747 * a dequeue/enqueue event is a NOP)
749 if (se
!= cfs_rq
->curr
)
750 update_stats_wait_start(cfs_rq
, se
);
754 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
756 schedstat_set(se
->statistics
.wait_max
, max(se
->statistics
.wait_max
,
757 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
));
758 schedstat_set(se
->statistics
.wait_count
, se
->statistics
.wait_count
+ 1);
759 schedstat_set(se
->statistics
.wait_sum
, se
->statistics
.wait_sum
+
760 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
761 #ifdef CONFIG_SCHEDSTATS
762 if (entity_is_task(se
)) {
763 trace_sched_stat_wait(task_of(se
),
764 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
767 schedstat_set(se
->statistics
.wait_start
, 0);
771 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
774 * Mark the end of the wait period if dequeueing a
777 if (se
!= cfs_rq
->curr
)
778 update_stats_wait_end(cfs_rq
, se
);
782 * We are picking a new current task - update its stats:
785 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
788 * We are starting a new run period:
790 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
793 /**************************************************
794 * Scheduling class queueing methods:
797 #ifdef CONFIG_NUMA_BALANCING
799 * Approximate time to scan a full NUMA task in ms. The task scan period is
800 * calculated based on the tasks virtual memory size and
801 * numa_balancing_scan_size.
803 unsigned int sysctl_numa_balancing_scan_period_min
= 1000;
804 unsigned int sysctl_numa_balancing_scan_period_max
= 60000;
806 /* Portion of address space to scan in MB */
807 unsigned int sysctl_numa_balancing_scan_size
= 256;
809 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
810 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
812 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
814 unsigned long rss
= 0;
815 unsigned long nr_scan_pages
;
818 * Calculations based on RSS as non-present and empty pages are skipped
819 * by the PTE scanner and NUMA hinting faults should be trapped based
822 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
823 rss
= get_mm_rss(p
->mm
);
827 rss
= round_up(rss
, nr_scan_pages
);
828 return rss
/ nr_scan_pages
;
831 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
832 #define MAX_SCAN_WINDOW 2560
834 static unsigned int task_scan_min(struct task_struct
*p
)
836 unsigned int scan_size
= ACCESS_ONCE(sysctl_numa_balancing_scan_size
);
837 unsigned int scan
, floor
;
838 unsigned int windows
= 1;
840 if (scan_size
< MAX_SCAN_WINDOW
)
841 windows
= MAX_SCAN_WINDOW
/ scan_size
;
842 floor
= 1000 / windows
;
844 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
845 return max_t(unsigned int, floor
, scan
);
848 static unsigned int task_scan_max(struct task_struct
*p
)
850 unsigned int smin
= task_scan_min(p
);
853 /* Watch for min being lower than max due to floor calculations */
854 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
855 return max(smin
, smax
);
858 static void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
860 rq
->nr_numa_running
+= (p
->numa_preferred_nid
!= -1);
861 rq
->nr_preferred_running
+= (p
->numa_preferred_nid
== task_node(p
));
864 static void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
866 rq
->nr_numa_running
-= (p
->numa_preferred_nid
!= -1);
867 rq
->nr_preferred_running
-= (p
->numa_preferred_nid
== task_node(p
));
873 spinlock_t lock
; /* nr_tasks, tasks */
878 nodemask_t active_nodes
;
879 unsigned long total_faults
;
881 * Faults_cpu is used to decide whether memory should move
882 * towards the CPU. As a consequence, these stats are weighted
883 * more by CPU use than by memory faults.
885 unsigned long *faults_cpu
;
886 unsigned long faults
[0];
889 /* Shared or private faults. */
890 #define NR_NUMA_HINT_FAULT_TYPES 2
892 /* Memory and CPU locality */
893 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
895 /* Averaged statistics, and temporary buffers. */
896 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
898 pid_t
task_numa_group_id(struct task_struct
*p
)
900 return p
->numa_group
? p
->numa_group
->gid
: 0;
904 * The averaged statistics, shared & private, memory & cpu,
905 * occupy the first half of the array. The second half of the
906 * array is for current counters, which are averaged into the
907 * first set by task_numa_placement.
909 static inline int task_faults_idx(enum numa_faults_stats s
, int nid
, int priv
)
911 return NR_NUMA_HINT_FAULT_TYPES
* (s
* nr_node_ids
+ nid
) + priv
;
914 static inline unsigned long task_faults(struct task_struct
*p
, int nid
)
919 return p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
920 p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
923 static inline unsigned long group_faults(struct task_struct
*p
, int nid
)
928 return p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
929 p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
932 static inline unsigned long group_faults_cpu(struct numa_group
*group
, int nid
)
934 return group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
935 group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 1)];
938 /* Handle placement on systems where not all nodes are directly connected. */
939 static unsigned long score_nearby_nodes(struct task_struct
*p
, int nid
,
940 int maxdist
, bool task
)
942 unsigned long score
= 0;
946 * All nodes are directly connected, and the same distance
947 * from each other. No need for fancy placement algorithms.
949 if (sched_numa_topology_type
== NUMA_DIRECT
)
953 * This code is called for each node, introducing N^2 complexity,
954 * which should be ok given the number of nodes rarely exceeds 8.
956 for_each_online_node(node
) {
957 unsigned long faults
;
958 int dist
= node_distance(nid
, node
);
961 * The furthest away nodes in the system are not interesting
962 * for placement; nid was already counted.
964 if (dist
== sched_max_numa_distance
|| node
== nid
)
968 * On systems with a backplane NUMA topology, compare groups
969 * of nodes, and move tasks towards the group with the most
970 * memory accesses. When comparing two nodes at distance
971 * "hoplimit", only nodes closer by than "hoplimit" are part
972 * of each group. Skip other nodes.
974 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
978 /* Add up the faults from nearby nodes. */
980 faults
= task_faults(p
, node
);
982 faults
= group_faults(p
, node
);
985 * On systems with a glueless mesh NUMA topology, there are
986 * no fixed "groups of nodes". Instead, nodes that are not
987 * directly connected bounce traffic through intermediate
988 * nodes; a numa_group can occupy any set of nodes.
989 * The further away a node is, the less the faults count.
990 * This seems to result in good task placement.
992 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
993 faults
*= (sched_max_numa_distance
- dist
);
994 faults
/= (sched_max_numa_distance
- LOCAL_DISTANCE
);
1004 * These return the fraction of accesses done by a particular task, or
1005 * task group, on a particular numa node. The group weight is given a
1006 * larger multiplier, in order to group tasks together that are almost
1007 * evenly spread out between numa nodes.
1009 static inline unsigned long task_weight(struct task_struct
*p
, int nid
,
1012 unsigned long faults
, total_faults
;
1014 if (!p
->numa_faults
)
1017 total_faults
= p
->total_numa_faults
;
1022 faults
= task_faults(p
, nid
);
1023 faults
+= score_nearby_nodes(p
, nid
, dist
, true);
1025 return 1000 * faults
/ total_faults
;
1028 static inline unsigned long group_weight(struct task_struct
*p
, int nid
,
1031 unsigned long faults
, total_faults
;
1036 total_faults
= p
->numa_group
->total_faults
;
1041 faults
= group_faults(p
, nid
);
1042 faults
+= score_nearby_nodes(p
, nid
, dist
, false);
1044 return 1000 * faults
/ total_faults
;
1047 bool should_numa_migrate_memory(struct task_struct
*p
, struct page
* page
,
1048 int src_nid
, int dst_cpu
)
1050 struct numa_group
*ng
= p
->numa_group
;
1051 int dst_nid
= cpu_to_node(dst_cpu
);
1052 int last_cpupid
, this_cpupid
;
1054 this_cpupid
= cpu_pid_to_cpupid(dst_cpu
, current
->pid
);
1057 * Multi-stage node selection is used in conjunction with a periodic
1058 * migration fault to build a temporal task<->page relation. By using
1059 * a two-stage filter we remove short/unlikely relations.
1061 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1062 * a task's usage of a particular page (n_p) per total usage of this
1063 * page (n_t) (in a given time-span) to a probability.
1065 * Our periodic faults will sample this probability and getting the
1066 * same result twice in a row, given these samples are fully
1067 * independent, is then given by P(n)^2, provided our sample period
1068 * is sufficiently short compared to the usage pattern.
1070 * This quadric squishes small probabilities, making it less likely we
1071 * act on an unlikely task<->page relation.
1073 last_cpupid
= page_cpupid_xchg_last(page
, this_cpupid
);
1074 if (!cpupid_pid_unset(last_cpupid
) &&
1075 cpupid_to_nid(last_cpupid
) != dst_nid
)
1078 /* Always allow migrate on private faults */
1079 if (cpupid_match_pid(p
, last_cpupid
))
1082 /* A shared fault, but p->numa_group has not been set up yet. */
1087 * Do not migrate if the destination is not a node that
1088 * is actively used by this numa group.
1090 if (!node_isset(dst_nid
, ng
->active_nodes
))
1094 * Source is a node that is not actively used by this
1095 * numa group, while the destination is. Migrate.
1097 if (!node_isset(src_nid
, ng
->active_nodes
))
1101 * Both source and destination are nodes in active
1102 * use by this numa group. Maximize memory bandwidth
1103 * by migrating from more heavily used groups, to less
1104 * heavily used ones, spreading the load around.
1105 * Use a 1/4 hysteresis to avoid spurious page movement.
1107 return group_faults(p
, dst_nid
) < (group_faults(p
, src_nid
) * 3 / 4);
1110 static unsigned long weighted_cpuload(const int cpu
);
1111 static unsigned long source_load(int cpu
, int type
);
1112 static unsigned long target_load(int cpu
, int type
);
1113 static unsigned long capacity_of(int cpu
);
1114 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
);
1116 /* Cached statistics for all CPUs within a node */
1118 unsigned long nr_running
;
1121 /* Total compute capacity of CPUs on a node */
1122 unsigned long compute_capacity
;
1124 /* Approximate capacity in terms of runnable tasks on a node */
1125 unsigned long task_capacity
;
1126 int has_free_capacity
;
1130 * XXX borrowed from update_sg_lb_stats
1132 static void update_numa_stats(struct numa_stats
*ns
, int nid
)
1134 int smt
, cpu
, cpus
= 0;
1135 unsigned long capacity
;
1137 memset(ns
, 0, sizeof(*ns
));
1138 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
1139 struct rq
*rq
= cpu_rq(cpu
);
1141 ns
->nr_running
+= rq
->nr_running
;
1142 ns
->load
+= weighted_cpuload(cpu
);
1143 ns
->compute_capacity
+= capacity_of(cpu
);
1149 * If we raced with hotplug and there are no CPUs left in our mask
1150 * the @ns structure is NULL'ed and task_numa_compare() will
1151 * not find this node attractive.
1153 * We'll either bail at !has_free_capacity, or we'll detect a huge
1154 * imbalance and bail there.
1159 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1160 smt
= DIV_ROUND_UP(SCHED_CAPACITY_SCALE
* cpus
, ns
->compute_capacity
);
1161 capacity
= cpus
/ smt
; /* cores */
1163 ns
->task_capacity
= min_t(unsigned, capacity
,
1164 DIV_ROUND_CLOSEST(ns
->compute_capacity
, SCHED_CAPACITY_SCALE
));
1165 ns
->has_free_capacity
= (ns
->nr_running
< ns
->task_capacity
);
1168 struct task_numa_env
{
1169 struct task_struct
*p
;
1171 int src_cpu
, src_nid
;
1172 int dst_cpu
, dst_nid
;
1174 struct numa_stats src_stats
, dst_stats
;
1179 struct task_struct
*best_task
;
1184 static void task_numa_assign(struct task_numa_env
*env
,
1185 struct task_struct
*p
, long imp
)
1188 put_task_struct(env
->best_task
);
1193 env
->best_imp
= imp
;
1194 env
->best_cpu
= env
->dst_cpu
;
1197 static bool load_too_imbalanced(long src_load
, long dst_load
,
1198 struct task_numa_env
*env
)
1201 long orig_src_load
, orig_dst_load
;
1202 long src_capacity
, dst_capacity
;
1205 * The load is corrected for the CPU capacity available on each node.
1208 * ------------ vs ---------
1209 * src_capacity dst_capacity
1211 src_capacity
= env
->src_stats
.compute_capacity
;
1212 dst_capacity
= env
->dst_stats
.compute_capacity
;
1214 /* We care about the slope of the imbalance, not the direction. */
1215 if (dst_load
< src_load
)
1216 swap(dst_load
, src_load
);
1218 /* Is the difference below the threshold? */
1219 imb
= dst_load
* src_capacity
* 100 -
1220 src_load
* dst_capacity
* env
->imbalance_pct
;
1225 * The imbalance is above the allowed threshold.
1226 * Compare it with the old imbalance.
1228 orig_src_load
= env
->src_stats
.load
;
1229 orig_dst_load
= env
->dst_stats
.load
;
1231 if (orig_dst_load
< orig_src_load
)
1232 swap(orig_dst_load
, orig_src_load
);
1234 old_imb
= orig_dst_load
* src_capacity
* 100 -
1235 orig_src_load
* dst_capacity
* env
->imbalance_pct
;
1237 /* Would this change make things worse? */
1238 return (imb
> old_imb
);
1242 * This checks if the overall compute and NUMA accesses of the system would
1243 * be improved if the source tasks was migrated to the target dst_cpu taking
1244 * into account that it might be best if task running on the dst_cpu should
1245 * be exchanged with the source task
1247 static void task_numa_compare(struct task_numa_env
*env
,
1248 long taskimp
, long groupimp
)
1250 struct rq
*src_rq
= cpu_rq(env
->src_cpu
);
1251 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1252 struct task_struct
*cur
;
1253 long src_load
, dst_load
;
1255 long imp
= env
->p
->numa_group
? groupimp
: taskimp
;
1257 int dist
= env
->dist
;
1261 raw_spin_lock_irq(&dst_rq
->lock
);
1264 * No need to move the exiting task, and this ensures that ->curr
1265 * wasn't reaped and thus get_task_struct() in task_numa_assign()
1266 * is safe under RCU read lock.
1267 * Note that rcu_read_lock() itself can't protect from the final
1268 * put_task_struct() after the last schedule().
1270 if ((cur
->flags
& PF_EXITING
) || is_idle_task(cur
))
1272 raw_spin_unlock_irq(&dst_rq
->lock
);
1275 * Because we have preemption enabled we can get migrated around and
1276 * end try selecting ourselves (current == env->p) as a swap candidate.
1282 * "imp" is the fault differential for the source task between the
1283 * source and destination node. Calculate the total differential for
1284 * the source task and potential destination task. The more negative
1285 * the value is, the more rmeote accesses that would be expected to
1286 * be incurred if the tasks were swapped.
1289 /* Skip this swap candidate if cannot move to the source cpu */
1290 if (!cpumask_test_cpu(env
->src_cpu
, tsk_cpus_allowed(cur
)))
1294 * If dst and source tasks are in the same NUMA group, or not
1295 * in any group then look only at task weights.
1297 if (cur
->numa_group
== env
->p
->numa_group
) {
1298 imp
= taskimp
+ task_weight(cur
, env
->src_nid
, dist
) -
1299 task_weight(cur
, env
->dst_nid
, dist
);
1301 * Add some hysteresis to prevent swapping the
1302 * tasks within a group over tiny differences.
1304 if (cur
->numa_group
)
1308 * Compare the group weights. If a task is all by
1309 * itself (not part of a group), use the task weight
1312 if (cur
->numa_group
)
1313 imp
+= group_weight(cur
, env
->src_nid
, dist
) -
1314 group_weight(cur
, env
->dst_nid
, dist
);
1316 imp
+= task_weight(cur
, env
->src_nid
, dist
) -
1317 task_weight(cur
, env
->dst_nid
, dist
);
1321 if (imp
<= env
->best_imp
&& moveimp
<= env
->best_imp
)
1325 /* Is there capacity at our destination? */
1326 if (env
->src_stats
.nr_running
<= env
->src_stats
.task_capacity
&&
1327 !env
->dst_stats
.has_free_capacity
)
1333 /* Balance doesn't matter much if we're running a task per cpu */
1334 if (imp
> env
->best_imp
&& src_rq
->nr_running
== 1 &&
1335 dst_rq
->nr_running
== 1)
1339 * In the overloaded case, try and keep the load balanced.
1342 load
= task_h_load(env
->p
);
1343 dst_load
= env
->dst_stats
.load
+ load
;
1344 src_load
= env
->src_stats
.load
- load
;
1346 if (moveimp
> imp
&& moveimp
> env
->best_imp
) {
1348 * If the improvement from just moving env->p direction is
1349 * better than swapping tasks around, check if a move is
1350 * possible. Store a slightly smaller score than moveimp,
1351 * so an actually idle CPU will win.
1353 if (!load_too_imbalanced(src_load
, dst_load
, env
)) {
1360 if (imp
<= env
->best_imp
)
1364 load
= task_h_load(cur
);
1369 if (load_too_imbalanced(src_load
, dst_load
, env
))
1373 * One idle CPU per node is evaluated for a task numa move.
1374 * Call select_idle_sibling to maybe find a better one.
1377 env
->dst_cpu
= select_idle_sibling(env
->p
, env
->dst_cpu
);
1380 task_numa_assign(env
, cur
, imp
);
1385 static void task_numa_find_cpu(struct task_numa_env
*env
,
1386 long taskimp
, long groupimp
)
1390 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1391 /* Skip this CPU if the source task cannot migrate */
1392 if (!cpumask_test_cpu(cpu
, tsk_cpus_allowed(env
->p
)))
1396 task_numa_compare(env
, taskimp
, groupimp
);
1400 static int task_numa_migrate(struct task_struct
*p
)
1402 struct task_numa_env env
= {
1405 .src_cpu
= task_cpu(p
),
1406 .src_nid
= task_node(p
),
1408 .imbalance_pct
= 112,
1414 struct sched_domain
*sd
;
1415 unsigned long taskweight
, groupweight
;
1417 long taskimp
, groupimp
;
1420 * Pick the lowest SD_NUMA domain, as that would have the smallest
1421 * imbalance and would be the first to start moving tasks about.
1423 * And we want to avoid any moving of tasks about, as that would create
1424 * random movement of tasks -- counter the numa conditions we're trying
1428 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1430 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1434 * Cpusets can break the scheduler domain tree into smaller
1435 * balance domains, some of which do not cross NUMA boundaries.
1436 * Tasks that are "trapped" in such domains cannot be migrated
1437 * elsewhere, so there is no point in (re)trying.
1439 if (unlikely(!sd
)) {
1440 p
->numa_preferred_nid
= task_node(p
);
1444 env
.dst_nid
= p
->numa_preferred_nid
;
1445 dist
= env
.dist
= node_distance(env
.src_nid
, env
.dst_nid
);
1446 taskweight
= task_weight(p
, env
.src_nid
, dist
);
1447 groupweight
= group_weight(p
, env
.src_nid
, dist
);
1448 update_numa_stats(&env
.src_stats
, env
.src_nid
);
1449 taskimp
= task_weight(p
, env
.dst_nid
, dist
) - taskweight
;
1450 groupimp
= group_weight(p
, env
.dst_nid
, dist
) - groupweight
;
1451 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1453 /* Try to find a spot on the preferred nid. */
1454 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1457 * Look at other nodes in these cases:
1458 * - there is no space available on the preferred_nid
1459 * - the task is part of a numa_group that is interleaved across
1460 * multiple NUMA nodes; in order to better consolidate the group,
1461 * we need to check other locations.
1463 if (env
.best_cpu
== -1 || (p
->numa_group
&&
1464 nodes_weight(p
->numa_group
->active_nodes
) > 1)) {
1465 for_each_online_node(nid
) {
1466 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
1469 dist
= node_distance(env
.src_nid
, env
.dst_nid
);
1470 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
1472 taskweight
= task_weight(p
, env
.src_nid
, dist
);
1473 groupweight
= group_weight(p
, env
.src_nid
, dist
);
1476 /* Only consider nodes where both task and groups benefit */
1477 taskimp
= task_weight(p
, nid
, dist
) - taskweight
;
1478 groupimp
= group_weight(p
, nid
, dist
) - groupweight
;
1479 if (taskimp
< 0 && groupimp
< 0)
1484 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1485 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1490 * If the task is part of a workload that spans multiple NUMA nodes,
1491 * and is migrating into one of the workload's active nodes, remember
1492 * this node as the task's preferred numa node, so the workload can
1494 * A task that migrated to a second choice node will be better off
1495 * trying for a better one later. Do not set the preferred node here.
1497 if (p
->numa_group
) {
1498 if (env
.best_cpu
== -1)
1503 if (node_isset(nid
, p
->numa_group
->active_nodes
))
1504 sched_setnuma(p
, env
.dst_nid
);
1507 /* No better CPU than the current one was found. */
1508 if (env
.best_cpu
== -1)
1512 * Reset the scan period if the task is being rescheduled on an
1513 * alternative node to recheck if the tasks is now properly placed.
1515 p
->numa_scan_period
= task_scan_min(p
);
1517 if (env
.best_task
== NULL
) {
1518 ret
= migrate_task_to(p
, env
.best_cpu
);
1520 trace_sched_stick_numa(p
, env
.src_cpu
, env
.best_cpu
);
1524 ret
= migrate_swap(p
, env
.best_task
);
1526 trace_sched_stick_numa(p
, env
.src_cpu
, task_cpu(env
.best_task
));
1527 put_task_struct(env
.best_task
);
1531 /* Attempt to migrate a task to a CPU on the preferred node. */
1532 static void numa_migrate_preferred(struct task_struct
*p
)
1534 unsigned long interval
= HZ
;
1536 /* This task has no NUMA fault statistics yet */
1537 if (unlikely(p
->numa_preferred_nid
== -1 || !p
->numa_faults
))
1540 /* Periodically retry migrating the task to the preferred node */
1541 interval
= min(interval
, msecs_to_jiffies(p
->numa_scan_period
) / 16);
1542 p
->numa_migrate_retry
= jiffies
+ interval
;
1544 /* Success if task is already running on preferred CPU */
1545 if (task_node(p
) == p
->numa_preferred_nid
)
1548 /* Otherwise, try migrate to a CPU on the preferred node */
1549 task_numa_migrate(p
);
1553 * Find the nodes on which the workload is actively running. We do this by
1554 * tracking the nodes from which NUMA hinting faults are triggered. This can
1555 * be different from the set of nodes where the workload's memory is currently
1558 * The bitmask is used to make smarter decisions on when to do NUMA page
1559 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1560 * are added when they cause over 6/16 of the maximum number of faults, but
1561 * only removed when they drop below 3/16.
1563 static void update_numa_active_node_mask(struct numa_group
*numa_group
)
1565 unsigned long faults
, max_faults
= 0;
1568 for_each_online_node(nid
) {
1569 faults
= group_faults_cpu(numa_group
, nid
);
1570 if (faults
> max_faults
)
1571 max_faults
= faults
;
1574 for_each_online_node(nid
) {
1575 faults
= group_faults_cpu(numa_group
, nid
);
1576 if (!node_isset(nid
, numa_group
->active_nodes
)) {
1577 if (faults
> max_faults
* 6 / 16)
1578 node_set(nid
, numa_group
->active_nodes
);
1579 } else if (faults
< max_faults
* 3 / 16)
1580 node_clear(nid
, numa_group
->active_nodes
);
1585 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1586 * increments. The more local the fault statistics are, the higher the scan
1587 * period will be for the next scan window. If local/(local+remote) ratio is
1588 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1589 * the scan period will decrease. Aim for 70% local accesses.
1591 #define NUMA_PERIOD_SLOTS 10
1592 #define NUMA_PERIOD_THRESHOLD 7
1595 * Increase the scan period (slow down scanning) if the majority of
1596 * our memory is already on our local node, or if the majority of
1597 * the page accesses are shared with other processes.
