2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/cpuidle.h>
27 #include <linux/slab.h>
28 #include <linux/profile.h>
29 #include <linux/interrupt.h>
30 #include <linux/mempolicy.h>
31 #include <linux/migrate.h>
32 #include <linux/task_work.h>
34 #include <trace/events/sched.h>
39 * Targeted preemption latency for CPU-bound tasks:
40 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
42 * NOTE: this latency value is not the same as the concept of
43 * 'timeslice length' - timeslices in CFS are of variable length
44 * and have no persistent notion like in traditional, time-slice
45 * based scheduling concepts.
47 * (to see the precise effective timeslice length of your workload,
48 * run vmstat and monitor the context-switches (cs) field)
50 unsigned int sysctl_sched_latency
= 6000000ULL;
51 unsigned int normalized_sysctl_sched_latency
= 6000000ULL;
54 * The initial- and re-scaling of tunables is configurable
55 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
58 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
59 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
60 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
62 enum sched_tunable_scaling sysctl_sched_tunable_scaling
63 = SCHED_TUNABLESCALING_LOG
;
66 * Minimal preemption granularity for CPU-bound tasks:
67 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
69 unsigned int sysctl_sched_min_granularity
= 750000ULL;
70 unsigned int normalized_sysctl_sched_min_granularity
= 750000ULL;
73 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
75 static unsigned int sched_nr_latency
= 8;
78 * After fork, child runs first. If set to 0 (default) then
79 * parent will (try to) run first.
81 unsigned int sysctl_sched_child_runs_first __read_mostly
;
84 * SCHED_OTHER wake-up granularity.
85 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
87 * This option delays the preemption effects of decoupled workloads
88 * and reduces their over-scheduling. Synchronous workloads will still
89 * have immediate wakeup/sleep latencies.
91 unsigned int sysctl_sched_wakeup_granularity
= 1000000UL;
92 unsigned int normalized_sysctl_sched_wakeup_granularity
= 1000000UL;
94 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
97 * The exponential sliding window over which load is averaged for shares
101 unsigned int __read_mostly sysctl_sched_shares_window
= 10000000UL;
103 #ifdef CONFIG_CFS_BANDWIDTH
105 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
106 * each time a cfs_rq requests quota.
108 * Note: in the case that the slice exceeds the runtime remaining (either due
109 * to consumption or the quota being specified to be smaller than the slice)
110 * we will always only issue the remaining available time.
112 * default: 5 msec, units: microseconds
114 unsigned int sysctl_sched_cfs_bandwidth_slice
= 5000UL;
117 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
123 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
129 static inline void update_load_set(struct load_weight
*lw
, unsigned long w
)
136 * Increase the granularity value when there are more CPUs,
137 * because with more CPUs the 'effective latency' as visible
138 * to users decreases. But the relationship is not linear,
139 * so pick a second-best guess by going with the log2 of the
142 * This idea comes from the SD scheduler of Con Kolivas:
144 static int get_update_sysctl_factor(void)
146 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
149 switch (sysctl_sched_tunable_scaling
) {
150 case SCHED_TUNABLESCALING_NONE
:
153 case SCHED_TUNABLESCALING_LINEAR
:
156 case SCHED_TUNABLESCALING_LOG
:
158 factor
= 1 + ilog2(cpus
);
165 static void update_sysctl(void)
167 unsigned int factor
= get_update_sysctl_factor();
169 #define SET_SYSCTL(name) \
170 (sysctl_##name = (factor) * normalized_sysctl_##name)
171 SET_SYSCTL(sched_min_granularity
);
172 SET_SYSCTL(sched_latency
);
173 SET_SYSCTL(sched_wakeup_granularity
);
177 void sched_init_granularity(void)
182 #define WMULT_CONST (~0U)
183 #define WMULT_SHIFT 32
185 static void __update_inv_weight(struct load_weight
*lw
)
189 if (likely(lw
->inv_weight
))
192 w
= scale_load_down(lw
->weight
);
194 if (BITS_PER_LONG
> 32 && unlikely(w
>= WMULT_CONST
))
196 else if (unlikely(!w
))
197 lw
->inv_weight
= WMULT_CONST
;
199 lw
->inv_weight
= WMULT_CONST
/ w
;
203 * delta_exec * weight / lw.weight
205 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
207 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
208 * we're guaranteed shift stays positive because inv_weight is guaranteed to
209 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
211 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
212 * weight/lw.weight <= 1, and therefore our shift will also be positive.
214 static u64
__calc_delta(u64 delta_exec
, unsigned long weight
, struct load_weight
*lw
)
216 u64 fact
= scale_load_down(weight
);
217 int shift
= WMULT_SHIFT
;
219 __update_inv_weight(lw
);
221 if (unlikely(fact
>> 32)) {
228 /* hint to use a 32x32->64 mul */
229 fact
= (u64
)(u32
)fact
* lw
->inv_weight
;
236 return mul_u64_u32_shr(delta_exec
, fact
, shift
);
240 const struct sched_class fair_sched_class
;
242 /**************************************************************
243 * CFS operations on generic schedulable entities:
246 #ifdef CONFIG_FAIR_GROUP_SCHED
248 /* cpu runqueue to which this cfs_rq is attached */
249 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
254 /* An entity is a task if it doesn't "own" a runqueue */
255 #define entity_is_task(se) (!se->my_q)
257 static inline struct task_struct
*task_of(struct sched_entity
*se
)
259 #ifdef CONFIG_SCHED_DEBUG
260 WARN_ON_ONCE(!entity_is_task(se
));
262 return container_of(se
, struct task_struct
, se
);
265 /* Walk up scheduling entities hierarchy */
266 #define for_each_sched_entity(se) \
267 for (; se; se = se->parent)
269 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
274 /* runqueue on which this entity is (to be) queued */
275 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
280 /* runqueue "owned" by this group */
281 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
286 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
289 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
291 if (!cfs_rq
->on_list
) {
293 * Ensure we either appear before our parent (if already
294 * enqueued) or force our parent to appear after us when it is
295 * enqueued. The fact that we always enqueue bottom-up
296 * reduces this to two cases.
298 if (cfs_rq
->tg
->parent
&&
299 cfs_rq
->tg
->parent
->cfs_rq
[cpu_of(rq_of(cfs_rq
))]->on_list
) {
300 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
,
301 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
303 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
304 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
308 /* We should have no load, but we need to update last_decay. */
309 update_cfs_rq_blocked_load(cfs_rq
, 0);
313 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
315 if (cfs_rq
->on_list
) {
316 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
321 /* Iterate thr' all leaf cfs_rq's on a runqueue */
322 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
323 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
325 /* Do the two (enqueued) entities belong to the same group ? */
326 static inline struct cfs_rq
*
327 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
329 if (se
->cfs_rq
== pse
->cfs_rq
)
335 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
341 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
343 int se_depth
, pse_depth
;
346 * preemption test can be made between sibling entities who are in the
347 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
348 * both tasks until we find their ancestors who are siblings of common
352 /* First walk up until both entities are at same depth */
353 se_depth
= (*se
)->depth
;
354 pse_depth
= (*pse
)->depth
;
356 while (se_depth
> pse_depth
) {
358 *se
= parent_entity(*se
);
361 while (pse_depth
> se_depth
) {
363 *pse
= parent_entity(*pse
);
366 while (!is_same_group(*se
, *pse
)) {
367 *se
= parent_entity(*se
);
368 *pse
= parent_entity(*pse
);
372 #else /* !CONFIG_FAIR_GROUP_SCHED */
374 static inline struct task_struct
*task_of(struct sched_entity
*se
)
376 return container_of(se
, struct task_struct
, se
);
379 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
381 return container_of(cfs_rq
, struct rq
, cfs
);
384 #define entity_is_task(se) 1
386 #define for_each_sched_entity(se) \
387 for (; se; se = NULL)
389 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
391 return &task_rq(p
)->cfs
;
394 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
396 struct task_struct
*p
= task_of(se
);
397 struct rq
*rq
= task_rq(p
);
402 /* runqueue "owned" by this group */
403 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
408 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
412 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
416 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
417 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
419 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
425 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
429 #endif /* CONFIG_FAIR_GROUP_SCHED */
431 static __always_inline
432 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
);
434 /**************************************************************
435 * Scheduling class tree data structure manipulation methods:
438 static inline u64
max_vruntime(u64 max_vruntime
, u64 vruntime
)
440 s64 delta
= (s64
)(vruntime
- max_vruntime
);
442 max_vruntime
= vruntime
;
447 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
449 s64 delta
= (s64
)(vruntime
- min_vruntime
);
451 min_vruntime
= vruntime
;
456 static inline int entity_before(struct sched_entity
*a
,
457 struct sched_entity
*b
)
459 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
462 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
464 u64 vruntime
= cfs_rq
->min_vruntime
;
467 vruntime
= cfs_rq
->curr
->vruntime
;
469 if (cfs_rq
->rb_leftmost
) {
470 struct sched_entity
*se
= rb_entry(cfs_rq
->rb_leftmost
,
475 vruntime
= se
->vruntime
;
477 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
480 /* ensure we never gain time by being placed backwards. */
481 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
484 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
489 * Enqueue an entity into the rb-tree:
491 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
493 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
494 struct rb_node
*parent
= NULL
;
495 struct sched_entity
*entry
;
499 * Find the right place in the rbtree:
503 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
505 * We dont care about collisions. Nodes with
506 * the same key stay together.
