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
23 #include <linux/sched.h>
24 #include <linux/latencytop.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 unsigned int get_update_sysctl_factor(void)
146 unsigned int cpus
= min_t(unsigned 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 inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
288 if (!cfs_rq
->on_list
) {
290 * Ensure we either appear before our parent (if already
291 * enqueued) or force our parent to appear after us when it is
292 * enqueued. The fact that we always enqueue bottom-up
293 * reduces this to two cases.
295 if (cfs_rq
->tg
->parent
&&
296 cfs_rq
->tg
->parent
->cfs_rq
[cpu_of(rq_of(cfs_rq
))]->on_list
) {
297 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
,
298 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
300 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
301 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
308 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
310 if (cfs_rq
->on_list
) {
311 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
316 /* Iterate thr' all leaf cfs_rq's on a runqueue */
317 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
318 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
320 /* Do the two (enqueued) entities belong to the same group ? */
321 static inline struct cfs_rq
*
322 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
324 if (se
->cfs_rq
== pse
->cfs_rq
)
330 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
336 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
338 int se_depth
, pse_depth
;
341 * preemption test can be made between sibling entities who are in the
342 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
343 * both tasks until we find their ancestors who are siblings of common
347 /* First walk up until both entities are at same depth */
348 se_depth
= (*se
)->depth
;
349 pse_depth
= (*pse
)->depth
;
351 while (se_depth
> pse_depth
) {
353 *se
= parent_entity(*se
);
356 while (pse_depth
> se_depth
) {
358 *pse
= parent_entity(*pse
);
361 while (!is_same_group(*se
, *pse
)) {
362 *se
= parent_entity(*se
);
363 *pse
= parent_entity(*pse
);
367 #else /* !CONFIG_FAIR_GROUP_SCHED */
369 static inline struct task_struct
*task_of(struct sched_entity
*se
)
371 return container_of(se
, struct task_struct
, se
);
374 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
376 return container_of(cfs_rq
, struct rq
, cfs
);
379 #define entity_is_task(se) 1
381 #define for_each_sched_entity(se) \
382 for (; se; se = NULL)
384 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
386 return &task_rq(p
)->cfs
;
389 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
391 struct task_struct
*p
= task_of(se
);
392 struct rq
*rq
= task_rq(p
);
397 /* runqueue "owned" by this group */
398 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
403 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
407 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
411 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
412 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
414 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
420 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
424 #endif /* CONFIG_FAIR_GROUP_SCHED */
426 static __always_inline
427 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
);
429 /**************************************************************
430 * Scheduling class tree data structure manipulation methods:
433 static inline u64
max_vruntime(u64 max_vruntime
, u64 vruntime
)
435 s64 delta
= (s64
)(vruntime
- max_vruntime
);
437 max_vruntime
= vruntime
;
442 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
444 s64 delta
= (s64
)(vruntime
- min_vruntime
);
446 min_vruntime
= vruntime
;
451 static inline int entity_before(struct sched_entity
*a
,
452 struct sched_entity
*b
)
454 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
457 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
459 u64 vruntime
= cfs_rq
->min_vruntime
;
462 vruntime
= cfs_rq
->curr
->vruntime
;
464 if (cfs_rq
->rb_leftmost
) {
465 struct sched_entity
*se
= rb_entry(cfs_rq
->rb_leftmost
,
470 vruntime
= se
->vruntime
;
472 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
475 /* ensure we never gain time by being placed backwards. */
476 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
479 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
484 * Enqueue an entity into the rb-tree:
486 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
488 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
489 struct rb_node
*parent
= NULL
;
490 struct sched_entity
*entry
;
494 * Find the right place in the rbtree:
498 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
500 * We dont care about collisions. Nodes with
501 * the same key stay together.
503 if (entity_before(se
, entry
)) {
504 link
= &parent
->rb_left
;
506 link
= &parent
->rb_right
;
512 * Maintain a cache of leftmost tree entries (it is frequently
516 cfs_rq
->rb_leftmost
= &se
->run_node
;
518 rb_link_node(&se
->run_node
, parent
, link
);
519 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
522 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
524 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
525 struct rb_node
*next_node
;
527 next_node
= rb_next(&se
->run_node
);
528 cfs_rq
->rb_leftmost
= next_node
;
531 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
534 struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
)
536 struct rb_node
*left
= cfs_rq
->rb_leftmost
;
541 return rb_entry(left
, struct sched_entity
, run_node
);
544 static struct sched_entity
*__pick_next_entity(struct sched_entity
*se
)
546 struct rb_node
*next
= rb_next(&se
->run_node
);
551 return rb_entry(next
, struct sched_entity
, run_node
);
554 #ifdef CONFIG_SCHED_DEBUG
555 struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
557 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
562 return rb_entry(last
, struct sched_entity
, run_node
);
565 /**************************************************************
566 * Scheduling class statistics methods:
569 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
570 void __user
*buffer
, size_t *lenp
,
573 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
574 unsigned int factor
= get_update_sysctl_factor();
579 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
580 sysctl_sched_min_granularity
);
582 #define WRT_SYSCTL(name) \
583 (normalized_sysctl_##name = sysctl_##name / (factor))
584 WRT_SYSCTL(sched_min_granularity
);
585 WRT_SYSCTL(sched_latency
);
586 WRT_SYSCTL(sched_wakeup_granularity
);
596 static inline u64
calc_delta_fair(u64 delta
, struct sched_entity
*se
)
598 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
599 delta
= __calc_delta(delta
, NICE_0_LOAD
, &se
->load
);
605 * The idea is to set a period in which each task runs once.
607 * When there are too many tasks (sched_nr_latency) we have to stretch
608 * this period because otherwise the slices get too small.
610 * p = (nr <= nl) ? l : l*nr/nl
612 static u64
__sched_period(unsigned long nr_running
)
614 if (unlikely(nr_running
> sched_nr_latency
))
615 return nr_running
* sysctl_sched_min_granularity
;
617 return sysctl_sched_latency
;
621 * We calculate the wall-time slice from the period by taking a part
622 * proportional to the weight.
626 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
628 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
630 for_each_sched_entity(se
) {
631 struct load_weight
*load
;
632 struct load_weight lw
;
634 cfs_rq
= cfs_rq_of(se
);
635 load
= &cfs_rq
->load
;
637 if (unlikely(!se
->on_rq
)) {
640 update_load_add(&lw
, se
->load
.weight
);
643 slice
= __calc_delta(slice
, se
->load
.weight
, load
);
649 * We calculate the vruntime slice of a to-be-inserted task.
653 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
655 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
659 static int select_idle_sibling(struct task_struct
*p
, int cpu
);
660 static unsigned long task_h_load(struct task_struct
*p
);
663 * We choose a half-life close to 1 scheduling period.
664 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
665 * dependent on this value.
667 #define LOAD_AVG_PERIOD 32
668 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
669 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
671 /* Give new sched_entity start runnable values to heavy its load in infant time */
672 void init_entity_runnable_average(struct sched_entity
*se
)
674 struct sched_avg
*sa
= &se
->avg
;
676 sa
->last_update_time
= 0;
678 * sched_avg's period_contrib should be strictly less then 1024, so
679 * we give it 1023 to make sure it is almost a period (1024us), and
680 * will definitely be update (after enqueue).
682 sa
->period_contrib
= 1023;
683 sa
->load_avg
= scale_load_down(se
->load
.weight
);
684 sa
->load_sum
= sa
->load_avg
* LOAD_AVG_MAX
;
685 sa
->util_avg
= scale_load_down(SCHED_LOAD_SCALE
);
686 sa
->util_sum
= sa
->util_avg
* LOAD_AVG_MAX
;
687 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
690 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq
*cfs_rq
);
691 static inline unsigned long cfs_rq_load_avg(struct cfs_rq
*cfs_rq
);
693 void init_entity_runnable_average(struct sched_entity
*se
)
699 * Update the current task's runtime statistics.
701 static void update_curr(struct cfs_rq
*cfs_rq
)
703 struct sched_entity
*curr
= cfs_rq
->curr
;
704 u64 now
= rq_clock_task(rq_of(cfs_rq
));
710 delta_exec
= now
- curr
->exec_start
;
711 if (unlikely((s64
)delta_exec
<= 0))
714 curr
->exec_start
= now
;
716 schedstat_set(curr
->statistics
.exec_max
,
717 max(delta_exec
, curr
->statistics
.exec_max
));
719 curr
->sum_exec_runtime
+= delta_exec
;
720 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
722 curr
->vruntime
+= calc_delta_fair(delta_exec
, curr
);
723 update_min_vruntime(cfs_rq
);
725 if (entity_is_task(curr
)) {
726 struct task_struct
*curtask
= task_of(curr
);
728 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
729 cpuacct_charge(curtask
, delta_exec
);
730 account_group_exec_runtime(curtask
, delta_exec
);
733 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
736 static void update_curr_fair(struct rq
*rq
)
738 update_curr(cfs_rq_of(&rq
->curr
->se
));
741 #ifdef CONFIG_SCHEDSTATS
743 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
745 u64 wait_start
= rq_clock(rq_of(cfs_rq
));
747 if (entity_is_task(se
) && task_on_rq_migrating(task_of(se
)) &&
748 likely(wait_start
> se
->statistics
.wait_start
))
749 wait_start
-= se
->statistics
.wait_start
;
751 se
->statistics
.wait_start
= wait_start
;
755 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
757 struct task_struct
*p
;
760 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
;
762 if (entity_is_task(se
)) {
764 if (task_on_rq_migrating(p
)) {
766 * Preserve migrating task's wait time so wait_start
767 * time stamp can be adjusted to accumulate wait time
768 * prior to migration.
770 se
->statistics
.wait_start
= delta
;
773 trace_sched_stat_wait(p
, delta
);
776 se
->statistics
.wait_max
= max(se
->statistics
.wait_max
, delta
);
777 se
->statistics
.wait_count
++;
778 se
->statistics
.wait_sum
+= delta
;
779 se
->statistics
.wait_start
= 0;
783 * Task is being enqueued - update stats:
786 update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
789 * Are we enqueueing a waiting task? (for current tasks
790 * a dequeue/enqueue event is a NOP)
792 if (se
!= cfs_rq
->curr
)
793 update_stats_wait_start(cfs_rq
, se
);
797 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
800 * Mark the end of the wait period if dequeueing a
803 if (se
!= cfs_rq
->curr
)
804 update_stats_wait_end(cfs_rq
, se
);
806 if (flags
& DEQUEUE_SLEEP
) {
807 if (entity_is_task(se
)) {
808 struct task_struct
*tsk
= task_of(se
);
810 if (tsk
->state
& TASK_INTERRUPTIBLE
)
811 se
->statistics
.sleep_start
= rq_clock(rq_of(cfs_rq
));
812 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
813 se
->statistics
.block_start
= rq_clock(rq_of(cfs_rq
));
820 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
825 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
830 update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
835 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
841 * We are picking a new current task - update its stats:
844 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
847 * We are starting a new run period:
849 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
852 /**************************************************
853 * Scheduling class queueing methods:
856 #ifdef CONFIG_NUMA_BALANCING
858 * Approximate time to scan a full NUMA task in ms. The task scan period is
859 * calculated based on the tasks virtual memory size and
860 * numa_balancing_scan_size.
862 unsigned int sysctl_numa_balancing_scan_period_min
= 1000;
863 unsigned int sysctl_numa_balancing_scan_period_max
= 60000;
865 /* Portion of address space to scan in MB */
866 unsigned int sysctl_numa_balancing_scan_size
= 256;
868 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
869 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
871 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
873 unsigned long rss
= 0;
874 unsigned long nr_scan_pages
;
877 * Calculations based on RSS as non-present and empty pages are skipped
878 * by the PTE scanner and NUMA hinting faults should be trapped based
881 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
882 rss
= get_mm_rss(p
->mm
);
886 rss
= round_up(rss
, nr_scan_pages
);
887 return rss
/ nr_scan_pages
;
890 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
891 #define MAX_SCAN_WINDOW 2560
893 static unsigned int task_scan_min(struct task_struct
*p
)
895 unsigned int scan_size
= READ_ONCE(sysctl_numa_balancing_scan_size
);
896 unsigned int scan
, floor
;
897 unsigned int windows
= 1;
899 if (scan_size
< MAX_SCAN_WINDOW
)
900 windows
= MAX_SCAN_WINDOW
/ scan_size
;
901 floor
= 1000 / windows
;
903 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
904 return max_t(unsigned int, floor
, scan
);
907 static unsigned int task_scan_max(struct task_struct
*p
)
909 unsigned int smin
= task_scan_min(p
);
912 /* Watch for min being lower than max due to floor calculations */
913 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
914 return max(smin
, smax
);
917 static void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
919 rq
->nr_numa_running
+= (p
->numa_preferred_nid
!= -1);
920 rq
->nr_preferred_running
+= (p
->numa_preferred_nid
== task_node(p
));
923 static void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
925 rq
->nr_numa_running
-= (p
->numa_preferred_nid
!= -1);
926 rq
->nr_preferred_running
-= (p
->numa_preferred_nid
== task_node(p
));
932 spinlock_t lock
; /* nr_tasks, tasks */
938 unsigned long total_faults
;
939 unsigned long max_faults_cpu
;
941 * Faults_cpu is used to decide whether memory should move
942 * towards the CPU. As a consequence, these stats are weighted
943 * more by CPU use than by memory faults.
945 unsigned long *faults_cpu
;
946 unsigned long faults
[0];
949 /* Shared or private faults. */
950 #define NR_NUMA_HINT_FAULT_TYPES 2
952 /* Memory and CPU locality */
953 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
955 /* Averaged statistics, and temporary buffers. */
956 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
958 pid_t
task_numa_group_id(struct task_struct
*p
)
960 return p
->numa_group
? p
->numa_group
->gid
: 0;
964 * The averaged statistics, shared & private, memory & cpu,
965 * occupy the first half of the array. The second half of the
966 * array is for current counters, which are averaged into the
967 * first set by task_numa_placement.
969 static inline int task_faults_idx(enum numa_faults_stats s
, int nid
, int priv
)
971 return NR_NUMA_HINT_FAULT_TYPES
* (s
* nr_node_ids
+ nid
) + priv
;
974 static inline unsigned long task_faults(struct task_struct
*p
, int nid
)
979 return p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
980 p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
983 static inline unsigned long group_faults(struct task_struct
*p
, int nid
)
988 return p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
989 p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
992 static inline unsigned long group_faults_cpu(struct numa_group
*group
, int nid
)
994 return group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
995 group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 1)];
999 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1000 * considered part of a numa group's pseudo-interleaving set. Migrations
1001 * between these nodes are slowed down, to allow things to settle down.
1003 #define ACTIVE_NODE_FRACTION 3
1005 static bool numa_is_active_node(int nid
, struct numa_group
*ng
)
1007 return group_faults_cpu(ng
, nid
) * ACTIVE_NODE_FRACTION
> ng
->max_faults_cpu
;
1010 /* Handle placement on systems where not all nodes are directly connected. */
1011 static unsigned long score_nearby_nodes(struct task_struct
*p
, int nid
,
1012 int maxdist
, bool task
)
1014 unsigned long score
= 0;
1018 * All nodes are directly connected, and the same distance
1019 * from each other. No need for fancy placement algorithms.
1021 if (sched_numa_topology_type
== NUMA_DIRECT
)
1025 * This code is called for each node, introducing N^2 complexity,
1026 * which should be ok given the number of nodes rarely exceeds 8.
1028 for_each_online_node(node
) {
1029 unsigned long faults
;
1030 int dist
= node_distance(nid
, node
);
1033 * The furthest away nodes in the system are not interesting
1034 * for placement; nid was already counted.
1036 if (dist
== sched_max_numa_distance
|| node
== nid
)
1040 * On systems with a backplane NUMA topology, compare groups
1041 * of nodes, and move tasks towards the group with the most
1042 * memory accesses. When comparing two nodes at distance
1043 * "hoplimit", only nodes closer by than "hoplimit" are part
1044 * of each group. Skip other nodes.
1046 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
1050 /* Add up the faults from nearby nodes. */
1052 faults
= task_faults(p
, node
);
1054 faults
= group_faults(p
, node
);
1057 * On systems with a glueless mesh NUMA topology, there are
1058 * no fixed "groups of nodes". Instead, nodes that are not
1059 * directly connected bounce traffic through intermediate
1060 * nodes; a numa_group can occupy any set of nodes.
1061 * The further away a node is, the less the faults count.
1062 * This seems to result in good task placement.
1064 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
1065 faults
*= (sched_max_numa_distance
- dist
);
1066 faults
/= (sched_max_numa_distance
- LOCAL_DISTANCE
);
1076 * These return the fraction of accesses done by a particular task, or
1077 * task group, on a particular numa node. The group weight is given a
1078 * larger multiplier, in order to group tasks together that are almost
1079 * evenly spread out between numa nodes.
1081 static inline unsigned long task_weight(struct task_struct
*p
, int nid
,
1084 unsigned long faults
, total_faults
;
1086 if (!p
->numa_faults
)
1089 total_faults
= p
->total_numa_faults
;
1094 faults
= task_faults(p
, nid
);
1095 faults
+= score_nearby_nodes(p
, nid
, dist
, true);
1097 return 1000 * faults
/ total_faults
;
1100 static inline unsigned long group_weight(struct task_struct
*p
, int nid
,
1103 unsigned long faults
, total_faults
;
1108 total_faults
= p
->numa_group
->total_faults
;
1113 faults
= group_faults(p
, nid
);
1114 faults
+= score_nearby_nodes(p
, nid
, dist
, false);
1116 return 1000 * faults
/ total_faults
;
1119 bool should_numa_migrate_memory(struct task_struct
*p
, struct page
* page
,
1120 int src_nid
, int dst_cpu
)
1122 struct numa_group
*ng
= p
->numa_group
;
1123 int dst_nid
= cpu_to_node(dst_cpu
);
1124 int last_cpupid
, this_cpupid
;
1126 this_cpupid
= cpu_pid_to_cpupid(dst_cpu
, current
->pid
);
1129 * Multi-stage node selection is used in conjunction with a periodic
1130 * migration fault to build a temporal task<->page relation. By using
1131 * a two-stage filter we remove short/unlikely relations.
1133 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1134 * a task's usage of a particular page (n_p) per total usage of this
1135 * page (n_t) (in a given time-span) to a probability.
1137 * Our periodic faults will sample this probability and getting the
1138 * same result twice in a row, given these samples are fully
1139 * independent, is then given by P(n)^2, provided our sample period
1140 * is sufficiently short compared to the usage pattern.
1142 * This quadric squishes small probabilities, making it less likely we
1143 * act on an unlikely task<->page relation.
1145 last_cpupid
= page_cpupid_xchg_last(page
, this_cpupid
);
1146 if (!cpupid_pid_unset(last_cpupid
) &&
1147 cpupid_to_nid(last_cpupid
) != dst_nid
)
1150 /* Always allow migrate on private faults */
1151 if (cpupid_match_pid(p
, last_cpupid
))
1154 /* A shared fault, but p->numa_group has not been set up yet. */
1159 * Destination node is much more heavily used than the source
1160 * node? Allow migration.
1162 if (group_faults_cpu(ng
, dst_nid
) > group_faults_cpu(ng
, src_nid
) *
1163 ACTIVE_NODE_FRACTION
)
1167 * Distribute memory according to CPU & memory use on each node,
1168 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1170 * faults_cpu(dst) 3 faults_cpu(src)
1171 * --------------- * - > ---------------
1172 * faults_mem(dst) 4 faults_mem(src)
1174 return group_faults_cpu(ng
, dst_nid
) * group_faults(p
, src_nid
) * 3 >
1175 group_faults_cpu(ng
, src_nid
) * group_faults(p
, dst_nid
) * 4;
1178 static unsigned long weighted_cpuload(const int cpu
);
1179 static unsigned long source_load(int cpu
, int type
);
1180 static unsigned long target_load(int cpu
, int type
);
1181 static unsigned long capacity_of(int cpu
);
1182 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
);
1184 /* Cached statistics for all CPUs within a node */
1186 unsigned long nr_running
;
1189 /* Total compute capacity of CPUs on a node */
1190 unsigned long compute_capacity
;
1192 /* Approximate capacity in terms of runnable tasks on a node */
1193 unsigned long task_capacity
;
1194 int has_free_capacity
;
1198 * XXX borrowed from update_sg_lb_stats
1200 static void update_numa_stats(struct numa_stats
*ns
, int nid
)
1202 int smt
, cpu
, cpus
= 0;
1203 unsigned long capacity
;
1205 memset(ns
, 0, sizeof(*ns
));
1206 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
1207 struct rq
*rq
= cpu_rq(cpu
);
1209 ns
->nr_running
+= rq
->nr_running
;
1210 ns
->load
+= weighted_cpuload(cpu
);
1211 ns
->compute_capacity
+= capacity_of(cpu
);
1217 * If we raced with hotplug and there are no CPUs left in our mask
1218 * the @ns structure is NULL'ed and task_numa_compare() will
1219 * not find this node attractive.
1221 * We'll either bail at !has_free_capacity, or we'll detect a huge
1222 * imbalance and bail there.
1227 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1228 smt
= DIV_ROUND_UP(SCHED_CAPACITY_SCALE
* cpus
, ns
->compute_capacity
);
1229 capacity
= cpus
/ smt
; /* cores */
1231 ns
->task_capacity
= min_t(unsigned, capacity
,
1232 DIV_ROUND_CLOSEST(ns
->compute_capacity
, SCHED_CAPACITY_SCALE
));
1233 ns
->has_free_capacity
= (ns
->nr_running
< ns
->task_capacity
);
1236 struct task_numa_env
{
1237 struct task_struct
*p
;
1239 int src_cpu
, src_nid
;
1240 int dst_cpu
, dst_nid
;
1242 struct numa_stats src_stats
, dst_stats
;
1247 struct task_struct
*best_task
;
1252 static void task_numa_assign(struct task_numa_env
*env
,
1253 struct task_struct
*p
, long imp
)
1256 put_task_struct(env
->best_task
);
1259 env
->best_imp
= imp
;
1260 env
->best_cpu
= env
->dst_cpu
;
1263 static bool load_too_imbalanced(long src_load
, long dst_load
,
1264 struct task_numa_env
*env
)
1267 long orig_src_load
, orig_dst_load
;
1268 long src_capacity
, dst_capacity
;
1271 * The load is corrected for the CPU capacity available on each node.
1274 * ------------ vs ---------
1275 * src_capacity dst_capacity
1277 src_capacity
= env
->src_stats
.compute_capacity
;
1278 dst_capacity
= env
->dst_stats
.compute_capacity
;
1280 /* We care about the slope of the imbalance, not the direction. */
1281 if (dst_load
< src_load
)
1282 swap(dst_load
, src_load
);
1284 /* Is the difference below the threshold? */
1285 imb
= dst_load
* src_capacity
* 100 -
1286 src_load
* dst_capacity
* env
->imbalance_pct
;
1291 * The imbalance is above the allowed threshold.
1292 * Compare it with the old imbalance.
1294 orig_src_load
= env
->src_stats
.load
;
1295 orig_dst_load
= env
->dst_stats
.load
;
1297 if (orig_dst_load
< orig_src_load
)
1298 swap(orig_dst_load
, orig_src_load
);
1300 old_imb
= orig_dst_load
* src_capacity
* 100 -
1301 orig_src_load
* dst_capacity
* env
->imbalance_pct
;
1303 /* Would this change make things worse? */
1304 return (imb
> old_imb
);
1308 * This checks if the overall compute and NUMA accesses of the system would
1309 * be improved if the source tasks was migrated to the target dst_cpu taking
1310 * into account that it might be best if task running on the dst_cpu should
1311 * be exchanged with the source task
1313 static void task_numa_compare(struct task_numa_env
*env
,
1314 long taskimp
, long groupimp
)
1316 struct rq
*src_rq
= cpu_rq(env
->src_cpu
);
1317 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1318 struct task_struct
*cur
;
1319 long src_load
, dst_load
;
1321 long imp
= env
->p
->numa_group
? groupimp
: taskimp
;
1323 int dist
= env
->dist
;
1324 bool assigned
= false;
1328 raw_spin_lock_irq(&dst_rq
->lock
);
1331 * No need to move the exiting task or idle task.
1333 if ((cur
->flags
& PF_EXITING
) || is_idle_task(cur
))
1337 * The task_struct must be protected here to protect the
1338 * p->numa_faults access in the task_weight since the
1339 * numa_faults could already be freed in the following path:
1340 * finish_task_switch()
1341 * --> put_task_struct()
1342 * --> __put_task_struct()
1343 * --> task_numa_free()
1345 get_task_struct(cur
);
1348 raw_spin_unlock_irq(&dst_rq
->lock
);
1351 * Because we have preemption enabled we can get migrated around and
1352 * end try selecting ourselves (current == env->p) as a swap candidate.
1358 * "imp" is the fault differential for the source task between the
1359 * source and destination node. Calculate the total differential for
1360 * the source task and potential destination task. The more negative
1361 * the value is, the more rmeote accesses that would be expected to
1362 * be incurred if the tasks were swapped.
1365 /* Skip this swap candidate if cannot move to the source cpu */
1366 if (!cpumask_test_cpu(env
->src_cpu
, tsk_cpus_allowed(cur
)))
1370 * If dst and source tasks are in the same NUMA group, or not
1371 * in any group then look only at task weights.
1373 if (cur
->numa_group
== env
->p
->numa_group
) {
1374 imp
= taskimp
+ task_weight(cur
, env
->src_nid
, dist
) -
1375 task_weight(cur
, env
->dst_nid
, dist
);
1377 * Add some hysteresis to prevent swapping the
1378 * tasks within a group over tiny differences.
1380 if (cur
->numa_group
)
1384 * Compare the group weights. If a task is all by
1385 * itself (not part of a group), use the task weight
1388 if (cur
->numa_group
)
1389 imp
+= group_weight(cur
, env
->src_nid
, dist
) -
1390 group_weight(cur
, env
->dst_nid
, dist
);
1392 imp
+= task_weight(cur
, env
->src_nid
, dist
) -
1393 task_weight(cur
, env
->dst_nid
, dist
);
1397 if (imp
<= env
->best_imp
&& moveimp
<= env
->best_imp
)
1401 /* Is there capacity at our destination? */
1402 if (env
->src_stats
.nr_running
<= env
->src_stats
.task_capacity
&&
1403 !env
->dst_stats
.has_free_capacity
)
1409 /* Balance doesn't matter much if we're running a task per cpu */
1410 if (imp
> env
->best_imp
&& src_rq
->nr_running
== 1 &&
1411 dst_rq
->nr_running
== 1)
1415 * In the overloaded case, try and keep the load balanced.