1598 * Otherwise, decrease the scan period.
1600 static void update_task_scan_period(struct task_struct
*p
,
1601 unsigned long shared
, unsigned long private)
1603 unsigned int period_slot
;
1607 unsigned long remote
= p
->numa_faults_locality
[0];
1608 unsigned long local
= p
->numa_faults_locality
[1];
1611 * If there were no record hinting faults then either the task is
1612 * completely idle or all activity is areas that are not of interest
1613 * to automatic numa balancing. Scan slower
1615 if (local
+ shared
== 0) {
1616 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1617 p
->numa_scan_period
<< 1);
1619 p
->mm
->numa_next_scan
= jiffies
+
1620 msecs_to_jiffies(p
->numa_scan_period
);
1626 * Prepare to scale scan period relative to the current period.
1627 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1628 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1629 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1631 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
1632 ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
1633 if (ratio
>= NUMA_PERIOD_THRESHOLD
) {
1634 int slot
= ratio
- NUMA_PERIOD_THRESHOLD
;
1637 diff
= slot
* period_slot
;
1639 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
1642 * Scale scan rate increases based on sharing. There is an
1643 * inverse relationship between the degree of sharing and
1644 * the adjustment made to the scanning period. Broadly
1645 * speaking the intent is that there is little point
1646 * scanning faster if shared accesses dominate as it may
1647 * simply bounce migrations uselessly
1649 ratio
= DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS
, (private + shared
+ 1));
1650 diff
= (diff
* ratio
) / NUMA_PERIOD_SLOTS
;
1653 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
1654 task_scan_min(p
), task_scan_max(p
));
1655 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1659 * Get the fraction of time the task has been running since the last
1660 * NUMA placement cycle. The scheduler keeps similar statistics, but
1661 * decays those on a 32ms period, which is orders of magnitude off
1662 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1663 * stats only if the task is so new there are no NUMA statistics yet.
1665 static u64
numa_get_avg_runtime(struct task_struct
*p
, u64
*period
)
1667 u64 runtime
, delta
, now
;
1668 /* Use the start of this time slice to avoid calculations. */
1669 now
= p
->se
.exec_start
;
1670 runtime
= p
->se
.sum_exec_runtime
;
1672 if (p
->last_task_numa_placement
) {
1673 delta
= runtime
- p
->last_sum_exec_runtime
;
1674 *period
= now
- p
->last_task_numa_placement
;
1676 delta
= p
->se
.avg
.runnable_avg_sum
;
1677 *period
= p
->se
.avg
.runnable_avg_period
;
1680 p
->last_sum_exec_runtime
= runtime
;
1681 p
->last_task_numa_placement
= now
;
1687 * Determine the preferred nid for a task in a numa_group. This needs to
1688 * be done in a way that produces consistent results with group_weight,
1689 * otherwise workloads might not converge.
1691 static int preferred_group_nid(struct task_struct
*p
, int nid
)
1696 /* Direct connections between all NUMA nodes. */
1697 if (sched_numa_topology_type
== NUMA_DIRECT
)
1701 * On a system with glueless mesh NUMA topology, group_weight
1702 * scores nodes according to the number of NUMA hinting faults on
1703 * both the node itself, and on nearby nodes.
1705 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
1706 unsigned long score
, max_score
= 0;
1707 int node
, max_node
= nid
;
1709 dist
= sched_max_numa_distance
;
1711 for_each_online_node(node
) {
1712 score
= group_weight(p
, node
, dist
);
1713 if (score
> max_score
) {
1722 * Finding the preferred nid in a system with NUMA backplane
1723 * interconnect topology is more involved. The goal is to locate
1724 * tasks from numa_groups near each other in the system, and
1725 * untangle workloads from different sides of the system. This requires
1726 * searching down the hierarchy of node groups, recursively searching
1727 * inside the highest scoring group of nodes. The nodemask tricks
1728 * keep the complexity of the search down.
1730 nodes
= node_online_map
;
1731 for (dist
= sched_max_numa_distance
; dist
> LOCAL_DISTANCE
; dist
--) {
1732 unsigned long max_faults
= 0;
1733 nodemask_t max_group
;
1736 /* Are there nodes at this distance from each other? */
1737 if (!find_numa_distance(dist
))
1740 for_each_node_mask(a
, nodes
) {
1741 unsigned long faults
= 0;
1742 nodemask_t this_group
;
1743 nodes_clear(this_group
);
1745 /* Sum group's NUMA faults; includes a==b case. */
1746 for_each_node_mask(b
, nodes
) {
1747 if (node_distance(a
, b
) < dist
) {
1748 faults
+= group_faults(p
, b
);
1749 node_set(b
, this_group
);
1750 node_clear(b
, nodes
);
1754 /* Remember the top group. */
1755 if (faults
> max_faults
) {
1756 max_faults
= faults
;
1757 max_group
= this_group
;
1759 * subtle: at the smallest distance there is
1760 * just one node left in each "group", the
1761 * winner is the preferred nid.
1766 /* Next round, evaluate the nodes within max_group. */
1772 static void task_numa_placement(struct task_struct
*p
)
1774 int seq
, nid
, max_nid
= -1, max_group_nid
= -1;
1775 unsigned long max_faults
= 0, max_group_faults
= 0;
1776 unsigned long fault_types
[2] = { 0, 0 };
1777 unsigned long total_faults
;
1778 u64 runtime
, period
;
1779 spinlock_t
*group_lock
= NULL
;
1781 seq
= ACCESS_ONCE(p
->mm
->numa_scan_seq
);
1782 if (p
->numa_scan_seq
== seq
)
1784 p
->numa_scan_seq
= seq
;
1785 p
->numa_scan_period_max
= task_scan_max(p
);
1787 total_faults
= p
->numa_faults_locality
[0] +
1788 p
->numa_faults_locality
[1];
1789 runtime
= numa_get_avg_runtime(p
, &period
);
1791 /* If the task is part of a group prevent parallel updates to group stats */
1792 if (p
->numa_group
) {
1793 group_lock
= &p
->numa_group
->lock
;
1794 spin_lock_irq(group_lock
);
1797 /* Find the node with the highest number of faults */
1798 for_each_online_node(nid
) {
1799 /* Keep track of the offsets in numa_faults array */
1800 int mem_idx
, membuf_idx
, cpu_idx
, cpubuf_idx
;
1801 unsigned long faults
= 0, group_faults
= 0;
1804 for (priv
= 0; priv
< NR_NUMA_HINT_FAULT_TYPES
; priv
++) {
1805 long diff
, f_diff
, f_weight
;
1807 mem_idx
= task_faults_idx(NUMA_MEM
, nid
, priv
);
1808 membuf_idx
= task_faults_idx(NUMA_MEMBUF
, nid
, priv
);
1809 cpu_idx
= task_faults_idx(NUMA_CPU
, nid
, priv
);
1810 cpubuf_idx
= task_faults_idx(NUMA_CPUBUF
, nid
, priv
);
1812 /* Decay existing window, copy faults since last scan */
1813 diff
= p
->numa_faults
[membuf_idx
] - p
->numa_faults
[mem_idx
] / 2;
1814 fault_types
[priv
] += p
->numa_faults
[membuf_idx
];
1815 p
->numa_faults
[membuf_idx
] = 0;
1818 * Normalize the faults_from, so all tasks in a group
1819 * count according to CPU use, instead of by the raw
1820 * number of faults. Tasks with little runtime have
1821 * little over-all impact on throughput, and thus their
1822 * faults are less important.
1824 f_weight
= div64_u64(runtime
<< 16, period
+ 1);
1825 f_weight
= (f_weight
* p
->numa_faults
[cpubuf_idx
]) /
1827 f_diff
= f_weight
- p
->numa_faults
[cpu_idx
] / 2;
1828 p
->numa_faults
[cpubuf_idx
] = 0;
1830 p
->numa_faults
[mem_idx
] += diff
;
1831 p
->numa_faults
[cpu_idx
] += f_diff
;
1832 faults
+= p
->numa_faults
[mem_idx
];
1833 p
->total_numa_faults
+= diff
;
1834 if (p
->numa_group
) {
1836 * safe because we can only change our own group
1838 * mem_idx represents the offset for a given
1839 * nid and priv in a specific region because it
1840 * is at the beginning of the numa_faults array.
1842 p
->numa_group
->faults
[mem_idx
] += diff
;
1843 p
->numa_group
->faults_cpu
[mem_idx
] += f_diff
;
1844 p
->numa_group
->total_faults
+= diff
;
1845 group_faults
+= p
->numa_group
->faults
[mem_idx
];
1849 if (faults
> max_faults
) {
1850 max_faults
= faults
;
1854 if (group_faults
> max_group_faults
) {
1855 max_group_faults
= group_faults
;
1856 max_group_nid
= nid
;
1860 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
1862 if (p
->numa_group
) {
1863 update_numa_active_node_mask(p
->numa_group
);
1864 spin_unlock_irq(group_lock
);
1865 max_nid
= preferred_group_nid(p
, max_group_nid
);
1869 /* Set the new preferred node */
1870 if (max_nid
!= p
->numa_preferred_nid
)
1871 sched_setnuma(p
, max_nid
);
1873 if (task_node(p
) != p
->numa_preferred_nid
)
1874 numa_migrate_preferred(p
);
1878 static inline int get_numa_group(struct numa_group
*grp
)
1880 return atomic_inc_not_zero(&grp
->refcount
);
1883 static inline void put_numa_group(struct numa_group
*grp
)
1885 if (atomic_dec_and_test(&grp
->refcount
))
1886 kfree_rcu(grp
, rcu
);
1889 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
1892 struct numa_group
*grp
, *my_grp
;
1893 struct task_struct
*tsk
;
1895 int cpu
= cpupid_to_cpu(cpupid
);
1898 if (unlikely(!p
->numa_group
)) {
1899 unsigned int size
= sizeof(struct numa_group
) +
1900 4*nr_node_ids
*sizeof(unsigned long);
1902 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
1906 atomic_set(&grp
->refcount
, 1);
1907 spin_lock_init(&grp
->lock
);
1909 /* Second half of the array tracks nids where faults happen */
1910 grp
->faults_cpu
= grp
->faults
+ NR_NUMA_HINT_FAULT_TYPES
*
1913 node_set(task_node(current
), grp
->active_nodes
);
1915 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
1916 grp
->faults
[i
] = p
->numa_faults
[i
];
1918 grp
->total_faults
= p
->total_numa_faults
;
1921 rcu_assign_pointer(p
->numa_group
, grp
);
1925 tsk
= ACCESS_ONCE(cpu_rq(cpu
)->curr
);
1927 if (!cpupid_match_pid(tsk
, cpupid
))
1930 grp
= rcu_dereference(tsk
->numa_group
);
1934 my_grp
= p
->numa_group
;
1939 * Only join the other group if its bigger; if we're the bigger group,
1940 * the other task will join us.
1942 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
1946 * Tie-break on the grp address.
1948 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
1951 /* Always join threads in the same process. */
1952 if (tsk
->mm
== current
->mm
)
1955 /* Simple filter to avoid false positives due to PID collisions */
1956 if (flags
& TNF_SHARED
)
1959 /* Update priv based on whether false sharing was detected */
1962 if (join
&& !get_numa_group(grp
))
1970 BUG_ON(irqs_disabled());
1971 double_lock_irq(&my_grp
->lock
, &grp
->lock
);
1973 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++) {
1974 my_grp
->faults
[i
] -= p
->numa_faults
[i
];
1975 grp
->faults
[i
] += p
->numa_faults
[i
];
1977 my_grp
->total_faults
-= p
->total_numa_faults
;
1978 grp
->total_faults
+= p
->total_numa_faults
;
1983 spin_unlock(&my_grp
->lock
);
1984 spin_unlock_irq(&grp
->lock
);
1986 rcu_assign_pointer(p
->numa_group
, grp
);
1988 put_numa_group(my_grp
);
1996 void task_numa_free(struct task_struct
*p
)
1998 struct numa_group
*grp
= p
->numa_group
;
1999 void *numa_faults
= p
->numa_faults
;
2000 unsigned long flags
;
2004 spin_lock_irqsave(&grp
->lock
, flags
);
2005 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2006 grp
->faults
[i
] -= p
->numa_faults
[i
];
2007 grp
->total_faults
-= p
->total_numa_faults
;
2010 spin_unlock_irqrestore(&grp
->lock
, flags
);
2011 RCU_INIT_POINTER(p
->numa_group
, NULL
);
2012 put_numa_group(grp
);
2015 p
->numa_faults
= NULL
;
2020 * Got a PROT_NONE fault for a page on @node.
2022 void task_numa_fault(int last_cpupid
, int mem_node
, int pages
, int flags
)
2024 struct task_struct
*p
= current
;
2025 bool migrated
= flags
& TNF_MIGRATED
;
2026 int cpu_node
= task_node(current
);
2027 int local
= !!(flags
& TNF_FAULT_LOCAL
);
2030 if (!numabalancing_enabled
)
2033 /* for example, ksmd faulting in a user's mm */
2037 /* Allocate buffer to track faults on a per-node basis */
2038 if (unlikely(!p
->numa_faults
)) {
2039 int size
= sizeof(*p
->numa_faults
) *
2040 NR_NUMA_HINT_FAULT_BUCKETS
* nr_node_ids
;
2042 p
->numa_faults
= kzalloc(size
, GFP_KERNEL
|__GFP_NOWARN
);
2043 if (!p
->numa_faults
)
2046 p
->total_numa_faults
= 0;
2047 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
2051 * First accesses are treated as private, otherwise consider accesses
2052 * to be private if the accessing pid has not changed
2054 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
2057 priv
= cpupid_match_pid(p
, last_cpupid
);
2058 if (!priv
&& !(flags
& TNF_NO_GROUP
))
2059 task_numa_group(p
, last_cpupid
, flags
, &priv
);
2063 * If a workload spans multiple NUMA nodes, a shared fault that
2064 * occurs wholly within the set of nodes that the workload is
2065 * actively using should be counted as local. This allows the
2066 * scan rate to slow down when a workload has settled down.
2068 if (!priv
&& !local
&& p
->numa_group
&&
2069 node_isset(cpu_node
, p
->numa_group
->active_nodes
) &&
2070 node_isset(mem_node
, p
->numa_group
->active_nodes
))
2073 task_numa_placement(p
);
2076 * Retry task to preferred node migration periodically, in case it
2077 * case it previously failed, or the scheduler moved us.
2079 if (time_after(jiffies
, p
->numa_migrate_retry
))
2080 numa_migrate_preferred(p
);
2083 p
->numa_pages_migrated
+= pages
;
2085 p
->numa_faults
[task_faults_idx(NUMA_MEMBUF
, mem_node
, priv
)] += pages
;
2086 p
->numa_faults
[task_faults_idx(NUMA_CPUBUF
, cpu_node
, priv
)] += pages
;
2087 p
->numa_faults_locality
[local
] += pages
;
2090 static void reset_ptenuma_scan(struct task_struct
*p
)
2092 ACCESS_ONCE(p
->mm
->numa_scan_seq
)++;
2093 p
->mm
->numa_scan_offset
= 0;
2097 * The expensive part of numa migration is done from task_work context.
2098 * Triggered from task_tick_numa().
2100 void task_numa_work(struct callback_head
*work
)
2102 unsigned long migrate
, next_scan
, now
= jiffies
;
2103 struct task_struct
*p
= current
;
2104 struct mm_struct
*mm
= p
->mm
;
2105 struct vm_area_struct
*vma
;
2106 unsigned long start
, end
;
2107 unsigned long nr_pte_updates
= 0;
2110 WARN_ON_ONCE(p
!= container_of(work
, struct task_struct
, numa_work
));
2112 work
->next
= work
; /* protect against double add */
2114 * Who cares about NUMA placement when they're dying.
2116 * NOTE: make sure not to dereference p->mm before this check,
2117 * exit_task_work() happens _after_ exit_mm() so we could be called
2118 * without p->mm even though we still had it when we enqueued this
2121 if (p
->flags
& PF_EXITING
)
2124 if (!mm
->numa_next_scan
) {
2125 mm
->numa_next_scan
= now
+
2126 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2130 * Enforce maximal scan/migration frequency..
2132 migrate
= mm
->numa_next_scan
;
2133 if (time_before(now
, migrate
))
2136 if (p
->numa_scan_period
== 0) {
2137 p
->numa_scan_period_max
= task_scan_max(p
);
2138 p
->numa_scan_period
= task_scan_min(p
);
2141 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
2142 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
2146 * Delay this task enough that another task of this mm will likely win
2147 * the next time around.
2149 p
->node_stamp
+= 2 * TICK_NSEC
;
2151 start
= mm
->numa_scan_offset
;
2152 pages
= sysctl_numa_balancing_scan_size
;
2153 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
2157 down_read(&mm
->mmap_sem
);
2158 vma
= find_vma(mm
, start
);
2160 reset_ptenuma_scan(p
);
2164 for (; vma
; vma
= vma
->vm_next
) {
2165 if (!vma_migratable(vma
) || !vma_policy_mof(vma
))
2169 * Shared library pages mapped by multiple processes are not
2170 * migrated as it is expected they are cache replicated. Avoid
2171 * hinting faults in read-only file-backed mappings or the vdso
2172 * as migrating the pages will be of marginal benefit.
2175 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
2179 * Skip inaccessible VMAs to avoid any confusion between
2180 * PROT_NONE and NUMA hinting ptes
2182 if (!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)))
2186 start
= max(start
, vma
->vm_start
);
2187 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
2188 end
= min(end
, vma
->vm_end
);
2189 nr_pte_updates
+= change_prot_numa(vma
, start
, end
);
2192 * Scan sysctl_numa_balancing_scan_size but ensure that
2193 * at least one PTE is updated so that unused virtual
2194 * address space is quickly skipped.
2197 pages
-= (end
- start
) >> PAGE_SHIFT
;
2204 } while (end
!= vma
->vm_end
);
2209 * It is possible to reach the end of the VMA list but the last few
2210 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2211 * would find the !migratable VMA on the next scan but not reset the
2212 * scanner to the start so check it now.
2215 mm
->numa_scan_offset
= start
;
2217 reset_ptenuma_scan(p
);
2218 up_read(&mm
->mmap_sem
);
2222 * Drive the periodic memory faults..
2224 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2226 struct callback_head
*work
= &curr
->numa_work
;
2230 * We don't care about NUMA placement if we don't have memory.
2232 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
2236 * Using runtime rather than walltime has the dual advantage that
2237 * we (mostly) drive the selection from busy threads and that the
2238 * task needs to have done some actual work before we bother with
2241 now
= curr
->se
.sum_exec_runtime
;
2242 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
2244 if (now
- curr
->node_stamp
> period
) {
2245 if (!curr
->node_stamp
)
2246 curr
->numa_scan_period
= task_scan_min(curr
);
2247 curr
->node_stamp
+= period
;
2249 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
2250 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
2251 task_work_add(curr
, work
, true);
2256 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2260 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
2264 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
2267 #endif /* CONFIG_NUMA_BALANCING */
2270 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2272 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
2273 if (!parent_entity(se
))
2274 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2276 if (entity_is_task(se
)) {
2277 struct rq
*rq
= rq_of(cfs_rq
);
2279 account_numa_enqueue(rq
, task_of(se
));
2280 list_add(&se
->group_node
, &rq
->cfs_tasks
);
2283 cfs_rq
->nr_running
++;
2287 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2289 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
2290 if (!parent_entity(se
))
2291 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2292 if (entity_is_task(se
)) {
2293 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
2294 list_del_init(&se
->group_node
);
2296 cfs_rq
->nr_running
--;
2299 #ifdef CONFIG_FAIR_GROUP_SCHED
2301 static inline long calc_tg_weight(struct task_group
*tg
, struct cfs_rq
*cfs_rq
)
2306 * Use this CPU's actual weight instead of the last load_contribution
2307 * to gain a more accurate current total weight. See
2308 * update_cfs_rq_load_contribution().
2310 tg_weight
= atomic_long_read(&tg
->load_avg
);
2311 tg_weight
-= cfs_rq
->tg_load_contrib
;
2312 tg_weight
+= cfs_rq
->load
.weight
;
2317 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2319 long tg_weight
, load
, shares
;
2321 tg_weight
= calc_tg_weight(tg
, cfs_rq
);
2322 load
= cfs_rq
->load
.weight
;
2324 shares
= (tg
->shares
* load
);
2326 shares
/= tg_weight
;
2328 if (shares
< MIN_SHARES
)
2329 shares
= MIN_SHARES
;
2330 if (shares
> tg
->shares
)
2331 shares
= tg
->shares
;
2335 # else /* CONFIG_SMP */
2336 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2340 # endif /* CONFIG_SMP */
2341 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
2342 unsigned long weight
)
2345 /* commit outstanding execution time */
2346 if (cfs_rq
->curr
== se
)
2347 update_curr(cfs_rq
);
2348 account_entity_dequeue(cfs_rq
, se
);
2351 update_load_set(&se
->load
, weight
);
2354 account_entity_enqueue(cfs_rq
, se
);
2357 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
2359 static void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2361 struct task_group
*tg
;
2362 struct sched_entity
*se
;
2366 se
= tg
->se
[cpu_of(rq_of(cfs_rq
))];
2367 if (!se
|| throttled_hierarchy(cfs_rq
))
2370 if (likely(se
->load
.weight
== tg
->shares
))
2373 shares
= calc_cfs_shares(cfs_rq
, tg
);
2375 reweight_entity(cfs_rq_of(se
), se
, shares
);
2377 #else /* CONFIG_FAIR_GROUP_SCHED */
2378 static inline void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2381 #endif /* CONFIG_FAIR_GROUP_SCHED */
2385 * We choose a half-life close to 1 scheduling period.
2386 * Note: The tables below are dependent on this value.
2388 #define LOAD_AVG_PERIOD 32
2389 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2390 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2392 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2393 static const u32 runnable_avg_yN_inv
[] = {
2394 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2395 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2396 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2397 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2398 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2399 0x85aac367, 0x82cd8698,
2403 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2404 * over-estimates when re-combining.
2406 static const u32 runnable_avg_yN_sum
[] = {
2407 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2408 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2409 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2414 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2416 static __always_inline u64
decay_load(u64 val
, u64 n
)
2418 unsigned int local_n
;
2422 else if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
2425 /* after bounds checking we can collapse to 32-bit */
2429 * As y^PERIOD = 1/2, we can combine
2430 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2431 * With a look-up table which covers y^n (n<PERIOD)
2433 * To achieve constant time decay_load.
2435 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
2436 val
>>= local_n
/ LOAD_AVG_PERIOD
;
2437 local_n
%= LOAD_AVG_PERIOD
;
2440 val
*= runnable_avg_yN_inv
[local_n
];
2441 /* We don't use SRR here since we always want to round down. */
2446 * For updates fully spanning n periods, the contribution to runnable
2447 * average will be: \Sum 1024*y^n
2449 * We can compute this reasonably efficiently by combining:
2450 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2452 static u32
__compute_runnable_contrib(u64 n
)
2456 if (likely(n
<= LOAD_AVG_PERIOD
))
2457 return runnable_avg_yN_sum
[n
];
2458 else if (unlikely(n
>= LOAD_AVG_MAX_N
))
2459 return LOAD_AVG_MAX
;
2461 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2463 contrib
/= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2464 contrib
+= runnable_avg_yN_sum
[LOAD_AVG_PERIOD
];
2466 n
-= LOAD_AVG_PERIOD
;
2467 } while (n
> LOAD_AVG_PERIOD
);
2469 contrib
= decay_load(contrib
, n
);
2470 return contrib
+ runnable_avg_yN_sum
[n
];
2474 * We can represent the historical contribution to runnable average as the
2475 * coefficients of a geometric series. To do this we sub-divide our runnable
2476 * history into segments of approximately 1ms (1024us); label the segment that
2477 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2479 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2481 * (now) (~1ms ago) (~2ms ago)
2483 * Let u_i denote the fraction of p_i that the entity was runnable.