508 if (entity_before(se
, entry
)) {
509 link
= &parent
->rb_left
;
511 link
= &parent
->rb_right
;
517 * Maintain a cache of leftmost tree entries (it is frequently
521 cfs_rq
->rb_leftmost
= &se
->run_node
;
523 rb_link_node(&se
->run_node
, parent
, link
);
524 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
527 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
529 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
530 struct rb_node
*next_node
;
532 next_node
= rb_next(&se
->run_node
);
533 cfs_rq
->rb_leftmost
= next_node
;
536 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
539 struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
)
541 struct rb_node
*left
= cfs_rq
->rb_leftmost
;
546 return rb_entry(left
, struct sched_entity
, run_node
);
549 static struct sched_entity
*__pick_next_entity(struct sched_entity
*se
)
551 struct rb_node
*next
= rb_next(&se
->run_node
);
556 return rb_entry(next
, struct sched_entity
, run_node
);
559 #ifdef CONFIG_SCHED_DEBUG
560 struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
562 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
567 return rb_entry(last
, struct sched_entity
, run_node
);
570 /**************************************************************
571 * Scheduling class statistics methods:
574 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
575 void __user
*buffer
, size_t *lenp
,
578 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
579 int factor
= get_update_sysctl_factor();
584 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
585 sysctl_sched_min_granularity
);
587 #define WRT_SYSCTL(name) \
588 (normalized_sysctl_##name = sysctl_##name / (factor))
589 WRT_SYSCTL(sched_min_granularity
);
590 WRT_SYSCTL(sched_latency
);
591 WRT_SYSCTL(sched_wakeup_granularity
);
601 static inline u64
calc_delta_fair(u64 delta
, struct sched_entity
*se
)
603 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
604 delta
= __calc_delta(delta
, NICE_0_LOAD
, &se
->load
);
610 * The idea is to set a period in which each task runs once.
612 * When there are too many tasks (sched_nr_latency) we have to stretch
613 * this period because otherwise the slices get too small.
615 * p = (nr <= nl) ? l : l*nr/nl
617 static u64
__sched_period(unsigned long nr_running
)
619 u64 period
= sysctl_sched_latency
;
620 unsigned long nr_latency
= sched_nr_latency
;
622 if (unlikely(nr_running
> nr_latency
)) {
623 period
= sysctl_sched_min_granularity
;
624 period
*= nr_running
;
631 * We calculate the wall-time slice from the period by taking a part
632 * proportional to the weight.
636 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
638 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
640 for_each_sched_entity(se
) {
641 struct load_weight
*load
;
642 struct load_weight lw
;
644 cfs_rq
= cfs_rq_of(se
);
645 load
= &cfs_rq
->load
;
647 if (unlikely(!se
->on_rq
)) {
650 update_load_add(&lw
, se
->load
.weight
);
653 slice
= __calc_delta(slice
, se
->load
.weight
, load
);
659 * We calculate the vruntime slice of a to-be-inserted task.
663 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
665 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
669 static int select_idle_sibling(struct task_struct
*p
, int cpu
);
670 static unsigned long task_h_load(struct task_struct
*p
);
672 static inline void __update_task_entity_contrib(struct sched_entity
*se
);
674 /* Give new task start runnable values to heavy its load in infant time */
675 void init_task_runnable_average(struct task_struct
*p
)
679 p
->se
.avg
.decay_count
= 0;
680 slice
= sched_slice(task_cfs_rq(p
), &p
->se
) >> 10;
681 p
->se
.avg
.runnable_avg_sum
= slice
;
682 p
->se
.avg
.runnable_avg_period
= slice
;
683 __update_task_entity_contrib(&p
->se
);
686 void init_task_runnable_average(struct task_struct
*p
)
692 * Update the current task's runtime statistics.
694 static void update_curr(struct cfs_rq
*cfs_rq
)
696 struct sched_entity
*curr
= cfs_rq
->curr
;
697 u64 now
= rq_clock_task(rq_of(cfs_rq
));
703 delta_exec
= now
- curr
->exec_start
;
704 if (unlikely((s64
)delta_exec
<= 0))
707 curr
->exec_start
= now
;
709 schedstat_set(curr
->statistics
.exec_max
,
710 max(delta_exec
, curr
->statistics
.exec_max
));
712 curr
->sum_exec_runtime
+= delta_exec
;
713 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
715 curr
->vruntime
+= calc_delta_fair(delta_exec
, curr
);
716 update_min_vruntime(cfs_rq
);
718 if (entity_is_task(curr
)) {
719 struct task_struct
*curtask
= task_of(curr
);
721 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
722 cpuacct_charge(curtask
, delta_exec
);
723 account_group_exec_runtime(curtask
, delta_exec
);
726 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
729 static void update_curr_fair(struct rq
*rq
)
731 update_curr(cfs_rq_of(&rq
->curr
->se
));
735 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
737 schedstat_set(se
->statistics
.wait_start
, rq_clock(rq_of(cfs_rq
)));
741 * Task is being enqueued - update stats:
743 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
746 * Are we enqueueing a waiting task? (for current tasks
747 * a dequeue/enqueue event is a NOP)
749 if (se
!= cfs_rq
->curr
)
750 update_stats_wait_start(cfs_rq
, se
);
754 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
756 schedstat_set(se
->statistics
.wait_max
, max(se
->statistics
.wait_max
,
757 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
));
758 schedstat_set(se
->statistics
.wait_count
, se
->statistics
.wait_count
+ 1);
759 schedstat_set(se
->statistics
.wait_sum
, se
->statistics
.wait_sum
+
760 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
761 #ifdef CONFIG_SCHEDSTATS
762 if (entity_is_task(se
)) {
763 trace_sched_stat_wait(task_of(se
),
764 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
767 schedstat_set(se
->statistics
.wait_start
, 0);
771 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
774 * Mark the end of the wait period if dequeueing a
777 if (se
!= cfs_rq
->curr
)
778 update_stats_wait_end(cfs_rq
, se
);
782 * We are picking a new current task - update its stats:
785 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
788 * We are starting a new run period:
790 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
793 /**************************************************
794 * Scheduling class queueing methods:
797 #ifdef CONFIG_NUMA_BALANCING
799 * Approximate time to scan a full NUMA task in ms. The task scan period is
800 * calculated based on the tasks virtual memory size and
801 * numa_balancing_scan_size.
803 unsigned int sysctl_numa_balancing_scan_period_min
= 1000;
804 unsigned int sysctl_numa_balancing_scan_period_max
= 60000;
806 /* Portion of address space to scan in MB */
807 unsigned int sysctl_numa_balancing_scan_size
= 256;
809 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
810 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
812 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
814 unsigned long rss
= 0;
815 unsigned long nr_scan_pages
;
818 * Calculations based on RSS as non-present and empty pages are skipped
819 * by the PTE scanner and NUMA hinting faults should be trapped based
822 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
823 rss
= get_mm_rss(p
->mm
);
827 rss
= round_up(rss
, nr_scan_pages
);
828 return rss
/ nr_scan_pages
;
831 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
832 #define MAX_SCAN_WINDOW 2560
834 static unsigned int task_scan_min(struct task_struct
*p
)
836 unsigned int scan_size
= ACCESS_ONCE(sysctl_numa_balancing_scan_size
);
837 unsigned int scan
, floor
;
838 unsigned int windows
= 1;
840 if (scan_size
< MAX_SCAN_WINDOW
)
841 windows
= MAX_SCAN_WINDOW
/ scan_size
;
842 floor
= 1000 / windows
;
844 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
845 return max_t(unsigned int, floor
, scan
);
848 static unsigned int task_scan_max(struct task_struct
*p
)
850 unsigned int smin
= task_scan_min(p
);
853 /* Watch for min being lower than max due to floor calculations */
854 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
855 return max(smin
, smax
);
858 static void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
860 rq
->nr_numa_running
+= (p
->numa_preferred_nid
!= -1);
861 rq
->nr_preferred_running
+= (p
->numa_preferred_nid
== task_node(p
));
864 static void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
866 rq
->nr_numa_running
-= (p
->numa_preferred_nid
!= -1);
867 rq
->nr_preferred_running
-= (p
->numa_preferred_nid
== task_node(p
));
873 spinlock_t lock
; /* nr_tasks, tasks */
876 struct list_head task_list
;
879 nodemask_t active_nodes
;
880 unsigned long total_faults
;
882 * Faults_cpu is used to decide whether memory should move
883 * towards the CPU. As a consequence, these stats are weighted
884 * more by CPU use than by memory faults.
886 unsigned long *faults_cpu
;
887 unsigned long faults
[0];
890 /* Shared or private faults. */
891 #define NR_NUMA_HINT_FAULT_TYPES 2
893 /* Memory and CPU locality */
894 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
896 /* Averaged statistics, and temporary buffers. */
897 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
899 pid_t
task_numa_group_id(struct task_struct
*p
)
901 return p
->numa_group
? p
->numa_group
->gid
: 0;
904 static inline int task_faults_idx(int nid
, int priv
)
906 return NR_NUMA_HINT_FAULT_TYPES
* nid
+ priv
;
909 static inline unsigned long task_faults(struct task_struct
*p
, int nid
)
911 if (!p
->numa_faults_memory
)
914 return p
->numa_faults_memory
[task_faults_idx(nid
, 0)] +
915 p
->numa_faults_memory
[task_faults_idx(nid
, 1)];
918 static inline unsigned long group_faults(struct task_struct
*p
, int nid
)
923 return p
->numa_group
->faults
[task_faults_idx(nid
, 0)] +
924 p
->numa_group
->faults
[task_faults_idx(nid
, 1)];
927 static inline unsigned long group_faults_cpu(struct numa_group
*group
, int nid
)
929 return group
->faults_cpu
[task_faults_idx(nid
, 0)] +
930 group
->faults_cpu
[task_faults_idx(nid
, 1)];
934 * These return the fraction of accesses done by a particular task, or
935 * task group, on a particular numa node. The group weight is given a
936 * larger multiplier, in order to group tasks together that are almost
937 * evenly spread out between numa nodes.
939 static inline unsigned long task_weight(struct task_struct
*p
, int nid
)
941 unsigned long total_faults
;
943 if (!p
->numa_faults_memory
)
946 total_faults
= p
->total_numa_faults
;
951 return 1000 * task_faults(p
, nid
) / total_faults
;
954 static inline unsigned long group_weight(struct task_struct
*p
, int nid
)
956 if (!p
->numa_group
|| !p
->numa_group
->total_faults
)
959 return 1000 * group_faults(p
, nid
) / p
->numa_group
->total_faults
;
962 bool should_numa_migrate_memory(struct task_struct
*p
, struct page
* page
,
963 int src_nid
, int dst_cpu
)
965 struct numa_group
*ng
= p
->numa_group
;
966 int dst_nid
= cpu_to_node(dst_cpu
);
967 int last_cpupid
, this_cpupid
;
969 this_cpupid
= cpu_pid_to_cpupid(dst_cpu
, current
->pid
);
972 * Multi-stage node selection is used in conjunction with a periodic
973 * migration fault to build a temporal task<->page relation. By using
974 * a two-stage filter we remove short/unlikely relations.