1418 load
= task_h_load(env
->p
);
1419 dst_load
= env
->dst_stats
.load
+ load
;
1420 src_load
= env
->src_stats
.load
- load
;
1422 if (moveimp
> imp
&& moveimp
> env
->best_imp
) {
1424 * If the improvement from just moving env->p direction is
1425 * better than swapping tasks around, check if a move is
1426 * possible. Store a slightly smaller score than moveimp,
1427 * so an actually idle CPU will win.
1429 if (!load_too_imbalanced(src_load
, dst_load
, env
)) {
1431 put_task_struct(cur
);
1437 if (imp
<= env
->best_imp
)
1441 load
= task_h_load(cur
);
1446 if (load_too_imbalanced(src_load
, dst_load
, env
))
1450 * One idle CPU per node is evaluated for a task numa move.
1451 * Call select_idle_sibling to maybe find a better one.
1454 env
->dst_cpu
= select_idle_sibling(env
->p
, env
->dst_cpu
);
1458 task_numa_assign(env
, cur
, imp
);
1462 * The dst_rq->curr isn't assigned. The protection for task_struct is
1465 if (cur
&& !assigned
)
1466 put_task_struct(cur
);
1469 static void task_numa_find_cpu(struct task_numa_env
*env
,
1470 long taskimp
, long groupimp
)
1474 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1475 /* Skip this CPU if the source task cannot migrate */
1476 if (!cpumask_test_cpu(cpu
, tsk_cpus_allowed(env
->p
)))
1480 task_numa_compare(env
, taskimp
, groupimp
);
1484 /* Only move tasks to a NUMA node less busy than the current node. */
1485 static bool numa_has_capacity(struct task_numa_env
*env
)
1487 struct numa_stats
*src
= &env
->src_stats
;
1488 struct numa_stats
*dst
= &env
->dst_stats
;
1490 if (src
->has_free_capacity
&& !dst
->has_free_capacity
)
1494 * Only consider a task move if the source has a higher load
1495 * than the destination, corrected for CPU capacity on each node.
1497 * src->load dst->load
1498 * --------------------- vs ---------------------
1499 * src->compute_capacity dst->compute_capacity
1501 if (src
->load
* dst
->compute_capacity
* env
->imbalance_pct
>
1503 dst
->load
* src
->compute_capacity
* 100)
1509 static int task_numa_migrate(struct task_struct
*p
)
1511 struct task_numa_env env
= {
1514 .src_cpu
= task_cpu(p
),
1515 .src_nid
= task_node(p
),
1517 .imbalance_pct
= 112,
1523 struct sched_domain
*sd
;
1524 unsigned long taskweight
, groupweight
;
1526 long taskimp
, groupimp
;
1529 * Pick the lowest SD_NUMA domain, as that would have the smallest
1530 * imbalance and would be the first to start moving tasks about.
1532 * And we want to avoid any moving of tasks about, as that would create
1533 * random movement of tasks -- counter the numa conditions we're trying
1537 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1539 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1543 * Cpusets can break the scheduler domain tree into smaller
1544 * balance domains, some of which do not cross NUMA boundaries.
1545 * Tasks that are "trapped" in such domains cannot be migrated
1546 * elsewhere, so there is no point in (re)trying.
1548 if (unlikely(!sd
)) {
1549 p
->numa_preferred_nid
= task_node(p
);
1553 env
.dst_nid
= p
->numa_preferred_nid
;
1554 dist
= env
.dist
= node_distance(env
.src_nid
, env
.dst_nid
);
1555 taskweight
= task_weight(p
, env
.src_nid
, dist
);
1556 groupweight
= group_weight(p
, env
.src_nid
, dist
);
1557 update_numa_stats(&env
.src_stats
, env
.src_nid
);
1558 taskimp
= task_weight(p
, env
.dst_nid
, dist
) - taskweight
;
1559 groupimp
= group_weight(p
, env
.dst_nid
, dist
) - groupweight
;
1560 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1562 /* Try to find a spot on the preferred nid. */
1563 if (numa_has_capacity(&env
))
1564 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1567 * Look at other nodes in these cases:
1568 * - there is no space available on the preferred_nid
1569 * - the task is part of a numa_group that is interleaved across
1570 * multiple NUMA nodes; in order to better consolidate the group,
1571 * we need to check other locations.
1573 if (env
.best_cpu
== -1 || (p
->numa_group
&& p
->numa_group
->active_nodes
> 1)) {
1574 for_each_online_node(nid
) {
1575 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
1578 dist
= node_distance(env
.src_nid
, env
.dst_nid
);
1579 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
1581 taskweight
= task_weight(p
, env
.src_nid
, dist
);
1582 groupweight
= group_weight(p
, env
.src_nid
, dist
);
1585 /* Only consider nodes where both task and groups benefit */
1586 taskimp
= task_weight(p
, nid
, dist
) - taskweight
;
1587 groupimp
= group_weight(p
, nid
, dist
) - groupweight
;
1588 if (taskimp
< 0 && groupimp
< 0)
1593 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1594 if (numa_has_capacity(&env
))
1595 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1600 * If the task is part of a workload that spans multiple NUMA nodes,
1601 * and is migrating into one of the workload's active nodes, remember
1602 * this node as the task's preferred numa node, so the workload can
1604 * A task that migrated to a second choice node will be better off
1605 * trying for a better one later. Do not set the preferred node here.
1607 if (p
->numa_group
) {
1608 struct numa_group
*ng
= p
->numa_group
;
1610 if (env
.best_cpu
== -1)
1615 if (ng
->active_nodes
> 1 && numa_is_active_node(env
.dst_nid
, ng
))
1616 sched_setnuma(p
, env
.dst_nid
);
1619 /* No better CPU than the current one was found. */
1620 if (env
.best_cpu
== -1)
1624 * Reset the scan period if the task is being rescheduled on an
1625 * alternative node to recheck if the tasks is now properly placed.
1627 p
->numa_scan_period
= task_scan_min(p
);
1629 if (env
.best_task
== NULL
) {
1630 ret
= migrate_task_to(p
, env
.best_cpu
);
1632 trace_sched_stick_numa(p
, env
.src_cpu
, env
.best_cpu
);
1636 ret
= migrate_swap(p
, env
.best_task
);
1638 trace_sched_stick_numa(p
, env
.src_cpu
, task_cpu(env
.best_task
));
1639 put_task_struct(env
.best_task
);
1643 /* Attempt to migrate a task to a CPU on the preferred node. */
1644 static void numa_migrate_preferred(struct task_struct
*p
)
1646 unsigned long interval
= HZ
;
1648 /* This task has no NUMA fault statistics yet */
1649 if (unlikely(p
->numa_preferred_nid
== -1 || !p
->numa_faults
))
1652 /* Periodically retry migrating the task to the preferred node */
1653 interval
= min(interval
, msecs_to_jiffies(p
->numa_scan_period
) / 16);
1654 p
->numa_migrate_retry
= jiffies
+ interval
;
1656 /* Success if task is already running on preferred CPU */
1657 if (task_node(p
) == p
->numa_preferred_nid
)
1660 /* Otherwise, try migrate to a CPU on the preferred node */
1661 task_numa_migrate(p
);
1665 * Find out how many nodes on the workload is actively running on. Do this by
1666 * tracking the nodes from which NUMA hinting faults are triggered. This can
1667 * be different from the set of nodes where the workload's memory is currently
1670 static void numa_group_count_active_nodes(struct numa_group
*numa_group
)
1672 unsigned long faults
, max_faults
= 0;
1673 int nid
, active_nodes
= 0;
1675 for_each_online_node(nid
) {
1676 faults
= group_faults_cpu(numa_group
, nid
);
1677 if (faults
> max_faults
)
1678 max_faults
= faults
;
1681 for_each_online_node(nid
) {
1682 faults
= group_faults_cpu(numa_group
, nid
);
1683 if (faults
* ACTIVE_NODE_FRACTION
> max_faults
)
1687 numa_group
->max_faults_cpu
= max_faults
;
1688 numa_group
->active_nodes
= active_nodes
;
1692 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1693 * increments. The more local the fault statistics are, the higher the scan
1694 * period will be for the next scan window. If local/(local+remote) ratio is
1695 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1696 * the scan period will decrease. Aim for 70% local accesses.
1698 #define NUMA_PERIOD_SLOTS 10
1699 #define NUMA_PERIOD_THRESHOLD 7
1702 * Increase the scan period (slow down scanning) if the majority of
1703 * our memory is already on our local node, or if the majority of
1704 * the page accesses are shared with other processes.
1705 * Otherwise, decrease the scan period.
1707 static void update_task_scan_period(struct task_struct
*p
,
1708 unsigned long shared
, unsigned long private)
1710 unsigned int period_slot
;
1714 unsigned long remote
= p
->numa_faults_locality
[0];
1715 unsigned long local
= p
->numa_faults_locality
[1];
1718 * If there were no record hinting faults then either the task is
1719 * completely idle or all activity is areas that are not of interest
1720 * to automatic numa balancing. Related to that, if there were failed
1721 * migration then it implies we are migrating too quickly or the local
1722 * node is overloaded. In either case, scan slower
1724 if (local
+ shared
== 0 || p
->numa_faults_locality
[2]) {
1725 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1726 p
->numa_scan_period
<< 1);
1728 p
->mm
->numa_next_scan
= jiffies
+
1729 msecs_to_jiffies(p
->numa_scan_period
);
1735 * Prepare to scale scan period relative to the current period.
1736 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1737 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1738 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1740 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
1741 ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
1742 if (ratio
>= NUMA_PERIOD_THRESHOLD
) {
1743 int slot
= ratio
- NUMA_PERIOD_THRESHOLD
;
1746 diff
= slot
* period_slot
;
1748 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
1751 * Scale scan rate increases based on sharing. There is an
1752 * inverse relationship between the degree of sharing and
1753 * the adjustment made to the scanning period. Broadly
1754 * speaking the intent is that there is little point
1755 * scanning faster if shared accesses dominate as it may
1756 * simply bounce migrations uselessly
1758 ratio
= DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS
, (private + shared
+ 1));
1759 diff
= (diff
* ratio
) / NUMA_PERIOD_SLOTS
;
1762 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
1763 task_scan_min(p
), task_scan_max(p
));
1764 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1768 * Get the fraction of time the task has been running since the last
1769 * NUMA placement cycle. The scheduler keeps similar statistics, but
1770 * decays those on a 32ms period, which is orders of magnitude off
1771 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1772 * stats only if the task is so new there are no NUMA statistics yet.
1774 static u64
numa_get_avg_runtime(struct task_struct
*p
, u64
*period
)
1776 u64 runtime
, delta
, now
;
1777 /* Use the start of this time slice to avoid calculations. */
1778 now
= p
->se
.exec_start
;
1779 runtime
= p
->se
.sum_exec_runtime
;
1781 if (p
->last_task_numa_placement
) {
1782 delta
= runtime
- p
->last_sum_exec_runtime
;
1783 *period
= now
- p
->last_task_numa_placement
;
1785 delta
= p
->se
.avg
.load_sum
/ p
->se
.load
.weight
;
1786 *period
= LOAD_AVG_MAX
;
1789 p
->last_sum_exec_runtime
= runtime
;
1790 p
->last_task_numa_placement
= now
;
1796 * Determine the preferred nid for a task in a numa_group. This needs to
1797 * be done in a way that produces consistent results with group_weight,
1798 * otherwise workloads might not converge.
1800 static int preferred_group_nid(struct task_struct
*p
, int nid
)
1805 /* Direct connections between all NUMA nodes. */
1806 if (sched_numa_topology_type
== NUMA_DIRECT
)
1810 * On a system with glueless mesh NUMA topology, group_weight
1811 * scores nodes according to the number of NUMA hinting faults on
1812 * both the node itself, and on nearby nodes.
1814 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
1815 unsigned long score
, max_score
= 0;
1816 int node
, max_node
= nid
;
1818 dist
= sched_max_numa_distance
;
1820 for_each_online_node(node
) {
1821 score
= group_weight(p
, node
, dist
);
1822 if (score
> max_score
) {
1831 * Finding the preferred nid in a system with NUMA backplane
1832 * interconnect topology is more involved. The goal is to locate
1833 * tasks from numa_groups near each other in the system, and
1834 * untangle workloads from different sides of the system. This requires
1835 * searching down the hierarchy of node groups, recursively searching
1836 * inside the highest scoring group of nodes. The nodemask tricks
1837 * keep the complexity of the search down.
1839 nodes
= node_online_map
;
1840 for (dist
= sched_max_numa_distance
; dist
> LOCAL_DISTANCE
; dist
--) {
1841 unsigned long max_faults
= 0;
1842 nodemask_t max_group
= NODE_MASK_NONE
;
1845 /* Are there nodes at this distance from each other? */
1846 if (!find_numa_distance(dist
))
1849 for_each_node_mask(a
, nodes
) {
1850 unsigned long faults
= 0;
1851 nodemask_t this_group
;
1852 nodes_clear(this_group
);
1854 /* Sum group's NUMA faults; includes a==b case. */
1855 for_each_node_mask(b
, nodes
) {
1856 if (node_distance(a
, b
) < dist
) {
1857 faults
+= group_faults(p
, b
);
1858 node_set(b
, this_group
);
1859 node_clear(b
, nodes
);
1863 /* Remember the top group. */
1864 if (faults
> max_faults
) {
1865 max_faults
= faults
;
1866 max_group
= this_group
;
1868 * subtle: at the smallest distance there is
1869 * just one node left in each "group", the
1870 * winner is the preferred nid.
1875 /* Next round, evaluate the nodes within max_group. */
1883 static void task_numa_placement(struct task_struct
*p
)
1885 int seq
, nid
, max_nid
= -1, max_group_nid
= -1;
1886 unsigned long max_faults
= 0, max_group_faults
= 0;
1887 unsigned long fault_types
[2] = { 0, 0 };
1888 unsigned long total_faults
;
1889 u64 runtime
, period
;
1890 spinlock_t
*group_lock
= NULL
;
1893 * The p->mm->numa_scan_seq field gets updated without
1894 * exclusive access. Use READ_ONCE() here to ensure
1895 * that the field is read in a single access:
1897 seq
= READ_ONCE(p
->mm
->numa_scan_seq
);
1898 if (p
->numa_scan_seq
== seq
)
1900 p
->numa_scan_seq
= seq
;
1901 p
->numa_scan_period_max
= task_scan_max(p
);
1903 total_faults
= p
->numa_faults_locality
[0] +
1904 p
->numa_faults_locality
[1];
1905 runtime
= numa_get_avg_runtime(p
, &period
);
1907 /* If the task is part of a group prevent parallel updates to group stats */
1908 if (p
->numa_group
) {
1909 group_lock
= &p
->numa_group
->lock
;
1910 spin_lock_irq(group_lock
);
1913 /* Find the node with the highest number of faults */
1914 for_each_online_node(nid
) {
1915 /* Keep track of the offsets in numa_faults array */
1916 int mem_idx
, membuf_idx
, cpu_idx
, cpubuf_idx
;
1917 unsigned long faults
= 0, group_faults
= 0;
1920 for (priv
= 0; priv
< NR_NUMA_HINT_FAULT_TYPES
; priv
++) {
1921 long diff
, f_diff
, f_weight
;
1923 mem_idx
= task_faults_idx(NUMA_MEM
, nid
, priv
);
1924 membuf_idx
= task_faults_idx(NUMA_MEMBUF
, nid
, priv
);
1925 cpu_idx
= task_faults_idx(NUMA_CPU
, nid
, priv
);
1926 cpubuf_idx
= task_faults_idx(NUMA_CPUBUF
, nid
, priv
);
1928 /* Decay existing window, copy faults since last scan */
1929 diff
= p
->numa_faults
[membuf_idx
] - p
->numa_faults
[mem_idx
] / 2;
1930 fault_types
[priv
] += p
->numa_faults
[membuf_idx
];
1931 p
->numa_faults
[membuf_idx
] = 0;
1934 * Normalize the faults_from, so all tasks in a group
1935 * count according to CPU use, instead of by the raw
1936 * number of faults. Tasks with little runtime have
1937 * little over-all impact on throughput, and thus their
1938 * faults are less important.
1940 f_weight
= div64_u64(runtime
<< 16, period
+ 1);
1941 f_weight
= (f_weight
* p
->numa_faults
[cpubuf_idx
]) /
1943 f_diff
= f_weight
- p
->numa_faults
[cpu_idx
] / 2;
1944 p
->numa_faults
[cpubuf_idx
] = 0;
1946 p
->numa_faults
[mem_idx
] += diff
;
1947 p
->numa_faults
[cpu_idx
] += f_diff
;
1948 faults
+= p
->numa_faults
[mem_idx
];
1949 p
->total_numa_faults
+= diff
;
1950 if (p
->numa_group
) {
1952 * safe because we can only change our own group
1954 * mem_idx represents the offset for a given
1955 * nid and priv in a specific region because it
1956 * is at the beginning of the numa_faults array.
1958 p
->numa_group
->faults
[mem_idx
] += diff
;
1959 p
->numa_group
->faults_cpu
[mem_idx
] += f_diff
;
1960 p
->numa_group
->total_faults
+= diff
;
1961 group_faults
+= p
->numa_group
->faults
[mem_idx
];
1965 if (faults
> max_faults
) {
1966 max_faults
= faults
;
1970 if (group_faults
> max_group_faults
) {
1971 max_group_faults
= group_faults
;
1972 max_group_nid
= nid
;
1976 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
1978 if (p
->numa_group
) {
1979 numa_group_count_active_nodes(p
->numa_group
);
1980 spin_unlock_irq(group_lock
);
1981 max_nid
= preferred_group_nid(p
, max_group_nid
);
1985 /* Set the new preferred node */
1986 if (max_nid
!= p
->numa_preferred_nid
)
1987 sched_setnuma(p
, max_nid
);
1989 if (task_node(p
) != p
->numa_preferred_nid
)
1990 numa_migrate_preferred(p
);
1994 static inline int get_numa_group(struct numa_group
*grp
)
1996 return atomic_inc_not_zero(&grp
->refcount
);
1999 static inline void put_numa_group(struct numa_group
*grp
)
2001 if (atomic_dec_and_test(&grp
->refcount
))
2002 kfree_rcu(grp
, rcu
);
2005 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
2008 struct numa_group
*grp
, *my_grp
;
2009 struct task_struct
*tsk
;
2011 int cpu
= cpupid_to_cpu(cpupid
);
2014 if (unlikely(!p
->numa_group
)) {
2015 unsigned int size
= sizeof(struct numa_group
) +
2016 4*nr_node_ids
*sizeof(unsigned long);
2018 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
2022 atomic_set(&grp
->refcount
, 1);
2023 grp
->active_nodes
= 1;
2024 grp
->max_faults_cpu
= 0;
2025 spin_lock_init(&grp
->lock
);
2027 /* Second half of the array tracks nids where faults happen */
2028 grp
->faults_cpu
= grp
->faults
+ NR_NUMA_HINT_FAULT_TYPES
*
2031 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2032 grp
->faults
[i
] = p
->numa_faults
[i
];
2034 grp
->total_faults
= p
->total_numa_faults
;
2037 rcu_assign_pointer(p
->numa_group
, grp
);
2041 tsk
= READ_ONCE(cpu_rq(cpu
)->curr
);
2043 if (!cpupid_match_pid(tsk
, cpupid
))
2046 grp
= rcu_dereference(tsk
->numa_group
);
2050 my_grp
= p
->numa_group
;
2055 * Only join the other group if its bigger; if we're the bigger group,
2056 * the other task will join us.
2058 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
2062 * Tie-break on the grp address.
2064 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
2067 /* Always join threads in the same process. */
2068 if (tsk
->mm
== current
->mm
)
2071 /* Simple filter to avoid false positives due to PID collisions */
2072 if (flags
& TNF_SHARED
)
2075 /* Update priv based on whether false sharing was detected */
2078 if (join
&& !get_numa_group(grp
))
2086 BUG_ON(irqs_disabled());
2087 double_lock_irq(&my_grp
->lock
, &grp
->lock
);
2089 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++) {
2090 my_grp
->faults
[i
] -= p
->numa_faults
[i
];
2091 grp
->faults
[i
] += p
->numa_faults
[i
];
2093 my_grp
->total_faults
-= p
->total_numa_faults
;
2094 grp
->total_faults
+= p
->total_numa_faults
;
2099 spin_unlock(&my_grp
->lock
);
2100 spin_unlock_irq(&grp
->lock
);
2102 rcu_assign_pointer(p
->numa_group
, grp
);
2104 put_numa_group(my_grp
);
2112 void task_numa_free(struct task_struct
*p
)
2114 struct numa_group
*grp
= p
->numa_group
;
2115 void *numa_faults
= p
->numa_faults
;
2116 unsigned long flags
;
2120 spin_lock_irqsave(&grp
->lock
, flags
);
2121 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2122 grp
->faults
[i
] -= p
->numa_faults
[i
];
2123 grp
->total_faults
-= p
->total_numa_faults
;
2126 spin_unlock_irqrestore(&grp
->lock
, flags
);
2127 RCU_INIT_POINTER(p
->numa_group
, NULL
);
2128 put_numa_group(grp
);
2131 p
->numa_faults
= NULL
;
2136 * Got a PROT_NONE fault for a page on @node.
2138 void task_numa_fault(int last_cpupid
, int mem_node
, int pages
, int flags
)
2140 struct task_struct
*p
= current
;
2141 bool migrated
= flags
& TNF_MIGRATED
;
2142 int cpu_node
= task_node(current
);
2143 int local
= !!(flags
& TNF_FAULT_LOCAL
);
2144 struct numa_group
*ng
;
2147 if (!static_branch_likely(&sched_numa_balancing
))
2150 /* for example, ksmd faulting in a user's mm */
2154 /* Allocate buffer to track faults on a per-node basis */
2155 if (unlikely(!p
->numa_faults
)) {
2156 int size
= sizeof(*p
->numa_faults
) *
2157 NR_NUMA_HINT_FAULT_BUCKETS
* nr_node_ids
;
2159 p
->numa_faults
= kzalloc(size
, GFP_KERNEL
|__GFP_NOWARN
);
2160 if (!p
->numa_faults
)
2163 p
->total_numa_faults
= 0;
2164 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
2168 * First accesses are treated as private, otherwise consider accesses
2169 * to be private if the accessing pid has not changed
2171 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
2174 priv
= cpupid_match_pid(p
, last_cpupid
);
2175 if (!priv
&& !(flags
& TNF_NO_GROUP
))
2176 task_numa_group(p
, last_cpupid
, flags
, &priv
);
2180 * If a workload spans multiple NUMA nodes, a shared fault that
2181 * occurs wholly within the set of nodes that the workload is
2182 * actively using should be counted as local. This allows the
2183 * scan rate to slow down when a workload has settled down.
2186 if (!priv
&& !local
&& ng
&& ng
->active_nodes
> 1 &&
2187 numa_is_active_node(cpu_node
, ng
) &&
2188 numa_is_active_node(mem_node
, ng
))
2191 task_numa_placement(p
);
2194 * Retry task to preferred node migration periodically, in case it
2195 * case it previously failed, or the scheduler moved us.
2197 if (time_after(jiffies
, p
->numa_migrate_retry
))
2198 numa_migrate_preferred(p
);
2201 p
->numa_pages_migrated
+= pages
;
2202 if (flags
& TNF_MIGRATE_FAIL
)
2203 p
->numa_faults_locality
[2] += pages
;
2205 p
->numa_faults
[task_faults_idx(NUMA_MEMBUF
, mem_node
, priv
)] += pages
;
2206 p
->numa_faults
[task_faults_idx(NUMA_CPUBUF
, cpu_node
, priv
)] += pages
;
2207 p
->numa_faults_locality
[local
] += pages
;
2210 static void reset_ptenuma_scan(struct task_struct
*p
)
2213 * We only did a read acquisition of the mmap sem, so
2214 * p->mm->numa_scan_seq is written to without exclusive access
2215 * and the update is not guaranteed to be atomic. That's not
2216 * much of an issue though, since this is just used for
2217 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2218 * expensive, to avoid any form of compiler optimizations:
2220 WRITE_ONCE(p
->mm
->numa_scan_seq
, READ_ONCE(p
->mm
->numa_scan_seq
) + 1);
2221 p
->mm
->numa_scan_offset
= 0;
2225 * The expensive part of numa migration is done from task_work context.
2226 * Triggered from task_tick_numa().
2228 void task_numa_work(struct callback_head
*work
)
2230 unsigned long migrate
, next_scan
, now
= jiffies
;
2231 struct task_struct
*p
= current
;
2232 struct mm_struct
*mm
= p
->mm
;
2233 u64 runtime
= p
->se
.sum_exec_runtime
;
2234 struct vm_area_struct
*vma
;
2235 unsigned long start
, end
;
2236 unsigned long nr_pte_updates
= 0;
2237 long pages
, virtpages
;
2239 WARN_ON_ONCE(p
!= container_of(work
, struct task_struct
, numa_work
));
2241 work
->next
= work
; /* protect against double add */
2243 * Who cares about NUMA placement when they're dying.
2245 * NOTE: make sure not to dereference p->mm before this check,
2246 * exit_task_work() happens _after_ exit_mm() so we could be called
2247 * without p->mm even though we still had it when we enqueued this
2250 if (p
->flags
& PF_EXITING
)
2253 if (!mm
->numa_next_scan
) {
2254 mm
->numa_next_scan
= now
+
2255 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2259 * Enforce maximal scan/migration frequency..
2261 migrate
= mm
->numa_next_scan
;
2262 if (time_before(now
, migrate
))
2265 if (p
->numa_scan_period
== 0) {
2266 p
->numa_scan_period_max
= task_scan_max(p
);
2267 p
->numa_scan_period
= task_scan_min(p
);
2270 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
2271 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
2275 * Delay this task enough that another task of this mm will likely win
2276 * the next time around.