2485 * We then designate the fractions u_i as our co-efficients, yielding the
2486 * following representation of historical load:
2487 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2489 * We choose y based on the with of a reasonably scheduling period, fixing:
2492 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2493 * approximately half as much as the contribution to load within the last ms
2496 * When a period "rolls over" and we have new u_0`, multiplying the previous
2497 * sum again by y is sufficient to update:
2498 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2499 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2501 static __always_inline
int __update_entity_runnable_avg(u64 now
,
2502 struct sched_avg
*sa
,
2506 u32 runnable_contrib
;
2507 int delta_w
, decayed
= 0;
2509 delta
= now
- sa
->last_runnable_update
;
2511 * This should only happen when time goes backwards, which it
2512 * unfortunately does during sched clock init when we swap over to TSC.
2514 if ((s64
)delta
< 0) {
2515 sa
->last_runnable_update
= now
;
2520 * Use 1024ns as the unit of measurement since it's a reasonable
2521 * approximation of 1us and fast to compute.
2526 sa
->last_runnable_update
= now
;
2528 /* delta_w is the amount already accumulated against our next period */
2529 delta_w
= sa
->runnable_avg_period
% 1024;
2530 if (delta
+ delta_w
>= 1024) {
2531 /* period roll-over */
2535 * Now that we know we're crossing a period boundary, figure
2536 * out how much from delta we need to complete the current
2537 * period and accrue it.
2539 delta_w
= 1024 - delta_w
;
2541 sa
->runnable_avg_sum
+= delta_w
;
2542 sa
->runnable_avg_period
+= delta_w
;
2546 /* Figure out how many additional periods this update spans */
2547 periods
= delta
/ 1024;
2550 sa
->runnable_avg_sum
= decay_load(sa
->runnable_avg_sum
,
2552 sa
->runnable_avg_period
= decay_load(sa
->runnable_avg_period
,
2555 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2556 runnable_contrib
= __compute_runnable_contrib(periods
);
2558 sa
->runnable_avg_sum
+= runnable_contrib
;
2559 sa
->runnable_avg_period
+= runnable_contrib
;
2562 /* Remainder of delta accrued against u_0` */
2564 sa
->runnable_avg_sum
+= delta
;
2565 sa
->runnable_avg_period
+= delta
;
2570 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2571 static inline u64
__synchronize_entity_decay(struct sched_entity
*se
)
2573 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2574 u64 decays
= atomic64_read(&cfs_rq
->decay_counter
);
2576 decays
-= se
->avg
.decay_count
;
2580 se
->avg
.load_avg_contrib
= decay_load(se
->avg
.load_avg_contrib
, decays
);
2581 se
->avg
.decay_count
= 0;
2586 #ifdef CONFIG_FAIR_GROUP_SCHED
2587 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2590 struct task_group
*tg
= cfs_rq
->tg
;
2593 tg_contrib
= cfs_rq
->runnable_load_avg
+ cfs_rq
->blocked_load_avg
;
2594 tg_contrib
-= cfs_rq
->tg_load_contrib
;
2599 if (force_update
|| abs(tg_contrib
) > cfs_rq
->tg_load_contrib
/ 8) {
2600 atomic_long_add(tg_contrib
, &tg
->load_avg
);
2601 cfs_rq
->tg_load_contrib
+= tg_contrib
;
2606 * Aggregate cfs_rq runnable averages into an equivalent task_group
2607 * representation for computing load contributions.
2609 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2610 struct cfs_rq
*cfs_rq
)
2612 struct task_group
*tg
= cfs_rq
->tg
;
2615 /* The fraction of a cpu used by this cfs_rq */
2616 contrib
= div_u64((u64
)sa
->runnable_avg_sum
<< NICE_0_SHIFT
,
2617 sa
->runnable_avg_period
+ 1);
2618 contrib
-= cfs_rq
->tg_runnable_contrib
;
2620 if (abs(contrib
) > cfs_rq
->tg_runnable_contrib
/ 64) {
2621 atomic_add(contrib
, &tg
->runnable_avg
);
2622 cfs_rq
->tg_runnable_contrib
+= contrib
;
2626 static inline void __update_group_entity_contrib(struct sched_entity
*se
)
2628 struct cfs_rq
*cfs_rq
= group_cfs_rq(se
);
2629 struct task_group
*tg
= cfs_rq
->tg
;
2634 contrib
= cfs_rq
->tg_load_contrib
* tg
->shares
;
2635 se
->avg
.load_avg_contrib
= div_u64(contrib
,
2636 atomic_long_read(&tg
->load_avg
) + 1);
2639 * For group entities we need to compute a correction term in the case
2640 * that they are consuming <1 cpu so that we would contribute the same
2641 * load as a task of equal weight.
2643 * Explicitly co-ordinating this measurement would be expensive, but
2644 * fortunately the sum of each cpus contribution forms a usable
2645 * lower-bound on the true value.
2647 * Consider the aggregate of 2 contributions. Either they are disjoint
2648 * (and the sum represents true value) or they are disjoint and we are
2649 * understating by the aggregate of their overlap.
2651 * Extending this to N cpus, for a given overlap, the maximum amount we
2652 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2653 * cpus that overlap for this interval and w_i is the interval width.
2655 * On a small machine; the first term is well-bounded which bounds the
2656 * total error since w_i is a subset of the period. Whereas on a
2657 * larger machine, while this first term can be larger, if w_i is the
2658 * of consequential size guaranteed to see n_i*w_i quickly converge to
2659 * our upper bound of 1-cpu.
2661 runnable_avg
= atomic_read(&tg
->runnable_avg
);
2662 if (runnable_avg
< NICE_0_LOAD
) {
2663 se
->avg
.load_avg_contrib
*= runnable_avg
;
2664 se
->avg
.load_avg_contrib
>>= NICE_0_SHIFT
;
2668 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
)
2670 __update_entity_runnable_avg(rq_clock_task(rq
), &rq
->avg
, runnable
);
2671 __update_tg_runnable_avg(&rq
->avg
, &rq
->cfs
);
2673 #else /* CONFIG_FAIR_GROUP_SCHED */
2674 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2675 int force_update
) {}
2676 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2677 struct cfs_rq
*cfs_rq
) {}
2678 static inline void __update_group_entity_contrib(struct sched_entity
*se
) {}
2679 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2680 #endif /* CONFIG_FAIR_GROUP_SCHED */
2682 static inline void __update_task_entity_contrib(struct sched_entity
*se
)
2686 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2687 contrib
= se
->avg
.runnable_avg_sum
* scale_load_down(se
->load
.weight
);
2688 contrib
/= (se
->avg
.runnable_avg_period
+ 1);
2689 se
->avg
.load_avg_contrib
= scale_load(contrib
);
2692 /* Compute the current contribution to load_avg by se, return any delta */
2693 static long __update_entity_load_avg_contrib(struct sched_entity
*se
)
2695 long old_contrib
= se
->avg
.load_avg_contrib
;
2697 if (entity_is_task(se
)) {
2698 __update_task_entity_contrib(se
);
2700 __update_tg_runnable_avg(&se
->avg
, group_cfs_rq(se
));
2701 __update_group_entity_contrib(se
);
2704 return se
->avg
.load_avg_contrib
- old_contrib
;
2707 static inline void subtract_blocked_load_contrib(struct cfs_rq
*cfs_rq
,
2710 if (likely(load_contrib
< cfs_rq
->blocked_load_avg
))
2711 cfs_rq
->blocked_load_avg
-= load_contrib
;
2713 cfs_rq
->blocked_load_avg
= 0;
2716 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
2718 /* Update a sched_entity's runnable average */
2719 static inline void update_entity_load_avg(struct sched_entity
*se
,
2722 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2727 * For a group entity we need to use their owned cfs_rq_clock_task() in
2728 * case they are the parent of a throttled hierarchy.
2730 if (entity_is_task(se
))
2731 now
= cfs_rq_clock_task(cfs_rq
);
2733 now
= cfs_rq_clock_task(group_cfs_rq(se
));
2735 if (!__update_entity_runnable_avg(now
, &se
->avg
, se
->on_rq
))
2738 contrib_delta
= __update_entity_load_avg_contrib(se
);
2744 cfs_rq
->runnable_load_avg
+= contrib_delta
;
2746 subtract_blocked_load_contrib(cfs_rq
, -contrib_delta
);
2750 * Decay the load contributed by all blocked children and account this so that
2751 * their contribution may appropriately discounted when they wake up.
2753 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
, int force_update
)
2755 u64 now
= cfs_rq_clock_task(cfs_rq
) >> 20;
2758 decays
= now
- cfs_rq
->last_decay
;
2759 if (!decays
&& !force_update
)
2762 if (atomic_long_read(&cfs_rq
->removed_load
)) {
2763 unsigned long removed_load
;
2764 removed_load
= atomic_long_xchg(&cfs_rq
->removed_load
, 0);
2765 subtract_blocked_load_contrib(cfs_rq
, removed_load
);
2769 cfs_rq
->blocked_load_avg
= decay_load(cfs_rq
->blocked_load_avg
,
2771 atomic64_add(decays
, &cfs_rq
->decay_counter
);
2772 cfs_rq
->last_decay
= now
;
2775 __update_cfs_rq_tg_load_contrib(cfs_rq
, force_update
);
2778 /* Add the load generated by se into cfs_rq's child load-average */
2779 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2780 struct sched_entity
*se
,
2784 * We track migrations using entity decay_count <= 0, on a wake-up
2785 * migration we use a negative decay count to track the remote decays
2786 * accumulated while sleeping.
2788 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2789 * are seen by enqueue_entity_load_avg() as a migration with an already
2790 * constructed load_avg_contrib.
2792 if (unlikely(se
->avg
.decay_count
<= 0)) {
2793 se
->avg
.last_runnable_update
= rq_clock_task(rq_of(cfs_rq
));
2794 if (se
->avg
.decay_count
) {
2796 * In a wake-up migration we have to approximate the
2797 * time sleeping. This is because we can't synchronize
2798 * clock_task between the two cpus, and it is not
2799 * guaranteed to be read-safe. Instead, we can
2800 * approximate this using our carried decays, which are
2801 * explicitly atomically readable.
2803 se
->avg
.last_runnable_update
-= (-se
->avg
.decay_count
)
2805 update_entity_load_avg(se
, 0);
2806 /* Indicate that we're now synchronized and on-rq */
2807 se
->avg
.decay_count
= 0;
2811 __synchronize_entity_decay(se
);
2814 /* migrated tasks did not contribute to our blocked load */
2816 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
2817 update_entity_load_avg(se
, 0);
2820 cfs_rq
->runnable_load_avg
+= se
->avg
.load_avg_contrib
;
2821 /* we force update consideration on load-balancer moves */
2822 update_cfs_rq_blocked_load(cfs_rq
, !wakeup
);
2826 * Remove se's load from this cfs_rq child load-average, if the entity is
2827 * transitioning to a blocked state we track its projected decay using
2830 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2831 struct sched_entity
*se
,
2834 update_entity_load_avg(se
, 1);
2835 /* we force update consideration on load-balancer moves */
2836 update_cfs_rq_blocked_load(cfs_rq
, !sleep
);
2838 cfs_rq
->runnable_load_avg
-= se
->avg
.load_avg_contrib
;
2840 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
2841 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
2842 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2846 * Update the rq's load with the elapsed running time before entering
2847 * idle. if the last scheduled task is not a CFS task, idle_enter will
2848 * be the only way to update the runnable statistic.
2850 void idle_enter_fair(struct rq
*this_rq
)
2852 update_rq_runnable_avg(this_rq
, 1);
2856 * Update the rq's load with the elapsed idle time before a task is
2857 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2858 * be the only way to update the runnable statistic.
2860 void idle_exit_fair(struct rq
*this_rq
)
2862 update_rq_runnable_avg(this_rq
, 0);
2865 static int idle_balance(struct rq
*this_rq
);
2867 #else /* CONFIG_SMP */
2869 static inline void update_entity_load_avg(struct sched_entity
*se
,
2870 int update_cfs_rq
) {}
2871 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2872 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2873 struct sched_entity
*se
,
2875 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2876 struct sched_entity
*se
,
2878 static inline void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
2879 int force_update
) {}
2881 static inline int idle_balance(struct rq
*rq
)
2886 #endif /* CONFIG_SMP */
2888 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2890 #ifdef CONFIG_SCHEDSTATS
2891 struct task_struct
*tsk
= NULL
;
2893 if (entity_is_task(se
))
2896 if (se
->statistics
.sleep_start
) {
2897 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.sleep_start
;
2902 if (unlikely(delta
> se
->statistics
.sleep_max
))
2903 se
->statistics
.sleep_max
= delta
;
2905 se
->statistics
.sleep_start
= 0;
2906 se
->statistics
.sum_sleep_runtime
+= delta
;
2909 account_scheduler_latency(tsk
, delta
>> 10, 1);
2910 trace_sched_stat_sleep(tsk
, delta
);
2913 if (se
->statistics
.block_start
) {
2914 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.block_start
;
2919 if (unlikely(delta
> se
->statistics
.block_max
))
2920 se
->statistics
.block_max
= delta
;
2922 se
->statistics
.block_start
= 0;
2923 se
->statistics
.sum_sleep_runtime
+= delta
;
2926 if (tsk
->in_iowait
) {
2927 se
->statistics
.iowait_sum
+= delta
;
2928 se
->statistics
.iowait_count
++;
2929 trace_sched_stat_iowait(tsk
, delta
);
2932 trace_sched_stat_blocked(tsk
, delta
);
2935 * Blocking time is in units of nanosecs, so shift by
2936 * 20 to get a milliseconds-range estimation of the
2937 * amount of time that the task spent sleeping:
2939 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
2940 profile_hits(SLEEP_PROFILING
,
2941 (void *)get_wchan(tsk
),
2944 account_scheduler_latency(tsk
, delta
>> 10, 0);
2950 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2952 #ifdef CONFIG_SCHED_DEBUG
2953 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
2958 if (d
> 3*sysctl_sched_latency
)
2959 schedstat_inc(cfs_rq
, nr_spread_over
);
2964 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
2966 u64 vruntime
= cfs_rq
->min_vruntime
;
2969 * The 'current' period is already promised to the current tasks,
2970 * however the extra weight of the new task will slow them down a
2971 * little, place the new task so that it fits in the slot that
2972 * stays open at the end.
2974 if (initial
&& sched_feat(START_DEBIT
))
2975 vruntime
+= sched_vslice(cfs_rq
, se
);
2977 /* sleeps up to a single latency don't count. */
2979 unsigned long thresh
= sysctl_sched_latency
;
2982 * Halve their sleep time's effect, to allow
2983 * for a gentler effect of sleepers:
2985 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
2991 /* ensure we never gain time by being placed backwards. */
2992 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
2995 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
2998 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3001 * Update the normalized vruntime before updating min_vruntime
3002 * through calling update_curr().
3004 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
3005 se
->vruntime
+= cfs_rq
->min_vruntime
;
3008 * Update run-time statistics of the 'current'.
3010 update_curr(cfs_rq
);
3011 enqueue_entity_load_avg(cfs_rq
, se
, flags
& ENQUEUE_WAKEUP
);
3012 account_entity_enqueue(cfs_rq
, se
);
3013 update_cfs_shares(cfs_rq
);
3015 if (flags
& ENQUEUE_WAKEUP
) {
3016 place_entity(cfs_rq
, se
, 0);
3017 enqueue_sleeper(cfs_rq
, se
);
3020 update_stats_enqueue(cfs_rq
, se
);
3021 check_spread(cfs_rq
, se
);
3022 if (se
!= cfs_rq
->curr
)
3023 __enqueue_entity(cfs_rq
, se
);
3026 if (cfs_rq
->nr_running
== 1) {
3027 list_add_leaf_cfs_rq(cfs_rq
);
3028 check_enqueue_throttle(cfs_rq
);
3032 static void __clear_buddies_last(struct sched_entity
*se
)
3034 for_each_sched_entity(se
) {
3035 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3036 if (cfs_rq
->last
!= se
)
3039 cfs_rq
->last
= NULL
;
3043 static void __clear_buddies_next(struct sched_entity
*se
)
3045 for_each_sched_entity(se
) {
3046 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3047 if (cfs_rq
->next
!= se
)
3050 cfs_rq
->next
= NULL
;
3054 static void __clear_buddies_skip(struct sched_entity
*se
)
3056 for_each_sched_entity(se
) {
3057 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3058 if (cfs_rq
->skip
!= se
)
3061 cfs_rq
->skip
= NULL
;
3065 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3067 if (cfs_rq
->last
== se
)
3068 __clear_buddies_last(se
);
3070 if (cfs_rq
->next
== se
)
3071 __clear_buddies_next(se
);
3073 if (cfs_rq
->skip
== se
)
3074 __clear_buddies_skip(se
);
3077 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3080 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3083 * Update run-time statistics of the 'current'.
3085 update_curr(cfs_rq
);
3086 dequeue_entity_load_avg(cfs_rq
, se
, flags
& DEQUEUE_SLEEP
);
3088 update_stats_dequeue(cfs_rq
, se
);
3089 if (flags
& DEQUEUE_SLEEP
) {
3090 #ifdef CONFIG_SCHEDSTATS
3091 if (entity_is_task(se
)) {
3092 struct task_struct
*tsk
= task_of(se
);
3094 if (tsk
->state
& TASK_INTERRUPTIBLE
)
3095 se
->statistics
.sleep_start
= rq_clock(rq_of(cfs_rq
));
3096 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
3097 se
->statistics
.block_start
= rq_clock(rq_of(cfs_rq
));
3102 clear_buddies(cfs_rq
, se
);
3104 if (se
!= cfs_rq
->curr
)
3105 __dequeue_entity(cfs_rq
, se
);
3107 account_entity_dequeue(cfs_rq
, se
);
3110 * Normalize the entity after updating the min_vruntime because the
3111 * update can refer to the ->curr item and we need to reflect this
3112 * movement in our normalized position.
3114 if (!(flags
& DEQUEUE_SLEEP
))
3115 se
->vruntime
-= cfs_rq
->min_vruntime
;
3117 /* return excess runtime on last dequeue */
3118 return_cfs_rq_runtime(cfs_rq
);
3120 update_min_vruntime(cfs_rq
);
3121 update_cfs_shares(cfs_rq
);
3125 * Preempt the current task with a newly woken task if needed:
3128 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
3130 unsigned long ideal_runtime
, delta_exec
;
3131 struct sched_entity
*se
;
3134 ideal_runtime
= sched_slice(cfs_rq
, curr
);
3135 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
3136 if (delta_exec
> ideal_runtime
) {
3137 resched_curr(rq_of(cfs_rq
));
3139 * The current task ran long enough, ensure it doesn't get
3140 * re-elected due to buddy favours.
3142 clear_buddies(cfs_rq
, curr
);
3147 * Ensure that a task that missed wakeup preemption by a
3148 * narrow margin doesn't have to wait for a full slice.
3149 * This also mitigates buddy induced latencies under load.
3151 if (delta_exec
< sysctl_sched_min_granularity
)
3154 se
= __pick_first_entity(cfs_rq
);
3155 delta
= curr
->vruntime
- se
->vruntime
;
3160 if (delta
> ideal_runtime
)
3161 resched_curr(rq_of(cfs_rq
));
3165 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3167 /* 'current' is not kept within the tree. */
3170 * Any task has to be enqueued before it get to execute on
3171 * a CPU. So account for the time it spent waiting on the
3174 update_stats_wait_end(cfs_rq
, se
);
3175 __dequeue_entity(cfs_rq
, se
);
3178 update_stats_curr_start(cfs_rq
, se
);
3180 #ifdef CONFIG_SCHEDSTATS
3182 * Track our maximum slice length, if the CPU's load is at
3183 * least twice that of our own weight (i.e. dont track it
3184 * when there are only lesser-weight tasks around):
3186 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
3187 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
3188 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
3191 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
3195 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
3198 * Pick the next process, keeping these things in mind, in this order:
3199 * 1) keep things fair between processes/task groups
3200 * 2) pick the "next" process, since someone really wants that to run
3201 * 3) pick the "last" process, for cache locality
3202 * 4) do not run the "skip" process, if something else is available
3204 static struct sched_entity
*
3205 pick_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
3207 struct sched_entity
*left
= __pick_first_entity(cfs_rq
);
3208 struct sched_entity
*se
;
3211 * If curr is set we have to see if its left of the leftmost entity
3212 * still in the tree, provided there was anything in the tree at all.
3214 if (!left
|| (curr
&& entity_before(curr
, left
)))
3217 se
= left
; /* ideally we run the leftmost entity */
3220 * Avoid running the skip buddy, if running something else can
3221 * be done without getting too unfair.
3223 if (cfs_rq
->skip
== se
) {
3224 struct sched_entity
*second
;
3227 second
= __pick_first_entity(cfs_rq
);
3229 second
= __pick_next_entity(se
);
3230 if (!second
|| (curr
&& entity_before(curr
, second
)))
3234 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
3239 * Prefer last buddy, try to return the CPU to a preempted task.
3241 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
3245 * Someone really wants this to run. If it's not unfair, run it.
3247 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
3250 clear_buddies(cfs_rq
, se
);
3255 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3257 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
3260 * If still on the runqueue then deactivate_task()
3261 * was not called and update_curr() has to be done:
3264 update_curr(cfs_rq
);
3266 /* throttle cfs_rqs exceeding runtime */
3267 check_cfs_rq_runtime(cfs_rq
);
3269 check_spread(cfs_rq
, prev
);
3271 update_stats_wait_start(cfs_rq
, prev
);
3272 /* Put 'current' back into the tree. */
3273 __enqueue_entity(cfs_rq
, prev
);
3274 /* in !on_rq case, update occurred at dequeue */
3275 update_entity_load_avg(prev
, 1);
3277 cfs_rq
->curr
= NULL
;
3281 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
3284 * Update run-time statistics of the 'current'.
3286 update_curr(cfs_rq
);
3289 * Ensure that runnable average is periodically updated.
3291 update_entity_load_avg(curr
, 1);
3292 update_cfs_rq_blocked_load(cfs_rq
, 1);
3293 update_cfs_shares(cfs_rq
);
3295 #ifdef CONFIG_SCHED_HRTICK
3297 * queued ticks are scheduled to match the slice, so don't bother
3298 * validating it and just reschedule.
3301 resched_curr(rq_of(cfs_rq
));
3305 * don't let the period tick interfere with the hrtick preemption
3307 if (!sched_feat(DOUBLE_TICK
) &&
3308 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
3312 if (cfs_rq
->nr_running
> 1)
3313 check_preempt_tick(cfs_rq
, curr
);
3317 /**************************************************
3318 * CFS bandwidth control machinery
3321 #ifdef CONFIG_CFS_BANDWIDTH
3323 #ifdef HAVE_JUMP_LABEL
3324 static struct static_key __cfs_bandwidth_used
;
3326 static inline bool cfs_bandwidth_used(void)
3328 return static_key_false(&__cfs_bandwidth_used
);
3331 void cfs_bandwidth_usage_inc(void)
3333 static_key_slow_inc(&__cfs_bandwidth_used
);
3336 void cfs_bandwidth_usage_dec(void)
3338 static_key_slow_dec(&__cfs_bandwidth_used
);
3340 #else /* HAVE_JUMP_LABEL */
3341 static bool cfs_bandwidth_used(void)
3346 void cfs_bandwidth_usage_inc(void) {}
3347 void cfs_bandwidth_usage_dec(void) {}
3348 #endif /* HAVE_JUMP_LABEL */
3351 * default period for cfs group bandwidth.