976 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
977 * a task's usage of a particular page (n_p) per total usage of this
978 * page (n_t) (in a given time-span) to a probability.
980 * Our periodic faults will sample this probability and getting the
981 * same result twice in a row, given these samples are fully
982 * independent, is then given by P(n)^2, provided our sample period
983 * is sufficiently short compared to the usage pattern.
985 * This quadric squishes small probabilities, making it less likely we
986 * act on an unlikely task<->page relation.
988 last_cpupid
= page_cpupid_xchg_last(page
, this_cpupid
);
989 if (!cpupid_pid_unset(last_cpupid
) &&
990 cpupid_to_nid(last_cpupid
) != dst_nid
)
993 /* Always allow migrate on private faults */
994 if (cpupid_match_pid(p
, last_cpupid
))
997 /* A shared fault, but p->numa_group has not been set up yet. */
1002 * Do not migrate if the destination is not a node that
1003 * is actively used by this numa group.
1005 if (!node_isset(dst_nid
, ng
->active_nodes
))
1009 * Source is a node that is not actively used by this
1010 * numa group, while the destination is. Migrate.
1012 if (!node_isset(src_nid
, ng
->active_nodes
))
1016 * Both source and destination are nodes in active
1017 * use by this numa group. Maximize memory bandwidth
1018 * by migrating from more heavily used groups, to less
1019 * heavily used ones, spreading the load around.
1020 * Use a 1/4 hysteresis to avoid spurious page movement.
1022 return group_faults(p
, dst_nid
) < (group_faults(p
, src_nid
) * 3 / 4);
1025 static unsigned long weighted_cpuload(const int cpu
);
1026 static unsigned long source_load(int cpu
, int type
);
1027 static unsigned long target_load(int cpu
, int type
);
1028 static unsigned long capacity_of(int cpu
);
1029 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
);
1031 /* Cached statistics for all CPUs within a node */
1033 unsigned long nr_running
;
1036 /* Total compute capacity of CPUs on a node */
1037 unsigned long compute_capacity
;
1039 /* Approximate capacity in terms of runnable tasks on a node */
1040 unsigned long task_capacity
;
1041 int has_free_capacity
;
1045 * XXX borrowed from update_sg_lb_stats
1047 static void update_numa_stats(struct numa_stats
*ns
, int nid
)
1049 int smt
, cpu
, cpus
= 0;
1050 unsigned long capacity
;
1052 memset(ns
, 0, sizeof(*ns
));
1053 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
1054 struct rq
*rq
= cpu_rq(cpu
);
1056 ns
->nr_running
+= rq
->nr_running
;
1057 ns
->load
+= weighted_cpuload(cpu
);
1058 ns
->compute_capacity
+= capacity_of(cpu
);
1064 * If we raced with hotplug and there are no CPUs left in our mask
1065 * the @ns structure is NULL'ed and task_numa_compare() will
1066 * not find this node attractive.
1068 * We'll either bail at !has_free_capacity, or we'll detect a huge
1069 * imbalance and bail there.
1074 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1075 smt
= DIV_ROUND_UP(SCHED_CAPACITY_SCALE
* cpus
, ns
->compute_capacity
);
1076 capacity
= cpus
/ smt
; /* cores */
1078 ns
->task_capacity
= min_t(unsigned, capacity
,
1079 DIV_ROUND_CLOSEST(ns
->compute_capacity
, SCHED_CAPACITY_SCALE
));
1080 ns
->has_free_capacity
= (ns
->nr_running
< ns
->task_capacity
);
1083 struct task_numa_env
{
1084 struct task_struct
*p
;
1086 int src_cpu
, src_nid
;
1087 int dst_cpu
, dst_nid
;
1089 struct numa_stats src_stats
, dst_stats
;
1093 struct task_struct
*best_task
;
1098 static void task_numa_assign(struct task_numa_env
*env
,
1099 struct task_struct
*p
, long imp
)
1102 put_task_struct(env
->best_task
);
1107 env
->best_imp
= imp
;
1108 env
->best_cpu
= env
->dst_cpu
;
1111 static bool load_too_imbalanced(long src_load
, long dst_load
,
1112 struct task_numa_env
*env
)
1115 long orig_src_load
, orig_dst_load
;
1116 long src_capacity
, dst_capacity
;
1119 * The load is corrected for the CPU capacity available on each node.
1122 * ------------ vs ---------
1123 * src_capacity dst_capacity
1125 src_capacity
= env
->src_stats
.compute_capacity
;
1126 dst_capacity
= env
->dst_stats
.compute_capacity
;
1128 /* We care about the slope of the imbalance, not the direction. */
1129 if (dst_load
< src_load
)
1130 swap(dst_load
, src_load
);
1132 /* Is the difference below the threshold? */
1133 imb
= dst_load
* src_capacity
* 100 -
1134 src_load
* dst_capacity
* env
->imbalance_pct
;
1139 * The imbalance is above the allowed threshold.
1140 * Compare it with the old imbalance.
1142 orig_src_load
= env
->src_stats
.load
;
1143 orig_dst_load
= env
->dst_stats
.load
;
1145 if (orig_dst_load
< orig_src_load
)
1146 swap(orig_dst_load
, orig_src_load
);
1148 old_imb
= orig_dst_load
* src_capacity
* 100 -
1149 orig_src_load
* dst_capacity
* env
->imbalance_pct
;
1151 /* Would this change make things worse? */
1152 return (imb
> old_imb
);
1156 * This checks if the overall compute and NUMA accesses of the system would
1157 * be improved if the source tasks was migrated to the target dst_cpu taking
1158 * into account that it might be best if task running on the dst_cpu should
1159 * be exchanged with the source task
1161 static void task_numa_compare(struct task_numa_env
*env
,
1162 long taskimp
, long groupimp
)
1164 struct rq
*src_rq
= cpu_rq(env
->src_cpu
);
1165 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1166 struct task_struct
*cur
;
1167 long src_load
, dst_load
;
1169 long imp
= env
->p
->numa_group
? groupimp
: taskimp
;
1174 raw_spin_lock_irq(&dst_rq
->lock
);
1177 * No need to move the exiting task, and this ensures that ->curr
1178 * wasn't reaped and thus get_task_struct() in task_numa_assign()
1179 * is safe under RCU read lock.
1180 * Note that rcu_read_lock() itself can't protect from the final
1181 * put_task_struct() after the last schedule().
1183 if ((cur
->flags
& PF_EXITING
) || is_idle_task(cur
))
1185 raw_spin_unlock_irq(&dst_rq
->lock
);
1188 * Because we have preemption enabled we can get migrated around and
1189 * end try selecting ourselves (current == env->p) as a swap candidate.
1195 * "imp" is the fault differential for the source task between the
1196 * source and destination node. Calculate the total differential for
1197 * the source task and potential destination task. The more negative
1198 * the value is, the more rmeote accesses that would be expected to
1199 * be incurred if the tasks were swapped.
1202 /* Skip this swap candidate if cannot move to the source cpu */
1203 if (!cpumask_test_cpu(env
->src_cpu
, tsk_cpus_allowed(cur
)))
1207 * If dst and source tasks are in the same NUMA group, or not
1208 * in any group then look only at task weights.
1210 if (cur
->numa_group
== env
->p
->numa_group
) {
1211 imp
= taskimp
+ task_weight(cur
, env
->src_nid
) -
1212 task_weight(cur
, env
->dst_nid
);
1214 * Add some hysteresis to prevent swapping the
1215 * tasks within a group over tiny differences.
1217 if (cur
->numa_group
)
1221 * Compare the group weights. If a task is all by
1222 * itself (not part of a group), use the task weight
1225 if (cur
->numa_group
)
1226 imp
+= group_weight(cur
, env
->src_nid
) -
1227 group_weight(cur
, env
->dst_nid
);
1229 imp
+= task_weight(cur
, env
->src_nid
) -
1230 task_weight(cur
, env
->dst_nid
);
1234 if (imp
<= env
->best_imp
&& moveimp
<= env
->best_imp
)
1238 /* Is there capacity at our destination? */
1239 if (env
->src_stats
.nr_running
<= env
->src_stats
.task_capacity
&&
1240 !env
->dst_stats
.has_free_capacity
)
1246 /* Balance doesn't matter much if we're running a task per cpu */
1247 if (imp
> env
->best_imp
&& src_rq
->nr_running
== 1 &&
1248 dst_rq
->nr_running
== 1)
1252 * In the overloaded case, try and keep the load balanced.
1255 load
= task_h_load(env
->p
);
1256 dst_load
= env
->dst_stats
.load
+ load
;
1257 src_load
= env
->src_stats
.load
- load
;
1259 if (moveimp
> imp
&& moveimp
> env
->best_imp
) {
1261 * If the improvement from just moving env->p direction is
1262 * better than swapping tasks around, check if a move is
1263 * possible. Store a slightly smaller score than moveimp,
1264 * so an actually idle CPU will win.
1266 if (!load_too_imbalanced(src_load
, dst_load
, env
)) {
1273 if (imp
<= env
->best_imp
)
1277 load
= task_h_load(cur
);
1282 if (load_too_imbalanced(src_load
, dst_load
, env
))
1286 * One idle CPU per node is evaluated for a task numa move.
1287 * Call select_idle_sibling to maybe find a better one.