2278 p
->node_stamp
+= 2 * TICK_NSEC
;
2280 start
= mm
->numa_scan_offset
;
2281 pages
= sysctl_numa_balancing_scan_size
;
2282 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
2283 virtpages
= pages
* 8; /* Scan up to this much virtual space */
2288 down_read(&mm
->mmap_sem
);
2289 vma
= find_vma(mm
, start
);
2291 reset_ptenuma_scan(p
);
2295 for (; vma
; vma
= vma
->vm_next
) {
2296 if (!vma_migratable(vma
) || !vma_policy_mof(vma
) ||
2297 is_vm_hugetlb_page(vma
) || (vma
->vm_flags
& VM_MIXEDMAP
)) {
2302 * Shared library pages mapped by multiple processes are not
2303 * migrated as it is expected they are cache replicated. Avoid
2304 * hinting faults in read-only file-backed mappings or the vdso
2305 * as migrating the pages will be of marginal benefit.
2308 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
2312 * Skip inaccessible VMAs to avoid any confusion between
2313 * PROT_NONE and NUMA hinting ptes
2315 if (!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)))
2319 start
= max(start
, vma
->vm_start
);
2320 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
2321 end
= min(end
, vma
->vm_end
);
2322 nr_pte_updates
= change_prot_numa(vma
, start
, end
);
2325 * Try to scan sysctl_numa_balancing_size worth of
2326 * hpages that have at least one present PTE that
2327 * is not already pte-numa. If the VMA contains
2328 * areas that are unused or already full of prot_numa
2329 * PTEs, scan up to virtpages, to skip through those
2333 pages
-= (end
- start
) >> PAGE_SHIFT
;
2334 virtpages
-= (end
- start
) >> PAGE_SHIFT
;
2337 if (pages
<= 0 || virtpages
<= 0)
2341 } while (end
!= vma
->vm_end
);
2346 * It is possible to reach the end of the VMA list but the last few
2347 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2348 * would find the !migratable VMA on the next scan but not reset the
2349 * scanner to the start so check it now.
2352 mm
->numa_scan_offset
= start
;
2354 reset_ptenuma_scan(p
);
2355 up_read(&mm
->mmap_sem
);
2358 * Make sure tasks use at least 32x as much time to run other code
2359 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2360 * Usually update_task_scan_period slows down scanning enough; on an
2361 * overloaded system we need to limit overhead on a per task basis.
2363 if (unlikely(p
->se
.sum_exec_runtime
!= runtime
)) {
2364 u64 diff
= p
->se
.sum_exec_runtime
- runtime
;
2365 p
->node_stamp
+= 32 * diff
;
2370 * Drive the periodic memory faults..
2372 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2374 struct callback_head
*work
= &curr
->numa_work
;
2378 * We don't care about NUMA placement if we don't have memory.
2380 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
2384 * Using runtime rather than walltime has the dual advantage that
2385 * we (mostly) drive the selection from busy threads and that the
2386 * task needs to have done some actual work before we bother with
2389 now
= curr
->se
.sum_exec_runtime
;
2390 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
2392 if (now
> curr
->node_stamp
+ period
) {
2393 if (!curr
->node_stamp
)
2394 curr
->numa_scan_period
= task_scan_min(curr
);
2395 curr
->node_stamp
+= period
;
2397 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
2398 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
2399 task_work_add(curr
, work
, true);
2404 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2408 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
2412 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
2415 #endif /* CONFIG_NUMA_BALANCING */
2418 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2420 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
2421 if (!parent_entity(se
))
2422 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2424 if (entity_is_task(se
)) {
2425 struct rq
*rq
= rq_of(cfs_rq
);
2427 account_numa_enqueue(rq
, task_of(se
));
2428 list_add(&se
->group_node
, &rq
->cfs_tasks
);
2431 cfs_rq
->nr_running
++;
2435 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2437 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
2438 if (!parent_entity(se
))
2439 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2440 if (entity_is_task(se
)) {
2441 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
2442 list_del_init(&se
->group_node
);
2444 cfs_rq
->nr_running
--;
2447 #ifdef CONFIG_FAIR_GROUP_SCHED
2449 static inline long calc_tg_weight(struct task_group
*tg
, struct cfs_rq
*cfs_rq
)
2454 * Use this CPU's real-time load instead of the last load contribution
2455 * as the updating of the contribution is delayed, and we will use the
2456 * the real-time load to calc the share. See update_tg_load_avg().
2458 tg_weight
= atomic_long_read(&tg
->load_avg
);
2459 tg_weight
-= cfs_rq
->tg_load_avg_contrib
;
2460 tg_weight
+= cfs_rq
->load
.weight
;
2465 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2467 long tg_weight
, load
, shares
;
2469 tg_weight
= calc_tg_weight(tg
, cfs_rq
);
2470 load
= cfs_rq
->load
.weight
;
2472 shares
= (tg
->shares
* load
);
2474 shares
/= tg_weight
;
2476 if (shares
< MIN_SHARES
)
2477 shares
= MIN_SHARES
;
2478 if (shares
> tg
->shares
)
2479 shares
= tg
->shares
;
2483 # else /* CONFIG_SMP */
2484 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2488 # endif /* CONFIG_SMP */
2489 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
2490 unsigned long weight
)
2493 /* commit outstanding execution time */
2494 if (cfs_rq
->curr
== se
)
2495 update_curr(cfs_rq
);
2496 account_entity_dequeue(cfs_rq
, se
);
2499 update_load_set(&se
->load
, weight
);
2502 account_entity_enqueue(cfs_rq
, se
);
2505 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
2507 static void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2509 struct task_group
*tg
;
2510 struct sched_entity
*se
;
2514 se
= tg
->se
[cpu_of(rq_of(cfs_rq
))];
2515 if (!se
|| throttled_hierarchy(cfs_rq
))
2518 if (likely(se
->load
.weight
== tg
->shares
))
2521 shares
= calc_cfs_shares(cfs_rq
, tg
);
2523 reweight_entity(cfs_rq_of(se
), se
, shares
);
2525 #else /* CONFIG_FAIR_GROUP_SCHED */
2526 static inline void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2529 #endif /* CONFIG_FAIR_GROUP_SCHED */
2532 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2533 static const u32 runnable_avg_yN_inv
[] = {
2534 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2535 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2536 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2537 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2538 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2539 0x85aac367, 0x82cd8698,
2543 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2544 * over-estimates when re-combining.
2546 static const u32 runnable_avg_yN_sum
[] = {
2547 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2548 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2549 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2554 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2556 static __always_inline u64
decay_load(u64 val
, u64 n
)
2558 unsigned int local_n
;
2562 else if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
2565 /* after bounds checking we can collapse to 32-bit */
2569 * As y^PERIOD = 1/2, we can combine
2570 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2571 * With a look-up table which covers y^n (n<PERIOD)
2573 * To achieve constant time decay_load.
2575 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
2576 val
>>= local_n
/ LOAD_AVG_PERIOD
;
2577 local_n
%= LOAD_AVG_PERIOD
;
2580 val
= mul_u64_u32_shr(val
, runnable_avg_yN_inv
[local_n
], 32);
2585 * For updates fully spanning n periods, the contribution to runnable
2586 * average will be: \Sum 1024*y^n
2588 * We can compute this reasonably efficiently by combining:
2589 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2591 static u32
__compute_runnable_contrib(u64 n
)
2595 if (likely(n
<= LOAD_AVG_PERIOD
))
2596 return runnable_avg_yN_sum
[n
];
2597 else if (unlikely(n
>= LOAD_AVG_MAX_N
))
2598 return LOAD_AVG_MAX
;
2600 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2602 contrib
/= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2603 contrib
+= runnable_avg_yN_sum
[LOAD_AVG_PERIOD
];
2605 n
-= LOAD_AVG_PERIOD
;
2606 } while (n
> LOAD_AVG_PERIOD
);
2608 contrib
= decay_load(contrib
, n
);
2609 return contrib
+ runnable_avg_yN_sum
[n
];
2612 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2613 #error "load tracking assumes 2^10 as unit"
2616 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2619 * We can represent the historical contribution to runnable average as the
2620 * coefficients of a geometric series. To do this we sub-divide our runnable
2621 * history into segments of approximately 1ms (1024us); label the segment that
2622 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2624 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2626 * (now) (~1ms ago) (~2ms ago)
2628 * Let u_i denote the fraction of p_i that the entity was runnable.
2630 * We then designate the fractions u_i as our co-efficients, yielding the
2631 * following representation of historical load:
2632 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2634 * We choose y based on the with of a reasonably scheduling period, fixing:
2637 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2638 * approximately half as much as the contribution to load within the last ms
2641 * When a period "rolls over" and we have new u_0`, multiplying the previous
2642 * sum again by y is sufficient to update:
2643 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2644 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2646 static __always_inline
int
2647 __update_load_avg(u64 now
, int cpu
, struct sched_avg
*sa
,
2648 unsigned long weight
, int running
, struct cfs_rq
*cfs_rq
)
2650 u64 delta
, scaled_delta
, periods
;
2652 unsigned int delta_w
, scaled_delta_w
, decayed
= 0;
2653 unsigned long scale_freq
, scale_cpu
;
2655 delta
= now
- sa
->last_update_time
;
2657 * This should only happen when time goes backwards, which it
2658 * unfortunately does during sched clock init when we swap over to TSC.
2660 if ((s64
)delta
< 0) {
2661 sa
->last_update_time
= now
;
2666 * Use 1024ns as the unit of measurement since it's a reasonable
2667 * approximation of 1us and fast to compute.
2672 sa
->last_update_time
= now
;
2674 scale_freq
= arch_scale_freq_capacity(NULL
, cpu
);
2675 scale_cpu
= arch_scale_cpu_capacity(NULL
, cpu
);
2677 /* delta_w is the amount already accumulated against our next period */
2678 delta_w
= sa
->period_contrib
;
2679 if (delta
+ delta_w
>= 1024) {
2682 /* how much left for next period will start over, we don't know yet */
2683 sa
->period_contrib
= 0;
2686 * Now that we know we're crossing a period boundary, figure
2687 * out how much from delta we need to complete the current
2688 * period and accrue it.
2690 delta_w
= 1024 - delta_w
;
2691 scaled_delta_w
= cap_scale(delta_w
, scale_freq
);
2693 sa
->load_sum
+= weight
* scaled_delta_w
;
2695 cfs_rq
->runnable_load_sum
+=
2696 weight
* scaled_delta_w
;
2700 sa
->util_sum
+= scaled_delta_w
* scale_cpu
;
2704 /* Figure out how many additional periods this update spans */
2705 periods
= delta
/ 1024;
2708 sa
->load_sum
= decay_load(sa
->load_sum
, periods
+ 1);
2710 cfs_rq
->runnable_load_sum
=
2711 decay_load(cfs_rq
->runnable_load_sum
, periods
+ 1);
2713 sa
->util_sum
= decay_load((u64
)(sa
->util_sum
), periods
+ 1);
2715 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2716 contrib
= __compute_runnable_contrib(periods
);
2717 contrib
= cap_scale(contrib
, scale_freq
);
2719 sa
->load_sum
+= weight
* contrib
;
2721 cfs_rq
->runnable_load_sum
+= weight
* contrib
;
2724 sa
->util_sum
+= contrib
* scale_cpu
;
2727 /* Remainder of delta accrued against u_0` */
2728 scaled_delta
= cap_scale(delta
, scale_freq
);
2730 sa
->load_sum
+= weight
* scaled_delta
;
2732 cfs_rq
->runnable_load_sum
+= weight
* scaled_delta
;
2735 sa
->util_sum
+= scaled_delta
* scale_cpu
;
2737 sa
->period_contrib
+= delta
;
2740 sa
->load_avg
= div_u64(sa
->load_sum
, LOAD_AVG_MAX
);
2742 cfs_rq
->runnable_load_avg
=
2743 div_u64(cfs_rq
->runnable_load_sum
, LOAD_AVG_MAX
);
2745 sa
->util_avg
= sa
->util_sum
/ LOAD_AVG_MAX
;
2751 #ifdef CONFIG_FAIR_GROUP_SCHED
2753 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2754 * and effective_load (which is not done because it is too costly).
2756 static inline void update_tg_load_avg(struct cfs_rq
*cfs_rq
, int force
)
2758 long delta
= cfs_rq
->avg
.load_avg
- cfs_rq
->tg_load_avg_contrib
;
2761 * No need to update load_avg for root_task_group as it is not used.
2763 if (cfs_rq
->tg
== &root_task_group
)
2766 if (force
|| abs(delta
) > cfs_rq
->tg_load_avg_contrib
/ 64) {
2767 atomic_long_add(delta
, &cfs_rq
->tg
->load_avg
);
2768 cfs_rq
->tg_load_avg_contrib
= cfs_rq
->avg
.load_avg
;
2773 * Called within set_task_rq() right before setting a task's cpu. The
2774 * caller only guarantees p->pi_lock is held; no other assumptions,
2775 * including the state of rq->lock, should be made.
2777 void set_task_rq_fair(struct sched_entity
*se
,
2778 struct cfs_rq
*prev
, struct cfs_rq
*next
)
2780 if (!sched_feat(ATTACH_AGE_LOAD
))
2784 * We are supposed to update the task to "current" time, then its up to
2785 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
2786 * getting what current time is, so simply throw away the out-of-date
2787 * time. This will result in the wakee task is less decayed, but giving
2788 * the wakee more load sounds not bad.
2790 if (se
->avg
.last_update_time
&& prev
) {
2791 u64 p_last_update_time
;
2792 u64 n_last_update_time
;
2794 #ifndef CONFIG_64BIT
2795 u64 p_last_update_time_copy
;
2796 u64 n_last_update_time_copy
;
2799 p_last_update_time_copy
= prev
->load_last_update_time_copy
;
2800 n_last_update_time_copy
= next
->load_last_update_time_copy
;
2804 p_last_update_time
= prev
->avg
.last_update_time
;
2805 n_last_update_time
= next
->avg
.last_update_time
;
2807 } while (p_last_update_time
!= p_last_update_time_copy
||
2808 n_last_update_time
!= n_last_update_time_copy
);
2810 p_last_update_time
= prev
->avg
.last_update_time
;
2811 n_last_update_time
= next
->avg
.last_update_time
;
2813 __update_load_avg(p_last_update_time
, cpu_of(rq_of(prev
)),
2814 &se
->avg
, 0, 0, NULL
);
2815 se
->avg
.last_update_time
= n_last_update_time
;
2818 #else /* CONFIG_FAIR_GROUP_SCHED */
2819 static inline void update_tg_load_avg(struct cfs_rq
*cfs_rq
, int force
) {}
2820 #endif /* CONFIG_FAIR_GROUP_SCHED */
2822 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
2824 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2825 static inline int update_cfs_rq_load_avg(u64 now
, struct cfs_rq
*cfs_rq
)
2827 struct sched_avg
*sa
= &cfs_rq
->avg
;
2828 int decayed
, removed
= 0;
2830 if (atomic_long_read(&cfs_rq
->removed_load_avg
)) {
2831 s64 r
= atomic_long_xchg(&cfs_rq
->removed_load_avg
, 0);
2832 sa
->load_avg
= max_t(long, sa
->load_avg
- r
, 0);
2833 sa
->load_sum
= max_t(s64
, sa
->load_sum
- r
* LOAD_AVG_MAX
, 0);
2837 if (atomic_long_read(&cfs_rq
->removed_util_avg
)) {
2838 long r
= atomic_long_xchg(&cfs_rq
->removed_util_avg
, 0);
2839 sa
->util_avg
= max_t(long, sa
->util_avg
- r
, 0);
2840 sa
->util_sum
= max_t(s32
, sa
->util_sum
- r
* LOAD_AVG_MAX
, 0);
2843 decayed
= __update_load_avg(now
, cpu_of(rq_of(cfs_rq
)), sa
,
2844 scale_load_down(cfs_rq
->load
.weight
), cfs_rq
->curr
!= NULL
, cfs_rq
);
2846 #ifndef CONFIG_64BIT
2848 cfs_rq
->load_last_update_time_copy
= sa
->last_update_time
;
2851 return decayed
|| removed
;
2854 /* Update task and its cfs_rq load average */
2855 static inline void update_load_avg(struct sched_entity
*se
, int update_tg
)
2857 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2858 u64 now
= cfs_rq_clock_task(cfs_rq
);
2859 int cpu
= cpu_of(rq_of(cfs_rq
));
2862 * Track task load average for carrying it to new CPU after migrated, and
2863 * track group sched_entity load average for task_h_load calc in migration
2865 __update_load_avg(now
, cpu
, &se
->avg
,
2866 se
->on_rq
* scale_load_down(se
->load
.weight
),
2867 cfs_rq
->curr
== se
, NULL
);
2869 if (update_cfs_rq_load_avg(now
, cfs_rq
) && update_tg
)
2870 update_tg_load_avg(cfs_rq
, 0);
2873 static void attach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2875 if (!sched_feat(ATTACH_AGE_LOAD
))
2879 * If we got migrated (either between CPUs or between cgroups) we'll
2880 * have aged the average right before clearing @last_update_time.
2882 if (se
->avg
.last_update_time
) {
2883 __update_load_avg(cfs_rq
->avg
.last_update_time
, cpu_of(rq_of(cfs_rq
)),
2884 &se
->avg
, 0, 0, NULL
);
2887 * XXX: we could have just aged the entire load away if we've been
2888 * absent from the fair class for too long.
2893 se
->avg
.last_update_time
= cfs_rq
->avg
.last_update_time
;
2894 cfs_rq
->avg
.load_avg
+= se
->avg
.load_avg
;
2895 cfs_rq
->avg
.load_sum
+= se
->avg
.load_sum
;
2896 cfs_rq
->avg
.util_avg
+= se
->avg
.util_avg
;
2897 cfs_rq
->avg
.util_sum
+= se
->avg
.util_sum
;
2900 static void detach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2902 __update_load_avg(cfs_rq
->avg
.last_update_time
, cpu_of(rq_of(cfs_rq
)),
2903 &se
->avg
, se
->on_rq
* scale_load_down(se
->load
.weight
),
2904 cfs_rq
->curr
== se
, NULL
);
2906 cfs_rq
->avg
.load_avg
= max_t(long, cfs_rq
->avg
.load_avg
- se
->avg
.load_avg
, 0);
2907 cfs_rq
->avg
.load_sum
= max_t(s64
, cfs_rq
->avg
.load_sum
- se
->avg
.load_sum
, 0);
2908 cfs_rq
->avg
.util_avg
= max_t(long, cfs_rq
->avg
.util_avg
- se
->avg
.util_avg
, 0);
2909 cfs_rq
->avg
.util_sum
= max_t(s32
, cfs_rq
->avg
.util_sum
- se
->avg
.util_sum
, 0);
2912 /* Add the load generated by se into cfs_rq's load average */
2914 enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2916 struct sched_avg
*sa
= &se
->avg
;
2917 u64 now
= cfs_rq_clock_task(cfs_rq
);
2918 int migrated
, decayed
;
2920 migrated
= !sa
->last_update_time
;
2922 __update_load_avg(now
, cpu_of(rq_of(cfs_rq
)), sa
,
2923 se
->on_rq
* scale_load_down(se
->load
.weight
),
2924 cfs_rq
->curr
== se
, NULL
);
2927 decayed
= update_cfs_rq_load_avg(now
, cfs_rq
);
2929 cfs_rq
->runnable_load_avg
+= sa
->load_avg
;
2930 cfs_rq
->runnable_load_sum
+= sa
->load_sum
;
2933 attach_entity_load_avg(cfs_rq
, se
);
2935 if (decayed
|| migrated
)
2936 update_tg_load_avg(cfs_rq
, 0);
2939 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2941 dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2943 update_load_avg(se
, 1);
2945 cfs_rq
->runnable_load_avg
=
2946 max_t(long, cfs_rq
->runnable_load_avg
- se
->avg
.load_avg
, 0);
2947 cfs_rq
->runnable_load_sum
=
2948 max_t(s64
, cfs_rq
->runnable_load_sum
- se
->avg
.load_sum
, 0);
2951 #ifndef CONFIG_64BIT
2952 static inline u64
cfs_rq_last_update_time(struct cfs_rq
*cfs_rq
)
2954 u64 last_update_time_copy
;
2955 u64 last_update_time
;
2958 last_update_time_copy
= cfs_rq
->load_last_update_time_copy
;
2960 last_update_time
= cfs_rq
->avg
.last_update_time
;
2961 } while (last_update_time
!= last_update_time_copy
);
2963 return last_update_time
;
2966 static inline u64
cfs_rq_last_update_time(struct cfs_rq
*cfs_rq
)
2968 return cfs_rq
->avg
.last_update_time
;
2973 * Task first catches up with cfs_rq, and then subtract
2974 * itself from the cfs_rq (task must be off the queue now).
2976 void remove_entity_load_avg(struct sched_entity
*se
)
2978 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2979 u64 last_update_time
;
2982 * Newly created task or never used group entity should not be removed
2983 * from its (source) cfs_rq
2985 if (se
->avg
.last_update_time
== 0)
2988 last_update_time
= cfs_rq_last_update_time(cfs_rq
);
2990 __update_load_avg(last_update_time
, cpu_of(rq_of(cfs_rq
)), &se
->avg
, 0, 0, NULL
);
2991 atomic_long_add(se
->avg
.load_avg
, &cfs_rq
->removed_load_avg
);
2992 atomic_long_add(se
->avg
.util_avg
, &cfs_rq
->removed_util_avg
);
2995 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq
*cfs_rq
)
2997 return cfs_rq
->runnable_load_avg
;
3000 static inline unsigned long cfs_rq_load_avg(struct cfs_rq
*cfs_rq
)
3002 return cfs_rq
->avg
.load_avg
;
3005 static int idle_balance(struct rq
*this_rq
);
3007 #else /* CONFIG_SMP */
3009 static inline void update_load_avg(struct sched_entity
*se
, int update_tg
) {}
3011 enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
3013 dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
3014 static inline void remove_entity_load_avg(struct sched_entity
*se
) {}
3017 attach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
3019 detach_entity_load_avg(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
) {}
3021 static inline int idle_balance(struct rq
*rq
)
3026 #endif /* CONFIG_SMP */
3028 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3030 #ifdef CONFIG_SCHEDSTATS
3031 struct task_struct
*tsk
= NULL
;
3033 if (entity_is_task(se
))
3036 if (se
->statistics
.sleep_start
) {
3037 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.sleep_start
;
3042 if (unlikely(delta
> se
->statistics
.sleep_max
))
3043 se
->statistics
.sleep_max
= delta
;
3045 se
->statistics
.sleep_start
= 0;
3046 se
->statistics
.sum_sleep_runtime
+= delta
;
3049 account_scheduler_latency(tsk
, delta
>> 10, 1);
3050 trace_sched_stat_sleep(tsk
, delta
);
3053 if (se
->statistics
.block_start
) {
3054 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.block_start
;
3059 if (unlikely(delta
> se
->statistics
.block_max
))
3060 se
->statistics
.block_max
= delta
;
3062 se
->statistics
.block_start
= 0;
3063 se
->statistics
.sum_sleep_runtime
+= delta
;
3066 if (tsk
->in_iowait
) {
3067 se
->statistics
.iowait_sum
+= delta
;
3068 se
->statistics
.iowait_count
++;
3069 trace_sched_stat_iowait(tsk
, delta
);
3072 trace_sched_stat_blocked(tsk
, delta
);
3075 * Blocking time is in units of nanosecs, so shift by
3076 * 20 to get a milliseconds-range estimation of the
3077 * amount of time that the task spent sleeping:
3079 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
3080 profile_hits(SLEEP_PROFILING
,
3081 (void *)get_wchan(tsk
),
3084 account_scheduler_latency(tsk
, delta
>> 10, 0);
3090 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3092 #ifdef CONFIG_SCHED_DEBUG
3093 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
3098 if (d
> 3*sysctl_sched_latency
)
3099 schedstat_inc(cfs_rq
, nr_spread_over
);
3104 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
3106 u64 vruntime
= cfs_rq
->min_vruntime
;
3109 * The 'current' period is already promised to the current tasks,
3110 * however the extra weight of the new task will slow them down a
3111 * little, place the new task so that it fits in the slot that
3112 * stays open at the end.
3114 if (initial
&& sched_feat(START_DEBIT
))
3115 vruntime
+= sched_vslice(cfs_rq
, se
);
3117 /* sleeps up to a single latency don't count. */
3119 unsigned long thresh
= sysctl_sched_latency
;
3122 * Halve their sleep time's effect, to allow
3123 * for a gentler effect of sleepers:
3125 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
3131 /* ensure we never gain time by being placed backwards. */
3132 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
3135 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
3137 static inline void check_schedstat_required(void)
3139 #ifdef CONFIG_SCHEDSTATS
3140 if (schedstat_enabled())
3143 /* Force schedstat enabled if a dependent tracepoint is active */
3144 if (trace_sched_stat_wait_enabled() ||
3145 trace_sched_stat_sleep_enabled() ||
3146 trace_sched_stat_iowait_enabled() ||
3147 trace_sched_stat_blocked_enabled() ||
3148 trace_sched_stat_runtime_enabled()) {
3149 pr_warn_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3150 "stat_blocked and stat_runtime require the "
3151 "kernel parameter schedstats=enabled or "
3152 "kernel.sched_schedstats=1\n");
3158 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3161 * Update the normalized vruntime before updating min_vruntime
3162 * through calling update_curr().
3164 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
3165 se
->vruntime
+= cfs_rq
->min_vruntime
;
3168 * Update run-time statistics of the 'current'.