3352 * default: 0.1s, units: nanoseconds
3354 static inline u64
default_cfs_period(void)
3356 return 100000000ULL;
3359 static inline u64
sched_cfs_bandwidth_slice(void)
3361 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
3365 * Replenish runtime according to assigned quota and update expiration time.
3366 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3367 * additional synchronization around rq->lock.
3369 * requires cfs_b->lock
3371 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
3375 if (cfs_b
->quota
== RUNTIME_INF
)
3378 now
= sched_clock_cpu(smp_processor_id());
3379 cfs_b
->runtime
= cfs_b
->quota
;
3380 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
3383 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3385 return &tg
->cfs_bandwidth
;
3388 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3389 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3391 if (unlikely(cfs_rq
->throttle_count
))
3392 return cfs_rq
->throttled_clock_task
;
3394 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
3397 /* returns 0 on failure to allocate runtime */
3398 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3400 struct task_group
*tg
= cfs_rq
->tg
;
3401 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
3402 u64 amount
= 0, min_amount
, expires
;
3404 /* note: this is a positive sum as runtime_remaining <= 0 */
3405 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
3407 raw_spin_lock(&cfs_b
->lock
);
3408 if (cfs_b
->quota
== RUNTIME_INF
)
3409 amount
= min_amount
;
3412 * If the bandwidth pool has become inactive, then at least one
3413 * period must have elapsed since the last consumption.
3414 * Refresh the global state and ensure bandwidth timer becomes
3417 if (!cfs_b
->timer_active
) {
3418 __refill_cfs_bandwidth_runtime(cfs_b
);
3419 __start_cfs_bandwidth(cfs_b
, false);
3422 if (cfs_b
->runtime
> 0) {
3423 amount
= min(cfs_b
->runtime
, min_amount
);
3424 cfs_b
->runtime
-= amount
;
3428 expires
= cfs_b
->runtime_expires
;
3429 raw_spin_unlock(&cfs_b
->lock
);
3431 cfs_rq
->runtime_remaining
+= amount
;
3433 * we may have advanced our local expiration to account for allowed
3434 * spread between our sched_clock and the one on which runtime was
3437 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
3438 cfs_rq
->runtime_expires
= expires
;
3440 return cfs_rq
->runtime_remaining
> 0;
3444 * Note: This depends on the synchronization provided by sched_clock and the
3445 * fact that rq->clock snapshots this value.
3447 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3449 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3451 /* if the deadline is ahead of our clock, nothing to do */
3452 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
3455 if (cfs_rq
->runtime_remaining
< 0)
3459 * If the local deadline has passed we have to consider the
3460 * possibility that our sched_clock is 'fast' and the global deadline
3461 * has not truly expired.
3463 * Fortunately we can check determine whether this the case by checking
3464 * whether the global deadline has advanced. It is valid to compare
3465 * cfs_b->runtime_expires without any locks since we only care about
3466 * exact equality, so a partial write will still work.
3469 if (cfs_rq
->runtime_expires
!= cfs_b
->runtime_expires
) {
3470 /* extend local deadline, drift is bounded above by 2 ticks */
3471 cfs_rq
->runtime_expires
+= TICK_NSEC
;
3473 /* global deadline is ahead, expiration has passed */
3474 cfs_rq
->runtime_remaining
= 0;
3478 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3480 /* dock delta_exec before expiring quota (as it could span periods) */
3481 cfs_rq
->runtime_remaining
-= delta_exec
;
3482 expire_cfs_rq_runtime(cfs_rq
);
3484 if (likely(cfs_rq
->runtime_remaining
> 0))
3488 * if we're unable to extend our runtime we resched so that the active
3489 * hierarchy can be throttled
3491 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
3492 resched_curr(rq_of(cfs_rq
));
3495 static __always_inline
3496 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3498 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
3501 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
3504 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3506 return cfs_bandwidth_used() && cfs_rq
->throttled
;
3509 /* check whether cfs_rq, or any parent, is throttled */
3510 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3512 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
3516 * Ensure that neither of the group entities corresponding to src_cpu or
3517 * dest_cpu are members of a throttled hierarchy when performing group
3518 * load-balance operations.
3520 static inline int throttled_lb_pair(struct task_group
*tg
,
3521 int src_cpu
, int dest_cpu
)
3523 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
3525 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
3526 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
3528 return throttled_hierarchy(src_cfs_rq
) ||
3529 throttled_hierarchy(dest_cfs_rq
);
3532 /* updated child weight may affect parent so we have to do this bottom up */
3533 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
3535 struct rq
*rq
= data
;
3536 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3538 cfs_rq
->throttle_count
--;
3540 if (!cfs_rq
->throttle_count
) {
3541 /* adjust cfs_rq_clock_task() */
3542 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
3543 cfs_rq
->throttled_clock_task
;
3550 static int tg_throttle_down(struct task_group
*tg
, void *data
)
3552 struct rq
*rq
= data
;
3553 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3555 /* group is entering throttled state, stop time */
3556 if (!cfs_rq
->throttle_count
)
3557 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
3558 cfs_rq
->throttle_count
++;
3563 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
3565 struct rq
*rq
= rq_of(cfs_rq
);
3566 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3567 struct sched_entity
*se
;
3568 long task_delta
, dequeue
= 1;
3570 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
3572 /* freeze hierarchy runnable averages while throttled */
3574 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
3577 task_delta
= cfs_rq
->h_nr_running
;
3578 for_each_sched_entity(se
) {
3579 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
3580 /* throttled entity or throttle-on-deactivate */
3585 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
3586 qcfs_rq
->h_nr_running
-= task_delta
;
3588 if (qcfs_rq
->load
.weight
)
3593 sub_nr_running(rq
, task_delta
);
3595 cfs_rq
->throttled
= 1;
3596 cfs_rq
->throttled_clock
= rq_clock(rq
);
3597 raw_spin_lock(&cfs_b
->lock
);
3599 * Add to the _head_ of the list, so that an already-started
3600 * distribute_cfs_runtime will not see us
3602 list_add_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
3603 if (!cfs_b
->timer_active
)
3604 __start_cfs_bandwidth(cfs_b
, false);
3605 raw_spin_unlock(&cfs_b
->lock
);
3608 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
3610 struct rq
*rq
= rq_of(cfs_rq
);
3611 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3612 struct sched_entity
*se
;
3616 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
3618 cfs_rq
->throttled
= 0;
3620 update_rq_clock(rq
);
3622 raw_spin_lock(&cfs_b
->lock
);
3623 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
3624 list_del_rcu(&cfs_rq
->throttled_list
);
3625 raw_spin_unlock(&cfs_b
->lock
);
3627 /* update hierarchical throttle state */
3628 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
3630 if (!cfs_rq
->load
.weight
)
3633 task_delta
= cfs_rq
->h_nr_running
;
3634 for_each_sched_entity(se
) {
3638 cfs_rq
= cfs_rq_of(se
);
3640 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
3641 cfs_rq
->h_nr_running
+= task_delta
;
3643 if (cfs_rq_throttled(cfs_rq
))
3648 add_nr_running(rq
, task_delta
);
3650 /* determine whether we need to wake up potentially idle cpu */
3651 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
3655 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
3656 u64 remaining
, u64 expires
)
3658 struct cfs_rq
*cfs_rq
;
3660 u64 starting_runtime
= remaining
;
3663 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
3665 struct rq
*rq
= rq_of(cfs_rq
);
3667 raw_spin_lock(&rq
->lock
);
3668 if (!cfs_rq_throttled(cfs_rq
))
3671 runtime
= -cfs_rq
->runtime_remaining
+ 1;
3672 if (runtime
> remaining
)
3673 runtime
= remaining
;
3674 remaining
-= runtime
;
3676 cfs_rq
->runtime_remaining
+= runtime
;
3677 cfs_rq
->runtime_expires
= expires
;
3679 /* we check whether we're throttled above */
3680 if (cfs_rq
->runtime_remaining
> 0)
3681 unthrottle_cfs_rq(cfs_rq
);
3684 raw_spin_unlock(&rq
->lock
);
3691 return starting_runtime
- remaining
;
3695 * Responsible for refilling a task_group's bandwidth and unthrottling its
3696 * cfs_rqs as appropriate. If there has been no activity within the last
3697 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3698 * used to track this state.
3700 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
3702 u64 runtime
, runtime_expires
;
3705 /* no need to continue the timer with no bandwidth constraint */
3706 if (cfs_b
->quota
== RUNTIME_INF
)
3707 goto out_deactivate
;
3709 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3710 cfs_b
->nr_periods
+= overrun
;
3713 * idle depends on !throttled (for the case of a large deficit), and if
3714 * we're going inactive then everything else can be deferred
3716 if (cfs_b
->idle
&& !throttled
)
3717 goto out_deactivate
;
3720 * if we have relooped after returning idle once, we need to update our
3721 * status as actually running, so that other cpus doing
3722 * __start_cfs_bandwidth will stop trying to cancel us.
3724 cfs_b
->timer_active
= 1;
3726 __refill_cfs_bandwidth_runtime(cfs_b
);
3729 /* mark as potentially idle for the upcoming period */
3734 /* account preceding periods in which throttling occurred */
3735 cfs_b
->nr_throttled
+= overrun
;
3737 runtime_expires
= cfs_b
->runtime_expires
;
3740 * This check is repeated as we are holding onto the new bandwidth while
3741 * we unthrottle. This can potentially race with an unthrottled group
3742 * trying to acquire new bandwidth from the global pool. This can result
3743 * in us over-using our runtime if it is all used during this loop, but
3744 * only by limited amounts in that extreme case.
3746 while (throttled
&& cfs_b
->runtime
> 0) {
3747 runtime
= cfs_b
->runtime
;
3748 raw_spin_unlock(&cfs_b
->lock
);
3749 /* we can't nest cfs_b->lock while distributing bandwidth */
3750 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
3752 raw_spin_lock(&cfs_b
->lock
);
3754 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3756 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
3760 * While we are ensured activity in the period following an
3761 * unthrottle, this also covers the case in which the new bandwidth is
3762 * insufficient to cover the existing bandwidth deficit. (Forcing the
3763 * timer to remain active while there are any throttled entities.)
3770 cfs_b
->timer_active
= 0;
3774 /* a cfs_rq won't donate quota below this amount */
3775 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
3776 /* minimum remaining period time to redistribute slack quota */
3777 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
3778 /* how long we wait to gather additional slack before distributing */
3779 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
3782 * Are we near the end of the current quota period?
3784 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3785 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3786 * migrate_hrtimers, base is never cleared, so we are fine.
3788 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
3790 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
3793 /* if the call-back is running a quota refresh is already occurring */
3794 if (hrtimer_callback_running(refresh_timer
))
3797 /* is a quota refresh about to occur? */
3798 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
3799 if (remaining
< min_expire
)
3805 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
3807 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
3809 /* if there's a quota refresh soon don't bother with slack */
3810 if (runtime_refresh_within(cfs_b
, min_left
))
3813 start_bandwidth_timer(&cfs_b
->slack_timer
,
3814 ns_to_ktime(cfs_bandwidth_slack_period
));
3817 /* we know any runtime found here is valid as update_curr() precedes return */
3818 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3820 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3821 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
3823 if (slack_runtime
<= 0)
3826 raw_spin_lock(&cfs_b
->lock
);
3827 if (cfs_b
->quota
!= RUNTIME_INF
&&
3828 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
3829 cfs_b
->runtime
+= slack_runtime
;
3831 /* we are under rq->lock, defer unthrottling using a timer */
3832 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
3833 !list_empty(&cfs_b
->throttled_cfs_rq
))
3834 start_cfs_slack_bandwidth(cfs_b
);
3836 raw_spin_unlock(&cfs_b
->lock
);
3838 /* even if it's not valid for return we don't want to try again */
3839 cfs_rq
->runtime_remaining
-= slack_runtime
;
3842 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3844 if (!cfs_bandwidth_used())
3847 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
3850 __return_cfs_rq_runtime(cfs_rq
);
3854 * This is done with a timer (instead of inline with bandwidth return) since
3855 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3857 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
3859 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
3862 /* confirm we're still not at a refresh boundary */
3863 raw_spin_lock(&cfs_b
->lock
);
3864 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
3865 raw_spin_unlock(&cfs_b
->lock
);
3869 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
)
3870 runtime
= cfs_b
->runtime
;
3872 expires
= cfs_b
->runtime_expires
;
3873 raw_spin_unlock(&cfs_b
->lock
);
3878 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
3880 raw_spin_lock(&cfs_b
->lock
);
3881 if (expires
== cfs_b
->runtime_expires
)
3882 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
3883 raw_spin_unlock(&cfs_b
->lock
);
3887 * When a group wakes up we want to make sure that its quota is not already
3888 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3889 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3891 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
3893 if (!cfs_bandwidth_used())
3896 /* an active group must be handled by the update_curr()->put() path */
3897 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
3900 /* ensure the group is not already throttled */
3901 if (cfs_rq_throttled(cfs_rq
))
3904 /* update runtime allocation */
3905 account_cfs_rq_runtime(cfs_rq
, 0);
3906 if (cfs_rq
->runtime_remaining
<= 0)
3907 throttle_cfs_rq(cfs_rq
);
3910 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3911 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3913 if (!cfs_bandwidth_used())
3916 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
3920 * it's possible for a throttled entity to be forced into a running
3921 * state (e.g. set_curr_task), in this case we're finished.
3923 if (cfs_rq_throttled(cfs_rq
))
3926 throttle_cfs_rq(cfs_rq
);
3930 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
3932 struct cfs_bandwidth
*cfs_b
=
3933 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
3934 do_sched_cfs_slack_timer(cfs_b
);
3936 return HRTIMER_NORESTART
;
3939 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
3941 struct cfs_bandwidth
*cfs_b
=
3942 container_of(timer
, struct cfs_bandwidth
, period_timer
);
3947 raw_spin_lock(&cfs_b
->lock
);
3949 now
= hrtimer_cb_get_time(timer
);
3950 overrun
= hrtimer_forward(timer
, now
, cfs_b
->period
);
3955 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
3957 raw_spin_unlock(&cfs_b
->lock
);
3959 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
3962 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3964 raw_spin_lock_init(&cfs_b
->lock
);
3966 cfs_b
->quota
= RUNTIME_INF
;
3967 cfs_b
->period
= ns_to_ktime(default_cfs_period());
3969 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
3970 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3971 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
3972 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3973 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
3976 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3978 cfs_rq
->runtime_enabled
= 0;
3979 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
3982 /* requires cfs_b->lock, may release to reprogram timer */
3983 void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
, bool force
)
3986 * The timer may be active because we're trying to set a new bandwidth
3987 * period or because we're racing with the tear-down path
3988 * (timer_active==0 becomes visible before the hrtimer call-back
3989 * terminates). In either case we ensure that it's re-programmed
3991 while (unlikely(hrtimer_active(&cfs_b
->period_timer
)) &&
3992 hrtimer_try_to_cancel(&cfs_b
->period_timer
) < 0) {
3993 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3994 raw_spin_unlock(&cfs_b
->lock
);
3996 raw_spin_lock(&cfs_b
->lock
);
3997 /* if someone else restarted the timer then we're done */
3998 if (!force
&& cfs_b
->timer_active
)
4002 cfs_b
->timer_active
= 1;
4003 start_bandwidth_timer(&cfs_b
->period_timer
, cfs_b
->period
);
4006 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
4008 /* init_cfs_bandwidth() was not called */
4009 if (!cfs_b
->throttled_cfs_rq
.next
)
4012 hrtimer_cancel(&cfs_b
->period_timer
);
4013 hrtimer_cancel(&cfs_b
->slack_timer
);
4016 static void __maybe_unused
update_runtime_enabled(struct rq
*rq
)
4018 struct cfs_rq
*cfs_rq
;
4020 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
4021 struct cfs_bandwidth
*cfs_b
= &cfs_rq
->tg
->cfs_bandwidth
;
4023 raw_spin_lock(&cfs_b
->lock
);
4024 cfs_rq
->runtime_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
4025 raw_spin_unlock(&cfs_b
->lock
);
4029 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
4031 struct cfs_rq
*cfs_rq
;
4033 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
4034 if (!cfs_rq
->runtime_enabled
)
4038 * clock_task is not advancing so we just need to make sure
4039 * there's some valid quota amount
4041 cfs_rq
->runtime_remaining
= 1;
4043 * Offline rq is schedulable till cpu is completely disabled
4044 * in take_cpu_down(), so we prevent new cfs throttling here.
4046 cfs_rq
->runtime_enabled
= 0;
4048 if (cfs_rq_throttled(cfs_rq
))
4049 unthrottle_cfs_rq(cfs_rq
);
4053 #else /* CONFIG_CFS_BANDWIDTH */
4054 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
4056 return rq_clock_task(rq_of(cfs_rq
));
4059 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
4060 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) { return false; }
4061 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
4062 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
4064 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
4069 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
4074 static inline int throttled_lb_pair(struct task_group
*tg
,
4075 int src_cpu
, int dest_cpu
)
4080 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
4082 #ifdef CONFIG_FAIR_GROUP_SCHED
4083 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
4086 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
4090 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
4091 static inline void update_runtime_enabled(struct rq
*rq
) {}
4092 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
4094 #endif /* CONFIG_CFS_BANDWIDTH */
4096 /**************************************************
4097 * CFS operations on tasks:
4100 #ifdef CONFIG_SCHED_HRTICK
4101 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
4103 struct sched_entity
*se
= &p
->se
;
4104 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4106 WARN_ON(task_rq(p
) != rq
);
4108 if (cfs_rq
->nr_running
> 1) {
4109 u64 slice
= sched_slice(cfs_rq
, se
);
4110 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
4111 s64 delta
= slice
- ran
;
4118 hrtick_start(rq
, delta
);
4123 * called from enqueue/dequeue and updates the hrtick when the
4124 * current task is from our class and nr_running is low enough
4127 static void hrtick_update(struct rq
*rq
)
4129 struct task_struct
*curr
= rq
->curr
;
4131 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
4134 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
4135 hrtick_start_fair(rq
, curr
);
4137 #else /* !CONFIG_SCHED_HRTICK */
4139 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
4143 static inline void hrtick_update(struct rq
*rq
)
4149 * The enqueue_task method is called before nr_running is
4150 * increased. Here we update the fair scheduling stats and
4151 * then put the task into the rbtree:
4154 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
4156 struct cfs_rq
*cfs_rq
;
4157 struct sched_entity
*se
= &p
->se
;
4159 for_each_sched_entity(se
) {
4162 cfs_rq
= cfs_rq_of(se
);
4163 enqueue_entity(cfs_rq
, se
, flags
);
4166 * end evaluation on encountering a throttled cfs_rq
4168 * note: in the case of encountering a throttled cfs_rq we will
4169 * post the final h_nr_running increment below.
4171 if (cfs_rq_throttled(cfs_rq
))
4173 cfs_rq
->h_nr_running
++;
4175 flags
= ENQUEUE_WAKEUP
;
4178 for_each_sched_entity(se
) {
4179 cfs_rq
= cfs_rq_of(se
);
4180 cfs_rq
->h_nr_running
++;
4182 if (cfs_rq_throttled(cfs_rq
))
4185 update_cfs_shares(cfs_rq
);
4186 update_entity_load_avg(se
, 1);
4190 update_rq_runnable_avg(rq
, rq
->nr_running
);
4191 add_nr_running(rq
, 1);
4196 static void set_next_buddy(struct sched_entity
*se
);
4199 * The dequeue_task method is called before nr_running is
4200 * decreased. We remove the task from the rbtree and
4201 * update the fair scheduling stats:
4203 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
4205 struct cfs_rq
*cfs_rq
;
4206 struct sched_entity
*se
= &p
->se
;
4207 int task_sleep
= flags
& DEQUEUE_SLEEP
;
4209 for_each_sched_entity(se
) {
4210 cfs_rq
= cfs_rq_of(se
);
4211 dequeue_entity(cfs_rq
, se
, flags
);
4214 * end evaluation on encountering a throttled cfs_rq
4216 * note: in the case of encountering a throttled cfs_rq we will
4217 * post the final h_nr_running decrement below.
4219 if (cfs_rq_throttled(cfs_rq
))
4221 cfs_rq
->h_nr_running
--;
4223 /* Don't dequeue parent if it has other entities besides us */
4224 if (cfs_rq
->load
.weight
) {
4226 * Bias pick_next to pick a task from this cfs_rq, as
4227 * p is sleeping when it is within its sched_slice.
4229 if (task_sleep
&& parent_entity(se
))
4230 set_next_buddy(parent_entity(se
));
4232 /* avoid re-evaluating load for this entity */
4233 se
= parent_entity(se
);
4236 flags
|= DEQUEUE_SLEEP
;
4239 for_each_sched_entity(se
) {
4240 cfs_rq
= cfs_rq_of(se
);
4241 cfs_rq
->h_nr_running
--;
4243 if (cfs_rq_throttled(cfs_rq
))
4246 update_cfs_shares(cfs_rq
);
4247 update_entity_load_avg(se
, 1);
4251 sub_nr_running(rq
, 1);
4252 update_rq_runnable_avg(rq
, 1);
4258 /* Used instead of source_load when we know the type == 0 */
4259 static unsigned long weighted_cpuload(const int cpu
)
4261 return cpu_rq(cpu
)->cfs
.runnable_load_avg
;
4265 * Return a low guess at the load of a migration-source cpu weighted
4266 * according to the scheduling class and "nice" value.
4268 * We want to under-estimate the load of migration sources, to
4269 * balance conservatively.
4271 static unsigned long source_load(int cpu
, int type
)
4273 struct rq
*rq
= cpu_rq(cpu
);
4274 unsigned long total
= weighted_cpuload(cpu
);
4276 if (type
== 0 || !sched_feat(LB_BIAS
))
4279 return min(rq
->cpu_load
[type
-1], total
);
4283 * Return a high guess at the load of a migration-target cpu weighted
4284 * according to the scheduling class and "nice" value.
4286 static unsigned long target_load(int cpu
, int type
)
4288 struct rq
*rq
= cpu_rq(cpu
);
4289 unsigned long total
= weighted_cpuload(cpu
);
4291 if (type
== 0 || !sched_feat(LB_BIAS
))
4294 return max(rq
->cpu_load
[type
-1], total
);
4297 static unsigned long capacity_of(int cpu
)
4299 return cpu_rq(cpu
)->cpu_capacity
;
4302 static unsigned long cpu_avg_load_per_task(int cpu
)
4304 struct rq
*rq
= cpu_rq(cpu
);
4305 unsigned long nr_running
= ACCESS_ONCE(rq
->cfs
.h_nr_running
);
4306 unsigned long load_avg
= rq
->cfs
.runnable_load_avg
;
4309 return load_avg
/ nr_running
;
4314 static void record_wakee(struct task_struct
*p
)
4317 * Rough decay (wiping) for cost saving, don't worry
4318 * about the boundary, really active task won't care
4321 if (time_after(jiffies
, current
->wakee_flip_decay_ts
+ HZ
)) {
4322 current
->wakee_flips
>>= 1;
4323 current
->wakee_flip_decay_ts
= jiffies
;
4326 if (current
->last_wakee
!= p
) {
4327 current
->last_wakee
= p
;
4328 current
->wakee_flips
++;
4332 static void task_waking_fair(struct task_struct
*p
)
4334 struct sched_entity
*se
= &p
->se
;
4335 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4338 #ifndef CONFIG_64BIT
4339 u64 min_vruntime_copy
;
4342 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
4344 min_vruntime
= cfs_rq
->min_vruntime
;
4345 } while (min_vruntime
!= min_vruntime_copy
);
4347 min_vruntime
= cfs_rq
->min_vruntime
;
4350 se
->vruntime
-= min_vruntime
;
4354 #ifdef CONFIG_FAIR_GROUP_SCHED
4356 * effective_load() calculates the load change as seen from the root_task_group
4358 * Adding load to a group doesn't make a group heavier, but can cause movement
4359 * of group shares between cpus. Assuming the shares were perfectly aligned one
4360 * can calculate the shift in shares.