1290 env
->dst_cpu
= select_idle_sibling(env
->p
, env
->dst_cpu
);
1293 task_numa_assign(env
, cur
, imp
);
1298 static void task_numa_find_cpu(struct task_numa_env
*env
,
1299 long taskimp
, long groupimp
)
1303 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1304 /* Skip this CPU if the source task cannot migrate */
1305 if (!cpumask_test_cpu(cpu
, tsk_cpus_allowed(env
->p
)))
1309 task_numa_compare(env
, taskimp
, groupimp
);
1313 static int task_numa_migrate(struct task_struct
*p
)
1315 struct task_numa_env env
= {
1318 .src_cpu
= task_cpu(p
),
1319 .src_nid
= task_node(p
),
1321 .imbalance_pct
= 112,
1327 struct sched_domain
*sd
;
1328 unsigned long taskweight
, groupweight
;
1330 long taskimp
, groupimp
;
1333 * Pick the lowest SD_NUMA domain, as that would have the smallest
1334 * imbalance and would be the first to start moving tasks about.
1336 * And we want to avoid any moving of tasks about, as that would create
1337 * random movement of tasks -- counter the numa conditions we're trying
1341 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1343 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1347 * Cpusets can break the scheduler domain tree into smaller
1348 * balance domains, some of which do not cross NUMA boundaries.
1349 * Tasks that are "trapped" in such domains cannot be migrated
1350 * elsewhere, so there is no point in (re)trying.
1352 if (unlikely(!sd
)) {
1353 p
->numa_preferred_nid
= task_node(p
);
1357 taskweight
= task_weight(p
, env
.src_nid
);
1358 groupweight
= group_weight(p
, env
.src_nid
);
1359 update_numa_stats(&env
.src_stats
, env
.src_nid
);
1360 env
.dst_nid
= p
->numa_preferred_nid
;
1361 taskimp
= task_weight(p
, env
.dst_nid
) - taskweight
;
1362 groupimp
= group_weight(p
, env
.dst_nid
) - groupweight
;
1363 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1365 /* Try to find a spot on the preferred nid. */
1366 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1368 /* No space available on the preferred nid. Look elsewhere. */
1369 if (env
.best_cpu
== -1) {
1370 for_each_online_node(nid
) {
1371 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
1374 /* Only consider nodes where both task and groups benefit */
1375 taskimp
= task_weight(p
, nid
) - taskweight
;
1376 groupimp
= group_weight(p
, nid
) - groupweight
;
1377 if (taskimp
< 0 && groupimp
< 0)
1381 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1382 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1387 * If the task is part of a workload that spans multiple NUMA nodes,
1388 * and is migrating into one of the workload's active nodes, remember
1389 * this node as the task's preferred numa node, so the workload can
1391 * A task that migrated to a second choice node will be better off
1392 * trying for a better one later. Do not set the preferred node here.
1394 if (p
->numa_group
) {
1395 if (env
.best_cpu
== -1)
1400 if (node_isset(nid
, p
->numa_group
->active_nodes
))
1401 sched_setnuma(p
, env
.dst_nid
);
1404 /* No better CPU than the current one was found. */
1405 if (env
.best_cpu
== -1)
1409 * Reset the scan period if the task is being rescheduled on an
1410 * alternative node to recheck if the tasks is now properly placed.
1412 p
->numa_scan_period
= task_scan_min(p
);
1414 if (env
.best_task
== NULL
) {
1415 ret
= migrate_task_to(p
, env
.best_cpu
);
1417 trace_sched_stick_numa(p
, env
.src_cpu
, env
.best_cpu
);
1421 ret
= migrate_swap(p
, env
.best_task
);
1423 trace_sched_stick_numa(p
, env
.src_cpu
, task_cpu(env
.best_task
));
1424 put_task_struct(env
.best_task
);
1428 /* Attempt to migrate a task to a CPU on the preferred node. */
1429 static void numa_migrate_preferred(struct task_struct
*p
)
1431 unsigned long interval
= HZ
;
1433 /* This task has no NUMA fault statistics yet */
1434 if (unlikely(p
->numa_preferred_nid
== -1 || !p
->numa_faults_memory
))
1437 /* Periodically retry migrating the task to the preferred node */
1438 interval
= min(interval
, msecs_to_jiffies(p
->numa_scan_period
) / 16);
1439 p
->numa_migrate_retry
= jiffies
+ interval
;
1441 /* Success if task is already running on preferred CPU */
1442 if (task_node(p
) == p
->numa_preferred_nid
)
1445 /* Otherwise, try migrate to a CPU on the preferred node */
1446 task_numa_migrate(p
);
1450 * Find the nodes on which the workload is actively running. We do this by
1451 * tracking the nodes from which NUMA hinting faults are triggered. This can
1452 * be different from the set of nodes where the workload's memory is currently
1455 * The bitmask is used to make smarter decisions on when to do NUMA page
1456 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1457 * are added when they cause over 6/16 of the maximum number of faults, but
1458 * only removed when they drop below 3/16.
1460 static void update_numa_active_node_mask(struct numa_group
*numa_group
)
1462 unsigned long faults
, max_faults
= 0;
1465 for_each_online_node(nid
) {
1466 faults
= group_faults_cpu(numa_group
, nid
);
1467 if (faults
> max_faults
)
1468 max_faults
= faults
;
1471 for_each_online_node(nid
) {
1472 faults
= group_faults_cpu(numa_group
, nid
);
1473 if (!node_isset(nid
, numa_group
->active_nodes
)) {
1474 if (faults
> max_faults
* 6 / 16)
1475 node_set(nid
, numa_group
->active_nodes
);
1476 } else if (faults
< max_faults
* 3 / 16)
1477 node_clear(nid
, numa_group
->active_nodes
);
1482 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1483 * increments. The more local the fault statistics are, the higher the scan
1484 * period will be for the next scan window. If local/(local+remote) ratio is
1485 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1486 * the scan period will decrease. Aim for 70% local accesses.
1488 #define NUMA_PERIOD_SLOTS 10
1489 #define NUMA_PERIOD_THRESHOLD 7
1492 * Increase the scan period (slow down scanning) if the majority of
1493 * our memory is already on our local node, or if the majority of
1494 * the page accesses are shared with other processes.
1495 * Otherwise, decrease the scan period.
1497 static void update_task_scan_period(struct task_struct
*p
,
1498 unsigned long shared
, unsigned long private)
1500 unsigned int period_slot
;
1504 unsigned long remote
= p
->numa_faults_locality
[0];
1505 unsigned long local
= p
->numa_faults_locality
[1];
1508 * If there were no record hinting faults then either the task is
1509 * completely idle or all activity is areas that are not of interest
1510 * to automatic numa balancing. Scan slower
1512 if (local
+ shared
== 0) {
1513 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1514 p
->numa_scan_period
<< 1);
1516 p
->mm
->numa_next_scan
= jiffies
+
1517 msecs_to_jiffies(p
->numa_scan_period
);
1523 * Prepare to scale scan period relative to the current period.
1524 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1525 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1526 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1528 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
1529 ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
1530 if (ratio
>= NUMA_PERIOD_THRESHOLD
) {
1531 int slot
= ratio
- NUMA_PERIOD_THRESHOLD
;
1534 diff
= slot
* period_slot
;
1536 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
1539 * Scale scan rate increases based on sharing. There is an
1540 * inverse relationship between the degree of sharing and
1541 * the adjustment made to the scanning period. Broadly
1542 * speaking the intent is that there is little point
1543 * scanning faster if shared accesses dominate as it may
1544 * simply bounce migrations uselessly
1546 ratio
= DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS
, (private + shared
+ 1));
1547 diff
= (diff
* ratio
) / NUMA_PERIOD_SLOTS
;
1550 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
1551 task_scan_min(p
), task_scan_max(p
));
1552 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1556 * Get the fraction of time the task has been running since the last
1557 * NUMA placement cycle. The scheduler keeps similar statistics, but
1558 * decays those on a 32ms period, which is orders of magnitude off
1559 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1560 * stats only if the task is so new there are no NUMA statistics yet.
1562 static u64
numa_get_avg_runtime(struct task_struct
*p
, u64
*period
)
1564 u64 runtime
, delta
, now
;
1565 /* Use the start of this time slice to avoid calculations. */
1566 now
= p
->se
.exec_start
;
1567 runtime
= p
->se
.sum_exec_runtime
;
1569 if (p
->last_task_numa_placement
) {
1570 delta
= runtime
- p
->last_sum_exec_runtime
;
1571 *period
= now
- p
->last_task_numa_placement
;
1573 delta
= p
->se
.avg
.runnable_avg_sum
;
1574 *period
= p
->se
.avg
.runnable_avg_period
;
1577 p
->last_sum_exec_runtime
= runtime
;
1578 p
->last_task_numa_placement
= now
;
1583 static void task_numa_placement(struct task_struct
*p
)
1585 int seq
, nid
, max_nid
= -1, max_group_nid
= -1;
1586 unsigned long max_faults
= 0, max_group_faults
= 0;
1587 unsigned long fault_types
[2] = { 0, 0 };
1588 unsigned long total_faults
;
1589 u64 runtime
, period
;
1590 spinlock_t
*group_lock
= NULL
;
1592 seq
= ACCESS_ONCE(p
->mm
->numa_scan_seq
);
1593 if (p
->numa_scan_seq
== seq
)
1595 p
->numa_scan_seq
= seq
;
1596 p
->numa_scan_period_max
= task_scan_max(p
);
1598 total_faults
= p
->numa_faults_locality
[0] +
1599 p
->numa_faults_locality
[1];
1600 runtime
= numa_get_avg_runtime(p
, &period
);
1602 /* If the task is part of a group prevent parallel updates to group stats */
1603 if (p
->numa_group
) {
1604 group_lock
= &p
->numa_group
->lock
;
1605 spin_lock_irq(group_lock
);
1608 /* Find the node with the highest number of faults */
1609 for_each_online_node(nid
) {
1610 unsigned long faults
= 0, group_faults
= 0;
1613 for (priv
= 0; priv
< NR_NUMA_HINT_FAULT_TYPES
; priv
++) {
1614 long diff
, f_diff
, f_weight
;
1616 i
= task_faults_idx(nid
, priv
);
1618 /* Decay existing window, copy faults since last scan */
1619 diff
= p
->numa_faults_buffer_memory
[i
] - p
->numa_faults_memory
[i
] / 2;
1620 fault_types
[priv
] += p
->numa_faults_buffer_memory
[i
];
1621 p
->numa_faults_buffer_memory
[i
] = 0;
1624 * Normalize the faults_from, so all tasks in a group
1625 * count according to CPU use, instead of by the raw
1626 * number of faults. Tasks with little runtime have
1627 * little over-all impact on throughput, and thus their
1628 * faults are less important.