3170 update_curr(cfs_rq
);
3171 enqueue_entity_load_avg(cfs_rq
, se
);
3172 account_entity_enqueue(cfs_rq
, se
);
3173 update_cfs_shares(cfs_rq
);
3175 if (flags
& ENQUEUE_WAKEUP
) {
3176 place_entity(cfs_rq
, se
, 0);
3177 if (schedstat_enabled())
3178 enqueue_sleeper(cfs_rq
, se
);
3181 check_schedstat_required();
3182 if (schedstat_enabled()) {
3183 update_stats_enqueue(cfs_rq
, se
);
3184 check_spread(cfs_rq
, se
);
3186 if (se
!= cfs_rq
->curr
)
3187 __enqueue_entity(cfs_rq
, se
);
3190 if (cfs_rq
->nr_running
== 1) {
3191 list_add_leaf_cfs_rq(cfs_rq
);
3192 check_enqueue_throttle(cfs_rq
);
3196 static void __clear_buddies_last(struct sched_entity
*se
)
3198 for_each_sched_entity(se
) {
3199 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3200 if (cfs_rq
->last
!= se
)
3203 cfs_rq
->last
= NULL
;
3207 static void __clear_buddies_next(struct sched_entity
*se
)
3209 for_each_sched_entity(se
) {
3210 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3211 if (cfs_rq
->next
!= se
)
3214 cfs_rq
->next
= NULL
;
3218 static void __clear_buddies_skip(struct sched_entity
*se
)
3220 for_each_sched_entity(se
) {
3221 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3222 if (cfs_rq
->skip
!= se
)
3225 cfs_rq
->skip
= NULL
;
3229 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3231 if (cfs_rq
->last
== se
)
3232 __clear_buddies_last(se
);
3234 if (cfs_rq
->next
== se
)
3235 __clear_buddies_next(se
);
3237 if (cfs_rq
->skip
== se
)
3238 __clear_buddies_skip(se
);
3241 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3244 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3247 * Update run-time statistics of the 'current'.
3249 update_curr(cfs_rq
);
3250 dequeue_entity_load_avg(cfs_rq
, se
);
3252 if (schedstat_enabled())
3253 update_stats_dequeue(cfs_rq
, se
, flags
);
3255 clear_buddies(cfs_rq
, se
);
3257 if (se
!= cfs_rq
->curr
)
3258 __dequeue_entity(cfs_rq
, se
);
3260 account_entity_dequeue(cfs_rq
, se
);
3263 * Normalize the entity after updating the min_vruntime because the
3264 * update can refer to the ->curr item and we need to reflect this
3265 * movement in our normalized position.
3267 if (!(flags
& DEQUEUE_SLEEP
))
3268 se
->vruntime
-= cfs_rq
->min_vruntime
;
3270 /* return excess runtime on last dequeue */
3271 return_cfs_rq_runtime(cfs_rq
);
3273 update_min_vruntime(cfs_rq
);
3274 update_cfs_shares(cfs_rq
);
3278 * Preempt the current task with a newly woken task if needed:
3281 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
3283 unsigned long ideal_runtime
, delta_exec
;
3284 struct sched_entity
*se
;
3287 ideal_runtime
= sched_slice(cfs_rq
, curr
);
3288 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
3289 if (delta_exec
> ideal_runtime
) {
3290 resched_curr(rq_of(cfs_rq
));
3292 * The current task ran long enough, ensure it doesn't get
3293 * re-elected due to buddy favours.
3295 clear_buddies(cfs_rq
, curr
);
3300 * Ensure that a task that missed wakeup preemption by a
3301 * narrow margin doesn't have to wait for a full slice.
3302 * This also mitigates buddy induced latencies under load.
3304 if (delta_exec
< sysctl_sched_min_granularity
)
3307 se
= __pick_first_entity(cfs_rq
);
3308 delta
= curr
->vruntime
- se
->vruntime
;
3313 if (delta
> ideal_runtime
)
3314 resched_curr(rq_of(cfs_rq
));
3318 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3320 /* 'current' is not kept within the tree. */
3323 * Any task has to be enqueued before it get to execute on
3324 * a CPU. So account for the time it spent waiting on the
3327 if (schedstat_enabled())
3328 update_stats_wait_end(cfs_rq
, se
);
3329 __dequeue_entity(cfs_rq
, se
);
3330 update_load_avg(se
, 1);
3333 update_stats_curr_start(cfs_rq
, se
);
3335 #ifdef CONFIG_SCHEDSTATS
3337 * Track our maximum slice length, if the CPU's load is at
3338 * least twice that of our own weight (i.e. dont track it
3339 * when there are only lesser-weight tasks around):
3341 if (schedstat_enabled() && rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
3342 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
3343 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
3346 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
3350 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
3353 * Pick the next process, keeping these things in mind, in this order:
3354 * 1) keep things fair between processes/task groups
3355 * 2) pick the "next" process, since someone really wants that to run
3356 * 3) pick the "last" process, for cache locality
3357 * 4) do not run the "skip" process, if something else is available
3359 static struct sched_entity
*
3360 pick_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
3362 struct sched_entity
*left
= __pick_first_entity(cfs_rq
);
3363 struct sched_entity
*se
;
3366 * If curr is set we have to see if its left of the leftmost entity
3367 * still in the tree, provided there was anything in the tree at all.
3369 if (!left
|| (curr
&& entity_before(curr
, left
)))
3372 se
= left
; /* ideally we run the leftmost entity */
3375 * Avoid running the skip buddy, if running something else can
3376 * be done without getting too unfair.
3378 if (cfs_rq
->skip
== se
) {
3379 struct sched_entity
*second
;
3382 second
= __pick_first_entity(cfs_rq
);
3384 second
= __pick_next_entity(se
);
3385 if (!second
|| (curr
&& entity_before(curr
, second
)))
3389 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
3394 * Prefer last buddy, try to return the CPU to a preempted task.
3396 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
3400 * Someone really wants this to run. If it's not unfair, run it.
3402 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
3405 clear_buddies(cfs_rq
, se
);
3410 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3412 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
3415 * If still on the runqueue then deactivate_task()
3416 * was not called and update_curr() has to be done:
3419 update_curr(cfs_rq
);
3421 /* throttle cfs_rqs exceeding runtime */
3422 check_cfs_rq_runtime(cfs_rq
);
3424 if (schedstat_enabled()) {
3425 check_spread(cfs_rq
, prev
);
3427 update_stats_wait_start(cfs_rq
, prev
);
3431 /* Put 'current' back into the tree. */
3432 __enqueue_entity(cfs_rq
, prev
);
3433 /* in !on_rq case, update occurred at dequeue */
3434 update_load_avg(prev
, 0);
3436 cfs_rq
->curr
= NULL
;
3440 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
3443 * Update run-time statistics of the 'current'.
3445 update_curr(cfs_rq
);
3448 * Ensure that runnable average is periodically updated.
3450 update_load_avg(curr
, 1);
3451 update_cfs_shares(cfs_rq
);
3453 #ifdef CONFIG_SCHED_HRTICK
3455 * queued ticks are scheduled to match the slice, so don't bother
3456 * validating it and just reschedule.
3459 resched_curr(rq_of(cfs_rq
));
3463 * don't let the period tick interfere with the hrtick preemption
3465 if (!sched_feat(DOUBLE_TICK
) &&
3466 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
3470 if (cfs_rq
->nr_running
> 1)
3471 check_preempt_tick(cfs_rq
, curr
);
3475 /**************************************************
3476 * CFS bandwidth control machinery
3479 #ifdef CONFIG_CFS_BANDWIDTH
3481 #ifdef HAVE_JUMP_LABEL
3482 static struct static_key __cfs_bandwidth_used
;
3484 static inline bool cfs_bandwidth_used(void)
3486 return static_key_false(&__cfs_bandwidth_used
);
3489 void cfs_bandwidth_usage_inc(void)
3491 static_key_slow_inc(&__cfs_bandwidth_used
);
3494 void cfs_bandwidth_usage_dec(void)
3496 static_key_slow_dec(&__cfs_bandwidth_used
);
3498 #else /* HAVE_JUMP_LABEL */
3499 static bool cfs_bandwidth_used(void)
3504 void cfs_bandwidth_usage_inc(void) {}
3505 void cfs_bandwidth_usage_dec(void) {}
3506 #endif /* HAVE_JUMP_LABEL */
3509 * default period for cfs group bandwidth.
3510 * default: 0.1s, units: nanoseconds
3512 static inline u64
default_cfs_period(void)
3514 return 100000000ULL;
3517 static inline u64
sched_cfs_bandwidth_slice(void)
3519 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
3523 * Replenish runtime according to assigned quota and update expiration time.
3524 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3525 * additional synchronization around rq->lock.
3527 * requires cfs_b->lock
3529 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
3533 if (cfs_b
->quota
== RUNTIME_INF
)
3536 now
= sched_clock_cpu(smp_processor_id());
3537 cfs_b
->runtime
= cfs_b
->quota
;
3538 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
3541 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3543 return &tg
->cfs_bandwidth
;
3546 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3547 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3549 if (unlikely(cfs_rq
->throttle_count
))
3550 return cfs_rq
->throttled_clock_task
;
3552 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
3555 /* returns 0 on failure to allocate runtime */
3556 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3558 struct task_group
*tg
= cfs_rq
->tg
;
3559 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
3560 u64 amount
= 0, min_amount
, expires
;
3562 /* note: this is a positive sum as runtime_remaining <= 0 */
3563 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
3565 raw_spin_lock(&cfs_b
->lock
);
3566 if (cfs_b
->quota
== RUNTIME_INF
)
3567 amount
= min_amount
;
3569 start_cfs_bandwidth(cfs_b
);
3571 if (cfs_b
->runtime
> 0) {
3572 amount
= min(cfs_b
->runtime
, min_amount
);
3573 cfs_b
->runtime
-= amount
;
3577 expires
= cfs_b
->runtime_expires
;
3578 raw_spin_unlock(&cfs_b
->lock
);
3580 cfs_rq
->runtime_remaining
+= amount
;
3582 * we may have advanced our local expiration to account for allowed
3583 * spread between our sched_clock and the one on which runtime was
3586 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
3587 cfs_rq
->runtime_expires
= expires
;
3589 return cfs_rq
->runtime_remaining
> 0;
3593 * Note: This depends on the synchronization provided by sched_clock and the
3594 * fact that rq->clock snapshots this value.
3596 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3598 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3600 /* if the deadline is ahead of our clock, nothing to do */
3601 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
3604 if (cfs_rq
->runtime_remaining
< 0)
3608 * If the local deadline has passed we have to consider the
3609 * possibility that our sched_clock is 'fast' and the global deadline
3610 * has not truly expired.
3612 * Fortunately we can check determine whether this the case by checking
3613 * whether the global deadline has advanced. It is valid to compare
3614 * cfs_b->runtime_expires without any locks since we only care about
3615 * exact equality, so a partial write will still work.
3618 if (cfs_rq
->runtime_expires
!= cfs_b
->runtime_expires
) {
3619 /* extend local deadline, drift is bounded above by 2 ticks */
3620 cfs_rq
->runtime_expires
+= TICK_NSEC
;
3622 /* global deadline is ahead, expiration has passed */
3623 cfs_rq
->runtime_remaining
= 0;
3627 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3629 /* dock delta_exec before expiring quota (as it could span periods) */
3630 cfs_rq
->runtime_remaining
-= delta_exec
;
3631 expire_cfs_rq_runtime(cfs_rq
);
3633 if (likely(cfs_rq
->runtime_remaining
> 0))
3637 * if we're unable to extend our runtime we resched so that the active
3638 * hierarchy can be throttled
3640 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
3641 resched_curr(rq_of(cfs_rq
));
3644 static __always_inline
3645 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3647 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
3650 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
3653 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3655 return cfs_bandwidth_used() && cfs_rq
->throttled
;
3658 /* check whether cfs_rq, or any parent, is throttled */
3659 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3661 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
3665 * Ensure that neither of the group entities corresponding to src_cpu or
3666 * dest_cpu are members of a throttled hierarchy when performing group
3667 * load-balance operations.
3669 static inline int throttled_lb_pair(struct task_group
*tg
,
3670 int src_cpu
, int dest_cpu
)
3672 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
3674 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
3675 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
3677 return throttled_hierarchy(src_cfs_rq
) ||
3678 throttled_hierarchy(dest_cfs_rq
);
3681 /* updated child weight may affect parent so we have to do this bottom up */
3682 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
3684 struct rq
*rq
= data
;
3685 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3687 cfs_rq
->throttle_count
--;
3689 if (!cfs_rq
->throttle_count
) {
3690 /* adjust cfs_rq_clock_task() */
3691 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
3692 cfs_rq
->throttled_clock_task
;
3699 static int tg_throttle_down(struct task_group
*tg
, void *data
)
3701 struct rq
*rq
= data
;
3702 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3704 /* group is entering throttled state, stop time */
3705 if (!cfs_rq
->throttle_count
)
3706 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
3707 cfs_rq
->throttle_count
++;
3712 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
3714 struct rq
*rq
= rq_of(cfs_rq
);
3715 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3716 struct sched_entity
*se
;
3717 long task_delta
, dequeue
= 1;
3720 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
3722 /* freeze hierarchy runnable averages while throttled */
3724 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
3727 task_delta
= cfs_rq
->h_nr_running
;
3728 for_each_sched_entity(se
) {
3729 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
3730 /* throttled entity or throttle-on-deactivate */
3735 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
3736 qcfs_rq
->h_nr_running
-= task_delta
;
3738 if (qcfs_rq
->load
.weight
)
3743 sub_nr_running(rq
, task_delta
);
3745 cfs_rq
->throttled
= 1;
3746 cfs_rq
->throttled_clock
= rq_clock(rq
);
3747 raw_spin_lock(&cfs_b
->lock
);
3748 empty
= list_empty(&cfs_b
->throttled_cfs_rq
);
3751 * Add to the _head_ of the list, so that an already-started
3752 * distribute_cfs_runtime will not see us
3754 list_add_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
3757 * If we're the first throttled task, make sure the bandwidth
3761 start_cfs_bandwidth(cfs_b
);
3763 raw_spin_unlock(&cfs_b
->lock
);
3766 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
3768 struct rq
*rq
= rq_of(cfs_rq
);
3769 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3770 struct sched_entity
*se
;
3774 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
3776 cfs_rq
->throttled
= 0;
3778 update_rq_clock(rq
);
3780 raw_spin_lock(&cfs_b
->lock
);
3781 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
3782 list_del_rcu(&cfs_rq
->throttled_list
);
3783 raw_spin_unlock(&cfs_b
->lock
);
3785 /* update hierarchical throttle state */
3786 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
3788 if (!cfs_rq
->load
.weight
)
3791 task_delta
= cfs_rq
->h_nr_running
;
3792 for_each_sched_entity(se
) {
3796 cfs_rq
= cfs_rq_of(se
);
3798 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
3799 cfs_rq
->h_nr_running
+= task_delta
;
3801 if (cfs_rq_throttled(cfs_rq
))
3806 add_nr_running(rq
, task_delta
);
3808 /* determine whether we need to wake up potentially idle cpu */
3809 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
3813 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
3814 u64 remaining
, u64 expires
)
3816 struct cfs_rq
*cfs_rq
;
3818 u64 starting_runtime
= remaining
;
3821 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
3823 struct rq
*rq
= rq_of(cfs_rq
);
3825 raw_spin_lock(&rq
->lock
);
3826 if (!cfs_rq_throttled(cfs_rq
))
3829 runtime
= -cfs_rq
->runtime_remaining
+ 1;
3830 if (runtime
> remaining
)
3831 runtime
= remaining
;
3832 remaining
-= runtime
;
3834 cfs_rq
->runtime_remaining
+= runtime
;
3835 cfs_rq
->runtime_expires
= expires
;
3837 /* we check whether we're throttled above */
3838 if (cfs_rq
->runtime_remaining
> 0)
3839 unthrottle_cfs_rq(cfs_rq
);
3842 raw_spin_unlock(&rq
->lock
);
3849 return starting_runtime
- remaining
;
3853 * Responsible for refilling a task_group's bandwidth and unthrottling its
3854 * cfs_rqs as appropriate. If there has been no activity within the last
3855 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3856 * used to track this state.
3858 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
3860 u64 runtime
, runtime_expires
;
3863 /* no need to continue the timer with no bandwidth constraint */
3864 if (cfs_b
->quota
== RUNTIME_INF
)
3865 goto out_deactivate
;
3867 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3868 cfs_b
->nr_periods
+= overrun
;
3871 * idle depends on !throttled (for the case of a large deficit), and if
3872 * we're going inactive then everything else can be deferred
3874 if (cfs_b
->idle
&& !throttled
)
3875 goto out_deactivate
;
3877 __refill_cfs_bandwidth_runtime(cfs_b
);
3880 /* mark as potentially idle for the upcoming period */
3885 /* account preceding periods in which throttling occurred */
3886 cfs_b
->nr_throttled
+= overrun
;
3888 runtime_expires
= cfs_b
->runtime_expires
;
3891 * This check is repeated as we are holding onto the new bandwidth while
3892 * we unthrottle. This can potentially race with an unthrottled group
3893 * trying to acquire new bandwidth from the global pool. This can result
3894 * in us over-using our runtime if it is all used during this loop, but
3895 * only by limited amounts in that extreme case.
3897 while (throttled
&& cfs_b
->runtime
> 0) {
3898 runtime
= cfs_b
->runtime
;
3899 raw_spin_unlock(&cfs_b
->lock
);
3900 /* we can't nest cfs_b->lock while distributing bandwidth */
3901 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
3903 raw_spin_lock(&cfs_b
->lock
);
3905 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3907 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
3911 * While we are ensured activity in the period following an
3912 * unthrottle, this also covers the case in which the new bandwidth is
3913 * insufficient to cover the existing bandwidth deficit. (Forcing the
3914 * timer to remain active while there are any throttled entities.)
3924 /* a cfs_rq won't donate quota below this amount */
3925 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
3926 /* minimum remaining period time to redistribute slack quota */
3927 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
3928 /* how long we wait to gather additional slack before distributing */
3929 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
3932 * Are we near the end of the current quota period?
3934 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3935 * hrtimer base being cleared by hrtimer_start. In the case of
3936 * migrate_hrtimers, base is never cleared, so we are fine.
3938 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
3940 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
3943 /* if the call-back is running a quota refresh is already occurring */
3944 if (hrtimer_callback_running(refresh_timer
))
3947 /* is a quota refresh about to occur? */
3948 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
3949 if (remaining
< min_expire
)
3955 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
3957 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
3959 /* if there's a quota refresh soon don't bother with slack */
3960 if (runtime_refresh_within(cfs_b
, min_left
))
3963 hrtimer_start(&cfs_b
->slack_timer
,
3964 ns_to_ktime(cfs_bandwidth_slack_period
),
3968 /* we know any runtime found here is valid as update_curr() precedes return */
3969 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3971 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3972 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
3974 if (slack_runtime
<= 0)
3977 raw_spin_lock(&cfs_b
->lock
);
3978 if (cfs_b
->quota
!= RUNTIME_INF
&&
3979 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
3980 cfs_b
->runtime
+= slack_runtime
;
3982 /* we are under rq->lock, defer unthrottling using a timer */
3983 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
3984 !list_empty(&cfs_b
->throttled_cfs_rq
))
3985 start_cfs_slack_bandwidth(cfs_b
);
3987 raw_spin_unlock(&cfs_b
->lock
);
3989 /* even if it's not valid for return we don't want to try again */
3990 cfs_rq
->runtime_remaining
-= slack_runtime
;
3993 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3995 if (!cfs_bandwidth_used())
3998 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
4001 __return_cfs_rq_runtime(cfs_rq
);
4005 * This is done with a timer (instead of inline with bandwidth return) since
4006 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4008 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
4010 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
4013 /* confirm we're still not at a refresh boundary */
4014 raw_spin_lock(&cfs_b
->lock
);
4015 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
4016 raw_spin_unlock(&cfs_b
->lock
);
4020 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
)
4021 runtime
= cfs_b
->runtime
;
4023 expires
= cfs_b
->runtime_expires
;
4024 raw_spin_unlock(&cfs_b
->lock
);
4029 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
4031 raw_spin_lock(&cfs_b
->lock
);
4032 if (expires
== cfs_b
->runtime_expires
)
4033 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
4034 raw_spin_unlock(&cfs_b
->lock
);
4038 * When a group wakes up we want to make sure that its quota is not already
4039 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4040 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4042 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
4044 if (!cfs_bandwidth_used())
4047 /* an active group must be handled by the update_curr()->put() path */
4048 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
4051 /* ensure the group is not already throttled */
4052 if (cfs_rq_throttled(cfs_rq
))
4055 /* update runtime allocation */
4056 account_cfs_rq_runtime(cfs_rq
, 0);
4057 if (cfs_rq
->runtime_remaining
<= 0)
4058 throttle_cfs_rq(cfs_rq
);
4061 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4062 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4064 if (!cfs_bandwidth_used())
4067 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
4071 * it's possible for a throttled entity to be forced into a running
4072 * state (e.g. set_curr_task), in this case we're finished.
4074 if (cfs_rq_throttled(cfs_rq
))
4077 throttle_cfs_rq(cfs_rq
);
4081 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
4083 struct cfs_bandwidth
*cfs_b
=
4084 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
4086 do_sched_cfs_slack_timer(cfs_b
);
4088 return HRTIMER_NORESTART
;
4091 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
4093 struct cfs_bandwidth
*cfs_b
=
4094 container_of(timer
, struct cfs_bandwidth
, period_timer
);
4098 raw_spin_lock(&cfs_b
->lock
);
4100 overrun
= hrtimer_forward_now(timer
, cfs_b
->period
);
4104 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
4107 cfs_b
->period_active
= 0;
4108 raw_spin_unlock(&cfs_b
->lock
);
4110 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
4113 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
4115 raw_spin_lock_init(&cfs_b
->lock
);
4117 cfs_b
->quota
= RUNTIME_INF
;
4118 cfs_b
->period
= ns_to_ktime(default_cfs_period());
4120 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
4121 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_ABS_PINNED
);
4122 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
4123 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
4124 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
4127 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4129 cfs_rq
->runtime_enabled
= 0;
4130 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
4133 void start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
4135 lockdep_assert_held(&cfs_b
->lock
);
4137 if (!cfs_b
->period_active
) {
4138 cfs_b
->period_active
= 1;
4139 hrtimer_forward_now(&cfs_b
->period_timer
, cfs_b
->period
);
4140 hrtimer_start_expires(&cfs_b
->period_timer
, HRTIMER_MODE_ABS_PINNED
);
4144 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
4146 /* init_cfs_bandwidth() was not called */
4147 if (!cfs_b
->throttled_cfs_rq
.next
)
4150 hrtimer_cancel(&cfs_b
->period_timer
);
4151 hrtimer_cancel(&cfs_b
->slack_timer
);
4154 static void __maybe_unused
update_runtime_enabled(struct rq
*rq
)
4156 struct cfs_rq
*cfs_rq
;
4158 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
4159 struct cfs_bandwidth
*cfs_b
= &cfs_rq
->tg
->cfs_bandwidth
;
4161 raw_spin_lock(&cfs_b
->lock
);
4162 cfs_rq
->runtime_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
4163 raw_spin_unlock(&cfs_b
->lock
);
4167 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
4169 struct cfs_rq
*cfs_rq
;
4171 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
4172 if (!cfs_rq
->runtime_enabled
)
4176 * clock_task is not advancing so we just need to make sure
4177 * there's some valid quota amount
4179 cfs_rq
->runtime_remaining
= 1;
4181 * Offline rq is schedulable till cpu is completely disabled
4182 * in take_cpu_down(), so we prevent new cfs throttling here.
4184 cfs_rq
->runtime_enabled
= 0;
4186 if (cfs_rq_throttled(cfs_rq
))
4187 unthrottle_cfs_rq(cfs_rq
);
4191 #else /* CONFIG_CFS_BANDWIDTH */
4192 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
4194 return rq_clock_task(rq_of(cfs_rq
));
4197 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
4198 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) { return false; }
4199 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
4200 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
4202 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
4207 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
4212 static inline int throttled_lb_pair(struct task_group
*tg
,
4213 int src_cpu
, int dest_cpu
)
4218 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
4220 #ifdef CONFIG_FAIR_GROUP_SCHED
4221 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
4224 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
4228 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
4229 static inline void update_runtime_enabled(struct rq
*rq
) {}
4230 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
4232 #endif /* CONFIG_CFS_BANDWIDTH */
4234 /**************************************************
4235 * CFS operations on tasks:
4238 #ifdef CONFIG_SCHED_HRTICK
4239 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
4241 struct sched_entity
*se
= &p
->se
;
4242 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4244 WARN_ON(task_rq(p
) != rq
);
4246 if (cfs_rq
->nr_running
> 1) {
4247 u64 slice
= sched_slice(cfs_rq
, se
);
4248 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
4249 s64 delta
= slice
- ran
;
4256 hrtick_start(rq
, delta
);
4261 * called from enqueue/dequeue and updates the hrtick when the
4262 * current task is from our class and nr_running is low enough
4265 static void hrtick_update(struct rq
*rq
)
4267 struct task_struct
*curr
= rq
->curr
;
4269 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
4272 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
4273 hrtick_start_fair(rq
, curr
);
4275 #else /* !CONFIG_SCHED_HRTICK */
4277 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
4281 static inline void hrtick_update(struct rq
*rq
)
4287 * The enqueue_task method is called before nr_running is
4288 * increased. Here we update the fair scheduling stats and
4289 * then put the task into the rbtree:
4292 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
4294 struct cfs_rq
*cfs_rq
;
4295 struct sched_entity
*se
= &p
->se
;
4297 for_each_sched_entity(se
) {
4300 cfs_rq
= cfs_rq_of(se
);
4301 enqueue_entity(cfs_rq
, se
, flags
);
4304 * end evaluation on encountering a throttled cfs_rq
4306 * note: in the case of encountering a throttled cfs_rq we will
4307 * post the final h_nr_running increment below.
4309 if (cfs_rq_throttled(cfs_rq
))
4311 cfs_rq
->h_nr_running
++;
4313 flags
= ENQUEUE_WAKEUP
;
4316 for_each_sched_entity(se
) {
4317 cfs_rq
= cfs_rq_of(se
);
4318 cfs_rq
->h_nr_running
++;
4320 if (cfs_rq_throttled(cfs_rq
))
4323 update_load_avg(se
, 1);
4324 update_cfs_shares(cfs_rq
);
4328 add_nr_running(rq
, 1);
4333 static void set_next_buddy(struct sched_entity
*se
);
4336 * The dequeue_task method is called before nr_running is
4337 * decreased. We remove the task from the rbtree and
4338 * update the fair scheduling stats:
4340 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
4342 struct cfs_rq
*cfs_rq
;
4343 struct sched_entity
*se
= &p
->se
;
4344 int task_sleep
= flags
& DEQUEUE_SLEEP
;
4346 for_each_sched_entity(se
) {
4347 cfs_rq
= cfs_rq_of(se
);
4348 dequeue_entity(cfs_rq
, se
, flags
);
4351 * end evaluation on encountering a throttled cfs_rq
4353 * note: in the case of encountering a throttled cfs_rq we will
4354 * post the final h_nr_running decrement below.