4362 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4363 * on this @cpu and results in a total addition (subtraction) of @wg to the
4364 * total group weight.
4366 * Given a runqueue weight distribution (rw_i) we can compute a shares
4367 * distribution (s_i) using:
4369 * s_i = rw_i / \Sum rw_j (1)
4371 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4372 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4373 * shares distribution (s_i):
4375 * rw_i = { 2, 4, 1, 0 }
4376 * s_i = { 2/7, 4/7, 1/7, 0 }
4378 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4379 * task used to run on and the CPU the waker is running on), we need to
4380 * compute the effect of waking a task on either CPU and, in case of a sync
4381 * wakeup, compute the effect of the current task going to sleep.
4383 * So for a change of @wl to the local @cpu with an overall group weight change
4384 * of @wl we can compute the new shares distribution (s'_i) using:
4386 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4388 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4389 * differences in waking a task to CPU 0. The additional task changes the
4390 * weight and shares distributions like:
4392 * rw'_i = { 3, 4, 1, 0 }
4393 * s'_i = { 3/8, 4/8, 1/8, 0 }
4395 * We can then compute the difference in effective weight by using:
4397 * dw_i = S * (s'_i - s_i) (3)
4399 * Where 'S' is the group weight as seen by its parent.
4401 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4402 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4403 * 4/7) times the weight of the group.
4405 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4407 struct sched_entity
*se
= tg
->se
[cpu
];
4409 if (!tg
->parent
) /* the trivial, non-cgroup case */
4412 for_each_sched_entity(se
) {
4418 * W = @wg + \Sum rw_j
4420 W
= wg
+ calc_tg_weight(tg
, se
->my_q
);
4425 w
= se
->my_q
->load
.weight
+ wl
;
4428 * wl = S * s'_i; see (2)
4431 wl
= (w
* (long)tg
->shares
) / W
;
4436 * Per the above, wl is the new se->load.weight value; since
4437 * those are clipped to [MIN_SHARES, ...) do so now. See
4438 * calc_cfs_shares().
4440 if (wl
< MIN_SHARES
)
4444 * wl = dw_i = S * (s'_i - s_i); see (3)
4446 wl
-= se
->load
.weight
;
4449 * Recursively apply this logic to all parent groups to compute
4450 * the final effective load change on the root group. Since
4451 * only the @tg group gets extra weight, all parent groups can
4452 * only redistribute existing shares. @wl is the shift in shares
4453 * resulting from this level per the above.
4462 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4469 static int wake_wide(struct task_struct
*p
)
4471 int factor
= this_cpu_read(sd_llc_size
);
4474 * Yeah, it's the switching-frequency, could means many wakee or
4475 * rapidly switch, use factor here will just help to automatically
4476 * adjust the loose-degree, so bigger node will lead to more pull.
4478 if (p
->wakee_flips
> factor
) {
4480 * wakee is somewhat hot, it needs certain amount of cpu
4481 * resource, so if waker is far more hot, prefer to leave
4484 if (current
->wakee_flips
> (factor
* p
->wakee_flips
))
4491 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
4493 s64 this_load
, load
;
4494 s64 this_eff_load
, prev_eff_load
;
4495 int idx
, this_cpu
, prev_cpu
;
4496 struct task_group
*tg
;
4497 unsigned long weight
;
4501 * If we wake multiple tasks be careful to not bounce
4502 * ourselves around too much.
4508 this_cpu
= smp_processor_id();
4509 prev_cpu
= task_cpu(p
);
4510 load
= source_load(prev_cpu
, idx
);
4511 this_load
= target_load(this_cpu
, idx
);
4514 * If sync wakeup then subtract the (maximum possible)
4515 * effect of the currently running task from the load
4516 * of the current CPU:
4519 tg
= task_group(current
);
4520 weight
= current
->se
.load
.weight
;
4522 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
4523 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
4527 weight
= p
->se
.load
.weight
;
4530 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4531 * due to the sync cause above having dropped this_load to 0, we'll
4532 * always have an imbalance, but there's really nothing you can do
4533 * about that, so that's good too.
4535 * Otherwise check if either cpus are near enough in load to allow this
4536 * task to be woken on this_cpu.
4538 this_eff_load
= 100;
4539 this_eff_load
*= capacity_of(prev_cpu
);
4541 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
4542 prev_eff_load
*= capacity_of(this_cpu
);
4544 if (this_load
> 0) {
4545 this_eff_load
*= this_load
+
4546 effective_load(tg
, this_cpu
, weight
, weight
);
4548 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
4551 balanced
= this_eff_load
<= prev_eff_load
;
4553 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
4558 schedstat_inc(sd
, ttwu_move_affine
);
4559 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
4565 * find_idlest_group finds and returns the least busy CPU group within the
4568 static struct sched_group
*
4569 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
4570 int this_cpu
, int sd_flag
)
4572 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
4573 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
4574 int load_idx
= sd
->forkexec_idx
;
4575 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
4577 if (sd_flag
& SD_BALANCE_WAKE
)
4578 load_idx
= sd
->wake_idx
;
4581 unsigned long load
, avg_load
;
4585 /* Skip over this group if it has no CPUs allowed */
4586 if (!cpumask_intersects(sched_group_cpus(group
),
4587 tsk_cpus_allowed(p
)))
4590 local_group
= cpumask_test_cpu(this_cpu
,
4591 sched_group_cpus(group
));
4593 /* Tally up the load of all CPUs in the group */
4596 for_each_cpu(i
, sched_group_cpus(group
)) {
4597 /* Bias balancing toward cpus of our domain */
4599 load
= source_load(i
, load_idx
);
4601 load
= target_load(i
, load_idx
);
4606 /* Adjust by relative CPU capacity of the group */
4607 avg_load
= (avg_load
* SCHED_CAPACITY_SCALE
) / group
->sgc
->capacity
;
4610 this_load
= avg_load
;
4611 } else if (avg_load
< min_load
) {
4612 min_load
= avg_load
;
4615 } while (group
= group
->next
, group
!= sd
->groups
);
4617 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
4623 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4626 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
4628 unsigned long load
, min_load
= ULONG_MAX
;
4629 unsigned int min_exit_latency
= UINT_MAX
;
4630 u64 latest_idle_timestamp
= 0;
4631 int least_loaded_cpu
= this_cpu
;
4632 int shallowest_idle_cpu
= -1;
4635 /* Traverse only the allowed CPUs */
4636 for_each_cpu_and(i
, sched_group_cpus(group
), tsk_cpus_allowed(p
)) {
4638 struct rq
*rq
= cpu_rq(i
);
4639 struct cpuidle_state
*idle
= idle_get_state(rq
);
4640 if (idle
&& idle
->exit_latency
< min_exit_latency
) {
4642 * We give priority to a CPU whose idle state
4643 * has the smallest exit latency irrespective
4644 * of any idle timestamp.
4646 min_exit_latency
= idle
->exit_latency
;
4647 latest_idle_timestamp
= rq
->idle_stamp
;
4648 shallowest_idle_cpu
= i
;
4649 } else if ((!idle
|| idle
->exit_latency
== min_exit_latency
) &&
4650 rq
->idle_stamp
> latest_idle_timestamp
) {
4652 * If equal or no active idle state, then
4653 * the most recently idled CPU might have
4656 latest_idle_timestamp
= rq
->idle_stamp
;
4657 shallowest_idle_cpu
= i
;
4659 } else if (shallowest_idle_cpu
== -1) {
4660 load
= weighted_cpuload(i
);
4661 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
4663 least_loaded_cpu
= i
;
4668 return shallowest_idle_cpu
!= -1 ? shallowest_idle_cpu
: least_loaded_cpu
;
4672 * Try and locate an idle CPU in the sched_domain.
4674 static int select_idle_sibling(struct task_struct
*p
, int target
)
4676 struct sched_domain
*sd
;
4677 struct sched_group
*sg
;
4678 int i
= task_cpu(p
);
4680 if (idle_cpu(target
))
4684 * If the prevous cpu is cache affine and idle, don't be stupid.
4686 if (i
!= target
&& cpus_share_cache(i
, target
) && idle_cpu(i
))
4690 * Otherwise, iterate the domains and find an elegible idle cpu.
4692 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
4693 for_each_lower_domain(sd
) {
4696 if (!cpumask_intersects(sched_group_cpus(sg
),
4697 tsk_cpus_allowed(p
)))
4700 for_each_cpu(i
, sched_group_cpus(sg
)) {
4701 if (i
== target
|| !idle_cpu(i
))
4705 target
= cpumask_first_and(sched_group_cpus(sg
),
4706 tsk_cpus_allowed(p
));
4710 } while (sg
!= sd
->groups
);
4717 * select_task_rq_fair: Select target runqueue for the waking task in domains
4718 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4719 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4721 * Balances load by selecting the idlest cpu in the idlest group, or under
4722 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4724 * Returns the target cpu number.
4726 * preempt must be disabled.
4729 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
4731 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
4732 int cpu
= smp_processor_id();
4734 int want_affine
= 0;
4735 int sync
= wake_flags
& WF_SYNC
;
4737 if (sd_flag
& SD_BALANCE_WAKE
)
4738 want_affine
= cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
));
4741 for_each_domain(cpu
, tmp
) {
4742 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
4746 * If both cpu and prev_cpu are part of this domain,
4747 * cpu is a valid SD_WAKE_AFFINE target.
4749 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
4750 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
4755 if (tmp
->flags
& sd_flag
)
4759 if (affine_sd
&& cpu
!= prev_cpu
&& wake_affine(affine_sd
, p
, sync
))
4762 if (sd_flag
& SD_BALANCE_WAKE
) {
4763 new_cpu
= select_idle_sibling(p
, prev_cpu
);
4768 struct sched_group
*group
;
4771 if (!(sd
->flags
& sd_flag
)) {
4776 group
= find_idlest_group(sd
, p
, cpu
, sd_flag
);
4782 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
4783 if (new_cpu
== -1 || new_cpu
== cpu
) {
4784 /* Now try balancing at a lower domain level of cpu */
4789 /* Now try balancing at a lower domain level of new_cpu */
4791 weight
= sd
->span_weight
;
4793 for_each_domain(cpu
, tmp
) {
4794 if (weight
<= tmp
->span_weight
)
4796 if (tmp
->flags
& sd_flag
)
4799 /* while loop will break here if sd == NULL */
4808 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4809 * cfs_rq_of(p) references at time of call are still valid and identify the
4810 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4811 * other assumptions, including the state of rq->lock, should be made.
4814 migrate_task_rq_fair(struct task_struct
*p
, int next_cpu
)
4816 struct sched_entity
*se
= &p
->se
;
4817 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4820 * Load tracking: accumulate removed load so that it can be processed
4821 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4822 * to blocked load iff they have a positive decay-count. It can never
4823 * be negative here since on-rq tasks have decay-count == 0.
4825 if (se
->avg
.decay_count
) {
4826 se
->avg
.decay_count
= -__synchronize_entity_decay(se
);
4827 atomic_long_add(se
->avg
.load_avg_contrib
,
4828 &cfs_rq
->removed_load
);
4831 /* We have migrated, no longer consider this task hot */
4834 #endif /* CONFIG_SMP */
4836 static unsigned long
4837 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
4839 unsigned long gran
= sysctl_sched_wakeup_granularity
;
4842 * Since its curr running now, convert the gran from real-time
4843 * to virtual-time in his units.
4845 * By using 'se' instead of 'curr' we penalize light tasks, so
4846 * they get preempted easier. That is, if 'se' < 'curr' then
4847 * the resulting gran will be larger, therefore penalizing the
4848 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4849 * be smaller, again penalizing the lighter task.
4851 * This is especially important for buddies when the leftmost
4852 * task is higher priority than the buddy.
4854 return calc_delta_fair(gran
, se
);
4858 * Should 'se' preempt 'curr'.
4872 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
4874 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
4879 gran
= wakeup_gran(curr
, se
);
4886 static void set_last_buddy(struct sched_entity
*se
)
4888 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4891 for_each_sched_entity(se
)
4892 cfs_rq_of(se
)->last
= se
;
4895 static void set_next_buddy(struct sched_entity
*se
)
4897 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4900 for_each_sched_entity(se
)
4901 cfs_rq_of(se
)->next
= se
;
4904 static void set_skip_buddy(struct sched_entity
*se
)
4906 for_each_sched_entity(se
)
4907 cfs_rq_of(se
)->skip
= se
;
4911 * Preempt the current task with a newly woken task if needed:
4913 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
4915 struct task_struct
*curr
= rq
->curr
;
4916 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
4917 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4918 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
4919 int next_buddy_marked
= 0;
4921 if (unlikely(se
== pse
))
4925 * This is possible from callers such as attach_tasks(), in which we
4926 * unconditionally check_prempt_curr() after an enqueue (which may have
4927 * lead to a throttle). This both saves work and prevents false
4928 * next-buddy nomination below.
4930 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
4933 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
4934 set_next_buddy(pse
);
4935 next_buddy_marked
= 1;
4939 * We can come here with TIF_NEED_RESCHED already set from new task
4942 * Note: this also catches the edge-case of curr being in a throttled
4943 * group (e.g. via set_curr_task), since update_curr() (in the
4944 * enqueue of curr) will have resulted in resched being set. This
4945 * prevents us from potentially nominating it as a false LAST_BUDDY
4948 if (test_tsk_need_resched(curr
))
4951 /* Idle tasks are by definition preempted by non-idle tasks. */
4952 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
4953 likely(p
->policy
!= SCHED_IDLE
))
4957 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4958 * is driven by the tick):
4960 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
4963 find_matching_se(&se
, &pse
);
4964 update_curr(cfs_rq_of(se
));
4966 if (wakeup_preempt_entity(se
, pse
) == 1) {
4968 * Bias pick_next to pick the sched entity that is
4969 * triggering this preemption.
4971 if (!next_buddy_marked
)
4972 set_next_buddy(pse
);
4981 * Only set the backward buddy when the current task is still
4982 * on the rq. This can happen when a wakeup gets interleaved
4983 * with schedule on the ->pre_schedule() or idle_balance()
4984 * point, either of which can * drop the rq lock.
4986 * Also, during early boot the idle thread is in the fair class,
4987 * for obvious reasons its a bad idea to schedule back to it.
4989 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
4992 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
4996 static struct task_struct
*
4997 pick_next_task_fair(struct rq
*rq
, struct task_struct
*prev
)
4999 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
5000 struct sched_entity
*se
;
5001 struct task_struct
*p
;
5005 #ifdef CONFIG_FAIR_GROUP_SCHED
5006 if (!cfs_rq
->nr_running
)
5009 if (prev
->sched_class
!= &fair_sched_class
)
5013 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5014 * likely that a next task is from the same cgroup as the current.
5016 * Therefore attempt to avoid putting and setting the entire cgroup
5017 * hierarchy, only change the part that actually changes.
5021 struct sched_entity
*curr
= cfs_rq
->curr
;
5024 * Since we got here without doing put_prev_entity() we also
5025 * have to consider cfs_rq->curr. If it is still a runnable
5026 * entity, update_curr() will update its vruntime, otherwise
5027 * forget we've ever seen it.
5029 if (curr
&& curr
->on_rq
)
5030 update_curr(cfs_rq
);
5035 * This call to check_cfs_rq_runtime() will do the throttle and
5036 * dequeue its entity in the parent(s). Therefore the 'simple'
5037 * nr_running test will indeed be correct.
5039 if (unlikely(check_cfs_rq_runtime(cfs_rq
)))
5042 se
= pick_next_entity(cfs_rq
, curr
);
5043 cfs_rq
= group_cfs_rq(se
);
5049 * Since we haven't yet done put_prev_entity and if the selected task
5050 * is a different task than we started out with, try and touch the
5051 * least amount of cfs_rqs.
5054 struct sched_entity
*pse
= &prev
->se
;
5056 while (!(cfs_rq
= is_same_group(se
, pse
))) {
5057 int se_depth
= se
->depth
;
5058 int pse_depth
= pse
->depth
;
5060 if (se_depth
<= pse_depth
) {
5061 put_prev_entity(cfs_rq_of(pse
), pse
);
5062 pse
= parent_entity(pse
);
5064 if (se_depth
>= pse_depth
) {
5065 set_next_entity(cfs_rq_of(se
), se
);
5066 se
= parent_entity(se
);
5070 put_prev_entity(cfs_rq
, pse
);
5071 set_next_entity(cfs_rq
, se
);
5074 if (hrtick_enabled(rq
))
5075 hrtick_start_fair(rq
, p
);
5082 if (!cfs_rq
->nr_running
)
5085 put_prev_task(rq
, prev
);
5088 se
= pick_next_entity(cfs_rq
, NULL
);
5089 set_next_entity(cfs_rq
, se
);
5090 cfs_rq
= group_cfs_rq(se
);
5095 if (hrtick_enabled(rq
))
5096 hrtick_start_fair(rq
, p
);
5101 new_tasks
= idle_balance(rq
);
5103 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5104 * possible for any higher priority task to appear. In that case we
5105 * must re-start the pick_next_entity() loop.
5117 * Account for a descheduled task:
5119 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
5121 struct sched_entity
*se
= &prev
->se
;
5122 struct cfs_rq
*cfs_rq
;
5124 for_each_sched_entity(se
) {
5125 cfs_rq
= cfs_rq_of(se
);
5126 put_prev_entity(cfs_rq
, se
);
5131 * sched_yield() is very simple
5133 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5135 static void yield_task_fair(struct rq
*rq
)
5137 struct task_struct
*curr
= rq
->curr
;
5138 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
5139 struct sched_entity
*se
= &curr
->se
;
5142 * Are we the only task in the tree?
5144 if (unlikely(rq
->nr_running
== 1))
5147 clear_buddies(cfs_rq
, se
);
5149 if (curr
->policy
!= SCHED_BATCH
) {
5150 update_rq_clock(rq
);
5152 * Update run-time statistics of the 'current'.
5154 update_curr(cfs_rq
);
5156 * Tell update_rq_clock() that we've just updated,
5157 * so we don't do microscopic update in schedule()
5158 * and double the fastpath cost.
5160 rq
->skip_clock_update
= 1;
5166 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
5168 struct sched_entity
*se
= &p
->se
;
5170 /* throttled hierarchies are not runnable */
5171 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
5174 /* Tell the scheduler that we'd really like pse to run next. */
5177 yield_task_fair(rq
);
5183 /**************************************************
5184 * Fair scheduling class load-balancing methods.
5188 * The purpose of load-balancing is to achieve the same basic fairness the
5189 * per-cpu scheduler provides, namely provide a proportional amount of compute
5190 * time to each task. This is expressed in the following equation:
5192 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5194 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5195 * W_i,0 is defined as:
5197 * W_i,0 = \Sum_j w_i,j (2)
5199 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5200 * is derived from the nice value as per prio_to_weight[].
5202 * The weight average is an exponential decay average of the instantaneous
5205 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5207 * C_i is the compute capacity of cpu i, typically it is the
5208 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5209 * can also include other factors [XXX].
5211 * To achieve this balance we define a measure of imbalance which follows
5212 * directly from (1):
5214 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5216 * We them move tasks around to minimize the imbalance. In the continuous
5217 * function space it is obvious this converges, in the discrete case we get
5218 * a few fun cases generally called infeasible weight scenarios.
5221 * - infeasible weights;
5222 * - local vs global optima in the discrete case. ]
5227 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5228 * for all i,j solution, we create a tree of cpus that follows the hardware
5229 * topology where each level pairs two lower groups (or better). This results
5230 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5231 * tree to only the first of the previous level and we decrease the frequency
5232 * of load-balance at each level inv. proportional to the number of cpus in
5238 * \Sum { --- * --- * 2^i } = O(n) (5)
5240 * `- size of each group
5241 * | | `- number of cpus doing load-balance
5243 * `- sum over all levels
5245 * Coupled with a limit on how many tasks we can migrate every balance pass,
5246 * this makes (5) the runtime complexity of the balancer.
5248 * An important property here is that each CPU is still (indirectly) connected
5249 * to every other cpu in at most O(log n) steps:
5251 * The adjacency matrix of the resulting graph is given by:
5254 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5257 * And you'll find that:
5259 * A^(log_2 n)_i,j != 0 for all i,j (7)
5261 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5262 * The task movement gives a factor of O(m), giving a convergence complexity
5265 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5270 * In order to avoid CPUs going idle while there's still work to do, new idle
5271 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5272 * tree itself instead of relying on other CPUs to bring it work.
5274 * This adds some complexity to both (5) and (8) but it reduces the total idle
5282 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5285 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5290 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5292 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5294 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5297 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5298 * rewrite all of this once again.]
5301 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
5303 enum fbq_type
{ regular
, remote
, all
};
5305 #define LBF_ALL_PINNED 0x01
5306 #define LBF_NEED_BREAK 0x02
5307 #define LBF_DST_PINNED 0x04
5308 #define LBF_SOME_PINNED 0x08
5311 struct sched_domain
*sd
;
5319 struct cpumask
*dst_grpmask
;
5321 enum cpu_idle_type idle
;
5323 /* The set of CPUs under consideration for load-balancing */
5324 struct cpumask
*cpus
;
5329 unsigned int loop_break
;
5330 unsigned int loop_max
;
5332 enum fbq_type fbq_type
;
5333 struct list_head tasks
;
5337 * Is this task likely cache-hot:
5339 static int task_hot(struct task_struct
*p
, struct lb_env
*env
)
5343 lockdep_assert_held(&env
->src_rq
->lock
);
5345 if (p
->sched_class
!= &fair_sched_class
)
5348 if (unlikely(p
->policy
== SCHED_IDLE
))
5352 * Buddy candidates are cache hot:
5354 if (sched_feat(CACHE_HOT_BUDDY
) && env
->dst_rq
->nr_running
&&
5355 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
5356 &p
->se
== cfs_rq_of(&p
->se
)->last
))
5359 if (sysctl_sched_migration_cost
== -1)
5361 if (sysctl_sched_migration_cost
== 0)
5364 delta
= rq_clock_task(env
->src_rq
) - p
->se
.exec_start
;
5366 return delta
< (s64
)sysctl_sched_migration_cost
;
5369 #ifdef CONFIG_NUMA_BALANCING
5370 /* Returns true if the destination node has incurred more faults */
5371 static bool migrate_improves_locality(struct task_struct
*p
, struct lb_env
*env
)
5373 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
5374 int src_nid
, dst_nid
;
5376 if (!sched_feat(NUMA_FAVOUR_HIGHER
) || !p
->numa_faults
||
5377 !(env
->sd
->flags
& SD_NUMA
)) {
5381 src_nid
= cpu_to_node(env
->src_cpu
);
5382 dst_nid
= cpu_to_node(env
->dst_cpu
);
5384 if (src_nid
== dst_nid
)
5388 /* Task is already in the group's interleave set. */
5389 if (node_isset(src_nid
, numa_group
->active_nodes
))
5392 /* Task is moving into the group's interleave set. */
5393 if (node_isset(dst_nid
, numa_group
->active_nodes
))
5396 return group_faults(p
, dst_nid
) > group_faults(p
, src_nid
);
5399 /* Encourage migration to the preferred node. */
5400 if (dst_nid
== p
->numa_preferred_nid
)
5403 return task_faults(p
, dst_nid
) > task_faults(p
, src_nid
);
5407 static bool migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
5409 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
5410 int src_nid
, dst_nid
;
5412 if (!sched_feat(NUMA
) || !sched_feat(NUMA_RESIST_LOWER
))
5415 if (!p
->numa_faults
|| !(env
->sd
->flags
& SD_NUMA
))
5418 src_nid
= cpu_to_node(env
->src_cpu
);
5419 dst_nid
= cpu_to_node(env
->dst_cpu
);
5421 if (src_nid
== dst_nid
)
5425 /* Task is moving within/into the group's interleave set. */
5426 if (node_isset(dst_nid
, numa_group
->active_nodes
))
5429 /* Task is moving out of the group's interleave set. */
5430 if (node_isset(src_nid
, numa_group
->active_nodes
))
5433 return group_faults(p
, dst_nid
) < group_faults(p
, src_nid
);
5436 /* Migrating away from the preferred node is always bad. */
5437 if (src_nid
== p
->numa_preferred_nid
)
5440 return task_faults(p
, dst_nid
) < task_faults(p
, src_nid
);
5444 static inline bool migrate_improves_locality(struct task_struct
*p
,
5450 static inline bool migrate_degrades_locality(struct task_struct
*p
,
5458 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5461 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
5463 int tsk_cache_hot
= 0;
5465 lockdep_assert_held(&env
->src_rq
->lock
);
5468 * We do not migrate tasks that are:
5469 * 1) throttled_lb_pair, or
5470 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5471 * 3) running (obviously), or
5472 * 4) are cache-hot on their current CPU.