1630 f_weight
= div64_u64(runtime
<< 16, period
+ 1);
1631 f_weight
= (f_weight
* p
->numa_faults_buffer_cpu
[i
]) /
1633 f_diff
= f_weight
- p
->numa_faults_cpu
[i
] / 2;
1634 p
->numa_faults_buffer_cpu
[i
] = 0;
1636 p
->numa_faults_memory
[i
] += diff
;
1637 p
->numa_faults_cpu
[i
] += f_diff
;
1638 faults
+= p
->numa_faults_memory
[i
];
1639 p
->total_numa_faults
+= diff
;
1640 if (p
->numa_group
) {
1641 /* safe because we can only change our own group */
1642 p
->numa_group
->faults
[i
] += diff
;
1643 p
->numa_group
->faults_cpu
[i
] += f_diff
;
1644 p
->numa_group
->total_faults
+= diff
;
1645 group_faults
+= p
->numa_group
->faults
[i
];
1649 if (faults
> max_faults
) {
1650 max_faults
= faults
;
1654 if (group_faults
> max_group_faults
) {
1655 max_group_faults
= group_faults
;
1656 max_group_nid
= nid
;
1660 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
1662 if (p
->numa_group
) {
1663 update_numa_active_node_mask(p
->numa_group
);
1664 spin_unlock_irq(group_lock
);
1665 max_nid
= max_group_nid
;
1669 /* Set the new preferred node */
1670 if (max_nid
!= p
->numa_preferred_nid
)
1671 sched_setnuma(p
, max_nid
);
1673 if (task_node(p
) != p
->numa_preferred_nid
)
1674 numa_migrate_preferred(p
);
1678 static inline int get_numa_group(struct numa_group
*grp
)
1680 return atomic_inc_not_zero(&grp
->refcount
);
1683 static inline void put_numa_group(struct numa_group
*grp
)
1685 if (atomic_dec_and_test(&grp
->refcount
))
1686 kfree_rcu(grp
, rcu
);
1689 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
1692 struct numa_group
*grp
, *my_grp
;
1693 struct task_struct
*tsk
;
1695 int cpu
= cpupid_to_cpu(cpupid
);
1698 if (unlikely(!p
->numa_group
)) {
1699 unsigned int size
= sizeof(struct numa_group
) +
1700 4*nr_node_ids
*sizeof(unsigned long);
1702 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
1706 atomic_set(&grp
->refcount
, 1);
1707 spin_lock_init(&grp
->lock
);
1708 INIT_LIST_HEAD(&grp
->task_list
);
1710 /* Second half of the array tracks nids where faults happen */
1711 grp
->faults_cpu
= grp
->faults
+ NR_NUMA_HINT_FAULT_TYPES
*
1714 node_set(task_node(current
), grp
->active_nodes
);
1716 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
1717 grp
->faults
[i
] = p
->numa_faults_memory
[i
];
1719 grp
->total_faults
= p
->total_numa_faults
;
1721 list_add(&p
->numa_entry
, &grp
->task_list
);
1723 rcu_assign_pointer(p
->numa_group
, grp
);
1727 tsk
= ACCESS_ONCE(cpu_rq(cpu
)->curr
);
1729 if (!cpupid_match_pid(tsk
, cpupid
))
1732 grp
= rcu_dereference(tsk
->numa_group
);
1736 my_grp
= p
->numa_group
;
1741 * Only join the other group if its bigger; if we're the bigger group,
1742 * the other task will join us.
1744 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
1748 * Tie-break on the grp address.
1750 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
1753 /* Always join threads in the same process. */
1754 if (tsk
->mm
== current
->mm
)
1757 /* Simple filter to avoid false positives due to PID collisions */
1758 if (flags
& TNF_SHARED
)
1761 /* Update priv based on whether false sharing was detected */
1764 if (join
&& !get_numa_group(grp
))
1772 BUG_ON(irqs_disabled());
1773 double_lock_irq(&my_grp
->lock
, &grp
->lock
);
1775 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++) {
1776 my_grp
->faults
[i
] -= p
->numa_faults_memory
[i
];
1777 grp
->faults
[i
] += p
->numa_faults_memory
[i
];
1779 my_grp
->total_faults
-= p
->total_numa_faults
;
1780 grp
->total_faults
+= p
->total_numa_faults
;
1782 list_move(&p
->numa_entry
, &grp
->task_list
);
1786 spin_unlock(&my_grp
->lock
);
1787 spin_unlock_irq(&grp
->lock
);
1789 rcu_assign_pointer(p
->numa_group
, grp
);
1791 put_numa_group(my_grp
);
1799 void task_numa_free(struct task_struct
*p
)
1801 struct numa_group
*grp
= p
->numa_group
;
1802 void *numa_faults
= p
->numa_faults_memory
;
1803 unsigned long flags
;
1807 spin_lock_irqsave(&grp
->lock
, flags
);
1808 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
1809 grp
->faults
[i
] -= p
->numa_faults_memory
[i
];
1810 grp
->total_faults
-= p
->total_numa_faults
;
1812 list_del(&p
->numa_entry
);
1814 spin_unlock_irqrestore(&grp
->lock
, flags
);
1815 RCU_INIT_POINTER(p
->numa_group
, NULL
);
1816 put_numa_group(grp
);
1819 p
->numa_faults_memory
= NULL
;
1820 p
->numa_faults_buffer_memory
= NULL
;
1821 p
->numa_faults_cpu
= NULL
;
1822 p
->numa_faults_buffer_cpu
= NULL
;
1827 * Got a PROT_NONE fault for a page on @node.
1829 void task_numa_fault(int last_cpupid
, int mem_node
, int pages
, int flags
)
1831 struct task_struct
*p
= current
;
1832 bool migrated
= flags
& TNF_MIGRATED
;
1833 int cpu_node
= task_node(current
);
1834 int local
= !!(flags
& TNF_FAULT_LOCAL
);
1837 if (!numabalancing_enabled
)
1840 /* for example, ksmd faulting in a user's mm */
1844 /* Allocate buffer to track faults on a per-node basis */
1845 if (unlikely(!p
->numa_faults_memory
)) {
1846 int size
= sizeof(*p
->numa_faults_memory
) *
1847 NR_NUMA_HINT_FAULT_BUCKETS
* nr_node_ids
;
1849 p
->numa_faults_memory
= kzalloc(size
, GFP_KERNEL
|__GFP_NOWARN
);
1850 if (!p
->numa_faults_memory
)
1853 BUG_ON(p
->numa_faults_buffer_memory
);
1855 * The averaged statistics, shared & private, memory & cpu,
1856 * occupy the first half of the array. The second half of the
1857 * array is for current counters, which are averaged into the
1858 * first set by task_numa_placement.
1860 p
->numa_faults_cpu
= p
->numa_faults_memory
+ (2 * nr_node_ids
);
1861 p
->numa_faults_buffer_memory
= p
->numa_faults_memory
+ (4 * nr_node_ids
);
1862 p
->numa_faults_buffer_cpu
= p
->numa_faults_memory
+ (6 * nr_node_ids
);
1863 p
->total_numa_faults
= 0;
1864 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1868 * First accesses are treated as private, otherwise consider accesses
1869 * to be private if the accessing pid has not changed
1871 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
1874 priv
= cpupid_match_pid(p
, last_cpupid
);
1875 if (!priv
&& !(flags
& TNF_NO_GROUP
))
1876 task_numa_group(p
, last_cpupid
, flags
, &priv
);
1880 * If a workload spans multiple NUMA nodes, a shared fault that
1881 * occurs wholly within the set of nodes that the workload is
1882 * actively using should be counted as local. This allows the
1883 * scan rate to slow down when a workload has settled down.
1885 if (!priv
&& !local
&& p
->numa_group
&&
1886 node_isset(cpu_node
, p
->numa_group
->active_nodes
) &&
1887 node_isset(mem_node
, p
->numa_group
->active_nodes
))
1890 task_numa_placement(p
);
1893 * Retry task to preferred node migration periodically, in case it
1894 * case it previously failed, or the scheduler moved us.
1896 if (time_after(jiffies
, p
->numa_migrate_retry
))
1897 numa_migrate_preferred(p
);
1900 p
->numa_pages_migrated
+= pages
;
1902 p
->numa_faults_buffer_memory
[task_faults_idx(mem_node
, priv
)] += pages
;
1903 p
->numa_faults_buffer_cpu
[task_faults_idx(cpu_node
, priv
)] += pages
;
1904 p
->numa_faults_locality
[local
] += pages
;
1907 static void reset_ptenuma_scan(struct task_struct
*p
)
1909 ACCESS_ONCE(p
->mm
->numa_scan_seq
)++;
1910 p
->mm
->numa_scan_offset
= 0;
1914 * The expensive part of numa migration is done from task_work context.
1915 * Triggered from task_tick_numa().
1917 void task_numa_work(struct callback_head
*work
)
1919 unsigned long migrate
, next_scan
, now
= jiffies
;
1920 struct task_struct
*p
= current
;
1921 struct mm_struct
*mm
= p
->mm
;
1922 struct vm_area_struct
*vma
;
1923 unsigned long start
, end
;
1924 unsigned long nr_pte_updates
= 0;
1927 WARN_ON_ONCE(p
!= container_of(work
, struct task_struct
, numa_work
));
1929 work
->next
= work
; /* protect against double add */
1931 * Who cares about NUMA placement when they're dying.
1933 * NOTE: make sure not to dereference p->mm before this check,
1934 * exit_task_work() happens _after_ exit_mm() so we could be called
1935 * without p->mm even though we still had it when we enqueued this
1938 if (p
->flags
& PF_EXITING
)
1941 if (!mm
->numa_next_scan
) {
1942 mm
->numa_next_scan
= now
+
1943 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
1947 * Enforce maximal scan/migration frequency..