4356 if (cfs_rq_throttled(cfs_rq
))
4358 cfs_rq
->h_nr_running
--;
4360 /* Don't dequeue parent if it has other entities besides us */
4361 if (cfs_rq
->load
.weight
) {
4363 * Bias pick_next to pick a task from this cfs_rq, as
4364 * p is sleeping when it is within its sched_slice.
4366 if (task_sleep
&& parent_entity(se
))
4367 set_next_buddy(parent_entity(se
));
4369 /* avoid re-evaluating load for this entity */
4370 se
= parent_entity(se
);
4373 flags
|= DEQUEUE_SLEEP
;
4376 for_each_sched_entity(se
) {
4377 cfs_rq
= cfs_rq_of(se
);
4378 cfs_rq
->h_nr_running
--;
4380 if (cfs_rq_throttled(cfs_rq
))
4383 update_load_avg(se
, 1);
4384 update_cfs_shares(cfs_rq
);
4388 sub_nr_running(rq
, 1);
4396 * per rq 'load' arrray crap; XXX kill this.
4400 * The exact cpuload calculated at every tick would be:
4402 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
4404 * If a cpu misses updates for n ticks (as it was idle) and update gets
4405 * called on the n+1-th tick when cpu may be busy, then we have:
4407 * load_n = (1 - 1/2^i)^n * load_0
4408 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
4410 * decay_load_missed() below does efficient calculation of
4412 * load' = (1 - 1/2^i)^n * load
4414 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
4415 * This allows us to precompute the above in said factors, thereby allowing the
4416 * reduction of an arbitrary n in O(log_2 n) steps. (See also
4417 * fixed_power_int())
4419 * The calculation is approximated on a 128 point scale.
4421 #define DEGRADE_SHIFT 7
4423 static const u8 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
4424 static const u8 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
4425 { 0, 0, 0, 0, 0, 0, 0, 0 },
4426 { 64, 32, 8, 0, 0, 0, 0, 0 },
4427 { 96, 72, 40, 12, 1, 0, 0, 0 },
4428 { 112, 98, 75, 43, 15, 1, 0, 0 },
4429 { 120, 112, 98, 76, 45, 16, 2, 0 }
4433 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4434 * would be when CPU is idle and so we just decay the old load without
4435 * adding any new load.
4437 static unsigned long
4438 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
4442 if (!missed_updates
)
4445 if (missed_updates
>= degrade_zero_ticks
[idx
])
4449 return load
>> missed_updates
;
4451 while (missed_updates
) {
4452 if (missed_updates
% 2)
4453 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
4455 missed_updates
>>= 1;
4462 * __update_cpu_load - update the rq->cpu_load[] statistics
4463 * @this_rq: The rq to update statistics for
4464 * @this_load: The current load
4465 * @pending_updates: The number of missed updates
4466 * @active: !0 for NOHZ_FULL
4468 * Update rq->cpu_load[] statistics. This function is usually called every
4469 * scheduler tick (TICK_NSEC).
4471 * This function computes a decaying average:
4473 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
4475 * Because of NOHZ it might not get called on every tick which gives need for
4476 * the @pending_updates argument.
4478 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
4479 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
4480 * = A * (A * load[i]_n-2 + B) + B
4481 * = A * (A * (A * load[i]_n-3 + B) + B) + B
4482 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
4483 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
4484 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
4485 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
4487 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
4488 * any change in load would have resulted in the tick being turned back on.
4490 * For regular NOHZ, this reduces to:
4492 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
4494 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
4495 * term. See the @active paramter.
4497 static void __update_cpu_load(struct rq
*this_rq
, unsigned long this_load
,
4498 unsigned long pending_updates
, int active
)
4500 unsigned long tickless_load
= active
? this_rq
->cpu_load
[0] : 0;
4503 this_rq
->nr_load_updates
++;
4505 /* Update our load: */
4506 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
4507 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
4508 unsigned long old_load
, new_load
;
4510 /* scale is effectively 1 << i now, and >> i divides by scale */
4512 old_load
= this_rq
->cpu_load
[i
];
4513 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
4514 if (tickless_load
) {
4515 old_load
-= decay_load_missed(tickless_load
, pending_updates
- 1, i
);
4517 * old_load can never be a negative value because a
4518 * decayed tickless_load cannot be greater than the
4519 * original tickless_load.
4521 old_load
+= tickless_load
;
4523 new_load
= this_load
;
4525 * Round up the averaging division if load is increasing. This
4526 * prevents us from getting stuck on 9 if the load is 10, for
4529 if (new_load
> old_load
)
4530 new_load
+= scale
- 1;
4532 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
4535 sched_avg_update(this_rq
);
4538 /* Used instead of source_load when we know the type == 0 */
4539 static unsigned long weighted_cpuload(const int cpu
)
4541 return cfs_rq_runnable_load_avg(&cpu_rq(cpu
)->cfs
);
4544 #ifdef CONFIG_NO_HZ_COMMON
4546 * There is no sane way to deal with nohz on smp when using jiffies because the
4547 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4548 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4550 * Therefore we cannot use the delta approach from the regular tick since that
4551 * would seriously skew the load calculation. However we'll make do for those
4552 * updates happening while idle (nohz_idle_balance) or coming out of idle
4553 * (tick_nohz_idle_exit).
4555 * This means we might still be one tick off for nohz periods.
4559 * Called from nohz_idle_balance() to update the load ratings before doing the
4562 static void update_idle_cpu_load(struct rq
*this_rq
)
4564 unsigned long curr_jiffies
= READ_ONCE(jiffies
);
4565 unsigned long load
= weighted_cpuload(cpu_of(this_rq
));
4566 unsigned long pending_updates
;
4569 * bail if there's load or we're actually up-to-date.
4571 if (load
|| curr_jiffies
== this_rq
->last_load_update_tick
)
4574 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
4575 this_rq
->last_load_update_tick
= curr_jiffies
;
4577 __update_cpu_load(this_rq
, load
, pending_updates
, 0);
4581 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4583 void update_cpu_load_nohz(int active
)
4585 struct rq
*this_rq
= this_rq();
4586 unsigned long curr_jiffies
= READ_ONCE(jiffies
);
4587 unsigned long load
= active
? weighted_cpuload(cpu_of(this_rq
)) : 0;
4588 unsigned long pending_updates
;
4590 if (curr_jiffies
== this_rq
->last_load_update_tick
)
4593 raw_spin_lock(&this_rq
->lock
);
4594 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
4595 if (pending_updates
) {
4596 this_rq
->last_load_update_tick
= curr_jiffies
;
4598 * In the regular NOHZ case, we were idle, this means load 0.
4599 * In the NOHZ_FULL case, we were non-idle, we should consider
4600 * its weighted load.
4602 __update_cpu_load(this_rq
, load
, pending_updates
, active
);
4604 raw_spin_unlock(&this_rq
->lock
);
4606 #endif /* CONFIG_NO_HZ */
4609 * Called from scheduler_tick()
4611 void update_cpu_load_active(struct rq
*this_rq
)
4613 unsigned long load
= weighted_cpuload(cpu_of(this_rq
));
4615 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4617 this_rq
->last_load_update_tick
= jiffies
;
4618 __update_cpu_load(this_rq
, load
, 1, 1);
4622 * Return a low guess at the load of a migration-source cpu weighted
4623 * according to the scheduling class and "nice" value.
4625 * We want to under-estimate the load of migration sources, to
4626 * balance conservatively.
4628 static unsigned long source_load(int cpu
, int type
)
4630 struct rq
*rq
= cpu_rq(cpu
);
4631 unsigned long total
= weighted_cpuload(cpu
);
4633 if (type
== 0 || !sched_feat(LB_BIAS
))
4636 return min(rq
->cpu_load
[type
-1], total
);
4640 * Return a high guess at the load of a migration-target cpu weighted
4641 * according to the scheduling class and "nice" value.
4643 static unsigned long target_load(int cpu
, int type
)
4645 struct rq
*rq
= cpu_rq(cpu
);
4646 unsigned long total
= weighted_cpuload(cpu
);
4648 if (type
== 0 || !sched_feat(LB_BIAS
))
4651 return max(rq
->cpu_load
[type
-1], total
);
4654 static unsigned long capacity_of(int cpu
)
4656 return cpu_rq(cpu
)->cpu_capacity
;
4659 static unsigned long capacity_orig_of(int cpu
)
4661 return cpu_rq(cpu
)->cpu_capacity_orig
;
4664 static unsigned long cpu_avg_load_per_task(int cpu
)
4666 struct rq
*rq
= cpu_rq(cpu
);
4667 unsigned long nr_running
= READ_ONCE(rq
->cfs
.h_nr_running
);
4668 unsigned long load_avg
= weighted_cpuload(cpu
);
4671 return load_avg
/ nr_running
;
4676 static void record_wakee(struct task_struct
*p
)
4679 * Rough decay (wiping) for cost saving, don't worry
4680 * about the boundary, really active task won't care
4683 if (time_after(jiffies
, current
->wakee_flip_decay_ts
+ HZ
)) {
4684 current
->wakee_flips
>>= 1;
4685 current
->wakee_flip_decay_ts
= jiffies
;
4688 if (current
->last_wakee
!= p
) {
4689 current
->last_wakee
= p
;
4690 current
->wakee_flips
++;
4694 static void task_waking_fair(struct task_struct
*p
)
4696 struct sched_entity
*se
= &p
->se
;
4697 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4700 #ifndef CONFIG_64BIT
4701 u64 min_vruntime_copy
;
4704 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
4706 min_vruntime
= cfs_rq
->min_vruntime
;
4707 } while (min_vruntime
!= min_vruntime_copy
);
4709 min_vruntime
= cfs_rq
->min_vruntime
;
4712 se
->vruntime
-= min_vruntime
;
4716 #ifdef CONFIG_FAIR_GROUP_SCHED
4718 * effective_load() calculates the load change as seen from the root_task_group
4720 * Adding load to a group doesn't make a group heavier, but can cause movement
4721 * of group shares between cpus. Assuming the shares were perfectly aligned one
4722 * can calculate the shift in shares.
4724 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4725 * on this @cpu and results in a total addition (subtraction) of @wg to the
4726 * total group weight.
4728 * Given a runqueue weight distribution (rw_i) we can compute a shares
4729 * distribution (s_i) using:
4731 * s_i = rw_i / \Sum rw_j (1)
4733 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4734 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4735 * shares distribution (s_i):
4737 * rw_i = { 2, 4, 1, 0 }
4738 * s_i = { 2/7, 4/7, 1/7, 0 }
4740 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4741 * task used to run on and the CPU the waker is running on), we need to
4742 * compute the effect of waking a task on either CPU and, in case of a sync
4743 * wakeup, compute the effect of the current task going to sleep.
4745 * So for a change of @wl to the local @cpu with an overall group weight change
4746 * of @wl we can compute the new shares distribution (s'_i) using:
4748 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4750 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4751 * differences in waking a task to CPU 0. The additional task changes the
4752 * weight and shares distributions like:
4754 * rw'_i = { 3, 4, 1, 0 }
4755 * s'_i = { 3/8, 4/8, 1/8, 0 }
4757 * We can then compute the difference in effective weight by using:
4759 * dw_i = S * (s'_i - s_i) (3)
4761 * Where 'S' is the group weight as seen by its parent.
4763 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4764 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4765 * 4/7) times the weight of the group.
4767 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4769 struct sched_entity
*se
= tg
->se
[cpu
];
4771 if (!tg
->parent
) /* the trivial, non-cgroup case */
4774 for_each_sched_entity(se
) {
4780 * W = @wg + \Sum rw_j
4782 W
= wg
+ calc_tg_weight(tg
, se
->my_q
);
4787 w
= cfs_rq_load_avg(se
->my_q
) + wl
;
4790 * wl = S * s'_i; see (2)
4793 wl
= (w
* (long)tg
->shares
) / W
;
4798 * Per the above, wl is the new se->load.weight value; since
4799 * those are clipped to [MIN_SHARES, ...) do so now. See
4800 * calc_cfs_shares().
4802 if (wl
< MIN_SHARES
)
4806 * wl = dw_i = S * (s'_i - s_i); see (3)
4808 wl
-= se
->avg
.load_avg
;
4811 * Recursively apply this logic to all parent groups to compute
4812 * the final effective load change on the root group. Since
4813 * only the @tg group gets extra weight, all parent groups can
4814 * only redistribute existing shares. @wl is the shift in shares
4815 * resulting from this level per the above.
4824 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4832 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
4833 * A waker of many should wake a different task than the one last awakened
4834 * at a frequency roughly N times higher than one of its wakees. In order
4835 * to determine whether we should let the load spread vs consolodating to
4836 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
4837 * partner, and a factor of lls_size higher frequency in the other. With
4838 * both conditions met, we can be relatively sure that the relationship is
4839 * non-monogamous, with partner count exceeding socket size. Waker/wakee
4840 * being client/server, worker/dispatcher, interrupt source or whatever is
4841 * irrelevant, spread criteria is apparent partner count exceeds socket size.
4843 static int wake_wide(struct task_struct
*p
)
4845 unsigned int master
= current
->wakee_flips
;
4846 unsigned int slave
= p
->wakee_flips
;
4847 int factor
= this_cpu_read(sd_llc_size
);
4850 swap(master
, slave
);
4851 if (slave
< factor
|| master
< slave
* factor
)
4856 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
4858 s64 this_load
, load
;
4859 s64 this_eff_load
, prev_eff_load
;
4860 int idx
, this_cpu
, prev_cpu
;
4861 struct task_group
*tg
;
4862 unsigned long weight
;
4866 this_cpu
= smp_processor_id();
4867 prev_cpu
= task_cpu(p
);
4868 load
= source_load(prev_cpu
, idx
);
4869 this_load
= target_load(this_cpu
, idx
);
4872 * If sync wakeup then subtract the (maximum possible)
4873 * effect of the currently running task from the load
4874 * of the current CPU:
4877 tg
= task_group(current
);
4878 weight
= current
->se
.avg
.load_avg
;
4880 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
4881 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
4885 weight
= p
->se
.avg
.load_avg
;
4888 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4889 * due to the sync cause above having dropped this_load to 0, we'll
4890 * always have an imbalance, but there's really nothing you can do
4891 * about that, so that's good too.
4893 * Otherwise check if either cpus are near enough in load to allow this
4894 * task to be woken on this_cpu.
4896 this_eff_load
= 100;
4897 this_eff_load
*= capacity_of(prev_cpu
);
4899 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
4900 prev_eff_load
*= capacity_of(this_cpu
);
4902 if (this_load
> 0) {
4903 this_eff_load
*= this_load
+
4904 effective_load(tg
, this_cpu
, weight
, weight
);
4906 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
4909 balanced
= this_eff_load
<= prev_eff_load
;
4911 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
4916 schedstat_inc(sd
, ttwu_move_affine
);
4917 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
4923 * find_idlest_group finds and returns the least busy CPU group within the
4926 static struct sched_group
*
4927 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
4928 int this_cpu
, int sd_flag
)
4930 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
4931 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
4932 int load_idx
= sd
->forkexec_idx
;
4933 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
4935 if (sd_flag
& SD_BALANCE_WAKE
)
4936 load_idx
= sd
->wake_idx
;
4939 unsigned long load
, avg_load
;
4943 /* Skip over this group if it has no CPUs allowed */
4944 if (!cpumask_intersects(sched_group_cpus(group
),
4945 tsk_cpus_allowed(p
)))
4948 local_group
= cpumask_test_cpu(this_cpu
,
4949 sched_group_cpus(group
));
4951 /* Tally up the load of all CPUs in the group */
4954 for_each_cpu(i
, sched_group_cpus(group
)) {
4955 /* Bias balancing toward cpus of our domain */
4957 load
= source_load(i
, load_idx
);
4959 load
= target_load(i
, load_idx
);
4964 /* Adjust by relative CPU capacity of the group */
4965 avg_load
= (avg_load
* SCHED_CAPACITY_SCALE
) / group
->sgc
->capacity
;
4968 this_load
= avg_load
;
4969 } else if (avg_load
< min_load
) {
4970 min_load
= avg_load
;
4973 } while (group
= group
->next
, group
!= sd
->groups
);
4975 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
4981 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4984 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
4986 unsigned long load
, min_load
= ULONG_MAX
;
4987 unsigned int min_exit_latency
= UINT_MAX
;
4988 u64 latest_idle_timestamp
= 0;
4989 int least_loaded_cpu
= this_cpu
;
4990 int shallowest_idle_cpu
= -1;
4993 /* Traverse only the allowed CPUs */
4994 for_each_cpu_and(i
, sched_group_cpus(group
), tsk_cpus_allowed(p
)) {
4996 struct rq
*rq
= cpu_rq(i
);
4997 struct cpuidle_state
*idle
= idle_get_state(rq
);
4998 if (idle
&& idle
->exit_latency
< min_exit_latency
) {
5000 * We give priority to a CPU whose idle state
5001 * has the smallest exit latency irrespective
5002 * of any idle timestamp.
5004 min_exit_latency
= idle
->exit_latency
;
5005 latest_idle_timestamp
= rq
->idle_stamp
;
5006 shallowest_idle_cpu
= i
;
5007 } else if ((!idle
|| idle
->exit_latency
== min_exit_latency
) &&
5008 rq
->idle_stamp
> latest_idle_timestamp
) {
5010 * If equal or no active idle state, then
5011 * the most recently idled CPU might have
5014 latest_idle_timestamp
= rq
->idle_stamp
;
5015 shallowest_idle_cpu
= i
;
5017 } else if (shallowest_idle_cpu
== -1) {
5018 load
= weighted_cpuload(i
);
5019 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
5021 least_loaded_cpu
= i
;
5026 return shallowest_idle_cpu
!= -1 ? shallowest_idle_cpu
: least_loaded_cpu
;
5030 * Try and locate an idle CPU in the sched_domain.
5032 static int select_idle_sibling(struct task_struct
*p
, int target
)
5034 struct sched_domain
*sd
;
5035 struct sched_group
*sg
;
5036 int i
= task_cpu(p
);
5038 if (idle_cpu(target
))
5042 * If the prevous cpu is cache affine and idle, don't be stupid.
5044 if (i
!= target
&& cpus_share_cache(i
, target
) && idle_cpu(i
))
5048 * Otherwise, iterate the domains and find an elegible idle cpu.
5050 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
5051 for_each_lower_domain(sd
) {
5054 if (!cpumask_intersects(sched_group_cpus(sg
),
5055 tsk_cpus_allowed(p
)))
5058 for_each_cpu(i
, sched_group_cpus(sg
)) {
5059 if (i
== target
|| !idle_cpu(i
))
5063 target
= cpumask_first_and(sched_group_cpus(sg
),
5064 tsk_cpus_allowed(p
));
5068 } while (sg
!= sd
->groups
);
5075 * cpu_util returns the amount of capacity of a CPU that is used by CFS
5076 * tasks. The unit of the return value must be the one of capacity so we can
5077 * compare the utilization with the capacity of the CPU that is available for
5078 * CFS task (ie cpu_capacity).
5080 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5081 * recent utilization of currently non-runnable tasks on a CPU. It represents
5082 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5083 * capacity_orig is the cpu_capacity available at the highest frequency
5084 * (arch_scale_freq_capacity()).
5085 * The utilization of a CPU converges towards a sum equal to or less than the
5086 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5087 * the running time on this CPU scaled by capacity_curr.
5089 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5090 * higher than capacity_orig because of unfortunate rounding in
5091 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5092 * the average stabilizes with the new running time. We need to check that the
5093 * utilization stays within the range of [0..capacity_orig] and cap it if
5094 * necessary. Without utilization capping, a group could be seen as overloaded
5095 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5096 * available capacity. We allow utilization to overshoot capacity_curr (but not
5097 * capacity_orig) as it useful for predicting the capacity required after task
5098 * migrations (scheduler-driven DVFS).
5100 static int cpu_util(int cpu
)
5102 unsigned long util
= cpu_rq(cpu
)->cfs
.avg
.util_avg
;
5103 unsigned long capacity
= capacity_orig_of(cpu
);
5105 return (util
>= capacity
) ? capacity
: util
;
5109 * select_task_rq_fair: Select target runqueue for the waking task in domains
5110 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5111 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5113 * Balances load by selecting the idlest cpu in the idlest group, or under
5114 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5116 * Returns the target cpu number.
5118 * preempt must be disabled.
5121 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
5123 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
5124 int cpu
= smp_processor_id();
5125 int new_cpu
= prev_cpu
;
5126 int want_affine
= 0;
5127 int sync
= wake_flags
& WF_SYNC
;
5129 if (sd_flag
& SD_BALANCE_WAKE
)
5130 want_affine
= !wake_wide(p
) && cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
));
5133 for_each_domain(cpu
, tmp
) {
5134 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
5138 * If both cpu and prev_cpu are part of this domain,
5139 * cpu is a valid SD_WAKE_AFFINE target.
5141 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
5142 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
5147 if (tmp
->flags
& sd_flag
)
5149 else if (!want_affine
)
5154 sd
= NULL
; /* Prefer wake_affine over balance flags */
5155 if (cpu
!= prev_cpu
&& wake_affine(affine_sd
, p
, sync
))
5160 if (sd_flag
& SD_BALANCE_WAKE
) /* XXX always ? */
5161 new_cpu
= select_idle_sibling(p
, new_cpu
);
5164 struct sched_group
*group
;
5167 if (!(sd
->flags
& sd_flag
)) {
5172 group
= find_idlest_group(sd
, p
, cpu
, sd_flag
);
5178 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
5179 if (new_cpu
== -1 || new_cpu
== cpu
) {
5180 /* Now try balancing at a lower domain level of cpu */
5185 /* Now try balancing at a lower domain level of new_cpu */
5187 weight
= sd
->span_weight
;
5189 for_each_domain(cpu
, tmp
) {
5190 if (weight
<= tmp
->span_weight
)
5192 if (tmp
->flags
& sd_flag
)
5195 /* while loop will break here if sd == NULL */
5203 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5204 * cfs_rq_of(p) references at time of call are still valid and identify the
5205 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
5207 static void migrate_task_rq_fair(struct task_struct
*p
)
5210 * We are supposed to update the task to "current" time, then its up to date
5211 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5212 * what current time is, so simply throw away the out-of-date time. This
5213 * will result in the wakee task is less decayed, but giving the wakee more
5214 * load sounds not bad.
5216 remove_entity_load_avg(&p
->se
);
5218 /* Tell new CPU we are migrated */
5219 p
->se
.avg
.last_update_time
= 0;
5221 /* We have migrated, no longer consider this task hot */
5222 p
->se
.exec_start
= 0;
5225 static void task_dead_fair(struct task_struct
*p
)
5227 remove_entity_load_avg(&p
->se
);
5229 #endif /* CONFIG_SMP */
5231 static unsigned long
5232 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
5234 unsigned long gran
= sysctl_sched_wakeup_granularity
;
5237 * Since its curr running now, convert the gran from real-time
5238 * to virtual-time in his units.
5240 * By using 'se' instead of 'curr' we penalize light tasks, so
5241 * they get preempted easier. That is, if 'se' < 'curr' then
5242 * the resulting gran will be larger, therefore penalizing the
5243 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5244 * be smaller, again penalizing the lighter task.
5246 * This is especially important for buddies when the leftmost
5247 * task is higher priority than the buddy.
5249 return calc_delta_fair(gran
, se
);
5253 * Should 'se' preempt 'curr'.
5267 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
5269 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
5274 gran
= wakeup_gran(curr
, se
);
5281 static void set_last_buddy(struct sched_entity
*se
)
5283 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
5286 for_each_sched_entity(se
)
5287 cfs_rq_of(se
)->last
= se
;
5290 static void set_next_buddy(struct sched_entity
*se
)
5292 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
5295 for_each_sched_entity(se
)
5296 cfs_rq_of(se
)->next
= se
;
5299 static void set_skip_buddy(struct sched_entity
*se
)
5301 for_each_sched_entity(se
)
5302 cfs_rq_of(se
)->skip
= se
;
5306 * Preempt the current task with a newly woken task if needed:
5308 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
5310 struct task_struct
*curr
= rq
->curr
;
5311 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
5312 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
5313 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
5314 int next_buddy_marked
= 0;
5316 if (unlikely(se
== pse
))
5320 * This is possible from callers such as attach_tasks(), in which we
5321 * unconditionally check_prempt_curr() after an enqueue (which may have
5322 * lead to a throttle). This both saves work and prevents false
5323 * next-buddy nomination below.
5325 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
5328 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
5329 set_next_buddy(pse
);
5330 next_buddy_marked
= 1;
5334 * We can come here with TIF_NEED_RESCHED already set from new task
5337 * Note: this also catches the edge-case of curr being in a throttled
5338 * group (e.g. via set_curr_task), since update_curr() (in the
5339 * enqueue of curr) will have resulted in resched being set. This
5340 * prevents us from potentially nominating it as a false LAST_BUDDY
5343 if (test_tsk_need_resched(curr
))
5346 /* Idle tasks are by definition preempted by non-idle tasks. */
5347 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
5348 likely(p
->policy
!= SCHED_IDLE
))
5352 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5353 * is driven by the tick):
5355 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
5358 find_matching_se(&se
, &pse
);
5359 update_curr(cfs_rq_of(se
));
5361 if (wakeup_preempt_entity(se
, pse
) == 1) {
5363 * Bias pick_next to pick the sched entity that is
5364 * triggering this preemption.
5366 if (!next_buddy_marked
)
5367 set_next_buddy(pse
);
5376 * Only set the backward buddy when the current task is still
5377 * on the rq. This can happen when a wakeup gets interleaved
5378 * with schedule on the ->pre_schedule() or idle_balance()
5379 * point, either of which can * drop the rq lock.
5381 * Also, during early boot the idle thread is in the fair class,
5382 * for obvious reasons its a bad idea to schedule back to it.
5384 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
5387 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
5391 static struct task_struct
*
5392 pick_next_task_fair(struct rq
*rq
, struct task_struct
*prev
)
5394 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
5395 struct sched_entity
*se
;
5396 struct task_struct
*p
;
5400 #ifdef CONFIG_FAIR_GROUP_SCHED
5401 if (!cfs_rq
->nr_running
)
5404 if (prev
->sched_class
!= &fair_sched_class
)
5408 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5409 * likely that a next task is from the same cgroup as the current.