5474 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
5477 if (!cpumask_test_cpu(env
->dst_cpu
, tsk_cpus_allowed(p
))) {
5480 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
5482 env
->flags
|= LBF_SOME_PINNED
;
5485 * Remember if this task can be migrated to any other cpu in
5486 * our sched_group. We may want to revisit it if we couldn't
5487 * meet load balance goals by pulling other tasks on src_cpu.
5489 * Also avoid computing new_dst_cpu if we have already computed
5490 * one in current iteration.
5492 if (!env
->dst_grpmask
|| (env
->flags
& LBF_DST_PINNED
))
5495 /* Prevent to re-select dst_cpu via env's cpus */
5496 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
5497 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
))) {
5498 env
->flags
|= LBF_DST_PINNED
;
5499 env
->new_dst_cpu
= cpu
;
5507 /* Record that we found atleast one task that could run on dst_cpu */
5508 env
->flags
&= ~LBF_ALL_PINNED
;
5510 if (task_running(env
->src_rq
, p
)) {
5511 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
5516 * Aggressive migration if:
5517 * 1) destination numa is preferred
5518 * 2) task is cache cold, or
5519 * 3) too many balance attempts have failed.
5521 tsk_cache_hot
= task_hot(p
, env
);
5523 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
5525 if (migrate_improves_locality(p
, env
) || !tsk_cache_hot
||
5526 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
5527 if (tsk_cache_hot
) {
5528 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
5529 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
5534 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
5539 * detach_task() -- detach the task for the migration specified in env
5541 static void detach_task(struct task_struct
*p
, struct lb_env
*env
)
5543 lockdep_assert_held(&env
->src_rq
->lock
);
5545 deactivate_task(env
->src_rq
, p
, 0);
5546 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
5547 set_task_cpu(p
, env
->dst_cpu
);
5551 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5552 * part of active balancing operations within "domain".
5554 * Returns a task if successful and NULL otherwise.
5556 static struct task_struct
*detach_one_task(struct lb_env
*env
)
5558 struct task_struct
*p
, *n
;
5560 lockdep_assert_held(&env
->src_rq
->lock
);
5562 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
5563 if (!can_migrate_task(p
, env
))
5566 detach_task(p
, env
);
5569 * Right now, this is only the second place where
5570 * lb_gained[env->idle] is updated (other is detach_tasks)
5571 * so we can safely collect stats here rather than
5572 * inside detach_tasks().
5574 schedstat_inc(env
->sd
, lb_gained
[env
->idle
]);
5580 static const unsigned int sched_nr_migrate_break
= 32;
5583 * detach_tasks() -- tries to detach up to imbalance weighted load from
5584 * busiest_rq, as part of a balancing operation within domain "sd".
5586 * Returns number of detached tasks if successful and 0 otherwise.
5588 static int detach_tasks(struct lb_env
*env
)
5590 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
5591 struct task_struct
*p
;
5595 lockdep_assert_held(&env
->src_rq
->lock
);
5597 if (env
->imbalance
<= 0)
5600 while (!list_empty(tasks
)) {
5601 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
5604 /* We've more or less seen every task there is, call it quits */
5605 if (env
->loop
> env
->loop_max
)
5608 /* take a breather every nr_migrate tasks */
5609 if (env
->loop
> env
->loop_break
) {
5610 env
->loop_break
+= sched_nr_migrate_break
;
5611 env
->flags
|= LBF_NEED_BREAK
;
5615 if (!can_migrate_task(p
, env
))
5618 load
= task_h_load(p
);
5620 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
5623 if ((load
/ 2) > env
->imbalance
)
5626 detach_task(p
, env
);
5627 list_add(&p
->se
.group_node
, &env
->tasks
);
5630 env
->imbalance
-= load
;
5632 #ifdef CONFIG_PREEMPT
5634 * NEWIDLE balancing is a source of latency, so preemptible
5635 * kernels will stop after the first task is detached to minimize
5636 * the critical section.
5638 if (env
->idle
== CPU_NEWLY_IDLE
)
5643 * We only want to steal up to the prescribed amount of
5646 if (env
->imbalance
<= 0)
5651 list_move_tail(&p
->se
.group_node
, tasks
);
5655 * Right now, this is one of only two places we collect this stat
5656 * so we can safely collect detach_one_task() stats here rather
5657 * than inside detach_one_task().
5659 schedstat_add(env
->sd
, lb_gained
[env
->idle
], detached
);
5665 * attach_task() -- attach the task detached by detach_task() to its new rq.
5667 static void attach_task(struct rq
*rq
, struct task_struct
*p
)
5669 lockdep_assert_held(&rq
->lock
);
5671 BUG_ON(task_rq(p
) != rq
);
5672 p
->on_rq
= TASK_ON_RQ_QUEUED
;
5673 activate_task(rq
, p
, 0);
5674 check_preempt_curr(rq
, p
, 0);
5678 * attach_one_task() -- attaches the task returned from detach_one_task() to
5681 static void attach_one_task(struct rq
*rq
, struct task_struct
*p
)
5683 raw_spin_lock(&rq
->lock
);
5685 raw_spin_unlock(&rq
->lock
);
5689 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5692 static void attach_tasks(struct lb_env
*env
)
5694 struct list_head
*tasks
= &env
->tasks
;
5695 struct task_struct
*p
;
5697 raw_spin_lock(&env
->dst_rq
->lock
);
5699 while (!list_empty(tasks
)) {
5700 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
5701 list_del_init(&p
->se
.group_node
);
5703 attach_task(env
->dst_rq
, p
);
5706 raw_spin_unlock(&env
->dst_rq
->lock
);
5709 #ifdef CONFIG_FAIR_GROUP_SCHED
5711 * update tg->load_weight by folding this cpu's load_avg
5713 static void __update_blocked_averages_cpu(struct task_group
*tg
, int cpu
)
5715 struct sched_entity
*se
= tg
->se
[cpu
];
5716 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu
];
5718 /* throttled entities do not contribute to load */
5719 if (throttled_hierarchy(cfs_rq
))
5722 update_cfs_rq_blocked_load(cfs_rq
, 1);
5725 update_entity_load_avg(se
, 1);
5727 * We pivot on our runnable average having decayed to zero for
5728 * list removal. This generally implies that all our children
5729 * have also been removed (modulo rounding error or bandwidth
5730 * control); however, such cases are rare and we can fix these
5733 * TODO: fix up out-of-order children on enqueue.
5735 if (!se
->avg
.runnable_avg_sum
&& !cfs_rq
->nr_running
)
5736 list_del_leaf_cfs_rq(cfs_rq
);
5738 struct rq
*rq
= rq_of(cfs_rq
);
5739 update_rq_runnable_avg(rq
, rq
->nr_running
);
5743 static void update_blocked_averages(int cpu
)
5745 struct rq
*rq
= cpu_rq(cpu
);
5746 struct cfs_rq
*cfs_rq
;
5747 unsigned long flags
;
5749 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5750 update_rq_clock(rq
);
5752 * Iterates the task_group tree in a bottom up fashion, see
5753 * list_add_leaf_cfs_rq() for details.
5755 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
5757 * Note: We may want to consider periodically releasing
5758 * rq->lock about these updates so that creating many task
5759 * groups does not result in continually extending hold time.
5761 __update_blocked_averages_cpu(cfs_rq
->tg
, rq
->cpu
);
5764 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5768 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5769 * This needs to be done in a top-down fashion because the load of a child
5770 * group is a fraction of its parents load.
5772 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
5774 struct rq
*rq
= rq_of(cfs_rq
);
5775 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
5776 unsigned long now
= jiffies
;
5779 if (cfs_rq
->last_h_load_update
== now
)
5782 cfs_rq
->h_load_next
= NULL
;
5783 for_each_sched_entity(se
) {
5784 cfs_rq
= cfs_rq_of(se
);
5785 cfs_rq
->h_load_next
= se
;
5786 if (cfs_rq
->last_h_load_update
== now
)
5791 cfs_rq
->h_load
= cfs_rq
->runnable_load_avg
;
5792 cfs_rq
->last_h_load_update
= now
;
5795 while ((se
= cfs_rq
->h_load_next
) != NULL
) {
5796 load
= cfs_rq
->h_load
;
5797 load
= div64_ul(load
* se
->avg
.load_avg_contrib
,
5798 cfs_rq
->runnable_load_avg
+ 1);
5799 cfs_rq
= group_cfs_rq(se
);
5800 cfs_rq
->h_load
= load
;
5801 cfs_rq
->last_h_load_update
= now
;
5805 static unsigned long task_h_load(struct task_struct
*p
)
5807 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
5809 update_cfs_rq_h_load(cfs_rq
);
5810 return div64_ul(p
->se
.avg
.load_avg_contrib
* cfs_rq
->h_load
,
5811 cfs_rq
->runnable_load_avg
+ 1);
5814 static inline void update_blocked_averages(int cpu
)
5818 static unsigned long task_h_load(struct task_struct
*p
)
5820 return p
->se
.avg
.load_avg_contrib
;
5824 /********** Helpers for find_busiest_group ************************/
5833 * sg_lb_stats - stats of a sched_group required for load_balancing
5835 struct sg_lb_stats
{
5836 unsigned long avg_load
; /*Avg load across the CPUs of the group */
5837 unsigned long group_load
; /* Total load over the CPUs of the group */
5838 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
5839 unsigned long load_per_task
;
5840 unsigned long group_capacity
;
5841 unsigned int sum_nr_running
; /* Nr tasks running in the group */
5842 unsigned int group_capacity_factor
;
5843 unsigned int idle_cpus
;
5844 unsigned int group_weight
;
5845 enum group_type group_type
;
5846 int group_has_free_capacity
;
5847 #ifdef CONFIG_NUMA_BALANCING
5848 unsigned int nr_numa_running
;
5849 unsigned int nr_preferred_running
;
5854 * sd_lb_stats - Structure to store the statistics of a sched_domain
5855 * during load balancing.
5857 struct sd_lb_stats
{
5858 struct sched_group
*busiest
; /* Busiest group in this sd */
5859 struct sched_group
*local
; /* Local group in this sd */
5860 unsigned long total_load
; /* Total load of all groups in sd */
5861 unsigned long total_capacity
; /* Total capacity of all groups in sd */
5862 unsigned long avg_load
; /* Average load across all groups in sd */
5864 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
5865 struct sg_lb_stats local_stat
; /* Statistics of the local group */
5868 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
5871 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5872 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5873 * We must however clear busiest_stat::avg_load because
5874 * update_sd_pick_busiest() reads this before assignment.
5876 *sds
= (struct sd_lb_stats
){
5880 .total_capacity
= 0UL,
5883 .sum_nr_running
= 0,
5884 .group_type
= group_other
,
5890 * get_sd_load_idx - Obtain the load index for a given sched domain.
5891 * @sd: The sched_domain whose load_idx is to be obtained.
5892 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5894 * Return: The load index.
5896 static inline int get_sd_load_idx(struct sched_domain
*sd
,
5897 enum cpu_idle_type idle
)
5903 load_idx
= sd
->busy_idx
;
5906 case CPU_NEWLY_IDLE
:
5907 load_idx
= sd
->newidle_idx
;
5910 load_idx
= sd
->idle_idx
;
5917 static unsigned long default_scale_capacity(struct sched_domain
*sd
, int cpu
)
5919 return SCHED_CAPACITY_SCALE
;
5922 unsigned long __weak
arch_scale_freq_capacity(struct sched_domain
*sd
, int cpu
)
5924 return default_scale_capacity(sd
, cpu
);
5927 static unsigned long default_scale_cpu_capacity(struct sched_domain
*sd
, int cpu
)
5929 if ((sd
->flags
& SD_SHARE_CPUCAPACITY
) && (sd
->span_weight
> 1))
5930 return sd
->smt_gain
/ sd
->span_weight
;
5932 return SCHED_CAPACITY_SCALE
;
5935 unsigned long __weak
arch_scale_cpu_capacity(struct sched_domain
*sd
, int cpu
)
5937 return default_scale_cpu_capacity(sd
, cpu
);
5940 static unsigned long scale_rt_capacity(int cpu
)
5942 struct rq
*rq
= cpu_rq(cpu
);
5943 u64 total
, available
, age_stamp
, avg
;
5947 * Since we're reading these variables without serialization make sure
5948 * we read them once before doing sanity checks on them.
5950 age_stamp
= ACCESS_ONCE(rq
->age_stamp
);
5951 avg
= ACCESS_ONCE(rq
->rt_avg
);
5953 delta
= rq_clock(rq
) - age_stamp
;
5954 if (unlikely(delta
< 0))
5957 total
= sched_avg_period() + delta
;
5959 if (unlikely(total
< avg
)) {
5960 /* Ensures that capacity won't end up being negative */
5963 available
= total
- avg
;
5966 if (unlikely((s64
)total
< SCHED_CAPACITY_SCALE
))
5967 total
= SCHED_CAPACITY_SCALE
;
5969 total
>>= SCHED_CAPACITY_SHIFT
;
5971 return div_u64(available
, total
);
5974 static void update_cpu_capacity(struct sched_domain
*sd
, int cpu
)
5976 unsigned long capacity
= SCHED_CAPACITY_SCALE
;
5977 struct sched_group
*sdg
= sd
->groups
;
5979 if (sched_feat(ARCH_CAPACITY
))
5980 capacity
*= arch_scale_cpu_capacity(sd
, cpu
);
5982 capacity
*= default_scale_cpu_capacity(sd
, cpu
);
5984 capacity
>>= SCHED_CAPACITY_SHIFT
;
5986 sdg
->sgc
->capacity_orig
= capacity
;
5988 if (sched_feat(ARCH_CAPACITY
))
5989 capacity
*= arch_scale_freq_capacity(sd
, cpu
);
5991 capacity
*= default_scale_capacity(sd
, cpu
);
5993 capacity
>>= SCHED_CAPACITY_SHIFT
;
5995 capacity
*= scale_rt_capacity(cpu
);
5996 capacity
>>= SCHED_CAPACITY_SHIFT
;
6001 cpu_rq(cpu
)->cpu_capacity
= capacity
;
6002 sdg
->sgc
->capacity
= capacity
;
6005 void update_group_capacity(struct sched_domain
*sd
, int cpu
)
6007 struct sched_domain
*child
= sd
->child
;
6008 struct sched_group
*group
, *sdg
= sd
->groups
;
6009 unsigned long capacity
, capacity_orig
;
6010 unsigned long interval
;
6012 interval
= msecs_to_jiffies(sd
->balance_interval
);
6013 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
6014 sdg
->sgc
->next_update
= jiffies
+ interval
;
6017 update_cpu_capacity(sd
, cpu
);
6021 capacity_orig
= capacity
= 0;
6023 if (child
->flags
& SD_OVERLAP
) {
6025 * SD_OVERLAP domains cannot assume that child groups
6026 * span the current group.
6029 for_each_cpu(cpu
, sched_group_cpus(sdg
)) {
6030 struct sched_group_capacity
*sgc
;
6031 struct rq
*rq
= cpu_rq(cpu
);
6034 * build_sched_domains() -> init_sched_groups_capacity()
6035 * gets here before we've attached the domains to the
6038 * Use capacity_of(), which is set irrespective of domains
6039 * in update_cpu_capacity().
6041 * This avoids capacity/capacity_orig from being 0 and
6042 * causing divide-by-zero issues on boot.
6044 * Runtime updates will correct capacity_orig.
6046 if (unlikely(!rq
->sd
)) {
6047 capacity_orig
+= capacity_of(cpu
);
6048 capacity
+= capacity_of(cpu
);
6052 sgc
= rq
->sd
->groups
->sgc
;
6053 capacity_orig
+= sgc
->capacity_orig
;
6054 capacity
+= sgc
->capacity
;
6058 * !SD_OVERLAP domains can assume that child groups
6059 * span the current group.
6062 group
= child
->groups
;
6064 capacity_orig
+= group
->sgc
->capacity_orig
;
6065 capacity
+= group
->sgc
->capacity
;
6066 group
= group
->next
;
6067 } while (group
!= child
->groups
);
6070 sdg
->sgc
->capacity_orig
= capacity_orig
;
6071 sdg
->sgc
->capacity
= capacity
;
6075 * Try and fix up capacity for tiny siblings, this is needed when
6076 * things like SD_ASYM_PACKING need f_b_g to select another sibling
6077 * which on its own isn't powerful enough.
6079 * See update_sd_pick_busiest() and check_asym_packing().
6082 fix_small_capacity(struct sched_domain
*sd
, struct sched_group
*group
)
6085 * Only siblings can have significantly less than SCHED_CAPACITY_SCALE
6087 if (!(sd
->flags
& SD_SHARE_CPUCAPACITY
))
6091 * If ~90% of the cpu_capacity is still there, we're good.
6093 if (group
->sgc
->capacity
* 32 > group
->sgc
->capacity_orig
* 29)
6100 * Group imbalance indicates (and tries to solve) the problem where balancing
6101 * groups is inadequate due to tsk_cpus_allowed() constraints.
6103 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6104 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6107 * { 0 1 2 3 } { 4 5 6 7 }
6110 * If we were to balance group-wise we'd place two tasks in the first group and
6111 * two tasks in the second group. Clearly this is undesired as it will overload
6112 * cpu 3 and leave one of the cpus in the second group unused.
6114 * The current solution to this issue is detecting the skew in the first group
6115 * by noticing the lower domain failed to reach balance and had difficulty
6116 * moving tasks due to affinity constraints.
6118 * When this is so detected; this group becomes a candidate for busiest; see
6119 * update_sd_pick_busiest(). And calculate_imbalance() and
6120 * find_busiest_group() avoid some of the usual balance conditions to allow it
6121 * to create an effective group imbalance.
6123 * This is a somewhat tricky proposition since the next run might not find the
6124 * group imbalance and decide the groups need to be balanced again. A most
6125 * subtle and fragile situation.
6128 static inline int sg_imbalanced(struct sched_group
*group
)
6130 return group
->sgc
->imbalance
;
6134 * Compute the group capacity factor.
6136 * Avoid the issue where N*frac(smt_capacity) >= 1 creates 'phantom' cores by
6137 * first dividing out the smt factor and computing the actual number of cores
6138 * and limit unit capacity with that.
6140 static inline int sg_capacity_factor(struct lb_env
*env
, struct sched_group
*group
)
6142 unsigned int capacity_factor
, smt
, cpus
;
6143 unsigned int capacity
, capacity_orig
;
6145 capacity
= group
->sgc
->capacity
;
6146 capacity_orig
= group
->sgc
->capacity_orig
;
6147 cpus
= group
->group_weight
;
6149 /* smt := ceil(cpus / capacity), assumes: 1 < smt_capacity < 2 */
6150 smt
= DIV_ROUND_UP(SCHED_CAPACITY_SCALE
* cpus
, capacity_orig
);
6151 capacity_factor
= cpus
/ smt
; /* cores */
6153 capacity_factor
= min_t(unsigned,
6154 capacity_factor
, DIV_ROUND_CLOSEST(capacity
, SCHED_CAPACITY_SCALE
));
6155 if (!capacity_factor
)
6156 capacity_factor
= fix_small_capacity(env
->sd
, group
);
6158 return capacity_factor
;
6161 static enum group_type
6162 group_classify(struct sched_group
*group
, struct sg_lb_stats
*sgs
)
6164 if (sgs
->sum_nr_running
> sgs
->group_capacity_factor
)
6165 return group_overloaded
;
6167 if (sg_imbalanced(group
))
6168 return group_imbalanced
;
6174 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6175 * @env: The load balancing environment.
6176 * @group: sched_group whose statistics are to be updated.
6177 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6178 * @local_group: Does group contain this_cpu.
6179 * @sgs: variable to hold the statistics for this group.
6180 * @overload: Indicate more than one runnable task for any CPU.
6182 static inline void update_sg_lb_stats(struct lb_env
*env
,
6183 struct sched_group
*group
, int load_idx
,
6184 int local_group
, struct sg_lb_stats
*sgs
,
6190 memset(sgs
, 0, sizeof(*sgs
));
6192 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
6193 struct rq
*rq
= cpu_rq(i
);
6195 /* Bias balancing toward cpus of our domain */
6197 load
= target_load(i
, load_idx
);
6199 load
= source_load(i
, load_idx
);
6201 sgs
->group_load
+= load
;
6202 sgs
->sum_nr_running
+= rq
->cfs
.h_nr_running
;
6204 if (rq
->nr_running
> 1)
6207 #ifdef CONFIG_NUMA_BALANCING
6208 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
6209 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
6211 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
6216 /* Adjust by relative CPU capacity of the group */
6217 sgs
->group_capacity
= group
->sgc
->capacity
;
6218 sgs
->avg_load
= (sgs
->group_load
*SCHED_CAPACITY_SCALE
) / sgs
->group_capacity
;
6220 if (sgs
->sum_nr_running
)
6221 sgs
->load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
6223 sgs
->group_weight
= group
->group_weight
;
6224 sgs
->group_capacity_factor
= sg_capacity_factor(env
, group
);
6225 sgs
->group_type
= group_classify(group
, sgs
);
6227 if (sgs
->group_capacity_factor
> sgs
->sum_nr_running
)
6228 sgs
->group_has_free_capacity
= 1;
6232 * update_sd_pick_busiest - return 1 on busiest group
6233 * @env: The load balancing environment.
6234 * @sds: sched_domain statistics
6235 * @sg: sched_group candidate to be checked for being the busiest
6236 * @sgs: sched_group statistics
6238 * Determine if @sg is a busier group than the previously selected
6241 * Return: %true if @sg is a busier group than the previously selected
6242 * busiest group. %false otherwise.
6244 static bool update_sd_pick_busiest(struct lb_env
*env
,
6245 struct sd_lb_stats
*sds
,
6246 struct sched_group
*sg
,
6247 struct sg_lb_stats
*sgs
)
6249 struct sg_lb_stats
*busiest
= &sds
->busiest_stat
;
6251 if (sgs
->group_type
> busiest
->group_type
)
6254 if (sgs
->group_type
< busiest
->group_type
)
6257 if (sgs
->avg_load
<= busiest
->avg_load
)
6260 /* This is the busiest node in its class. */
6261 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
6265 * ASYM_PACKING needs to move all the work to the lowest
6266 * numbered CPUs in the group, therefore mark all groups
6267 * higher than ourself as busy.
6269 if (sgs
->sum_nr_running
&& env
->dst_cpu
< group_first_cpu(sg
)) {
6273 if (group_first_cpu(sds
->busiest
) > group_first_cpu(sg
))
6280 #ifdef CONFIG_NUMA_BALANCING
6281 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
6283 if (sgs
->sum_nr_running
> sgs
->nr_numa_running
)
6285 if (sgs
->sum_nr_running
> sgs
->nr_preferred_running
)
6290 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
6292 if (rq
->nr_running
> rq
->nr_numa_running
)
6294 if (rq
->nr_running
> rq
->nr_preferred_running
)
6299 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
6304 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
6308 #endif /* CONFIG_NUMA_BALANCING */
6311 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6312 * @env: The load balancing environment.