1949 migrate
= mm
->numa_next_scan
;
1950 if (time_before(now
, migrate
))
1953 if (p
->numa_scan_period
== 0) {
1954 p
->numa_scan_period_max
= task_scan_max(p
);
1955 p
->numa_scan_period
= task_scan_min(p
);
1958 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
1959 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
1963 * Delay this task enough that another task of this mm will likely win
1964 * the next time around.
1966 p
->node_stamp
+= 2 * TICK_NSEC
;
1968 start
= mm
->numa_scan_offset
;
1969 pages
= sysctl_numa_balancing_scan_size
;
1970 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
1974 down_read(&mm
->mmap_sem
);
1975 vma
= find_vma(mm
, start
);
1977 reset_ptenuma_scan(p
);
1981 for (; vma
; vma
= vma
->vm_next
) {
1982 if (!vma_migratable(vma
) || !vma_policy_mof(vma
))
1986 * Shared library pages mapped by multiple processes are not
1987 * migrated as it is expected they are cache replicated. Avoid
1988 * hinting faults in read-only file-backed mappings or the vdso
1989 * as migrating the pages will be of marginal benefit.
1992 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
1996 * Skip inaccessible VMAs to avoid any confusion between
1997 * PROT_NONE and NUMA hinting ptes
1999 if (!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)))
2003 start
= max(start
, vma
->vm_start
);
2004 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
2005 end
= min(end
, vma
->vm_end
);
2006 nr_pte_updates
+= change_prot_numa(vma
, start
, end
);
2009 * Scan sysctl_numa_balancing_scan_size but ensure that
2010 * at least one PTE is updated so that unused virtual
2011 * address space is quickly skipped.
2014 pages
-= (end
- start
) >> PAGE_SHIFT
;
2021 } while (end
!= vma
->vm_end
);
2026 * It is possible to reach the end of the VMA list but the last few
2027 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2028 * would find the !migratable VMA on the next scan but not reset the
2029 * scanner to the start so check it now.
2032 mm
->numa_scan_offset
= start
;
2034 reset_ptenuma_scan(p
);
2035 up_read(&mm
->mmap_sem
);
2039 * Drive the periodic memory faults..
2041 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2043 struct callback_head
*work
= &curr
->numa_work
;
2047 * We don't care about NUMA placement if we don't have memory.
2049 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
2053 * Using runtime rather than walltime has the dual advantage that
2054 * we (mostly) drive the selection from busy threads and that the
2055 * task needs to have done some actual work before we bother with
2058 now
= curr
->se
.sum_exec_runtime
;
2059 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
2061 if (now
- curr
->node_stamp
> period
) {
2062 if (!curr
->node_stamp
)
2063 curr
->numa_scan_period
= task_scan_min(curr
);
2064 curr
->node_stamp
+= period
;
2066 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
2067 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
2068 task_work_add(curr
, work
, true);
2073 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2077 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
2081 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
2084 #endif /* CONFIG_NUMA_BALANCING */
2087 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2089 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
2090 if (!parent_entity(se
))
2091 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2093 if (entity_is_task(se
)) {
2094 struct rq
*rq
= rq_of(cfs_rq
);
2096 account_numa_enqueue(rq
, task_of(se
));
2097 list_add(&se
->group_node
, &rq
->cfs_tasks
);
2100 cfs_rq
->nr_running
++;
2104 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2106 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
2107 if (!parent_entity(se
))
2108 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2109 if (entity_is_task(se
)) {
2110 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
2111 list_del_init(&se
->group_node
);
2113 cfs_rq
->nr_running
--;
2116 #ifdef CONFIG_FAIR_GROUP_SCHED
2118 static inline long calc_tg_weight(struct task_group
*tg
, struct cfs_rq
*cfs_rq
)
2123 * Use this CPU's actual weight instead of the last load_contribution
2124 * to gain a more accurate current total weight. See
2125 * update_cfs_rq_load_contribution().
2127 tg_weight
= atomic_long_read(&tg
->load_avg
);
2128 tg_weight
-= cfs_rq
->tg_load_contrib
;
2129 tg_weight
+= cfs_rq
->load
.weight
;
2134 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2136 long tg_weight
, load
, shares
;
2138 tg_weight
= calc_tg_weight(tg
, cfs_rq
);
2139 load
= cfs_rq
->load
.weight
;
2141 shares
= (tg
->shares
* load
);
2143 shares
/= tg_weight
;
2145 if (shares
< MIN_SHARES
)
2146 shares
= MIN_SHARES
;
2147 if (shares
> tg
->shares
)
2148 shares
= tg
->shares
;
2152 # else /* CONFIG_SMP */
2153 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2157 # endif /* CONFIG_SMP */
2158 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
2159 unsigned long weight
)
2162 /* commit outstanding execution time */
2163 if (cfs_rq
->curr
== se
)
2164 update_curr(cfs_rq
);
2165 account_entity_dequeue(cfs_rq
, se
);
2168 update_load_set(&se
->load
, weight
);
2171 account_entity_enqueue(cfs_rq
, se
);
2174 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
2176 static void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2178 struct task_group
*tg
;
2179 struct sched_entity
*se
;
2183 se
= tg
->se
[cpu_of(rq_of(cfs_rq
))];
2184 if (!se
|| throttled_hierarchy(cfs_rq
))
2187 if (likely(se
->load
.weight
== tg
->shares
))
2190 shares
= calc_cfs_shares(cfs_rq
, tg
);
2192 reweight_entity(cfs_rq_of(se
), se
, shares
);
2194 #else /* CONFIG_FAIR_GROUP_SCHED */
2195 static inline void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2198 #endif /* CONFIG_FAIR_GROUP_SCHED */
2202 * We choose a half-life close to 1 scheduling period.
2203 * Note: The tables below are dependent on this value.
2205 #define LOAD_AVG_PERIOD 32
2206 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2207 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2209 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2210 static const u32 runnable_avg_yN_inv
[] = {
2211 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2212 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2213 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2214 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2215 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2216 0x85aac367, 0x82cd8698,
2220 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2221 * over-estimates when re-combining.
2223 static const u32 runnable_avg_yN_sum
[] = {
2224 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2225 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2226 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2231 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2233 static __always_inline u64
decay_load(u64 val
, u64 n
)
2235 unsigned int local_n
;
2239 else if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
2242 /* after bounds checking we can collapse to 32-bit */
2246 * As y^PERIOD = 1/2, we can combine
2247 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2248 * With a look-up table which covers y^n (n<PERIOD)
2250 * To achieve constant time decay_load.
2252 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
2253 val
>>= local_n
/ LOAD_AVG_PERIOD
;
2254 local_n
%= LOAD_AVG_PERIOD
;
2257 val
*= runnable_avg_yN_inv
[local_n
];
2258 /* We don't use SRR here since we always want to round down. */
2263 * For updates fully spanning n periods, the contribution to runnable
2264 * average will be: \Sum 1024*y^n
2266 * We can compute this reasonably efficiently by combining:
2267 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2269 static u32
__compute_runnable_contrib(u64 n
)
2273 if (likely(n
<= LOAD_AVG_PERIOD
))
2274 return runnable_avg_yN_sum
[n
];
2275 else if (unlikely(n
>= LOAD_AVG_MAX_N
))
2276 return LOAD_AVG_MAX
;
2278 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2280 contrib
/= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2281 contrib
+= runnable_avg_yN_sum
[LOAD_AVG_PERIOD
];
2283 n
-= LOAD_AVG_PERIOD
;
2284 } while (n
> LOAD_AVG_PERIOD
);
2286 contrib
= decay_load(contrib
, n
);
2287 return contrib
+ runnable_avg_yN_sum
[n
];
2291 * We can represent the historical contribution to runnable average as the
2292 * coefficients of a geometric series. To do this we sub-divide our runnable
2293 * history into segments of approximately 1ms (1024us); label the segment that
2294 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2296 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2298 * (now) (~1ms ago) (~2ms ago)
2300 * Let u_i denote the fraction of p_i that the entity was runnable.
2302 * We then designate the fractions u_i as our co-efficients, yielding the
2303 * following representation of historical load:
2304 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2306 * We choose y based on the with of a reasonably scheduling period, fixing:
2309 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2310 * approximately half as much as the contribution to load within the last ms
2313 * When a period "rolls over" and we have new u_0`, multiplying the previous
2314 * sum again by y is sufficient to update:
2315 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2316 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2318 static __always_inline
int __update_entity_runnable_avg(u64 now
,
2319 struct sched_avg
*sa
,
2323 u32 runnable_contrib
;
2324 int delta_w
, decayed
= 0;
2326 delta
= now
- sa
->last_runnable_update
;
2328 * This should only happen when time goes backwards, which it
2329 * unfortunately does during sched clock init when we swap over to TSC.
2331 if ((s64
)delta
< 0) {
2332 sa
->last_runnable_update
= now
;
2337 * Use 1024ns as the unit of measurement since it's a reasonable
2338 * approximation of 1us and fast to compute.
2343 sa
->last_runnable_update
= now
;
2345 /* delta_w is the amount already accumulated against our next period */
2346 delta_w
= sa
->runnable_avg_period
% 1024;
2347 if (delta
+ delta_w
>= 1024) {
2348 /* period roll-over */
2352 * Now that we know we're crossing a period boundary, figure
2353 * out how much from delta we need to complete the current
2354 * period and accrue it.