5411 * Therefore attempt to avoid putting and setting the entire cgroup
5412 * hierarchy, only change the part that actually changes.
5416 struct sched_entity
*curr
= cfs_rq
->curr
;
5419 * Since we got here without doing put_prev_entity() we also
5420 * have to consider cfs_rq->curr. If it is still a runnable
5421 * entity, update_curr() will update its vruntime, otherwise
5422 * forget we've ever seen it.
5426 update_curr(cfs_rq
);
5431 * This call to check_cfs_rq_runtime() will do the
5432 * throttle and dequeue its entity in the parent(s).
5433 * Therefore the 'simple' nr_running test will indeed
5436 if (unlikely(check_cfs_rq_runtime(cfs_rq
)))
5440 se
= pick_next_entity(cfs_rq
, curr
);
5441 cfs_rq
= group_cfs_rq(se
);
5447 * Since we haven't yet done put_prev_entity and if the selected task
5448 * is a different task than we started out with, try and touch the
5449 * least amount of cfs_rqs.
5452 struct sched_entity
*pse
= &prev
->se
;
5454 while (!(cfs_rq
= is_same_group(se
, pse
))) {
5455 int se_depth
= se
->depth
;
5456 int pse_depth
= pse
->depth
;
5458 if (se_depth
<= pse_depth
) {
5459 put_prev_entity(cfs_rq_of(pse
), pse
);
5460 pse
= parent_entity(pse
);
5462 if (se_depth
>= pse_depth
) {
5463 set_next_entity(cfs_rq_of(se
), se
);
5464 se
= parent_entity(se
);
5468 put_prev_entity(cfs_rq
, pse
);
5469 set_next_entity(cfs_rq
, se
);
5472 if (hrtick_enabled(rq
))
5473 hrtick_start_fair(rq
, p
);
5480 if (!cfs_rq
->nr_running
)
5483 put_prev_task(rq
, prev
);
5486 se
= pick_next_entity(cfs_rq
, NULL
);
5487 set_next_entity(cfs_rq
, se
);
5488 cfs_rq
= group_cfs_rq(se
);
5493 if (hrtick_enabled(rq
))
5494 hrtick_start_fair(rq
, p
);
5500 * This is OK, because current is on_cpu, which avoids it being picked
5501 * for load-balance and preemption/IRQs are still disabled avoiding
5502 * further scheduler activity on it and we're being very careful to
5503 * re-start the picking loop.
5505 lockdep_unpin_lock(&rq
->lock
);
5506 new_tasks
= idle_balance(rq
);
5507 lockdep_pin_lock(&rq
->lock
);
5509 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5510 * possible for any higher priority task to appear. In that case we
5511 * must re-start the pick_next_entity() loop.
5523 * Account for a descheduled task:
5525 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
5527 struct sched_entity
*se
= &prev
->se
;
5528 struct cfs_rq
*cfs_rq
;
5530 for_each_sched_entity(se
) {
5531 cfs_rq
= cfs_rq_of(se
);
5532 put_prev_entity(cfs_rq
, se
);
5537 * sched_yield() is very simple
5539 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5541 static void yield_task_fair(struct rq
*rq
)
5543 struct task_struct
*curr
= rq
->curr
;
5544 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
5545 struct sched_entity
*se
= &curr
->se
;
5548 * Are we the only task in the tree?
5550 if (unlikely(rq
->nr_running
== 1))
5553 clear_buddies(cfs_rq
, se
);
5555 if (curr
->policy
!= SCHED_BATCH
) {
5556 update_rq_clock(rq
);
5558 * Update run-time statistics of the 'current'.
5560 update_curr(cfs_rq
);
5562 * Tell update_rq_clock() that we've just updated,
5563 * so we don't do microscopic update in schedule()
5564 * and double the fastpath cost.
5566 rq_clock_skip_update(rq
, true);
5572 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
5574 struct sched_entity
*se
= &p
->se
;
5576 /* throttled hierarchies are not runnable */
5577 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
5580 /* Tell the scheduler that we'd really like pse to run next. */
5583 yield_task_fair(rq
);
5589 /**************************************************
5590 * Fair scheduling class load-balancing methods.
5594 * The purpose of load-balancing is to achieve the same basic fairness the
5595 * per-cpu scheduler provides, namely provide a proportional amount of compute
5596 * time to each task. This is expressed in the following equation:
5598 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5600 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5601 * W_i,0 is defined as:
5603 * W_i,0 = \Sum_j w_i,j (2)
5605 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5606 * is derived from the nice value as per prio_to_weight[].
5608 * The weight average is an exponential decay average of the instantaneous
5611 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5613 * C_i is the compute capacity of cpu i, typically it is the
5614 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5615 * can also include other factors [XXX].
5617 * To achieve this balance we define a measure of imbalance which follows
5618 * directly from (1):
5620 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5622 * We them move tasks around to minimize the imbalance. In the continuous
5623 * function space it is obvious this converges, in the discrete case we get
5624 * a few fun cases generally called infeasible weight scenarios.
5627 * - infeasible weights;
5628 * - local vs global optima in the discrete case. ]
5633 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5634 * for all i,j solution, we create a tree of cpus that follows the hardware
5635 * topology where each level pairs two lower groups (or better). This results
5636 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5637 * tree to only the first of the previous level and we decrease the frequency
5638 * of load-balance at each level inv. proportional to the number of cpus in
5644 * \Sum { --- * --- * 2^i } = O(n) (5)
5646 * `- size of each group
5647 * | | `- number of cpus doing load-balance
5649 * `- sum over all levels
5651 * Coupled with a limit on how many tasks we can migrate every balance pass,
5652 * this makes (5) the runtime complexity of the balancer.
5654 * An important property here is that each CPU is still (indirectly) connected
5655 * to every other cpu in at most O(log n) steps:
5657 * The adjacency matrix of the resulting graph is given by:
5660 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5663 * And you'll find that:
5665 * A^(log_2 n)_i,j != 0 for all i,j (7)
5667 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5668 * The task movement gives a factor of O(m), giving a convergence complexity
5671 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5676 * In order to avoid CPUs going idle while there's still work to do, new idle
5677 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5678 * tree itself instead of relying on other CPUs to bring it work.
5680 * This adds some complexity to both (5) and (8) but it reduces the total idle
5688 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5691 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5696 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5698 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5700 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5703 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5704 * rewrite all of this once again.]
5707 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
5709 enum fbq_type
{ regular
, remote
, all
};
5711 #define LBF_ALL_PINNED 0x01
5712 #define LBF_NEED_BREAK 0x02
5713 #define LBF_DST_PINNED 0x04
5714 #define LBF_SOME_PINNED 0x08
5717 struct sched_domain
*sd
;
5725 struct cpumask
*dst_grpmask
;
5727 enum cpu_idle_type idle
;
5729 /* The set of CPUs under consideration for load-balancing */
5730 struct cpumask
*cpus
;
5735 unsigned int loop_break
;
5736 unsigned int loop_max
;
5738 enum fbq_type fbq_type
;
5739 struct list_head tasks
;
5743 * Is this task likely cache-hot:
5745 static int task_hot(struct task_struct
*p
, struct lb_env
*env
)
5749 lockdep_assert_held(&env
->src_rq
->lock
);
5751 if (p
->sched_class
!= &fair_sched_class
)
5754 if (unlikely(p
->policy
== SCHED_IDLE
))
5758 * Buddy candidates are cache hot:
5760 if (sched_feat(CACHE_HOT_BUDDY
) && env
->dst_rq
->nr_running
&&
5761 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
5762 &p
->se
== cfs_rq_of(&p
->se
)->last
))
5765 if (sysctl_sched_migration_cost
== -1)
5767 if (sysctl_sched_migration_cost
== 0)
5770 delta
= rq_clock_task(env
->src_rq
) - p
->se
.exec_start
;
5772 return delta
< (s64
)sysctl_sched_migration_cost
;
5775 #ifdef CONFIG_NUMA_BALANCING
5777 * Returns 1, if task migration degrades locality
5778 * Returns 0, if task migration improves locality i.e migration preferred.
5779 * Returns -1, if task migration is not affected by locality.
5781 static int migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
5783 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
5784 unsigned long src_faults
, dst_faults
;
5785 int src_nid
, dst_nid
;
5787 if (!static_branch_likely(&sched_numa_balancing
))
5790 if (!p
->numa_faults
|| !(env
->sd
->flags
& SD_NUMA
))
5793 src_nid
= cpu_to_node(env
->src_cpu
);
5794 dst_nid
= cpu_to_node(env
->dst_cpu
);
5796 if (src_nid
== dst_nid
)
5799 /* Migrating away from the preferred node is always bad. */
5800 if (src_nid
== p
->numa_preferred_nid
) {
5801 if (env
->src_rq
->nr_running
> env
->src_rq
->nr_preferred_running
)
5807 /* Encourage migration to the preferred node. */
5808 if (dst_nid
== p
->numa_preferred_nid
)
5812 src_faults
= group_faults(p
, src_nid
);
5813 dst_faults
= group_faults(p
, dst_nid
);
5815 src_faults
= task_faults(p
, src_nid
);
5816 dst_faults
= task_faults(p
, dst_nid
);
5819 return dst_faults
< src_faults
;
5823 static inline int migrate_degrades_locality(struct task_struct
*p
,
5831 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5834 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
5838 lockdep_assert_held(&env
->src_rq
->lock
);
5841 * We do not migrate tasks that are:
5842 * 1) throttled_lb_pair, or
5843 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5844 * 3) running (obviously), or
5845 * 4) are cache-hot on their current CPU.
5847 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
5850 if (!cpumask_test_cpu(env
->dst_cpu
, tsk_cpus_allowed(p
))) {
5853 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
5855 env
->flags
|= LBF_SOME_PINNED
;
5858 * Remember if this task can be migrated to any other cpu in
5859 * our sched_group. We may want to revisit it if we couldn't
5860 * meet load balance goals by pulling other tasks on src_cpu.
5862 * Also avoid computing new_dst_cpu if we have already computed
5863 * one in current iteration.
5865 if (!env
->dst_grpmask
|| (env
->flags
& LBF_DST_PINNED
))
5868 /* Prevent to re-select dst_cpu via env's cpus */
5869 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
5870 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
))) {
5871 env
->flags
|= LBF_DST_PINNED
;
5872 env
->new_dst_cpu
= cpu
;
5880 /* Record that we found atleast one task that could run on dst_cpu */
5881 env
->flags
&= ~LBF_ALL_PINNED
;
5883 if (task_running(env
->src_rq
, p
)) {
5884 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
5889 * Aggressive migration if:
5890 * 1) destination numa is preferred
5891 * 2) task is cache cold, or
5892 * 3) too many balance attempts have failed.
5894 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
5895 if (tsk_cache_hot
== -1)
5896 tsk_cache_hot
= task_hot(p
, env
);
5898 if (tsk_cache_hot
<= 0 ||
5899 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
5900 if (tsk_cache_hot
== 1) {
5901 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
5902 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
5907 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
5912 * detach_task() -- detach the task for the migration specified in env
5914 static void detach_task(struct task_struct
*p
, struct lb_env
*env
)
5916 lockdep_assert_held(&env
->src_rq
->lock
);
5918 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
5919 deactivate_task(env
->src_rq
, p
, 0);
5920 set_task_cpu(p
, env
->dst_cpu
);
5924 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5925 * part of active balancing operations within "domain".
5927 * Returns a task if successful and NULL otherwise.
5929 static struct task_struct
*detach_one_task(struct lb_env
*env
)
5931 struct task_struct
*p
, *n
;
5933 lockdep_assert_held(&env
->src_rq
->lock
);
5935 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
5936 if (!can_migrate_task(p
, env
))
5939 detach_task(p
, env
);
5942 * Right now, this is only the second place where
5943 * lb_gained[env->idle] is updated (other is detach_tasks)
5944 * so we can safely collect stats here rather than
5945 * inside detach_tasks().
5947 schedstat_inc(env
->sd
, lb_gained
[env
->idle
]);
5953 static const unsigned int sched_nr_migrate_break
= 32;
5956 * detach_tasks() -- tries to detach up to imbalance weighted load from
5957 * busiest_rq, as part of a balancing operation within domain "sd".
5959 * Returns number of detached tasks if successful and 0 otherwise.
5961 static int detach_tasks(struct lb_env
*env
)
5963 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
5964 struct task_struct
*p
;
5968 lockdep_assert_held(&env
->src_rq
->lock
);
5970 if (env
->imbalance
<= 0)
5973 while (!list_empty(tasks
)) {
5975 * We don't want to steal all, otherwise we may be treated likewise,
5976 * which could at worst lead to a livelock crash.
5978 if (env
->idle
!= CPU_NOT_IDLE
&& env
->src_rq
->nr_running
<= 1)
5981 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
5984 /* We've more or less seen every task there is, call it quits */
5985 if (env
->loop
> env
->loop_max
)
5988 /* take a breather every nr_migrate tasks */
5989 if (env
->loop
> env
->loop_break
) {
5990 env
->loop_break
+= sched_nr_migrate_break
;
5991 env
->flags
|= LBF_NEED_BREAK
;
5995 if (!can_migrate_task(p
, env
))
5998 load
= task_h_load(p
);
6000 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
6003 if ((load
/ 2) > env
->imbalance
)
6006 detach_task(p
, env
);
6007 list_add(&p
->se
.group_node
, &env
->tasks
);
6010 env
->imbalance
-= load
;
6012 #ifdef CONFIG_PREEMPT
6014 * NEWIDLE balancing is a source of latency, so preemptible
6015 * kernels will stop after the first task is detached to minimize
6016 * the critical section.
6018 if (env
->idle
== CPU_NEWLY_IDLE
)
6023 * We only want to steal up to the prescribed amount of
6026 if (env
->imbalance
<= 0)
6031 list_move_tail(&p
->se
.group_node
, tasks
);
6035 * Right now, this is one of only two places we collect this stat
6036 * so we can safely collect detach_one_task() stats here rather
6037 * than inside detach_one_task().
6039 schedstat_add(env
->sd
, lb_gained
[env
->idle
], detached
);
6045 * attach_task() -- attach the task detached by detach_task() to its new rq.
6047 static void attach_task(struct rq
*rq
, struct task_struct
*p
)
6049 lockdep_assert_held(&rq
->lock
);
6051 BUG_ON(task_rq(p
) != rq
);
6052 activate_task(rq
, p
, 0);
6053 p
->on_rq
= TASK_ON_RQ_QUEUED
;
6054 check_preempt_curr(rq
, p
, 0);
6058 * attach_one_task() -- attaches the task returned from detach_one_task() to
6061 static void attach_one_task(struct rq
*rq
, struct task_struct
*p
)
6063 raw_spin_lock(&rq
->lock
);
6065 raw_spin_unlock(&rq
->lock
);
6069 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6072 static void attach_tasks(struct lb_env
*env
)
6074 struct list_head
*tasks
= &env
->tasks
;
6075 struct task_struct
*p
;
6077 raw_spin_lock(&env
->dst_rq
->lock
);
6079 while (!list_empty(tasks
)) {
6080 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
6081 list_del_init(&p
->se
.group_node
);
6083 attach_task(env
->dst_rq
, p
);
6086 raw_spin_unlock(&env
->dst_rq
->lock
);
6089 #ifdef CONFIG_FAIR_GROUP_SCHED
6090 static void update_blocked_averages(int cpu
)
6092 struct rq
*rq
= cpu_rq(cpu
);
6093 struct cfs_rq
*cfs_rq
;
6094 unsigned long flags
;
6096 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6097 update_rq_clock(rq
);
6100 * Iterates the task_group tree in a bottom up fashion, see
6101 * list_add_leaf_cfs_rq() for details.
6103 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
6104 /* throttled entities do not contribute to load */
6105 if (throttled_hierarchy(cfs_rq
))
6108 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq
), cfs_rq
))
6109 update_tg_load_avg(cfs_rq
, 0);
6111 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6115 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6116 * This needs to be done in a top-down fashion because the load of a child
6117 * group is a fraction of its parents load.
6119 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
6121 struct rq
*rq
= rq_of(cfs_rq
);
6122 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
6123 unsigned long now
= jiffies
;
6126 if (cfs_rq
->last_h_load_update
== now
)
6129 cfs_rq
->h_load_next
= NULL
;
6130 for_each_sched_entity(se
) {
6131 cfs_rq
= cfs_rq_of(se
);
6132 cfs_rq
->h_load_next
= se
;
6133 if (cfs_rq
->last_h_load_update
== now
)
6138 cfs_rq
->h_load
= cfs_rq_load_avg(cfs_rq
);
6139 cfs_rq
->last_h_load_update
= now
;
6142 while ((se
= cfs_rq
->h_load_next
) != NULL
) {
6143 load
= cfs_rq
->h_load
;
6144 load
= div64_ul(load
* se
->avg
.load_avg
,
6145 cfs_rq_load_avg(cfs_rq
) + 1);
6146 cfs_rq
= group_cfs_rq(se
);
6147 cfs_rq
->h_load
= load
;
6148 cfs_rq
->last_h_load_update
= now
;
6152 static unsigned long task_h_load(struct task_struct
*p
)
6154 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
6156 update_cfs_rq_h_load(cfs_rq
);
6157 return div64_ul(p
->se
.avg
.load_avg
* cfs_rq
->h_load
,
6158 cfs_rq_load_avg(cfs_rq
) + 1);
6161 static inline void update_blocked_averages(int cpu
)
6163 struct rq
*rq
= cpu_rq(cpu
);
6164 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
6165 unsigned long flags
;
6167 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6168 update_rq_clock(rq
);
6169 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq
), cfs_rq
);
6170 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6173 static unsigned long task_h_load(struct task_struct
*p
)
6175 return p
->se
.avg
.load_avg
;
6179 /********** Helpers for find_busiest_group ************************/
6188 * sg_lb_stats - stats of a sched_group required for load_balancing
6190 struct sg_lb_stats
{
6191 unsigned long avg_load
; /*Avg load across the CPUs of the group */
6192 unsigned long group_load
; /* Total load over the CPUs of the group */
6193 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
6194 unsigned long load_per_task
;
6195 unsigned long group_capacity
;
6196 unsigned long group_util
; /* Total utilization of the group */
6197 unsigned int sum_nr_running
; /* Nr tasks running in the group */
6198 unsigned int idle_cpus
;
6199 unsigned int group_weight
;
6200 enum group_type group_type
;
6201 int group_no_capacity
;
6202 #ifdef CONFIG_NUMA_BALANCING
6203 unsigned int nr_numa_running
;
6204 unsigned int nr_preferred_running
;
6209 * sd_lb_stats - Structure to store the statistics of a sched_domain
6210 * during load balancing.
6212 struct sd_lb_stats
{
6213 struct sched_group
*busiest
; /* Busiest group in this sd */
6214 struct sched_group
*local
; /* Local group in this sd */
6215 unsigned long total_load
; /* Total load of all groups in sd */
6216 unsigned long total_capacity
; /* Total capacity of all groups in sd */
6217 unsigned long avg_load
; /* Average load across all groups in sd */
6219 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
6220 struct sg_lb_stats local_stat
; /* Statistics of the local group */
6223 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
6226 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6227 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6228 * We must however clear busiest_stat::avg_load because
6229 * update_sd_pick_busiest() reads this before assignment.
6231 *sds
= (struct sd_lb_stats
){
6235 .total_capacity
= 0UL,
6238 .sum_nr_running
= 0,
6239 .group_type
= group_other
,
6245 * get_sd_load_idx - Obtain the load index for a given sched domain.
6246 * @sd: The sched_domain whose load_idx is to be obtained.
6247 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6249 * Return: The load index.
6251 static inline int get_sd_load_idx(struct sched_domain
*sd
,
6252 enum cpu_idle_type idle
)
6258 load_idx
= sd
->busy_idx
;
6261 case CPU_NEWLY_IDLE
:
6262 load_idx
= sd
->newidle_idx
;
6265 load_idx
= sd
->idle_idx
;
6272 static unsigned long scale_rt_capacity(int cpu
)
6274 struct rq
*rq
= cpu_rq(cpu
);
6275 u64 total
, used
, age_stamp
, avg
;
6279 * Since we're reading these variables without serialization make sure
6280 * we read them once before doing sanity checks on them.
6282 age_stamp
= READ_ONCE(rq
->age_stamp
);
6283 avg
= READ_ONCE(rq
->rt_avg
);
6284 delta
= __rq_clock_broken(rq
) - age_stamp
;
6286 if (unlikely(delta
< 0))
6289 total
= sched_avg_period() + delta
;
6291 used
= div_u64(avg
, total
);
6293 if (likely(used
< SCHED_CAPACITY_SCALE
))
6294 return SCHED_CAPACITY_SCALE
- used
;
6299 static void update_cpu_capacity(struct sched_domain
*sd
, int cpu
)
6301 unsigned long capacity
= arch_scale_cpu_capacity(sd
, cpu
);
6302 struct sched_group
*sdg
= sd
->groups
;
6304 cpu_rq(cpu
)->cpu_capacity_orig
= capacity
;
6306 capacity
*= scale_rt_capacity(cpu
);
6307 capacity
>>= SCHED_CAPACITY_SHIFT
;
6312 cpu_rq(cpu
)->cpu_capacity
= capacity
;
6313 sdg
->sgc
->capacity
= capacity
;
6316 void update_group_capacity(struct sched_domain
*sd
, int cpu
)
6318 struct sched_domain
*child
= sd
->child
;
6319 struct sched_group
*group
, *sdg
= sd
->groups
;
6320 unsigned long capacity
;
6321 unsigned long interval
;
6323 interval
= msecs_to_jiffies(sd
->balance_interval
);
6324 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
6325 sdg
->sgc
->next_update
= jiffies
+ interval
;
6328 update_cpu_capacity(sd
, cpu
);
6334 if (child
->flags
& SD_OVERLAP
) {
6336 * SD_OVERLAP domains cannot assume that child groups
6337 * span the current group.
6340 for_each_cpu(cpu
, sched_group_cpus(sdg
)) {
6341 struct sched_group_capacity
*sgc
;
6342 struct rq
*rq
= cpu_rq(cpu
);
6345 * build_sched_domains() -> init_sched_groups_capacity()
6346 * gets here before we've attached the domains to the
6349 * Use capacity_of(), which is set irrespective of domains
6350 * in update_cpu_capacity().
6352 * This avoids capacity from being 0 and
6353 * causing divide-by-zero issues on boot.
6355 if (unlikely(!rq
->sd
)) {
6356 capacity
+= capacity_of(cpu
);
6360 sgc
= rq
->sd
->groups
->sgc
;
6361 capacity
+= sgc
->capacity
;
6365 * !SD_OVERLAP domains can assume that child groups
6366 * span the current group.
6369 group
= child
->groups
;
6371 capacity
+= group
->sgc
->capacity
;
6372 group
= group
->next
;
6373 } while (group
!= child
->groups
);
6376 sdg
->sgc
->capacity
= capacity
;
6380 * Check whether the capacity of the rq has been noticeably reduced by side
6381 * activity. The imbalance_pct is used for the threshold.
6382 * Return true is the capacity is reduced
6385 check_cpu_capacity(struct rq
*rq
, struct sched_domain
*sd
)
6387 return ((rq
->cpu_capacity
* sd
->imbalance_pct
) <
6388 (rq
->cpu_capacity_orig
* 100));
6392 * Group imbalance indicates (and tries to solve) the problem where balancing
6393 * groups is inadequate due to tsk_cpus_allowed() constraints.
6395 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6396 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6399 * { 0 1 2 3 } { 4 5 6 7 }
6402 * If we were to balance group-wise we'd place two tasks in the first group and
6403 * two tasks in the second group. Clearly this is undesired as it will overload
6404 * cpu 3 and leave one of the cpus in the second group unused.
6406 * The current solution to this issue is detecting the skew in the first group
6407 * by noticing the lower domain failed to reach balance and had difficulty
6408 * moving tasks due to affinity constraints.
6410 * When this is so detected; this group becomes a candidate for busiest; see
6411 * update_sd_pick_busiest(). And calculate_imbalance() and
6412 * find_busiest_group() avoid some of the usual balance conditions to allow it
6413 * to create an effective group imbalance.
6415 * This is a somewhat tricky proposition since the next run might not find the
6416 * group imbalance and decide the groups need to be balanced again. A most
6417 * subtle and fragile situation.
6420 static inline int sg_imbalanced(struct sched_group
*group
)
6422 return group
->sgc
->imbalance
;
6426 * group_has_capacity returns true if the group has spare capacity that could
6427 * be used by some tasks.
6428 * We consider that a group has spare capacity if the * number of task is
6429 * smaller than the number of CPUs or if the utilization is lower than the
6430 * available capacity for CFS tasks.
6431 * For the latter, we use a threshold to stabilize the state, to take into
6432 * account the variance of the tasks' load and to return true if the available
6433 * capacity in meaningful for the load balancer.
6434 * As an example, an available capacity of 1% can appear but it doesn't make
6435 * any benefit for the load balance.
6438 group_has_capacity(struct lb_env
*env
, struct sg_lb_stats
*sgs
)
6440 if (sgs
->sum_nr_running
< sgs
->group_weight
)
6443 if ((sgs
->group_capacity
* 100) >
6444 (sgs
->group_util
* env
->sd
->imbalance_pct
))
6451 * group_is_overloaded returns true if the group has more tasks than it can
6453 * group_is_overloaded is not equals to !group_has_capacity because a group
6454 * with the exact right number of tasks, has no more spare capacity but is not
6455 * overloaded so both group_has_capacity and group_is_overloaded return
6459 group_is_overloaded(struct lb_env
*env
, struct sg_lb_stats
*sgs
)
6461 if (sgs
->sum_nr_running
<= sgs
->group_weight
)
6464 if ((sgs
->group_capacity
* 100) <
6465 (sgs
->group_util
* env
->sd
->imbalance_pct
))
6472 group_type
group_classify(struct sched_group
*group
,
6473 struct sg_lb_stats
*sgs
)
6475 if (sgs
->group_no_capacity
)
6476 return group_overloaded
;
6478 if (sg_imbalanced(group
))
6479 return group_imbalanced
;
6485 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6486 * @env: The load balancing environment.
6487 * @group: sched_group whose statistics are to be updated.
6488 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6489 * @local_group: Does group contain this_cpu.
6490 * @sgs: variable to hold the statistics for this group.