6313 * @sds: variable to hold the statistics for this sched_domain.
6315 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6317 struct sched_domain
*child
= env
->sd
->child
;
6318 struct sched_group
*sg
= env
->sd
->groups
;
6319 struct sg_lb_stats tmp_sgs
;
6320 int load_idx
, prefer_sibling
= 0;
6321 bool overload
= false;
6323 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
6326 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
6329 struct sg_lb_stats
*sgs
= &tmp_sgs
;
6332 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_cpus(sg
));
6335 sgs
= &sds
->local_stat
;
6337 if (env
->idle
!= CPU_NEWLY_IDLE
||
6338 time_after_eq(jiffies
, sg
->sgc
->next_update
))
6339 update_group_capacity(env
->sd
, env
->dst_cpu
);
6342 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, sgs
,
6349 * In case the child domain prefers tasks go to siblings
6350 * first, lower the sg capacity factor to one so that we'll try
6351 * and move all the excess tasks away. We lower the capacity
6352 * of a group only if the local group has the capacity to fit
6353 * these excess tasks, i.e. nr_running < group_capacity_factor. The
6354 * extra check prevents the case where you always pull from the
6355 * heaviest group when it is already under-utilized (possible
6356 * with a large weight task outweighs the tasks on the system).
6358 if (prefer_sibling
&& sds
->local
&&
6359 sds
->local_stat
.group_has_free_capacity
) {
6360 sgs
->group_capacity_factor
= min(sgs
->group_capacity_factor
, 1U);
6361 sgs
->group_type
= group_classify(sg
, sgs
);
6364 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
6366 sds
->busiest_stat
= *sgs
;
6370 /* Now, start updating sd_lb_stats */
6371 sds
->total_load
+= sgs
->group_load
;
6372 sds
->total_capacity
+= sgs
->group_capacity
;
6375 } while (sg
!= env
->sd
->groups
);
6377 if (env
->sd
->flags
& SD_NUMA
)
6378 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
6380 if (!env
->sd
->parent
) {
6381 /* update overload indicator if we are at root domain */
6382 if (env
->dst_rq
->rd
->overload
!= overload
)
6383 env
->dst_rq
->rd
->overload
= overload
;
6389 * check_asym_packing - Check to see if the group is packed into the
6392 * This is primarily intended to used at the sibling level. Some
6393 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6394 * case of POWER7, it can move to lower SMT modes only when higher
6395 * threads are idle. When in lower SMT modes, the threads will
6396 * perform better since they share less core resources. Hence when we
6397 * have idle threads, we want them to be the higher ones.
6399 * This packing function is run on idle threads. It checks to see if
6400 * the busiest CPU in this domain (core in the P7 case) has a higher
6401 * CPU number than the packing function is being run on. Here we are
6402 * assuming lower CPU number will be equivalent to lower a SMT thread
6405 * Return: 1 when packing is required and a task should be moved to
6406 * this CPU. The amount of the imbalance is returned in *imbalance.
6408 * @env: The load balancing environment.
6409 * @sds: Statistics of the sched_domain which is to be packed
6411 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6415 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
6421 busiest_cpu
= group_first_cpu(sds
->busiest
);
6422 if (env
->dst_cpu
> busiest_cpu
)
6425 env
->imbalance
= DIV_ROUND_CLOSEST(
6426 sds
->busiest_stat
.avg_load
* sds
->busiest_stat
.group_capacity
,
6427 SCHED_CAPACITY_SCALE
);
6433 * fix_small_imbalance - Calculate the minor imbalance that exists
6434 * amongst the groups of a sched_domain, during
6436 * @env: The load balancing environment.
6437 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6440 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6442 unsigned long tmp
, capa_now
= 0, capa_move
= 0;
6443 unsigned int imbn
= 2;
6444 unsigned long scaled_busy_load_per_task
;
6445 struct sg_lb_stats
*local
, *busiest
;
6447 local
= &sds
->local_stat
;
6448 busiest
= &sds
->busiest_stat
;
6450 if (!local
->sum_nr_running
)
6451 local
->load_per_task
= cpu_avg_load_per_task(env
->dst_cpu
);
6452 else if (busiest
->load_per_task
> local
->load_per_task
)
6455 scaled_busy_load_per_task
=
6456 (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
6457 busiest
->group_capacity
;
6459 if (busiest
->avg_load
+ scaled_busy_load_per_task
>=
6460 local
->avg_load
+ (scaled_busy_load_per_task
* imbn
)) {
6461 env
->imbalance
= busiest
->load_per_task
;
6466 * OK, we don't have enough imbalance to justify moving tasks,
6467 * however we may be able to increase total CPU capacity used by
6471 capa_now
+= busiest
->group_capacity
*
6472 min(busiest
->load_per_task
, busiest
->avg_load
);
6473 capa_now
+= local
->group_capacity
*
6474 min(local
->load_per_task
, local
->avg_load
);
6475 capa_now
/= SCHED_CAPACITY_SCALE
;
6477 /* Amount of load we'd subtract */
6478 if (busiest
->avg_load
> scaled_busy_load_per_task
) {
6479 capa_move
+= busiest
->group_capacity
*
6480 min(busiest
->load_per_task
,
6481 busiest
->avg_load
- scaled_busy_load_per_task
);
6484 /* Amount of load we'd add */
6485 if (busiest
->avg_load
* busiest
->group_capacity
<
6486 busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) {
6487 tmp
= (busiest
->avg_load
* busiest
->group_capacity
) /
6488 local
->group_capacity
;
6490 tmp
= (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
6491 local
->group_capacity
;
6493 capa_move
+= local
->group_capacity
*
6494 min(local
->load_per_task
, local
->avg_load
+ tmp
);
6495 capa_move
/= SCHED_CAPACITY_SCALE
;
6497 /* Move if we gain throughput */
6498 if (capa_move
> capa_now
)
6499 env
->imbalance
= busiest
->load_per_task
;
6503 * calculate_imbalance - Calculate the amount of imbalance present within the
6504 * groups of a given sched_domain during load balance.
6505 * @env: load balance environment
6506 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6508 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6510 unsigned long max_pull
, load_above_capacity
= ~0UL;
6511 struct sg_lb_stats
*local
, *busiest
;
6513 local
= &sds
->local_stat
;
6514 busiest
= &sds
->busiest_stat
;
6516 if (busiest
->group_type
== group_imbalanced
) {
6518 * In the group_imb case we cannot rely on group-wide averages
6519 * to ensure cpu-load equilibrium, look at wider averages. XXX
6521 busiest
->load_per_task
=
6522 min(busiest
->load_per_task
, sds
->avg_load
);
6526 * In the presence of smp nice balancing, certain scenarios can have
6527 * max load less than avg load(as we skip the groups at or below
6528 * its cpu_capacity, while calculating max_load..)
6530 if (busiest
->avg_load
<= sds
->avg_load
||
6531 local
->avg_load
>= sds
->avg_load
) {
6533 return fix_small_imbalance(env
, sds
);
6537 * If there aren't any idle cpus, avoid creating some.
6539 if (busiest
->group_type
== group_overloaded
&&
6540 local
->group_type
== group_overloaded
) {
6541 load_above_capacity
=
6542 (busiest
->sum_nr_running
- busiest
->group_capacity_factor
);
6544 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_CAPACITY_SCALE
);
6545 load_above_capacity
/= busiest
->group_capacity
;
6549 * We're trying to get all the cpus to the average_load, so we don't
6550 * want to push ourselves above the average load, nor do we wish to
6551 * reduce the max loaded cpu below the average load. At the same time,
6552 * we also don't want to reduce the group load below the group capacity
6553 * (so that we can implement power-savings policies etc). Thus we look
6554 * for the minimum possible imbalance.
6556 max_pull
= min(busiest
->avg_load
- sds
->avg_load
, load_above_capacity
);
6558 /* How much load to actually move to equalise the imbalance */
6559 env
->imbalance
= min(
6560 max_pull
* busiest
->group_capacity
,
6561 (sds
->avg_load
- local
->avg_load
) * local
->group_capacity
6562 ) / SCHED_CAPACITY_SCALE
;
6565 * if *imbalance is less than the average load per runnable task
6566 * there is no guarantee that any tasks will be moved so we'll have
6567 * a think about bumping its value to force at least one task to be
6570 if (env
->imbalance
< busiest
->load_per_task
)
6571 return fix_small_imbalance(env
, sds
);
6574 /******* find_busiest_group() helpers end here *********************/
6577 * find_busiest_group - Returns the busiest group within the sched_domain
6578 * if there is an imbalance. If there isn't an imbalance, and
6579 * the user has opted for power-savings, it returns a group whose
6580 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6581 * such a group exists.
6583 * Also calculates the amount of weighted load which should be moved
6584 * to restore balance.
6586 * @env: The load balancing environment.
6588 * Return: - The busiest group if imbalance exists.
6589 * - If no imbalance and user has opted for power-savings balance,
6590 * return the least loaded group whose CPUs can be
6591 * put to idle by rebalancing its tasks onto our group.
6593 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
6595 struct sg_lb_stats
*local
, *busiest
;
6596 struct sd_lb_stats sds
;
6598 init_sd_lb_stats(&sds
);
6601 * Compute the various statistics relavent for load balancing at
6604 update_sd_lb_stats(env
, &sds
);
6605 local
= &sds
.local_stat
;
6606 busiest
= &sds
.busiest_stat
;
6608 if ((env
->idle
== CPU_IDLE
|| env
->idle
== CPU_NEWLY_IDLE
) &&
6609 check_asym_packing(env
, &sds
))
6612 /* There is no busy sibling group to pull tasks from */
6613 if (!sds
.busiest
|| busiest
->sum_nr_running
== 0)
6616 sds
.avg_load
= (SCHED_CAPACITY_SCALE
* sds
.total_load
)
6617 / sds
.total_capacity
;
6620 * If the busiest group is imbalanced the below checks don't
6621 * work because they assume all things are equal, which typically
6622 * isn't true due to cpus_allowed constraints and the like.
6624 if (busiest
->group_type
== group_imbalanced
)
6627 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6628 if (env
->idle
== CPU_NEWLY_IDLE
&& local
->group_has_free_capacity
&&
6629 !busiest
->group_has_free_capacity
)
6633 * If the local group is busier than the selected busiest group
6634 * don't try and pull any tasks.
6636 if (local
->avg_load
>= busiest
->avg_load
)
6640 * Don't pull any tasks if this group is already above the domain
6643 if (local
->avg_load
>= sds
.avg_load
)
6646 if (env
->idle
== CPU_IDLE
) {
6648 * This cpu is idle. If the busiest group is not overloaded
6649 * and there is no imbalance between this and busiest group
6650 * wrt idle cpus, it is balanced. The imbalance becomes
6651 * significant if the diff is greater than 1 otherwise we
6652 * might end up to just move the imbalance on another group
6654 if ((busiest
->group_type
!= group_overloaded
) &&
6655 (local
->idle_cpus
<= (busiest
->idle_cpus
+ 1)))
6659 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6660 * imbalance_pct to be conservative.
6662 if (100 * busiest
->avg_load
<=
6663 env
->sd
->imbalance_pct
* local
->avg_load
)
6668 /* Looks like there is an imbalance. Compute it */
6669 calculate_imbalance(env
, &sds
);
6678 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6680 static struct rq
*find_busiest_queue(struct lb_env
*env
,
6681 struct sched_group
*group
)
6683 struct rq
*busiest
= NULL
, *rq
;
6684 unsigned long busiest_load
= 0, busiest_capacity
= 1;
6687 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
6688 unsigned long capacity
, capacity_factor
, wl
;
6692 rt
= fbq_classify_rq(rq
);
6695 * We classify groups/runqueues into three groups:
6696 * - regular: there are !numa tasks
6697 * - remote: there are numa tasks that run on the 'wrong' node
6698 * - all: there is no distinction
6700 * In order to avoid migrating ideally placed numa tasks,
6701 * ignore those when there's better options.
6703 * If we ignore the actual busiest queue to migrate another
6704 * task, the next balance pass can still reduce the busiest
6705 * queue by moving tasks around inside the node.
6707 * If we cannot move enough load due to this classification
6708 * the next pass will adjust the group classification and
6709 * allow migration of more tasks.
6711 * Both cases only affect the total convergence complexity.
6713 if (rt
> env
->fbq_type
)
6716 capacity
= capacity_of(i
);
6717 capacity_factor
= DIV_ROUND_CLOSEST(capacity
, SCHED_CAPACITY_SCALE
);
6718 if (!capacity_factor
)
6719 capacity_factor
= fix_small_capacity(env
->sd
, group
);
6721 wl
= weighted_cpuload(i
);
6724 * When comparing with imbalance, use weighted_cpuload()
6725 * which is not scaled with the cpu capacity.
6727 if (capacity_factor
&& rq
->nr_running
== 1 && wl
> env
->imbalance
)
6731 * For the load comparisons with the other cpu's, consider
6732 * the weighted_cpuload() scaled with the cpu capacity, so
6733 * that the load can be moved away from the cpu that is
6734 * potentially running at a lower capacity.
6736 * Thus we're looking for max(wl_i / capacity_i), crosswise
6737 * multiplication to rid ourselves of the division works out
6738 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6739 * our previous maximum.
6741 if (wl
* busiest_capacity
> busiest_load
* capacity
) {
6743 busiest_capacity
= capacity
;
6752 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6753 * so long as it is large enough.
6755 #define MAX_PINNED_INTERVAL 512
6757 /* Working cpumask for load_balance and load_balance_newidle. */
6758 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6760 static int need_active_balance(struct lb_env
*env
)
6762 struct sched_domain
*sd
= env
->sd
;
6764 if (env
->idle
== CPU_NEWLY_IDLE
) {
6767 * ASYM_PACKING needs to force migrate tasks from busy but
6768 * higher numbered CPUs in order to pack all tasks in the
6769 * lowest numbered CPUs.
6771 if ((sd
->flags
& SD_ASYM_PACKING
) && env
->src_cpu
> env
->dst_cpu
)
6775 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
6778 static int active_load_balance_cpu_stop(void *data
);
6780 static int should_we_balance(struct lb_env
*env
)
6782 struct sched_group
*sg
= env
->sd
->groups
;
6783 struct cpumask
*sg_cpus
, *sg_mask
;
6784 int cpu
, balance_cpu
= -1;
6787 * In the newly idle case, we will allow all the cpu's
6788 * to do the newly idle load balance.
6790 if (env
->idle
== CPU_NEWLY_IDLE
)
6793 sg_cpus
= sched_group_cpus(sg
);
6794 sg_mask
= sched_group_mask(sg
);
6795 /* Try to find first idle cpu */
6796 for_each_cpu_and(cpu
, sg_cpus
, env
->cpus
) {
6797 if (!cpumask_test_cpu(cpu
, sg_mask
) || !idle_cpu(cpu
))
6804 if (balance_cpu
== -1)
6805 balance_cpu
= group_balance_cpu(sg
);
6808 * First idle cpu or the first cpu(busiest) in this sched group
6809 * is eligible for doing load balancing at this and above domains.
6811 return balance_cpu
== env
->dst_cpu
;
6815 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6816 * tasks if there is an imbalance.
6818 static int load_balance(int this_cpu
, struct rq
*this_rq
,
6819 struct sched_domain
*sd
, enum cpu_idle_type idle
,
6820 int *continue_balancing
)
6822 int ld_moved
, cur_ld_moved
, active_balance
= 0;
6823 struct sched_domain
*sd_parent
= sd
->parent
;
6824 struct sched_group
*group
;
6826 unsigned long flags
;
6827 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(load_balance_mask
);
6829 struct lb_env env
= {
6831 .dst_cpu
= this_cpu
,
6833 .dst_grpmask
= sched_group_cpus(sd
->groups
),
6835 .loop_break
= sched_nr_migrate_break
,
6838 .tasks
= LIST_HEAD_INIT(env
.tasks
),
6842 * For NEWLY_IDLE load_balancing, we don't need to consider
6843 * other cpus in our group
6845 if (idle
== CPU_NEWLY_IDLE
)
6846 env
.dst_grpmask
= NULL
;
6848 cpumask_copy(cpus
, cpu_active_mask
);
6850 schedstat_inc(sd
, lb_count
[idle
]);
6853 if (!should_we_balance(&env
)) {
6854 *continue_balancing
= 0;
6858 group
= find_busiest_group(&env
);
6860 schedstat_inc(sd
, lb_nobusyg
[idle
]);
6864 busiest
= find_busiest_queue(&env
, group
);
6866 schedstat_inc(sd
, lb_nobusyq
[idle
]);
6870 BUG_ON(busiest
== env
.dst_rq
);
6872 schedstat_add(sd
, lb_imbalance
[idle
], env
.imbalance
);
6875 if (busiest
->nr_running
> 1) {
6877 * Attempt to move tasks. If find_busiest_group has found
6878 * an imbalance but busiest->nr_running <= 1, the group is
6879 * still unbalanced. ld_moved simply stays zero, so it is
6880 * correctly treated as an imbalance.
6882 env
.flags
|= LBF_ALL_PINNED
;
6883 env
.src_cpu
= busiest
->cpu
;
6884 env
.src_rq
= busiest
;
6885 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
6888 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
6891 * cur_ld_moved - load moved in current iteration
6892 * ld_moved - cumulative load moved across iterations
6894 cur_ld_moved
= detach_tasks(&env
);
6897 * We've detached some tasks from busiest_rq. Every
6898 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
6899 * unlock busiest->lock, and we are able to be sure
6900 * that nobody can manipulate the tasks in parallel.
6901 * See task_rq_lock() family for the details.
6904 raw_spin_unlock(&busiest
->lock
);
6908 ld_moved
+= cur_ld_moved
;
6911 local_irq_restore(flags
);
6913 if (env
.flags
& LBF_NEED_BREAK
) {
6914 env
.flags
&= ~LBF_NEED_BREAK
;
6919 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6920 * us and move them to an alternate dst_cpu in our sched_group
6921 * where they can run. The upper limit on how many times we
6922 * iterate on same src_cpu is dependent on number of cpus in our
6925 * This changes load balance semantics a bit on who can move
6926 * load to a given_cpu. In addition to the given_cpu itself
6927 * (or a ilb_cpu acting on its behalf where given_cpu is
6928 * nohz-idle), we now have balance_cpu in a position to move
6929 * load to given_cpu. In rare situations, this may cause
6930 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6931 * _independently_ and at _same_ time to move some load to
6932 * given_cpu) causing exceess load to be moved to given_cpu.
6933 * This however should not happen so much in practice and
6934 * moreover subsequent load balance cycles should correct the
6935 * excess load moved.
6937 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
6939 /* Prevent to re-select dst_cpu via env's cpus */
6940 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
6942 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
6943 env
.dst_cpu
= env
.new_dst_cpu
;
6944 env
.flags
&= ~LBF_DST_PINNED
;
6946 env
.loop_break
= sched_nr_migrate_break
;
6949 * Go back to "more_balance" rather than "redo" since we
6950 * need to continue with same src_cpu.
6956 * We failed to reach balance because of affinity.
6959 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
6961 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0)
6962 *group_imbalance
= 1;
6965 /* All tasks on this runqueue were pinned by CPU affinity */
6966 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
6967 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
6968 if (!cpumask_empty(cpus
)) {
6970 env
.loop_break
= sched_nr_migrate_break
;
6973 goto out_all_pinned
;
6978 schedstat_inc(sd
, lb_failed
[idle
]);
6980 * Increment the failure counter only on periodic balance.
6981 * We do not want newidle balance, which can be very
6982 * frequent, pollute the failure counter causing
6983 * excessive cache_hot migrations and active balances.
6985 if (idle
!= CPU_NEWLY_IDLE
)
6986 sd
->nr_balance_failed
++;
6988 if (need_active_balance(&env
)) {
6989 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
6991 /* don't kick the active_load_balance_cpu_stop,
6992 * if the curr task on busiest cpu can't be
6995 if (!cpumask_test_cpu(this_cpu
,
6996 tsk_cpus_allowed(busiest
->curr
))) {
6997 raw_spin_unlock_irqrestore(&busiest
->lock
,
6999 env
.flags
|= LBF_ALL_PINNED
;
7000 goto out_one_pinned
;
7004 * ->active_balance synchronizes accesses to
7005 * ->active_balance_work. Once set, it's cleared
7006 * only after active load balance is finished.
7008 if (!busiest
->active_balance
) {
7009 busiest
->active_balance
= 1;
7010 busiest
->push_cpu
= this_cpu
;
7013 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
7015 if (active_balance
) {
7016 stop_one_cpu_nowait(cpu_of(busiest
),
7017 active_load_balance_cpu_stop
, busiest
,
7018 &busiest
->active_balance_work
);
7022 * We've kicked active balancing, reset the failure
7025 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
7028 sd
->nr_balance_failed
= 0;
7030 if (likely(!active_balance
)) {
7031 /* We were unbalanced, so reset the balancing interval */
7032 sd
->balance_interval
= sd
->min_interval
;
7035 * If we've begun active balancing, start to back off. This
7036 * case may not be covered by the all_pinned logic if there
7037 * is only 1 task on the busy runqueue (because we don't call
7040 if (sd
->balance_interval
< sd
->max_interval
)
7041 sd
->balance_interval
*= 2;
7048 * We reach balance although we may have faced some affinity
7049 * constraints. Clear the imbalance flag if it was set.
7052 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
7054 if (*group_imbalance
)
7055 *group_imbalance
= 0;
7060 * We reach balance because all tasks are pinned at this level so
7061 * we can't migrate them. Let the imbalance flag set so parent level
7062 * can try to migrate them.
7064 schedstat_inc(sd
, lb_balanced
[idle
]);
7066 sd
->nr_balance_failed
= 0;
7069 /* tune up the balancing interval */
7070 if (((env
.flags
& LBF_ALL_PINNED
) &&
7071 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
7072 (sd
->balance_interval
< sd
->max_interval
))
7073 sd
->balance_interval
*= 2;
7080 static inline unsigned long
7081 get_sd_balance_interval(struct sched_domain
*sd
, int cpu_busy
)
7083 unsigned long interval
= sd
->balance_interval
;
7086 interval
*= sd
->busy_factor
;
7088 /* scale ms to jiffies */
7089 interval
= msecs_to_jiffies(interval
);
7090 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
7096 update_next_balance(struct sched_domain
*sd
, int cpu_busy
, unsigned long *next_balance
)
7098 unsigned long interval
, next
;
7100 interval
= get_sd_balance_interval(sd
, cpu_busy
);
7101 next
= sd
->last_balance
+ interval
;
7103 if (time_after(*next_balance
, next
))
7104 *next_balance
= next
;
7108 * idle_balance is called by schedule() if this_cpu is about to become
7109 * idle. Attempts to pull tasks from other CPUs.
7111 static int idle_balance(struct rq
*this_rq
)
7113 unsigned long next_balance
= jiffies
+ HZ
;
7114 int this_cpu
= this_rq
->cpu
;
7115 struct sched_domain
*sd
;
7116 int pulled_task
= 0;
7119 idle_enter_fair(this_rq
);
7122 * We must set idle_stamp _before_ calling idle_balance(), such that we
7123 * measure the duration of idle_balance() as idle time.
7125 this_rq
->idle_stamp
= rq_clock(this_rq
);
7127 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
||
7128 !this_rq
->rd
->overload
) {
7130 sd
= rcu_dereference_check_sched_domain(this_rq
->sd
);
7132 update_next_balance(sd
, 0, &next_balance
);
7139 * Drop the rq->lock, but keep IRQ/preempt disabled.