2356 delta_w
= 1024 - delta_w
;
2358 sa
->runnable_avg_sum
+= delta_w
;
2359 sa
->runnable_avg_period
+= delta_w
;
2363 /* Figure out how many additional periods this update spans */
2364 periods
= delta
/ 1024;
2367 sa
->runnable_avg_sum
= decay_load(sa
->runnable_avg_sum
,
2369 sa
->runnable_avg_period
= decay_load(sa
->runnable_avg_period
,
2372 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2373 runnable_contrib
= __compute_runnable_contrib(periods
);
2375 sa
->runnable_avg_sum
+= runnable_contrib
;
2376 sa
->runnable_avg_period
+= runnable_contrib
;
2379 /* Remainder of delta accrued against u_0` */
2381 sa
->runnable_avg_sum
+= delta
;
2382 sa
->runnable_avg_period
+= delta
;
2387 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2388 static inline u64
__synchronize_entity_decay(struct sched_entity
*se
)
2390 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2391 u64 decays
= atomic64_read(&cfs_rq
->decay_counter
);
2393 decays
-= se
->avg
.decay_count
;
2397 se
->avg
.load_avg_contrib
= decay_load(se
->avg
.load_avg_contrib
, decays
);
2398 se
->avg
.decay_count
= 0;
2403 #ifdef CONFIG_FAIR_GROUP_SCHED
2404 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2407 struct task_group
*tg
= cfs_rq
->tg
;
2410 tg_contrib
= cfs_rq
->runnable_load_avg
+ cfs_rq
->blocked_load_avg
;
2411 tg_contrib
-= cfs_rq
->tg_load_contrib
;
2416 if (force_update
|| abs(tg_contrib
) > cfs_rq
->tg_load_contrib
/ 8) {
2417 atomic_long_add(tg_contrib
, &tg
->load_avg
);
2418 cfs_rq
->tg_load_contrib
+= tg_contrib
;
2423 * Aggregate cfs_rq runnable averages into an equivalent task_group
2424 * representation for computing load contributions.
2426 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2427 struct cfs_rq
*cfs_rq
)
2429 struct task_group
*tg
= cfs_rq
->tg
;
2432 /* The fraction of a cpu used by this cfs_rq */
2433 contrib
= div_u64((u64
)sa
->runnable_avg_sum
<< NICE_0_SHIFT
,
2434 sa
->runnable_avg_period
+ 1);
2435 contrib
-= cfs_rq
->tg_runnable_contrib
;
2437 if (abs(contrib
) > cfs_rq
->tg_runnable_contrib
/ 64) {
2438 atomic_add(contrib
, &tg
->runnable_avg
);
2439 cfs_rq
->tg_runnable_contrib
+= contrib
;
2443 static inline void __update_group_entity_contrib(struct sched_entity
*se
)
2445 struct cfs_rq
*cfs_rq
= group_cfs_rq(se
);
2446 struct task_group
*tg
= cfs_rq
->tg
;
2451 contrib
= cfs_rq
->tg_load_contrib
* tg
->shares
;
2452 se
->avg
.load_avg_contrib
= div_u64(contrib
,
2453 atomic_long_read(&tg
->load_avg
) + 1);
2456 * For group entities we need to compute a correction term in the case
2457 * that they are consuming <1 cpu so that we would contribute the same
2458 * load as a task of equal weight.
2460 * Explicitly co-ordinating this measurement would be expensive, but
2461 * fortunately the sum of each cpus contribution forms a usable
2462 * lower-bound on the true value.
2464 * Consider the aggregate of 2 contributions. Either they are disjoint
2465 * (and the sum represents true value) or they are disjoint and we are
2466 * understating by the aggregate of their overlap.
2468 * Extending this to N cpus, for a given overlap, the maximum amount we
2469 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2470 * cpus that overlap for this interval and w_i is the interval width.
2472 * On a small machine; the first term is well-bounded which bounds the
2473 * total error since w_i is a subset of the period. Whereas on a
2474 * larger machine, while this first term can be larger, if w_i is the
2475 * of consequential size guaranteed to see n_i*w_i quickly converge to
2476 * our upper bound of 1-cpu.
2478 runnable_avg
= atomic_read(&tg
->runnable_avg
);
2479 if (runnable_avg
< NICE_0_LOAD
) {
2480 se
->avg
.load_avg_contrib
*= runnable_avg
;
2481 se
->avg
.load_avg_contrib
>>= NICE_0_SHIFT
;
2485 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
)
2487 __update_entity_runnable_avg(rq_clock_task(rq
), &rq
->avg
, runnable
);
2488 __update_tg_runnable_avg(&rq
->avg
, &rq
->cfs
);
2490 #else /* CONFIG_FAIR_GROUP_SCHED */
2491 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2492 int force_update
) {}
2493 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2494 struct cfs_rq
*cfs_rq
) {}
2495 static inline void __update_group_entity_contrib(struct sched_entity
*se
) {}
2496 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2497 #endif /* CONFIG_FAIR_GROUP_SCHED */
2499 static inline void __update_task_entity_contrib(struct sched_entity
*se
)
2503 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2504 contrib
= se
->avg
.runnable_avg_sum
* scale_load_down(se
->load
.weight
);
2505 contrib
/= (se
->avg
.runnable_avg_period
+ 1);
2506 se
->avg
.load_avg_contrib
= scale_load(contrib
);
2509 /* Compute the current contribution to load_avg by se, return any delta */
2510 static long __update_entity_load_avg_contrib(struct sched_entity
*se
)
2512 long old_contrib
= se
->avg
.load_avg_contrib
;
2514 if (entity_is_task(se
)) {
2515 __update_task_entity_contrib(se
);
2517 __update_tg_runnable_avg(&se
->avg
, group_cfs_rq(se
));
2518 __update_group_entity_contrib(se
);
2521 return se
->avg
.load_avg_contrib
- old_contrib
;
2524 static inline void subtract_blocked_load_contrib(struct cfs_rq
*cfs_rq
,
2527 if (likely(load_contrib
< cfs_rq
->blocked_load_avg
))
2528 cfs_rq
->blocked_load_avg
-= load_contrib
;
2530 cfs_rq
->blocked_load_avg
= 0;
2533 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
2535 /* Update a sched_entity's runnable average */
2536 static inline void update_entity_load_avg(struct sched_entity
*se
,
2539 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2544 * For a group entity we need to use their owned cfs_rq_clock_task() in
2545 * case they are the parent of a throttled hierarchy.
2547 if (entity_is_task(se
))
2548 now
= cfs_rq_clock_task(cfs_rq
);
2550 now
= cfs_rq_clock_task(group_cfs_rq(se
));
2552 if (!__update_entity_runnable_avg(now
, &se
->avg
, se
->on_rq
))
2555 contrib_delta
= __update_entity_load_avg_contrib(se
);
2561 cfs_rq
->runnable_load_avg
+= contrib_delta
;
2563 subtract_blocked_load_contrib(cfs_rq
, -contrib_delta
);
2567 * Decay the load contributed by all blocked children and account this so that
2568 * their contribution may appropriately discounted when they wake up.
2570 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
, int force_update
)
2572 u64 now
= cfs_rq_clock_task(cfs_rq
) >> 20;
2575 decays
= now
- cfs_rq
->last_decay
;
2576 if (!decays
&& !force_update
)
2579 if (atomic_long_read(&cfs_rq
->removed_load
)) {
2580 unsigned long removed_load
;
2581 removed_load
= atomic_long_xchg(&cfs_rq
->removed_load
, 0);
2582 subtract_blocked_load_contrib(cfs_rq
, removed_load
);
2586 cfs_rq
->blocked_load_avg
= decay_load(cfs_rq
->blocked_load_avg
,
2588 atomic64_add(decays
, &cfs_rq
->decay_counter
);
2589 cfs_rq
->last_decay
= now
;
2592 __update_cfs_rq_tg_load_contrib(cfs_rq
, force_update
);
2595 /* Add the load generated by se into cfs_rq's child load-average */
2596 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2597 struct sched_entity
*se
,
2601 * We track migrations using entity decay_count <= 0, on a wake-up
2602 * migration we use a negative decay count to track the remote decays
2603 * accumulated while sleeping.
2605 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2606 * are seen by enqueue_entity_load_avg() as a migration with an already
2607 * constructed load_avg_contrib.
2609 if (unlikely(se
->avg
.decay_count
<= 0)) {
2610 se
->avg
.last_runnable_update
= rq_clock_task(rq_of(cfs_rq
));
2611 if (se
->avg
.decay_count
) {
2613 * In a wake-up migration we have to approximate the
2614 * time sleeping. This is because we can't synchronize
2615 * clock_task between the two cpus, and it is not
2616 * guaranteed to be read-safe. Instead, we can
2617 * approximate this using our carried decays, which are
2618 * explicitly atomically readable.
2620 se
->avg
.last_runnable_update
-= (-se
->avg
.decay_count
)
2622 update_entity_load_avg(se
, 0);
2623 /* Indicate that we're now synchronized and on-rq */
2624 se
->avg
.decay_count
= 0;
2628 __synchronize_entity_decay(se
);
2631 /* migrated tasks did not contribute to our blocked load */
2633 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
2634 update_entity_load_avg(se
, 0);
2637 cfs_rq
->runnable_load_avg
+= se
->avg
.load_avg_contrib
;
2638 /* we force update consideration on load-balancer moves */
2639 update_cfs_rq_blocked_load(cfs_rq
, !wakeup
);
2643 * Remove se's load from this cfs_rq child load-average, if the entity is
2644 * transitioning to a blocked state we track its projected decay using
2647 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2648 struct sched_entity
*se
,
2651 update_entity_load_avg(se
, 1);
2652 /* we force update consideration on load-balancer moves */
2653 update_cfs_rq_blocked_load(cfs_rq
, !sleep
);
2655 cfs_rq
->runnable_load_avg
-= se
->avg
.load_avg_contrib
;
2657 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
2658 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
2659 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2663 * Update the rq's load with the elapsed running time before entering
2664 * idle. if the last scheduled task is not a CFS task, idle_enter will
2665 * be the only way to update the runnable statistic.
2667 void idle_enter_fair(struct rq
*this_rq
)
2669 update_rq_runnable_avg(this_rq
, 1);
2673 * Update the rq's load with the elapsed idle time before a task is
2674 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2675 * be the only way to update the runnable statistic.