6491 * @overload: Indicate more than one runnable task for any CPU.
6493 static inline void update_sg_lb_stats(struct lb_env
*env
,
6494 struct sched_group
*group
, int load_idx
,
6495 int local_group
, struct sg_lb_stats
*sgs
,
6501 memset(sgs
, 0, sizeof(*sgs
));
6503 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
6504 struct rq
*rq
= cpu_rq(i
);
6506 /* Bias balancing toward cpus of our domain */
6508 load
= target_load(i
, load_idx
);
6510 load
= source_load(i
, load_idx
);
6512 sgs
->group_load
+= load
;
6513 sgs
->group_util
+= cpu_util(i
);
6514 sgs
->sum_nr_running
+= rq
->cfs
.h_nr_running
;
6516 nr_running
= rq
->nr_running
;
6520 #ifdef CONFIG_NUMA_BALANCING
6521 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
6522 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
6524 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
6526 * No need to call idle_cpu() if nr_running is not 0
6528 if (!nr_running
&& idle_cpu(i
))
6532 /* Adjust by relative CPU capacity of the group */
6533 sgs
->group_capacity
= group
->sgc
->capacity
;
6534 sgs
->avg_load
= (sgs
->group_load
*SCHED_CAPACITY_SCALE
) / sgs
->group_capacity
;
6536 if (sgs
->sum_nr_running
)
6537 sgs
->load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
6539 sgs
->group_weight
= group
->group_weight
;
6541 sgs
->group_no_capacity
= group_is_overloaded(env
, sgs
);
6542 sgs
->group_type
= group_classify(group
, sgs
);
6546 * update_sd_pick_busiest - return 1 on busiest group
6547 * @env: The load balancing environment.
6548 * @sds: sched_domain statistics
6549 * @sg: sched_group candidate to be checked for being the busiest
6550 * @sgs: sched_group statistics
6552 * Determine if @sg is a busier group than the previously selected
6555 * Return: %true if @sg is a busier group than the previously selected
6556 * busiest group. %false otherwise.
6558 static bool update_sd_pick_busiest(struct lb_env
*env
,
6559 struct sd_lb_stats
*sds
,
6560 struct sched_group
*sg
,
6561 struct sg_lb_stats
*sgs
)
6563 struct sg_lb_stats
*busiest
= &sds
->busiest_stat
;
6565 if (sgs
->group_type
> busiest
->group_type
)
6568 if (sgs
->group_type
< busiest
->group_type
)
6571 if (sgs
->avg_load
<= busiest
->avg_load
)
6574 /* This is the busiest node in its class. */
6575 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
6579 * ASYM_PACKING needs to move all the work to the lowest
6580 * numbered CPUs in the group, therefore mark all groups
6581 * higher than ourself as busy.
6583 if (sgs
->sum_nr_running
&& env
->dst_cpu
< group_first_cpu(sg
)) {
6587 if (group_first_cpu(sds
->busiest
) > group_first_cpu(sg
))
6594 #ifdef CONFIG_NUMA_BALANCING
6595 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
6597 if (sgs
->sum_nr_running
> sgs
->nr_numa_running
)
6599 if (sgs
->sum_nr_running
> sgs
->nr_preferred_running
)
6604 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
6606 if (rq
->nr_running
> rq
->nr_numa_running
)
6608 if (rq
->nr_running
> rq
->nr_preferred_running
)
6613 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
6618 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
6622 #endif /* CONFIG_NUMA_BALANCING */
6625 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6626 * @env: The load balancing environment.
6627 * @sds: variable to hold the statistics for this sched_domain.
6629 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6631 struct sched_domain
*child
= env
->sd
->child
;
6632 struct sched_group
*sg
= env
->sd
->groups
;
6633 struct sg_lb_stats tmp_sgs
;
6634 int load_idx
, prefer_sibling
= 0;
6635 bool overload
= false;
6637 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
6640 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
6643 struct sg_lb_stats
*sgs
= &tmp_sgs
;
6646 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_cpus(sg
));
6649 sgs
= &sds
->local_stat
;
6651 if (env
->idle
!= CPU_NEWLY_IDLE
||
6652 time_after_eq(jiffies
, sg
->sgc
->next_update
))
6653 update_group_capacity(env
->sd
, env
->dst_cpu
);
6656 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, sgs
,
6663 * In case the child domain prefers tasks go to siblings
6664 * first, lower the sg capacity so that we'll try
6665 * and move all the excess tasks away. We lower the capacity
6666 * of a group only if the local group has the capacity to fit
6667 * these excess tasks. The extra check prevents the case where
6668 * you always pull from the heaviest group when it is already
6669 * under-utilized (possible with a large weight task outweighs
6670 * the tasks on the system).
6672 if (prefer_sibling
&& sds
->local
&&
6673 group_has_capacity(env
, &sds
->local_stat
) &&
6674 (sgs
->sum_nr_running
> 1)) {
6675 sgs
->group_no_capacity
= 1;
6676 sgs
->group_type
= group_classify(sg
, sgs
);
6679 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
6681 sds
->busiest_stat
= *sgs
;
6685 /* Now, start updating sd_lb_stats */
6686 sds
->total_load
+= sgs
->group_load
;
6687 sds
->total_capacity
+= sgs
->group_capacity
;
6690 } while (sg
!= env
->sd
->groups
);
6692 if (env
->sd
->flags
& SD_NUMA
)
6693 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
6695 if (!env
->sd
->parent
) {
6696 /* update overload indicator if we are at root domain */
6697 if (env
->dst_rq
->rd
->overload
!= overload
)
6698 env
->dst_rq
->rd
->overload
= overload
;
6704 * check_asym_packing - Check to see if the group is packed into the
6707 * This is primarily intended to used at the sibling level. Some
6708 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6709 * case of POWER7, it can move to lower SMT modes only when higher
6710 * threads are idle. When in lower SMT modes, the threads will
6711 * perform better since they share less core resources. Hence when we
6712 * have idle threads, we want them to be the higher ones.
6714 * This packing function is run on idle threads. It checks to see if
6715 * the busiest CPU in this domain (core in the P7 case) has a higher
6716 * CPU number than the packing function is being run on. Here we are
6717 * assuming lower CPU number will be equivalent to lower a SMT thread
6720 * Return: 1 when packing is required and a task should be moved to
6721 * this CPU. The amount of the imbalance is returned in *imbalance.
6723 * @env: The load balancing environment.
6724 * @sds: Statistics of the sched_domain which is to be packed
6726 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6730 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
6736 busiest_cpu
= group_first_cpu(sds
->busiest
);
6737 if (env
->dst_cpu
> busiest_cpu
)
6740 env
->imbalance
= DIV_ROUND_CLOSEST(
6741 sds
->busiest_stat
.avg_load
* sds
->busiest_stat
.group_capacity
,
6742 SCHED_CAPACITY_SCALE
);
6748 * fix_small_imbalance - Calculate the minor imbalance that exists
6749 * amongst the groups of a sched_domain, during
6751 * @env: The load balancing environment.
6752 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6755 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6757 unsigned long tmp
, capa_now
= 0, capa_move
= 0;
6758 unsigned int imbn
= 2;
6759 unsigned long scaled_busy_load_per_task
;
6760 struct sg_lb_stats
*local
, *busiest
;
6762 local
= &sds
->local_stat
;
6763 busiest
= &sds
->busiest_stat
;
6765 if (!local
->sum_nr_running
)
6766 local
->load_per_task
= cpu_avg_load_per_task(env
->dst_cpu
);
6767 else if (busiest
->load_per_task
> local
->load_per_task
)
6770 scaled_busy_load_per_task
=
6771 (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
6772 busiest
->group_capacity
;
6774 if (busiest
->avg_load
+ scaled_busy_load_per_task
>=
6775 local
->avg_load
+ (scaled_busy_load_per_task
* imbn
)) {
6776 env
->imbalance
= busiest
->load_per_task
;
6781 * OK, we don't have enough imbalance to justify moving tasks,
6782 * however we may be able to increase total CPU capacity used by
6786 capa_now
+= busiest
->group_capacity
*
6787 min(busiest
->load_per_task
, busiest
->avg_load
);
6788 capa_now
+= local
->group_capacity
*
6789 min(local
->load_per_task
, local
->avg_load
);
6790 capa_now
/= SCHED_CAPACITY_SCALE
;
6792 /* Amount of load we'd subtract */
6793 if (busiest
->avg_load
> scaled_busy_load_per_task
) {
6794 capa_move
+= busiest
->group_capacity
*
6795 min(busiest
->load_per_task
,
6796 busiest
->avg_load
- scaled_busy_load_per_task
);
6799 /* Amount of load we'd add */
6800 if (busiest
->avg_load
* busiest
->group_capacity
<
6801 busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) {
6802 tmp
= (busiest
->avg_load
* busiest
->group_capacity
) /
6803 local
->group_capacity
;
6805 tmp
= (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
6806 local
->group_capacity
;
6808 capa_move
+= local
->group_capacity
*
6809 min(local
->load_per_task
, local
->avg_load
+ tmp
);
6810 capa_move
/= SCHED_CAPACITY_SCALE
;
6812 /* Move if we gain throughput */
6813 if (capa_move
> capa_now
)
6814 env
->imbalance
= busiest
->load_per_task
;
6818 * calculate_imbalance - Calculate the amount of imbalance present within the
6819 * groups of a given sched_domain during load balance.
6820 * @env: load balance environment
6821 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6823 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6825 unsigned long max_pull
, load_above_capacity
= ~0UL;
6826 struct sg_lb_stats
*local
, *busiest
;
6828 local
= &sds
->local_stat
;
6829 busiest
= &sds
->busiest_stat
;
6831 if (busiest
->group_type
== group_imbalanced
) {
6833 * In the group_imb case we cannot rely on group-wide averages
6834 * to ensure cpu-load equilibrium, look at wider averages. XXX
6836 busiest
->load_per_task
=
6837 min(busiest
->load_per_task
, sds
->avg_load
);
6841 * In the presence of smp nice balancing, certain scenarios can have
6842 * max load less than avg load(as we skip the groups at or below
6843 * its cpu_capacity, while calculating max_load..)
6845 if (busiest
->avg_load
<= sds
->avg_load
||
6846 local
->avg_load
>= sds
->avg_load
) {
6848 return fix_small_imbalance(env
, sds
);
6852 * If there aren't any idle cpus, avoid creating some.
6854 if (busiest
->group_type
== group_overloaded
&&
6855 local
->group_type
== group_overloaded
) {
6856 load_above_capacity
= busiest
->sum_nr_running
*
6858 if (load_above_capacity
> busiest
->group_capacity
)
6859 load_above_capacity
-= busiest
->group_capacity
;
6861 load_above_capacity
= ~0UL;
6865 * We're trying to get all the cpus to the average_load, so we don't
6866 * want to push ourselves above the average load, nor do we wish to
6867 * reduce the max loaded cpu below the average load. At the same time,
6868 * we also don't want to reduce the group load below the group capacity
6869 * (so that we can implement power-savings policies etc). Thus we look
6870 * for the minimum possible imbalance.
6872 max_pull
= min(busiest
->avg_load
- sds
->avg_load
, load_above_capacity
);
6874 /* How much load to actually move to equalise the imbalance */
6875 env
->imbalance
= min(
6876 max_pull
* busiest
->group_capacity
,
6877 (sds
->avg_load
- local
->avg_load
) * local
->group_capacity
6878 ) / SCHED_CAPACITY_SCALE
;
6881 * if *imbalance is less than the average load per runnable task
6882 * there is no guarantee that any tasks will be moved so we'll have
6883 * a think about bumping its value to force at least one task to be
6886 if (env
->imbalance
< busiest
->load_per_task
)
6887 return fix_small_imbalance(env
, sds
);
6890 /******* find_busiest_group() helpers end here *********************/
6893 * find_busiest_group - Returns the busiest group within the sched_domain
6894 * if there is an imbalance. If there isn't an imbalance, and
6895 * the user has opted for power-savings, it returns a group whose
6896 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6897 * such a group exists.
6899 * Also calculates the amount of weighted load which should be moved
6900 * to restore balance.
6902 * @env: The load balancing environment.
6904 * Return: - The busiest group if imbalance exists.
6905 * - If no imbalance and user has opted for power-savings balance,
6906 * return the least loaded group whose CPUs can be
6907 * put to idle by rebalancing its tasks onto our group.
6909 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
6911 struct sg_lb_stats
*local
, *busiest
;
6912 struct sd_lb_stats sds
;
6914 init_sd_lb_stats(&sds
);
6917 * Compute the various statistics relavent for load balancing at
6920 update_sd_lb_stats(env
, &sds
);
6921 local
= &sds
.local_stat
;
6922 busiest
= &sds
.busiest_stat
;
6924 /* ASYM feature bypasses nice load balance check */
6925 if ((env
->idle
== CPU_IDLE
|| env
->idle
== CPU_NEWLY_IDLE
) &&
6926 check_asym_packing(env
, &sds
))
6929 /* There is no busy sibling group to pull tasks from */
6930 if (!sds
.busiest
|| busiest
->sum_nr_running
== 0)
6933 sds
.avg_load
= (SCHED_CAPACITY_SCALE
* sds
.total_load
)
6934 / sds
.total_capacity
;
6937 * If the busiest group is imbalanced the below checks don't
6938 * work because they assume all things are equal, which typically
6939 * isn't true due to cpus_allowed constraints and the like.
6941 if (busiest
->group_type
== group_imbalanced
)
6944 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6945 if (env
->idle
== CPU_NEWLY_IDLE
&& group_has_capacity(env
, local
) &&
6946 busiest
->group_no_capacity
)
6950 * If the local group is busier than the selected busiest group
6951 * don't try and pull any tasks.
6953 if (local
->avg_load
>= busiest
->avg_load
)
6957 * Don't pull any tasks if this group is already above the domain
6960 if (local
->avg_load
>= sds
.avg_load
)
6963 if (env
->idle
== CPU_IDLE
) {
6965 * This cpu is idle. If the busiest group is not overloaded
6966 * and there is no imbalance between this and busiest group
6967 * wrt idle cpus, it is balanced. The imbalance becomes
6968 * significant if the diff is greater than 1 otherwise we
6969 * might end up to just move the imbalance on another group
6971 if ((busiest
->group_type
!= group_overloaded
) &&
6972 (local
->idle_cpus
<= (busiest
->idle_cpus
+ 1)))
6976 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6977 * imbalance_pct to be conservative.
6979 if (100 * busiest
->avg_load
<=
6980 env
->sd
->imbalance_pct
* local
->avg_load
)
6985 /* Looks like there is an imbalance. Compute it */
6986 calculate_imbalance(env
, &sds
);
6995 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6997 static struct rq
*find_busiest_queue(struct lb_env
*env
,
6998 struct sched_group
*group
)
7000 struct rq
*busiest
= NULL
, *rq
;
7001 unsigned long busiest_load
= 0, busiest_capacity
= 1;
7004 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
7005 unsigned long capacity
, wl
;
7009 rt
= fbq_classify_rq(rq
);
7012 * We classify groups/runqueues into three groups:
7013 * - regular: there are !numa tasks
7014 * - remote: there are numa tasks that run on the 'wrong' node
7015 * - all: there is no distinction
7017 * In order to avoid migrating ideally placed numa tasks,
7018 * ignore those when there's better options.
7020 * If we ignore the actual busiest queue to migrate another
7021 * task, the next balance pass can still reduce the busiest
7022 * queue by moving tasks around inside the node.
7024 * If we cannot move enough load due to this classification
7025 * the next pass will adjust the group classification and
7026 * allow migration of more tasks.
7028 * Both cases only affect the total convergence complexity.
7030 if (rt
> env
->fbq_type
)
7033 capacity
= capacity_of(i
);
7035 wl
= weighted_cpuload(i
);
7038 * When comparing with imbalance, use weighted_cpuload()
7039 * which is not scaled with the cpu capacity.
7042 if (rq
->nr_running
== 1 && wl
> env
->imbalance
&&
7043 !check_cpu_capacity(rq
, env
->sd
))
7047 * For the load comparisons with the other cpu's, consider
7048 * the weighted_cpuload() scaled with the cpu capacity, so
7049 * that the load can be moved away from the cpu that is
7050 * potentially running at a lower capacity.
7052 * Thus we're looking for max(wl_i / capacity_i), crosswise
7053 * multiplication to rid ourselves of the division works out
7054 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7055 * our previous maximum.
7057 if (wl
* busiest_capacity
> busiest_load
* capacity
) {
7059 busiest_capacity
= capacity
;
7068 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7069 * so long as it is large enough.
7071 #define MAX_PINNED_INTERVAL 512
7073 /* Working cpumask for load_balance and load_balance_newidle. */
7074 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
7076 static int need_active_balance(struct lb_env
*env
)
7078 struct sched_domain
*sd
= env
->sd
;
7080 if (env
->idle
== CPU_NEWLY_IDLE
) {
7083 * ASYM_PACKING needs to force migrate tasks from busy but
7084 * higher numbered CPUs in order to pack all tasks in the
7085 * lowest numbered CPUs.
7087 if ((sd
->flags
& SD_ASYM_PACKING
) && env
->src_cpu
> env
->dst_cpu
)
7092 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7093 * It's worth migrating the task if the src_cpu's capacity is reduced
7094 * because of other sched_class or IRQs if more capacity stays
7095 * available on dst_cpu.
7097 if ((env
->idle
!= CPU_NOT_IDLE
) &&
7098 (env
->src_rq
->cfs
.h_nr_running
== 1)) {
7099 if ((check_cpu_capacity(env
->src_rq
, sd
)) &&
7100 (capacity_of(env
->src_cpu
)*sd
->imbalance_pct
< capacity_of(env
->dst_cpu
)*100))
7104 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
7107 static int active_load_balance_cpu_stop(void *data
);
7109 static int should_we_balance(struct lb_env
*env
)
7111 struct sched_group
*sg
= env
->sd
->groups
;
7112 struct cpumask
*sg_cpus
, *sg_mask
;
7113 int cpu
, balance_cpu
= -1;
7116 * In the newly idle case, we will allow all the cpu's
7117 * to do the newly idle load balance.
7119 if (env
->idle
== CPU_NEWLY_IDLE
)
7122 sg_cpus
= sched_group_cpus(sg
);
7123 sg_mask
= sched_group_mask(sg
);
7124 /* Try to find first idle cpu */
7125 for_each_cpu_and(cpu
, sg_cpus
, env
->cpus
) {
7126 if (!cpumask_test_cpu(cpu
, sg_mask
) || !idle_cpu(cpu
))
7133 if (balance_cpu
== -1)
7134 balance_cpu
= group_balance_cpu(sg
);
7137 * First idle cpu or the first cpu(busiest) in this sched group
7138 * is eligible for doing load balancing at this and above domains.
7140 return balance_cpu
== env
->dst_cpu
;
7144 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7145 * tasks if there is an imbalance.
7147 static int load_balance(int this_cpu
, struct rq
*this_rq
,
7148 struct sched_domain
*sd
, enum cpu_idle_type idle
,
7149 int *continue_balancing
)
7151 int ld_moved
, cur_ld_moved
, active_balance
= 0;
7152 struct sched_domain
*sd_parent
= sd
->parent
;
7153 struct sched_group
*group
;
7155 unsigned long flags
;
7156 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(load_balance_mask
);
7158 struct lb_env env
= {
7160 .dst_cpu
= this_cpu
,
7162 .dst_grpmask
= sched_group_cpus(sd
->groups
),
7164 .loop_break
= sched_nr_migrate_break
,
7167 .tasks
= LIST_HEAD_INIT(env
.tasks
),
7171 * For NEWLY_IDLE load_balancing, we don't need to consider
7172 * other cpus in our group
7174 if (idle
== CPU_NEWLY_IDLE
)
7175 env
.dst_grpmask
= NULL
;
7177 cpumask_copy(cpus
, cpu_active_mask
);
7179 schedstat_inc(sd
, lb_count
[idle
]);
7182 if (!should_we_balance(&env
)) {
7183 *continue_balancing
= 0;
7187 group
= find_busiest_group(&env
);
7189 schedstat_inc(sd
, lb_nobusyg
[idle
]);
7193 busiest
= find_busiest_queue(&env
, group
);
7195 schedstat_inc(sd
, lb_nobusyq
[idle
]);
7199 BUG_ON(busiest
== env
.dst_rq
);
7201 schedstat_add(sd
, lb_imbalance
[idle
], env
.imbalance
);
7203 env
.src_cpu
= busiest
->cpu
;
7204 env
.src_rq
= busiest
;
7207 if (busiest
->nr_running
> 1) {
7209 * Attempt to move tasks. If find_busiest_group has found
7210 * an imbalance but busiest->nr_running <= 1, the group is
7211 * still unbalanced. ld_moved simply stays zero, so it is
7212 * correctly treated as an imbalance.
7214 env
.flags
|= LBF_ALL_PINNED
;
7215 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
7218 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
7221 * cur_ld_moved - load moved in current iteration
7222 * ld_moved - cumulative load moved across iterations
7224 cur_ld_moved
= detach_tasks(&env
);
7227 * We've detached some tasks from busiest_rq. Every
7228 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7229 * unlock busiest->lock, and we are able to be sure
7230 * that nobody can manipulate the tasks in parallel.
7231 * See task_rq_lock() family for the details.
7234 raw_spin_unlock(&busiest
->lock
);
7238 ld_moved
+= cur_ld_moved
;
7241 local_irq_restore(flags
);
7243 if (env
.flags
& LBF_NEED_BREAK
) {
7244 env
.flags
&= ~LBF_NEED_BREAK
;
7249 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7250 * us and move them to an alternate dst_cpu in our sched_group
7251 * where they can run. The upper limit on how many times we
7252 * iterate on same src_cpu is dependent on number of cpus in our
7255 * This changes load balance semantics a bit on who can move
7256 * load to a given_cpu. In addition to the given_cpu itself
7257 * (or a ilb_cpu acting on its behalf where given_cpu is
7258 * nohz-idle), we now have balance_cpu in a position to move
7259 * load to given_cpu. In rare situations, this may cause
7260 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7261 * _independently_ and at _same_ time to move some load to
7262 * given_cpu) causing exceess load to be moved to given_cpu.
7263 * This however should not happen so much in practice and
7264 * moreover subsequent load balance cycles should correct the
7265 * excess load moved.
7267 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
7269 /* Prevent to re-select dst_cpu via env's cpus */
7270 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
7272 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
7273 env
.dst_cpu
= env
.new_dst_cpu
;
7274 env
.flags
&= ~LBF_DST_PINNED
;
7276 env
.loop_break
= sched_nr_migrate_break
;
7279 * Go back to "more_balance" rather than "redo" since we
7280 * need to continue with same src_cpu.
7286 * We failed to reach balance because of affinity.
7289 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
7291 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0)
7292 *group_imbalance
= 1;
7295 /* All tasks on this runqueue were pinned by CPU affinity */
7296 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
7297 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
7298 if (!cpumask_empty(cpus
)) {
7300 env
.loop_break
= sched_nr_migrate_break
;
7303 goto out_all_pinned
;
7308 schedstat_inc(sd
, lb_failed
[idle
]);
7310 * Increment the failure counter only on periodic balance.
7311 * We do not want newidle balance, which can be very
7312 * frequent, pollute the failure counter causing
7313 * excessive cache_hot migrations and active balances.
7315 if (idle
!= CPU_NEWLY_IDLE
)
7316 sd
->nr_balance_failed
++;
7318 if (need_active_balance(&env
)) {
7319 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
7321 /* don't kick the active_load_balance_cpu_stop,
7322 * if the curr task on busiest cpu can't be
7325 if (!cpumask_test_cpu(this_cpu
,
7326 tsk_cpus_allowed(busiest
->curr
))) {
7327 raw_spin_unlock_irqrestore(&busiest
->lock
,
7329 env
.flags
|= LBF_ALL_PINNED
;
7330 goto out_one_pinned
;
7334 * ->active_balance synchronizes accesses to
7335 * ->active_balance_work. Once set, it's cleared
7336 * only after active load balance is finished.
7338 if (!busiest
->active_balance
) {
7339 busiest
->active_balance
= 1;
7340 busiest
->push_cpu
= this_cpu
;
7343 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
7345 if (active_balance
) {
7346 stop_one_cpu_nowait(cpu_of(busiest
),
7347 active_load_balance_cpu_stop
, busiest
,
7348 &busiest
->active_balance_work
);
7352 * We've kicked active balancing, reset the failure
7355 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
7358 sd
->nr_balance_failed
= 0;
7360 if (likely(!active_balance
)) {
7361 /* We were unbalanced, so reset the balancing interval */
7362 sd
->balance_interval
= sd
->min_interval
;
7365 * If we've begun active balancing, start to back off. This
7366 * case may not be covered by the all_pinned logic if there
7367 * is only 1 task on the busy runqueue (because we don't call
7370 if (sd
->balance_interval
< sd
->max_interval
)
7371 sd
->balance_interval
*= 2;
7378 * We reach balance although we may have faced some affinity
7379 * constraints. Clear the imbalance flag if it was set.
7382 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
7384 if (*group_imbalance
)
7385 *group_imbalance
= 0;
7390 * We reach balance because all tasks are pinned at this level so
7391 * we can't migrate them. Let the imbalance flag set so parent level
7392 * can try to migrate them.
7394 schedstat_inc(sd
, lb_balanced
[idle
]);
7396 sd
->nr_balance_failed
= 0;
7399 /* tune up the balancing interval */
7400 if (((env
.flags
& LBF_ALL_PINNED
) &&
7401 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
7402 (sd
->balance_interval
< sd
->max_interval
))
7403 sd
->balance_interval
*= 2;
7410 static inline unsigned long
7411 get_sd_balance_interval(struct sched_domain
*sd
, int cpu_busy
)
7413 unsigned long interval
= sd
->balance_interval
;
7416 interval
*= sd
->busy_factor
;
7418 /* scale ms to jiffies */
7419 interval
= msecs_to_jiffies(interval
);
7420 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
7426 update_next_balance(struct sched_domain
*sd
, int cpu_busy
, unsigned long *next_balance
)
7428 unsigned long interval
, next
;
7430 interval
= get_sd_balance_interval(sd
, cpu_busy
);
7431 next
= sd
->last_balance
+ interval
;
7433 if (time_after(*next_balance
, next
))
7434 *next_balance
= next
;
7438 * idle_balance is called by schedule() if this_cpu is about to become
7439 * idle. Attempts to pull tasks from other CPUs.