7141 raw_spin_unlock(&this_rq
->lock
);
7143 update_blocked_averages(this_cpu
);
7145 for_each_domain(this_cpu
, sd
) {
7146 int continue_balancing
= 1;
7147 u64 t0
, domain_cost
;
7149 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7152 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
) {
7153 update_next_balance(sd
, 0, &next_balance
);
7157 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
7158 t0
= sched_clock_cpu(this_cpu
);
7160 pulled_task
= load_balance(this_cpu
, this_rq
,
7162 &continue_balancing
);
7164 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
7165 if (domain_cost
> sd
->max_newidle_lb_cost
)
7166 sd
->max_newidle_lb_cost
= domain_cost
;
7168 curr_cost
+= domain_cost
;
7171 update_next_balance(sd
, 0, &next_balance
);
7174 * Stop searching for tasks to pull if there are
7175 * now runnable tasks on this rq.
7177 if (pulled_task
|| this_rq
->nr_running
> 0)
7182 raw_spin_lock(&this_rq
->lock
);
7184 if (curr_cost
> this_rq
->max_idle_balance_cost
)
7185 this_rq
->max_idle_balance_cost
= curr_cost
;
7188 * While browsing the domains, we released the rq lock, a task could
7189 * have been enqueued in the meantime. Since we're not going idle,
7190 * pretend we pulled a task.
7192 if (this_rq
->cfs
.h_nr_running
&& !pulled_task
)
7196 /* Move the next balance forward */
7197 if (time_after(this_rq
->next_balance
, next_balance
))
7198 this_rq
->next_balance
= next_balance
;
7200 /* Is there a task of a high priority class? */
7201 if (this_rq
->nr_running
!= this_rq
->cfs
.h_nr_running
)
7205 idle_exit_fair(this_rq
);
7206 this_rq
->idle_stamp
= 0;
7213 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7214 * running tasks off the busiest CPU onto idle CPUs. It requires at
7215 * least 1 task to be running on each physical CPU where possible, and
7216 * avoids physical / logical imbalances.
7218 static int active_load_balance_cpu_stop(void *data
)
7220 struct rq
*busiest_rq
= data
;
7221 int busiest_cpu
= cpu_of(busiest_rq
);
7222 int target_cpu
= busiest_rq
->push_cpu
;
7223 struct rq
*target_rq
= cpu_rq(target_cpu
);
7224 struct sched_domain
*sd
;
7225 struct task_struct
*p
= NULL
;
7227 raw_spin_lock_irq(&busiest_rq
->lock
);
7229 /* make sure the requested cpu hasn't gone down in the meantime */
7230 if (unlikely(busiest_cpu
!= smp_processor_id() ||
7231 !busiest_rq
->active_balance
))
7234 /* Is there any task to move? */
7235 if (busiest_rq
->nr_running
<= 1)
7239 * This condition is "impossible", if it occurs
7240 * we need to fix it. Originally reported by
7241 * Bjorn Helgaas on a 128-cpu setup.
7243 BUG_ON(busiest_rq
== target_rq
);
7245 /* Search for an sd spanning us and the target CPU. */
7247 for_each_domain(target_cpu
, sd
) {
7248 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
7249 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
7254 struct lb_env env
= {
7256 .dst_cpu
= target_cpu
,
7257 .dst_rq
= target_rq
,
7258 .src_cpu
= busiest_rq
->cpu
,
7259 .src_rq
= busiest_rq
,
7263 schedstat_inc(sd
, alb_count
);
7265 p
= detach_one_task(&env
);
7267 schedstat_inc(sd
, alb_pushed
);
7269 schedstat_inc(sd
, alb_failed
);
7273 busiest_rq
->active_balance
= 0;
7274 raw_spin_unlock(&busiest_rq
->lock
);
7277 attach_one_task(target_rq
, p
);
7284 static inline int on_null_domain(struct rq
*rq
)
7286 return unlikely(!rcu_dereference_sched(rq
->sd
));
7289 #ifdef CONFIG_NO_HZ_COMMON
7291 * idle load balancing details
7292 * - When one of the busy CPUs notice that there may be an idle rebalancing
7293 * needed, they will kick the idle load balancer, which then does idle
7294 * load balancing for all the idle CPUs.
7297 cpumask_var_t idle_cpus_mask
;
7299 unsigned long next_balance
; /* in jiffy units */
7300 } nohz ____cacheline_aligned
;
7302 static inline int find_new_ilb(void)
7304 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
7306 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
7313 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7314 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7315 * CPU (if there is one).
7317 static void nohz_balancer_kick(void)
7321 nohz
.next_balance
++;
7323 ilb_cpu
= find_new_ilb();
7325 if (ilb_cpu
>= nr_cpu_ids
)
7328 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
7331 * Use smp_send_reschedule() instead of resched_cpu().
7332 * This way we generate a sched IPI on the target cpu which
7333 * is idle. And the softirq performing nohz idle load balance
7334 * will be run before returning from the IPI.
7336 smp_send_reschedule(ilb_cpu
);
7340 static inline void nohz_balance_exit_idle(int cpu
)
7342 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
7344 * Completely isolated CPUs don't ever set, so we must test.
7346 if (likely(cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))) {
7347 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
7348 atomic_dec(&nohz
.nr_cpus
);
7350 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
7354 static inline void set_cpu_sd_state_busy(void)
7356 struct sched_domain
*sd
;
7357 int cpu
= smp_processor_id();
7360 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7362 if (!sd
|| !sd
->nohz_idle
)
7366 atomic_inc(&sd
->groups
->sgc
->nr_busy_cpus
);
7371 void set_cpu_sd_state_idle(void)
7373 struct sched_domain
*sd
;
7374 int cpu
= smp_processor_id();
7377 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7379 if (!sd
|| sd
->nohz_idle
)
7383 atomic_dec(&sd
->groups
->sgc
->nr_busy_cpus
);
7389 * This routine will record that the cpu is going idle with tick stopped.
7390 * This info will be used in performing idle load balancing in the future.
7392 void nohz_balance_enter_idle(int cpu
)
7395 * If this cpu is going down, then nothing needs to be done.
7397 if (!cpu_active(cpu
))
7400 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
7404 * If we're a completely isolated CPU, we don't play.
7406 if (on_null_domain(cpu_rq(cpu
)))
7409 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
7410 atomic_inc(&nohz
.nr_cpus
);
7411 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
7414 static int sched_ilb_notifier(struct notifier_block
*nfb
,
7415 unsigned long action
, void *hcpu
)
7417 switch (action
& ~CPU_TASKS_FROZEN
) {
7419 nohz_balance_exit_idle(smp_processor_id());
7427 static DEFINE_SPINLOCK(balancing
);
7430 * Scale the max load_balance interval with the number of CPUs in the system.
7431 * This trades load-balance latency on larger machines for less cross talk.
7433 void update_max_interval(void)
7435 max_load_balance_interval
= HZ
*num_online_cpus()/10;
7439 * It checks each scheduling domain to see if it is due to be balanced,
7440 * and initiates a balancing operation if so.
7442 * Balancing parameters are set up in init_sched_domains.
7444 static void rebalance_domains(struct rq
*rq
, enum cpu_idle_type idle
)
7446 int continue_balancing
= 1;
7448 unsigned long interval
;
7449 struct sched_domain
*sd
;
7450 /* Earliest time when we have to do rebalance again */
7451 unsigned long next_balance
= jiffies
+ 60*HZ
;
7452 int update_next_balance
= 0;
7453 int need_serialize
, need_decay
= 0;
7456 update_blocked_averages(cpu
);
7459 for_each_domain(cpu
, sd
) {
7461 * Decay the newidle max times here because this is a regular
7462 * visit to all the domains. Decay ~1% per second.
7464 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
7465 sd
->max_newidle_lb_cost
=
7466 (sd
->max_newidle_lb_cost
* 253) / 256;
7467 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
7470 max_cost
+= sd
->max_newidle_lb_cost
;
7472 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7476 * Stop the load balance at this level. There is another
7477 * CPU in our sched group which is doing load balancing more
7480 if (!continue_balancing
) {
7486 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7488 need_serialize
= sd
->flags
& SD_SERIALIZE
;
7489 if (need_serialize
) {
7490 if (!spin_trylock(&balancing
))
7494 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
7495 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
7497 * The LBF_DST_PINNED logic could have changed
7498 * env->dst_cpu, so we can't know our idle
7499 * state even if we migrated tasks. Update it.
7501 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
7503 sd
->last_balance
= jiffies
;
7504 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7507 spin_unlock(&balancing
);
7509 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
7510 next_balance
= sd
->last_balance
+ interval
;
7511 update_next_balance
= 1;
7516 * Ensure the rq-wide value also decays but keep it at a
7517 * reasonable floor to avoid funnies with rq->avg_idle.
7519 rq
->max_idle_balance_cost
=
7520 max((u64
)sysctl_sched_migration_cost
, max_cost
);
7525 * next_balance will be updated only when there is a need.
7526 * When the cpu is attached to null domain for ex, it will not be
7529 if (likely(update_next_balance
))
7530 rq
->next_balance
= next_balance
;
7533 #ifdef CONFIG_NO_HZ_COMMON
7535 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7536 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7538 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
7540 int this_cpu
= this_rq
->cpu
;
7544 if (idle
!= CPU_IDLE
||
7545 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
7548 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
7549 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
7553 * If this cpu gets work to do, stop the load balancing
7554 * work being done for other cpus. Next load
7555 * balancing owner will pick it up.
7560 rq
= cpu_rq(balance_cpu
);
7563 * If time for next balance is due,
7566 if (time_after_eq(jiffies
, rq
->next_balance
)) {
7567 raw_spin_lock_irq(&rq
->lock
);
7568 update_rq_clock(rq
);
7569 update_idle_cpu_load(rq
);
7570 raw_spin_unlock_irq(&rq
->lock
);
7571 rebalance_domains(rq
, CPU_IDLE
);
7574 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
7575 this_rq
->next_balance
= rq
->next_balance
;
7577 nohz
.next_balance
= this_rq
->next_balance
;
7579 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
7583 * Current heuristic for kicking the idle load balancer in the presence
7584 * of an idle cpu is the system.
7585 * - This rq has more than one task.
7586 * - At any scheduler domain level, this cpu's scheduler group has multiple
7587 * busy cpu's exceeding the group's capacity.
7588 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7589 * domain span are idle.
7591 static inline int nohz_kick_needed(struct rq
*rq
)
7593 unsigned long now
= jiffies
;
7594 struct sched_domain
*sd
;
7595 struct sched_group_capacity
*sgc
;
7596 int nr_busy
, cpu
= rq
->cpu
;
7598 if (unlikely(rq
->idle_balance
))
7602 * We may be recently in ticked or tickless idle mode. At the first
7603 * busy tick after returning from idle, we will update the busy stats.
7605 set_cpu_sd_state_busy();
7606 nohz_balance_exit_idle(cpu
);
7609 * None are in tickless mode and hence no need for NOHZ idle load
7612 if (likely(!atomic_read(&nohz
.nr_cpus
)))
7615 if (time_before(now
, nohz
.next_balance
))
7618 if (rq
->nr_running
>= 2)
7622 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7625 sgc
= sd
->groups
->sgc
;
7626 nr_busy
= atomic_read(&sgc
->nr_busy_cpus
);
7629 goto need_kick_unlock
;
7632 sd
= rcu_dereference(per_cpu(sd_asym
, cpu
));
7634 if (sd
&& (cpumask_first_and(nohz
.idle_cpus_mask
,
7635 sched_domain_span(sd
)) < cpu
))
7636 goto need_kick_unlock
;
7647 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
) { }
7651 * run_rebalance_domains is triggered when needed from the scheduler tick.
7652 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7654 static void run_rebalance_domains(struct softirq_action
*h
)
7656 struct rq
*this_rq
= this_rq();
7657 enum cpu_idle_type idle
= this_rq
->idle_balance
?
7658 CPU_IDLE
: CPU_NOT_IDLE
;
7660 rebalance_domains(this_rq
, idle
);
7663 * If this cpu has a pending nohz_balance_kick, then do the
7664 * balancing on behalf of the other idle cpus whose ticks are
7667 nohz_idle_balance(this_rq
, idle
);
7671 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7673 void trigger_load_balance(struct rq
*rq
)
7675 /* Don't need to rebalance while attached to NULL domain */
7676 if (unlikely(on_null_domain(rq
)))
7679 if (time_after_eq(jiffies
, rq
->next_balance
))
7680 raise_softirq(SCHED_SOFTIRQ
);
7681 #ifdef CONFIG_NO_HZ_COMMON
7682 if (nohz_kick_needed(rq
))
7683 nohz_balancer_kick();
7687 static void rq_online_fair(struct rq
*rq
)
7691 update_runtime_enabled(rq
);
7694 static void rq_offline_fair(struct rq
*rq
)
7698 /* Ensure any throttled groups are reachable by pick_next_task */
7699 unthrottle_offline_cfs_rqs(rq
);
7702 #endif /* CONFIG_SMP */
7705 * scheduler tick hitting a task of our scheduling class:
7707 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
7709 struct cfs_rq
*cfs_rq
;
7710 struct sched_entity
*se
= &curr
->se
;
7712 for_each_sched_entity(se
) {
7713 cfs_rq
= cfs_rq_of(se
);
7714 entity_tick(cfs_rq
, se
, queued
);
7717 if (numabalancing_enabled
)
7718 task_tick_numa(rq
, curr
);
7720 update_rq_runnable_avg(rq
, 1);
7724 * called on fork with the child task as argument from the parent's context
7725 * - child not yet on the tasklist
7726 * - preemption disabled
7728 static void task_fork_fair(struct task_struct
*p
)
7730 struct cfs_rq
*cfs_rq
;
7731 struct sched_entity
*se
= &p
->se
, *curr
;
7732 int this_cpu
= smp_processor_id();
7733 struct rq
*rq
= this_rq();
7734 unsigned long flags
;
7736 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7738 update_rq_clock(rq
);
7740 cfs_rq
= task_cfs_rq(current
);
7741 curr
= cfs_rq
->curr
;
7744 * Not only the cpu but also the task_group of the parent might have
7745 * been changed after parent->se.parent,cfs_rq were copied to
7746 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7747 * of child point to valid ones.
7750 __set_task_cpu(p
, this_cpu
);
7753 update_curr(cfs_rq
);
7756 se
->vruntime
= curr
->vruntime
;
7757 place_entity(cfs_rq
, se
, 1);
7759 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
7761 * Upon rescheduling, sched_class::put_prev_task() will place
7762 * 'current' within the tree based on its new key value.
7764 swap(curr
->vruntime
, se
->vruntime
);
7768 se
->vruntime
-= cfs_rq
->min_vruntime
;
7770 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7774 * Priority of the task has changed. Check to see if we preempt
7778 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
7780 if (!task_on_rq_queued(p
))
7784 * Reschedule if we are currently running on this runqueue and
7785 * our priority decreased, or if we are not currently running on
7786 * this runqueue and our priority is higher than the current's
7788 if (rq
->curr
== p
) {
7789 if (p
->prio
> oldprio
)
7792 check_preempt_curr(rq
, p
, 0);
7795 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
7797 struct sched_entity
*se
= &p
->se
;
7798 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7801 * Ensure the task's vruntime is normalized, so that when it's
7802 * switched back to the fair class the enqueue_entity(.flags=0) will
7803 * do the right thing.
7805 * If it's queued, then the dequeue_entity(.flags=0) will already
7806 * have normalized the vruntime, if it's !queued, then only when
7807 * the task is sleeping will it still have non-normalized vruntime.
7809 if (!task_on_rq_queued(p
) && p
->state
!= TASK_RUNNING
) {
7811 * Fix up our vruntime so that the current sleep doesn't
7812 * cause 'unlimited' sleep bonus.
7814 place_entity(cfs_rq
, se
, 0);
7815 se
->vruntime
-= cfs_rq
->min_vruntime
;
7820 * Remove our load from contribution when we leave sched_fair
7821 * and ensure we don't carry in an old decay_count if we
7824 if (se
->avg
.decay_count
) {
7825 __synchronize_entity_decay(se
);
7826 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
7832 * We switched to the sched_fair class.
7834 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
7836 #ifdef CONFIG_FAIR_GROUP_SCHED
7837 struct sched_entity
*se
= &p
->se
;
7839 * Since the real-depth could have been changed (only FAIR
7840 * class maintain depth value), reset depth properly.
7842 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
7844 if (!task_on_rq_queued(p
))
7848 * We were most likely switched from sched_rt, so
7849 * kick off the schedule if running, otherwise just see
7850 * if we can still preempt the current task.
7855 check_preempt_curr(rq
, p
, 0);
7858 /* Account for a task changing its policy or group.
7860 * This routine is mostly called to set cfs_rq->curr field when a task
7861 * migrates between groups/classes.
7863 static void set_curr_task_fair(struct rq
*rq
)
7865 struct sched_entity
*se
= &rq
->curr
->se
;
7867 for_each_sched_entity(se
) {
7868 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7870 set_next_entity(cfs_rq
, se
);
7871 /* ensure bandwidth has been allocated on our new cfs_rq */
7872 account_cfs_rq_runtime(cfs_rq
, 0);
7876 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
7878 cfs_rq
->tasks_timeline
= RB_ROOT
;
7879 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7880 #ifndef CONFIG_64BIT
7881 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
7884 atomic64_set(&cfs_rq
->decay_counter
, 1);
7885 atomic_long_set(&cfs_rq
->removed_load
, 0);
7889 #ifdef CONFIG_FAIR_GROUP_SCHED
7890 static void task_move_group_fair(struct task_struct
*p
, int queued
)
7892 struct sched_entity
*se
= &p
->se
;
7893 struct cfs_rq
*cfs_rq
;
7896 * If the task was not on the rq at the time of this cgroup movement
7897 * it must have been asleep, sleeping tasks keep their ->vruntime
7898 * absolute on their old rq until wakeup (needed for the fair sleeper
7899 * bonus in place_entity()).
7901 * If it was on the rq, we've just 'preempted' it, which does convert
7902 * ->vruntime to a relative base.
7904 * Make sure both cases convert their relative position when migrating
7905 * to another cgroup's rq. This does somewhat interfere with the
7906 * fair sleeper stuff for the first placement, but who cares.
7909 * When !queued, vruntime of the task has usually NOT been normalized.
7910 * But there are some cases where it has already been normalized:
7912 * - Moving a forked child which is waiting for being woken up by
7913 * wake_up_new_task().
7914 * - Moving a task which has been woken up by try_to_wake_up() and
7915 * waiting for actually being woken up by sched_ttwu_pending().
7917 * To prevent boost or penalty in the new cfs_rq caused by delta
7918 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7920 if (!queued
&& (!se
->sum_exec_runtime
|| p
->state
== TASK_WAKING
))
7924 se
->vruntime
-= cfs_rq_of(se
)->min_vruntime
;
7925 set_task_rq(p
, task_cpu(p
));
7926 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
7928 cfs_rq
= cfs_rq_of(se
);
7929 se
->vruntime
+= cfs_rq
->min_vruntime
;
7932 * migrate_task_rq_fair() will have removed our previous
7933 * contribution, but we must synchronize for ongoing future
7936 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
7937 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
7942 void free_fair_sched_group(struct task_group
*tg
)
7946 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7948 for_each_possible_cpu(i
) {
7950 kfree(tg
->cfs_rq
[i
]);
7959 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7961 struct cfs_rq
*cfs_rq
;
7962 struct sched_entity
*se
;
7965 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
7968 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
7972 tg
->shares
= NICE_0_LOAD
;
7974 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7976 for_each_possible_cpu(i
) {
7977 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
7978 GFP_KERNEL
, cpu_to_node(i
));
7982 se
= kzalloc_node(sizeof(struct sched_entity
),
7983 GFP_KERNEL
, cpu_to_node(i
));
7987 init_cfs_rq(cfs_rq
);
7988 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
7999 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8001 struct rq
*rq
= cpu_rq(cpu
);
8002 unsigned long flags
;
8005 * Only empty task groups can be destroyed; so we can speculatively
8006 * check on_list without danger of it being re-added.
8008 if (!tg
->cfs_rq
[cpu
]->on_list
)
8011 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8012 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
8013 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8016 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8017 struct sched_entity
*se
, int cpu
,
8018 struct sched_entity
*parent
)
8020 struct rq
*rq
= cpu_rq(cpu
);
8024 init_cfs_rq_runtime(cfs_rq
);
8026 tg
->cfs_rq
[cpu
] = cfs_rq
;
8029 /* se could be NULL for root_task_group */
8034 se
->cfs_rq
= &rq
->cfs
;
8037 se
->cfs_rq
= parent
->my_q
;
8038 se
->depth
= parent
->depth
+ 1;
8042 /* guarantee group entities always have weight */
8043 update_load_set(&se
->load
, NICE_0_LOAD
);
8044 se
->parent
= parent
;
8047 static DEFINE_MUTEX(shares_mutex
);
8049 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8052 unsigned long flags
;
8055 * We can't change the weight of the root cgroup.
8060 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
8062 mutex_lock(&shares_mutex
);
8063 if (tg
->shares
== shares
)
8066 tg
->shares
= shares
;
8067 for_each_possible_cpu(i
) {
8068 struct rq
*rq
= cpu_rq(i
);
8069 struct sched_entity
*se
;
8072 /* Propagate contribution to hierarchy */
8073 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8075 /* Possible calls to update_curr() need rq clock */
8076 update_rq_clock(rq
);
8077 for_each_sched_entity(se
)
8078 update_cfs_shares(group_cfs_rq(se
));
8079 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8083 mutex_unlock(&shares_mutex
);
8086 #else /* CONFIG_FAIR_GROUP_SCHED */
8088 void free_fair_sched_group(struct task_group
*tg
) { }
8090 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8095 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
) { }
8097 #endif /* CONFIG_FAIR_GROUP_SCHED */
8100 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
8102 struct sched_entity
*se
= &task
->se
;
8103 unsigned int rr_interval
= 0;
8106 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8109 if (rq
->cfs
.load
.weight
)
8110 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
8116 * All the scheduling class methods:
8118 const struct sched_class fair_sched_class
= {
8119 .next
= &idle_sched_class
,
8120 .enqueue_task
= enqueue_task_fair
,
8121 .dequeue_task
= dequeue_task_fair
,
8122 .yield_task
= yield_task_fair
,
8123 .yield_to_task
= yield_to_task_fair
,
8125 .check_preempt_curr
= check_preempt_wakeup
,
8127 .pick_next_task
= pick_next_task_fair
,
8128 .put_prev_task
= put_prev_task_fair
,
8131 .select_task_rq
= select_task_rq_fair
,
8132 .migrate_task_rq
= migrate_task_rq_fair
,
8134 .rq_online
= rq_online_fair
,
8135 .rq_offline
= rq_offline_fair
,
8137 .task_waking
= task_waking_fair
,
8140 .set_curr_task
= set_curr_task_fair
,
8141 .task_tick
= task_tick_fair
,
8142 .task_fork
= task_fork_fair
,
8144 .prio_changed
= prio_changed_fair
,
8145 .switched_from
= switched_from_fair
,
8146 .switched_to
= switched_to_fair
,
8148 .get_rr_interval
= get_rr_interval_fair
,
8150 .update_curr
= update_curr_fair
,
8152 #ifdef CONFIG_FAIR_GROUP_SCHED
8153 .task_move_group
= task_move_group_fair
,
8157 #ifdef CONFIG_SCHED_DEBUG
8158 void print_cfs_stats(struct seq_file
*m
, int cpu
)
8160 struct cfs_rq
*cfs_rq
;
8163 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
8164 print_cfs_rq(m
, cpu
, cfs_rq
);
8169 __init
void init_sched_fair_class(void)
8172 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8174 #ifdef CONFIG_NO_HZ_COMMON
8175 nohz
.next_balance
= jiffies
;
8176 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
8177 cpu_notifier(sched_ilb_notifier
, 0);