7441 static int idle_balance(struct rq
*this_rq
)
7443 unsigned long next_balance
= jiffies
+ HZ
;
7444 int this_cpu
= this_rq
->cpu
;
7445 struct sched_domain
*sd
;
7446 int pulled_task
= 0;
7450 * We must set idle_stamp _before_ calling idle_balance(), such that we
7451 * measure the duration of idle_balance() as idle time.
7453 this_rq
->idle_stamp
= rq_clock(this_rq
);
7455 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
||
7456 !this_rq
->rd
->overload
) {
7458 sd
= rcu_dereference_check_sched_domain(this_rq
->sd
);
7460 update_next_balance(sd
, 0, &next_balance
);
7466 raw_spin_unlock(&this_rq
->lock
);
7468 update_blocked_averages(this_cpu
);
7470 for_each_domain(this_cpu
, sd
) {
7471 int continue_balancing
= 1;
7472 u64 t0
, domain_cost
;
7474 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7477 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
) {
7478 update_next_balance(sd
, 0, &next_balance
);
7482 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
7483 t0
= sched_clock_cpu(this_cpu
);
7485 pulled_task
= load_balance(this_cpu
, this_rq
,
7487 &continue_balancing
);
7489 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
7490 if (domain_cost
> sd
->max_newidle_lb_cost
)
7491 sd
->max_newidle_lb_cost
= domain_cost
;
7493 curr_cost
+= domain_cost
;
7496 update_next_balance(sd
, 0, &next_balance
);
7499 * Stop searching for tasks to pull if there are
7500 * now runnable tasks on this rq.
7502 if (pulled_task
|| this_rq
->nr_running
> 0)
7507 raw_spin_lock(&this_rq
->lock
);
7509 if (curr_cost
> this_rq
->max_idle_balance_cost
)
7510 this_rq
->max_idle_balance_cost
= curr_cost
;
7513 * While browsing the domains, we released the rq lock, a task could
7514 * have been enqueued in the meantime. Since we're not going idle,
7515 * pretend we pulled a task.
7517 if (this_rq
->cfs
.h_nr_running
&& !pulled_task
)
7521 /* Move the next balance forward */
7522 if (time_after(this_rq
->next_balance
, next_balance
))
7523 this_rq
->next_balance
= next_balance
;
7525 /* Is there a task of a high priority class? */
7526 if (this_rq
->nr_running
!= this_rq
->cfs
.h_nr_running
)
7530 this_rq
->idle_stamp
= 0;
7536 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7537 * running tasks off the busiest CPU onto idle CPUs. It requires at
7538 * least 1 task to be running on each physical CPU where possible, and
7539 * avoids physical / logical imbalances.
7541 static int active_load_balance_cpu_stop(void *data
)
7543 struct rq
*busiest_rq
= data
;
7544 int busiest_cpu
= cpu_of(busiest_rq
);
7545 int target_cpu
= busiest_rq
->push_cpu
;
7546 struct rq
*target_rq
= cpu_rq(target_cpu
);
7547 struct sched_domain
*sd
;
7548 struct task_struct
*p
= NULL
;
7550 raw_spin_lock_irq(&busiest_rq
->lock
);
7552 /* make sure the requested cpu hasn't gone down in the meantime */
7553 if (unlikely(busiest_cpu
!= smp_processor_id() ||
7554 !busiest_rq
->active_balance
))
7557 /* Is there any task to move? */
7558 if (busiest_rq
->nr_running
<= 1)
7562 * This condition is "impossible", if it occurs
7563 * we need to fix it. Originally reported by
7564 * Bjorn Helgaas on a 128-cpu setup.
7566 BUG_ON(busiest_rq
== target_rq
);
7568 /* Search for an sd spanning us and the target CPU. */
7570 for_each_domain(target_cpu
, sd
) {
7571 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
7572 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
7577 struct lb_env env
= {
7579 .dst_cpu
= target_cpu
,
7580 .dst_rq
= target_rq
,
7581 .src_cpu
= busiest_rq
->cpu
,
7582 .src_rq
= busiest_rq
,
7586 schedstat_inc(sd
, alb_count
);
7588 p
= detach_one_task(&env
);
7590 schedstat_inc(sd
, alb_pushed
);
7592 schedstat_inc(sd
, alb_failed
);
7596 busiest_rq
->active_balance
= 0;
7597 raw_spin_unlock(&busiest_rq
->lock
);
7600 attach_one_task(target_rq
, p
);
7607 static inline int on_null_domain(struct rq
*rq
)
7609 return unlikely(!rcu_dereference_sched(rq
->sd
));
7612 #ifdef CONFIG_NO_HZ_COMMON
7614 * idle load balancing details
7615 * - When one of the busy CPUs notice that there may be an idle rebalancing
7616 * needed, they will kick the idle load balancer, which then does idle
7617 * load balancing for all the idle CPUs.
7620 cpumask_var_t idle_cpus_mask
;
7622 unsigned long next_balance
; /* in jiffy units */
7623 } nohz ____cacheline_aligned
;
7625 static inline int find_new_ilb(void)
7627 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
7629 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
7636 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7637 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7638 * CPU (if there is one).
7640 static void nohz_balancer_kick(void)
7644 nohz
.next_balance
++;
7646 ilb_cpu
= find_new_ilb();
7648 if (ilb_cpu
>= nr_cpu_ids
)
7651 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
7654 * Use smp_send_reschedule() instead of resched_cpu().
7655 * This way we generate a sched IPI on the target cpu which
7656 * is idle. And the softirq performing nohz idle load balance
7657 * will be run before returning from the IPI.
7659 smp_send_reschedule(ilb_cpu
);
7663 static inline void nohz_balance_exit_idle(int cpu
)
7665 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
7667 * Completely isolated CPUs don't ever set, so we must test.
7669 if (likely(cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))) {
7670 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
7671 atomic_dec(&nohz
.nr_cpus
);
7673 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
7677 static inline void set_cpu_sd_state_busy(void)
7679 struct sched_domain
*sd
;
7680 int cpu
= smp_processor_id();
7683 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7685 if (!sd
|| !sd
->nohz_idle
)
7689 atomic_inc(&sd
->groups
->sgc
->nr_busy_cpus
);
7694 void set_cpu_sd_state_idle(void)
7696 struct sched_domain
*sd
;
7697 int cpu
= smp_processor_id();
7700 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7702 if (!sd
|| sd
->nohz_idle
)
7706 atomic_dec(&sd
->groups
->sgc
->nr_busy_cpus
);
7712 * This routine will record that the cpu is going idle with tick stopped.
7713 * This info will be used in performing idle load balancing in the future.
7715 void nohz_balance_enter_idle(int cpu
)
7718 * If this cpu is going down, then nothing needs to be done.
7720 if (!cpu_active(cpu
))
7723 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
7727 * If we're a completely isolated CPU, we don't play.
7729 if (on_null_domain(cpu_rq(cpu
)))
7732 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
7733 atomic_inc(&nohz
.nr_cpus
);
7734 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
7737 static int sched_ilb_notifier(struct notifier_block
*nfb
,
7738 unsigned long action
, void *hcpu
)
7740 switch (action
& ~CPU_TASKS_FROZEN
) {
7742 nohz_balance_exit_idle(smp_processor_id());
7750 static DEFINE_SPINLOCK(balancing
);
7753 * Scale the max load_balance interval with the number of CPUs in the system.
7754 * This trades load-balance latency on larger machines for less cross talk.
7756 void update_max_interval(void)
7758 max_load_balance_interval
= HZ
*num_online_cpus()/10;
7762 * It checks each scheduling domain to see if it is due to be balanced,
7763 * and initiates a balancing operation if so.
7765 * Balancing parameters are set up in init_sched_domains.
7767 static void rebalance_domains(struct rq
*rq
, enum cpu_idle_type idle
)
7769 int continue_balancing
= 1;
7771 unsigned long interval
;
7772 struct sched_domain
*sd
;
7773 /* Earliest time when we have to do rebalance again */
7774 unsigned long next_balance
= jiffies
+ 60*HZ
;
7775 int update_next_balance
= 0;
7776 int need_serialize
, need_decay
= 0;
7779 update_blocked_averages(cpu
);
7782 for_each_domain(cpu
, sd
) {
7784 * Decay the newidle max times here because this is a regular
7785 * visit to all the domains. Decay ~1% per second.
7787 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
7788 sd
->max_newidle_lb_cost
=
7789 (sd
->max_newidle_lb_cost
* 253) / 256;
7790 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
7793 max_cost
+= sd
->max_newidle_lb_cost
;
7795 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7799 * Stop the load balance at this level. There is another
7800 * CPU in our sched group which is doing load balancing more
7803 if (!continue_balancing
) {
7809 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7811 need_serialize
= sd
->flags
& SD_SERIALIZE
;
7812 if (need_serialize
) {
7813 if (!spin_trylock(&balancing
))
7817 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
7818 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
7820 * The LBF_DST_PINNED logic could have changed
7821 * env->dst_cpu, so we can't know our idle
7822 * state even if we migrated tasks. Update it.
7824 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
7826 sd
->last_balance
= jiffies
;
7827 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7830 spin_unlock(&balancing
);
7832 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
7833 next_balance
= sd
->last_balance
+ interval
;
7834 update_next_balance
= 1;
7839 * Ensure the rq-wide value also decays but keep it at a
7840 * reasonable floor to avoid funnies with rq->avg_idle.
7842 rq
->max_idle_balance_cost
=
7843 max((u64
)sysctl_sched_migration_cost
, max_cost
);
7848 * next_balance will be updated only when there is a need.
7849 * When the cpu is attached to null domain for ex, it will not be
7852 if (likely(update_next_balance
)) {
7853 rq
->next_balance
= next_balance
;
7855 #ifdef CONFIG_NO_HZ_COMMON
7857 * If this CPU has been elected to perform the nohz idle
7858 * balance. Other idle CPUs have already rebalanced with
7859 * nohz_idle_balance() and nohz.next_balance has been
7860 * updated accordingly. This CPU is now running the idle load
7861 * balance for itself and we need to update the
7862 * nohz.next_balance accordingly.
7864 if ((idle
== CPU_IDLE
) && time_after(nohz
.next_balance
, rq
->next_balance
))
7865 nohz
.next_balance
= rq
->next_balance
;
7870 #ifdef CONFIG_NO_HZ_COMMON
7872 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7873 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7875 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
7877 int this_cpu
= this_rq
->cpu
;
7880 /* Earliest time when we have to do rebalance again */
7881 unsigned long next_balance
= jiffies
+ 60*HZ
;
7882 int update_next_balance
= 0;
7884 if (idle
!= CPU_IDLE
||
7885 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
7888 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
7889 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
7893 * If this cpu gets work to do, stop the load balancing
7894 * work being done for other cpus. Next load
7895 * balancing owner will pick it up.
7900 rq
= cpu_rq(balance_cpu
);
7903 * If time for next balance is due,
7906 if (time_after_eq(jiffies
, rq
->next_balance
)) {
7907 raw_spin_lock_irq(&rq
->lock
);
7908 update_rq_clock(rq
);
7909 update_idle_cpu_load(rq
);
7910 raw_spin_unlock_irq(&rq
->lock
);
7911 rebalance_domains(rq
, CPU_IDLE
);
7914 if (time_after(next_balance
, rq
->next_balance
)) {
7915 next_balance
= rq
->next_balance
;
7916 update_next_balance
= 1;
7921 * next_balance will be updated only when there is a need.
7922 * When the CPU is attached to null domain for ex, it will not be
7925 if (likely(update_next_balance
))
7926 nohz
.next_balance
= next_balance
;
7928 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
7932 * Current heuristic for kicking the idle load balancer in the presence
7933 * of an idle cpu in the system.
7934 * - This rq has more than one task.
7935 * - This rq has at least one CFS task and the capacity of the CPU is
7936 * significantly reduced because of RT tasks or IRQs.
7937 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
7938 * multiple busy cpu.
7939 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7940 * domain span are idle.
7942 static inline bool nohz_kick_needed(struct rq
*rq
)
7944 unsigned long now
= jiffies
;
7945 struct sched_domain
*sd
;
7946 struct sched_group_capacity
*sgc
;
7947 int nr_busy
, cpu
= rq
->cpu
;
7950 if (unlikely(rq
->idle_balance
))
7954 * We may be recently in ticked or tickless idle mode. At the first
7955 * busy tick after returning from idle, we will update the busy stats.
7957 set_cpu_sd_state_busy();
7958 nohz_balance_exit_idle(cpu
);
7961 * None are in tickless mode and hence no need for NOHZ idle load
7964 if (likely(!atomic_read(&nohz
.nr_cpus
)))
7967 if (time_before(now
, nohz
.next_balance
))
7970 if (rq
->nr_running
>= 2)
7974 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7976 sgc
= sd
->groups
->sgc
;
7977 nr_busy
= atomic_read(&sgc
->nr_busy_cpus
);
7986 sd
= rcu_dereference(rq
->sd
);
7988 if ((rq
->cfs
.h_nr_running
>= 1) &&
7989 check_cpu_capacity(rq
, sd
)) {
7995 sd
= rcu_dereference(per_cpu(sd_asym
, cpu
));
7996 if (sd
&& (cpumask_first_and(nohz
.idle_cpus_mask
,
7997 sched_domain_span(sd
)) < cpu
)) {
8007 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
) { }
8011 * run_rebalance_domains is triggered when needed from the scheduler tick.
8012 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8014 static void run_rebalance_domains(struct softirq_action
*h
)
8016 struct rq
*this_rq
= this_rq();
8017 enum cpu_idle_type idle
= this_rq
->idle_balance
?
8018 CPU_IDLE
: CPU_NOT_IDLE
;
8021 * If this cpu has a pending nohz_balance_kick, then do the
8022 * balancing on behalf of the other idle cpus whose ticks are
8023 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8024 * give the idle cpus a chance to load balance. Else we may
8025 * load balance only within the local sched_domain hierarchy
8026 * and abort nohz_idle_balance altogether if we pull some load.
8028 nohz_idle_balance(this_rq
, idle
);
8029 rebalance_domains(this_rq
, idle
);
8033 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8035 void trigger_load_balance(struct rq
*rq
)
8037 /* Don't need to rebalance while attached to NULL domain */
8038 if (unlikely(on_null_domain(rq
)))
8041 if (time_after_eq(jiffies
, rq
->next_balance
))
8042 raise_softirq(SCHED_SOFTIRQ
);
8043 #ifdef CONFIG_NO_HZ_COMMON
8044 if (nohz_kick_needed(rq
))
8045 nohz_balancer_kick();
8049 static void rq_online_fair(struct rq
*rq
)
8053 update_runtime_enabled(rq
);
8056 static void rq_offline_fair(struct rq
*rq
)
8060 /* Ensure any throttled groups are reachable by pick_next_task */
8061 unthrottle_offline_cfs_rqs(rq
);
8064 #endif /* CONFIG_SMP */
8067 * scheduler tick hitting a task of our scheduling class:
8069 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
8071 struct cfs_rq
*cfs_rq
;
8072 struct sched_entity
*se
= &curr
->se
;
8074 for_each_sched_entity(se
) {
8075 cfs_rq
= cfs_rq_of(se
);
8076 entity_tick(cfs_rq
, se
, queued
);
8079 if (static_branch_unlikely(&sched_numa_balancing
))
8080 task_tick_numa(rq
, curr
);
8084 * called on fork with the child task as argument from the parent's context
8085 * - child not yet on the tasklist
8086 * - preemption disabled
8088 static void task_fork_fair(struct task_struct
*p
)
8090 struct cfs_rq
*cfs_rq
;
8091 struct sched_entity
*se
= &p
->se
, *curr
;
8092 int this_cpu
= smp_processor_id();
8093 struct rq
*rq
= this_rq();
8094 unsigned long flags
;
8096 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8098 update_rq_clock(rq
);
8100 cfs_rq
= task_cfs_rq(current
);
8101 curr
= cfs_rq
->curr
;
8104 * Not only the cpu but also the task_group of the parent might have
8105 * been changed after parent->se.parent,cfs_rq were copied to
8106 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
8107 * of child point to valid ones.
8110 __set_task_cpu(p
, this_cpu
);
8113 update_curr(cfs_rq
);
8116 se
->vruntime
= curr
->vruntime
;
8117 place_entity(cfs_rq
, se
, 1);
8119 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
8121 * Upon rescheduling, sched_class::put_prev_task() will place
8122 * 'current' within the tree based on its new key value.
8124 swap(curr
->vruntime
, se
->vruntime
);
8128 se
->vruntime
-= cfs_rq
->min_vruntime
;
8130 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8134 * Priority of the task has changed. Check to see if we preempt
8138 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
8140 if (!task_on_rq_queued(p
))
8144 * Reschedule if we are currently running on this runqueue and
8145 * our priority decreased, or if we are not currently running on
8146 * this runqueue and our priority is higher than the current's
8148 if (rq
->curr
== p
) {
8149 if (p
->prio
> oldprio
)
8152 check_preempt_curr(rq
, p
, 0);
8155 static inline bool vruntime_normalized(struct task_struct
*p
)
8157 struct sched_entity
*se
= &p
->se
;
8160 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8161 * the dequeue_entity(.flags=0) will already have normalized the
8168 * When !on_rq, vruntime of the task has usually NOT been normalized.
8169 * But there are some cases where it has already been normalized:
8171 * - A forked child which is waiting for being woken up by
8172 * wake_up_new_task().
8173 * - A task which has been woken up by try_to_wake_up() and
8174 * waiting for actually being woken up by sched_ttwu_pending().
8176 if (!se
->sum_exec_runtime
|| p
->state
== TASK_WAKING
)
8182 static void detach_task_cfs_rq(struct task_struct
*p
)
8184 struct sched_entity
*se
= &p
->se
;
8185 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
8187 if (!vruntime_normalized(p
)) {
8189 * Fix up our vruntime so that the current sleep doesn't
8190 * cause 'unlimited' sleep bonus.
8192 place_entity(cfs_rq
, se
, 0);
8193 se
->vruntime
-= cfs_rq
->min_vruntime
;
8196 /* Catch up with the cfs_rq and remove our load when we leave */
8197 detach_entity_load_avg(cfs_rq
, se
);
8200 static void attach_task_cfs_rq(struct task_struct
*p
)
8202 struct sched_entity
*se
= &p
->se
;
8203 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
8205 #ifdef CONFIG_FAIR_GROUP_SCHED
8207 * Since the real-depth could have been changed (only FAIR
8208 * class maintain depth value), reset depth properly.
8210 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
8213 /* Synchronize task with its cfs_rq */
8214 attach_entity_load_avg(cfs_rq
, se
);
8216 if (!vruntime_normalized(p
))
8217 se
->vruntime
+= cfs_rq
->min_vruntime
;
8220 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
8222 detach_task_cfs_rq(p
);
8225 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
8227 attach_task_cfs_rq(p
);
8229 if (task_on_rq_queued(p
)) {
8231 * We were most likely switched from sched_rt, so
8232 * kick off the schedule if running, otherwise just see
8233 * if we can still preempt the current task.
8238 check_preempt_curr(rq
, p
, 0);
8242 /* Account for a task changing its policy or group.
8244 * This routine is mostly called to set cfs_rq->curr field when a task
8245 * migrates between groups/classes.
8247 static void set_curr_task_fair(struct rq
*rq
)
8249 struct sched_entity
*se
= &rq
->curr
->se
;
8251 for_each_sched_entity(se
) {
8252 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
8254 set_next_entity(cfs_rq
, se
);
8255 /* ensure bandwidth has been allocated on our new cfs_rq */
8256 account_cfs_rq_runtime(cfs_rq
, 0);
8260 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
8262 cfs_rq
->tasks_timeline
= RB_ROOT
;
8263 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
8264 #ifndef CONFIG_64BIT
8265 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
8268 atomic_long_set(&cfs_rq
->removed_load_avg
, 0);
8269 atomic_long_set(&cfs_rq
->removed_util_avg
, 0);
8273 #ifdef CONFIG_FAIR_GROUP_SCHED
8274 static void task_move_group_fair(struct task_struct
*p
)
8276 detach_task_cfs_rq(p
);
8277 set_task_rq(p
, task_cpu(p
));
8280 /* Tell se's cfs_rq has been changed -- migrated */
8281 p
->se
.avg
.last_update_time
= 0;
8283 attach_task_cfs_rq(p
);
8286 void free_fair_sched_group(struct task_group
*tg
)
8290 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
8292 for_each_possible_cpu(i
) {
8294 kfree(tg
->cfs_rq
[i
]);
8303 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8305 struct cfs_rq
*cfs_rq
;
8306 struct sched_entity
*se
;
8309 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8312 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8316 tg
->shares
= NICE_0_LOAD
;
8318 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
8320 for_each_possible_cpu(i
) {
8321 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
8322 GFP_KERNEL
, cpu_to_node(i
));
8326 se
= kzalloc_node(sizeof(struct sched_entity
),
8327 GFP_KERNEL
, cpu_to_node(i
));
8331 init_cfs_rq(cfs_rq
);
8332 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
8333 init_entity_runnable_average(se
);
8344 void unregister_fair_sched_group(struct task_group
*tg
)
8346 unsigned long flags
;
8350 for_each_possible_cpu(cpu
) {
8352 remove_entity_load_avg(tg
->se
[cpu
]);
8355 * Only empty task groups can be destroyed; so we can speculatively
8356 * check on_list without danger of it being re-added.
8358 if (!tg
->cfs_rq
[cpu
]->on_list
)
8363 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8364 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
8365 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8369 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8370 struct sched_entity
*se
, int cpu
,
8371 struct sched_entity
*parent
)
8373 struct rq
*rq
= cpu_rq(cpu
);
8377 init_cfs_rq_runtime(cfs_rq
);
8379 tg
->cfs_rq
[cpu
] = cfs_rq
;
8382 /* se could be NULL for root_task_group */
8387 se
->cfs_rq
= &rq
->cfs
;
8390 se
->cfs_rq
= parent
->my_q
;
8391 se
->depth
= parent
->depth
+ 1;
8395 /* guarantee group entities always have weight */
8396 update_load_set(&se
->load
, NICE_0_LOAD
);
8397 se
->parent
= parent
;
8400 static DEFINE_MUTEX(shares_mutex
);
8402 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8405 unsigned long flags
;
8408 * We can't change the weight of the root cgroup.
8413 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
8415 mutex_lock(&shares_mutex
);
8416 if (tg
->shares
== shares
)
8419 tg
->shares
= shares
;
8420 for_each_possible_cpu(i
) {
8421 struct rq
*rq
= cpu_rq(i
);
8422 struct sched_entity
*se
;
8425 /* Propagate contribution to hierarchy */
8426 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8428 /* Possible calls to update_curr() need rq clock */
8429 update_rq_clock(rq
);
8430 for_each_sched_entity(se
)
8431 update_cfs_shares(group_cfs_rq(se
));
8432 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8436 mutex_unlock(&shares_mutex
);
8439 #else /* CONFIG_FAIR_GROUP_SCHED */
8441 void free_fair_sched_group(struct task_group
*tg
) { }
8443 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8448 void unregister_fair_sched_group(struct task_group
*tg
) { }
8450 #endif /* CONFIG_FAIR_GROUP_SCHED */
8453 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
8455 struct sched_entity
*se
= &task
->se
;
8456 unsigned int rr_interval
= 0;
8459 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8462 if (rq
->cfs
.load
.weight
)
8463 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
8469 * All the scheduling class methods:
8471 const struct sched_class fair_sched_class
= {
8472 .next
= &idle_sched_class
,
8473 .enqueue_task
= enqueue_task_fair
,
8474 .dequeue_task
= dequeue_task_fair
,
8475 .yield_task
= yield_task_fair
,
8476 .yield_to_task
= yield_to_task_fair
,
8478 .check_preempt_curr
= check_preempt_wakeup
,
8480 .pick_next_task
= pick_next_task_fair
,
8481 .put_prev_task
= put_prev_task_fair
,
8484 .select_task_rq
= select_task_rq_fair
,
8485 .migrate_task_rq
= migrate_task_rq_fair
,
8487 .rq_online
= rq_online_fair
,
8488 .rq_offline
= rq_offline_fair
,
8490 .task_waking
= task_waking_fair
,
8491 .task_dead
= task_dead_fair
,
8492 .set_cpus_allowed
= set_cpus_allowed_common
,
8495 .set_curr_task
= set_curr_task_fair
,
8496 .task_tick
= task_tick_fair
,
8497 .task_fork
= task_fork_fair
,
8499 .prio_changed
= prio_changed_fair
,
8500 .switched_from
= switched_from_fair
,
8501 .switched_to
= switched_to_fair
,
8503 .get_rr_interval
= get_rr_interval_fair
,
8505 .update_curr
= update_curr_fair
,
8507 #ifdef CONFIG_FAIR_GROUP_SCHED
8508 .task_move_group
= task_move_group_fair
,
8512 #ifdef CONFIG_SCHED_DEBUG
8513 void print_cfs_stats(struct seq_file
*m
, int cpu
)
8515 struct cfs_rq
*cfs_rq
;
8518 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
8519 print_cfs_rq(m
, cpu
, cfs_rq
);
8523 #ifdef CONFIG_NUMA_BALANCING
8524 void show_numa_stats(struct task_struct
*p
, struct seq_file
*m
)
8527 unsigned long tsf
= 0, tpf
= 0, gsf
= 0, gpf
= 0;
8529 for_each_online_node(node
) {
8530 if (p
->numa_faults
) {
8531 tsf
= p
->numa_faults
[task_faults_idx(NUMA_MEM
, node
, 0)];
8532 tpf
= p
->numa_faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
8534 if (p
->numa_group
) {
8535 gsf
= p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, node
, 0)],
8536 gpf
= p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, node
, 1)];
8538 print_numa_stats(m
, node
, tsf
, tpf
, gsf
, gpf
);
8541 #endif /* CONFIG_NUMA_BALANCING */
8542 #endif /* CONFIG_SCHED_DEBUG */
8544 __init
void init_sched_fair_class(void)
8547 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8549 #ifdef CONFIG_NO_HZ_COMMON
8550 nohz
.next_balance
= jiffies
;
8551 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
8552 cpu_notifier(sched_ilb_notifier
, 0);