2677 void idle_exit_fair(struct rq
*this_rq
)
2679 update_rq_runnable_avg(this_rq
, 0);
2682 static int idle_balance(struct rq
*this_rq
);
2684 #else /* CONFIG_SMP */
2686 static inline void update_entity_load_avg(struct sched_entity
*se
,
2687 int update_cfs_rq
) {}
2688 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2689 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2690 struct sched_entity
*se
,
2692 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2693 struct sched_entity
*se
,
2695 static inline void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
2696 int force_update
) {}
2698 static inline int idle_balance(struct rq
*rq
)
2703 #endif /* CONFIG_SMP */
2705 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2707 #ifdef CONFIG_SCHEDSTATS
2708 struct task_struct
*tsk
= NULL
;
2710 if (entity_is_task(se
))
2713 if (se
->statistics
.sleep_start
) {
2714 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.sleep_start
;
2719 if (unlikely(delta
> se
->statistics
.sleep_max
))
2720 se
->statistics
.sleep_max
= delta
;
2722 se
->statistics
.sleep_start
= 0;
2723 se
->statistics
.sum_sleep_runtime
+= delta
;
2726 account_scheduler_latency(tsk
, delta
>> 10, 1);
2727 trace_sched_stat_sleep(tsk
, delta
);
2730 if (se
->statistics
.block_start
) {
2731 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.block_start
;
2736 if (unlikely(delta
> se
->statistics
.block_max
))
2737 se
->statistics
.block_max
= delta
;
2739 se
->statistics
.block_start
= 0;
2740 se
->statistics
.sum_sleep_runtime
+= delta
;
2743 if (tsk
->in_iowait
) {
2744 se
->statistics
.iowait_sum
+= delta
;
2745 se
->statistics
.iowait_count
++;
2746 trace_sched_stat_iowait(tsk
, delta
);
2749 trace_sched_stat_blocked(tsk
, delta
);
2752 * Blocking time is in units of nanosecs, so shift by
2753 * 20 to get a milliseconds-range estimation of the
2754 * amount of time that the task spent sleeping:
2756 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
2757 profile_hits(SLEEP_PROFILING
,
2758 (void *)get_wchan(tsk
),
2761 account_scheduler_latency(tsk
, delta
>> 10, 0);
2767 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2769 #ifdef CONFIG_SCHED_DEBUG
2770 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
2775 if (d
> 3*sysctl_sched_latency
)
2776 schedstat_inc(cfs_rq
, nr_spread_over
);
2781 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
2783 u64 vruntime
= cfs_rq
->min_vruntime
;
2786 * The 'current' period is already promised to the current tasks,
2787 * however the extra weight of the new task will slow them down a
2788 * little, place the new task so that it fits in the slot that
2789 * stays open at the end.
2791 if (initial
&& sched_feat(START_DEBIT
))
2792 vruntime
+= sched_vslice(cfs_rq
, se
);
2794 /* sleeps up to a single latency don't count. */
2796 unsigned long thresh
= sysctl_sched_latency
;
2799 * Halve their sleep time's effect, to allow
2800 * for a gentler effect of sleepers:
2802 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
2808 /* ensure we never gain time by being placed backwards. */
2809 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
2812 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
2815 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2818 * Update the normalized vruntime before updating min_vruntime
2819 * through calling update_curr().
2821 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
2822 se
->vruntime
+= cfs_rq
->min_vruntime
;
2825 * Update run-time statistics of the 'current'.
2827 update_curr(cfs_rq
);
2828 enqueue_entity_load_avg(cfs_rq
, se
, flags
& ENQUEUE_WAKEUP
);
2829 account_entity_enqueue(cfs_rq
, se
);
2830 update_cfs_shares(cfs_rq
);
2832 if (flags
& ENQUEUE_WAKEUP
) {
2833 place_entity(cfs_rq
, se
, 0);
2834 enqueue_sleeper(cfs_rq
, se
);
2837 update_stats_enqueue(cfs_rq
, se
);
2838 check_spread(cfs_rq
, se
);
2839 if (se
!= cfs_rq
->curr
)
2840 __enqueue_entity(cfs_rq
, se
);
2843 if (cfs_rq
->nr_running
== 1) {
2844 list_add_leaf_cfs_rq(cfs_rq
);
2845 check_enqueue_throttle(cfs_rq
);
2849 static void __clear_buddies_last(struct sched_entity
*se
)
2851 for_each_sched_entity(se
) {
2852 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2853 if (cfs_rq
->last
!= se
)
2856 cfs_rq
->last
= NULL
;
2860 static void __clear_buddies_next(struct sched_entity
*se
)
2862 for_each_sched_entity(se
) {
2863 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2864 if (cfs_rq
->next
!= se
)
2867 cfs_rq
->next
= NULL
;
2871 static void __clear_buddies_skip(struct sched_entity
*se
)
2873 for_each_sched_entity(se
) {
2874 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2875 if (cfs_rq
->skip
!= se
)
2878 cfs_rq
->skip
= NULL
;
2882 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2884 if (cfs_rq
->last
== se
)
2885 __clear_buddies_last(se
);
2887 if (cfs_rq
->next
== se
)
2888 __clear_buddies_next(se
);
2890 if (cfs_rq
->skip
== se
)
2891 __clear_buddies_skip(se
);
2894 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
2897 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
2900 * Update run-time statistics of the 'current'.
2902 update_curr(cfs_rq
);
2903 dequeue_entity_load_avg(cfs_rq
, se
, flags
& DEQUEUE_SLEEP
);
2905 update_stats_dequeue(cfs_rq
, se
);
2906 if (flags
& DEQUEUE_SLEEP
) {
2907 #ifdef CONFIG_SCHEDSTATS
2908 if (entity_is_task(se
)) {
2909 struct task_struct
*tsk
= task_of(se
);
2911 if (tsk
->state
& TASK_INTERRUPTIBLE
)
2912 se
->statistics
.sleep_start
= rq_clock(rq_of(cfs_rq
));
2913 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
2914 se
->statistics
.block_start
= rq_clock(rq_of(cfs_rq
));
2919 clear_buddies(cfs_rq
, se
);
2921 if (se
!= cfs_rq
->curr
)
2922 __dequeue_entity(cfs_rq
, se
);
2924 account_entity_dequeue(cfs_rq
, se
);
2927 * Normalize the entity after updating the min_vruntime because the
2928 * update can refer to the ->curr item and we need to reflect this
2929 * movement in our normalized position.
2931 if (!(flags
& DEQUEUE_SLEEP
))
2932 se
->vruntime
-= cfs_rq
->min_vruntime
;
2934 /* return excess runtime on last dequeue */
2935 return_cfs_rq_runtime(cfs_rq
);
2937 update_min_vruntime(cfs_rq
);
2938 update_cfs_shares(cfs_rq
);
2942 * Preempt the current task with a newly woken task if needed:
2945 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
2947 unsigned long ideal_runtime
, delta_exec
;
2948 struct sched_entity
*se
;
2951 ideal_runtime
= sched_slice(cfs_rq
, curr
);
2952 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
2953 if (delta_exec
> ideal_runtime
) {
2954 resched_curr(rq_of(cfs_rq
));
2956 * The current task ran long enough, ensure it doesn't get
2957 * re-elected due to buddy favours.
2959 clear_buddies(cfs_rq
, curr
);
2964 * Ensure that a task that missed wakeup preemption by a
2965 * narrow margin doesn't have to wait for a full slice.
2966 * This also mitigates buddy induced latencies under load.
2968 if (delta_exec
< sysctl_sched_min_granularity
)
2971 se
= __pick_first_entity(cfs_rq
);
2972 delta
= curr
->vruntime
- se
->vruntime
;
2977 if (delta
> ideal_runtime
)
2978 resched_curr(rq_of(cfs_rq
));
2982 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2984 /* 'current' is not kept within the tree. */
2987 * Any task has to be enqueued before it get to execute on
2988 * a CPU. So account for the time it spent waiting on the
2991 update_stats_wait_end(cfs_rq
, se
);
2992 __dequeue_entity(cfs_rq
, se
);
2995 update_stats_curr_start(cfs_rq
, se
);
2997 #ifdef CONFIG_SCHEDSTATS
2999 * Track our maximum slice length, if the CPU's load is at
3000 * least twice that of our own weight (i.e. dont track it
3001 * when there are only lesser-weight tasks around):
3003 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
3004 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
3005 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
3008 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
3012 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
3015 * Pick the next process, keeping these things in mind, in this order:
3016 * 1) keep things fair between processes/task groups
3017 * 2) pick the "next" process, since someone really wants that to run
3018 * 3) pick the "last" process, for cache locality
3019 * 4) do not run the "skip" process, if something else is available
3021 static struct sched_entity
*
3022 pick_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
3024 struct sched_entity
*left
= __pick_first_entity(cfs_rq
);
3025 struct sched_entity
*se
;
3028 * If curr is set we have to see if its left of the leftmost entity
3029 * still in the tree, provided there was anything in the tree at all.
3031 if (!left
|| (curr
&& entity_before(curr
, left
)))
3034 se
= left
; /* ideally we run the leftmost entity */
3037 * Avoid running the skip buddy, if running something else can
3038 * be done without getting too unfair.
3040 if (cfs_rq
->skip
== se
) {
3041 struct sched_entity
*second
;
3044 second
= __pick_first_entity(cfs_rq
);
3046 second
= __pick_next_entity(se
);
3047 if (!second
|| (curr
&& entity_before(curr
, second
)))
3051 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
3056 * Prefer last buddy, try to return the CPU to a preempted task.
3058 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
3062 * Someone really wants this to run. If it's not unfair, run it.
3064 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
3067 clear_buddies(cfs_rq
, se
);
3072 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3074 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
3077 * If still on the runqueue then deactivate_task()
3078 * was not called and update_curr() has to be done:
3081 update_curr(cfs_rq
);
3083 /* throttle cfs_rqs exceeding runtime */
3084 check_cfs_rq_runtime(cfs_rq
);
3086 check_spread(cfs_rq
, prev
);
3088 update_stats_wait_start(cfs_rq
, prev
);
3089 /* Put 'current' back into the tree. */
3090 __enqueue_entity(cfs_rq
, prev
);
3091 /* in !on_rq case, update occurred at dequeue */
3092 update_entity_load_avg(prev
, 1);
3094 cfs_rq
->curr
= NULL
;
3098 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
3101 * Update run-time statistics of the 'current'.
3103 update_curr(cfs_rq
);
3106 * Ensure that runnable average is periodically updated.
3108 update_entity_load_avg(curr
, 1);
3109 update_cfs_rq_blocked_load(cfs_rq
, 1);
3110 update_cfs_shares(cfs_rq
);
3112 #ifdef CONFIG_SCHED_HRTICK
3114 * queued ticks are scheduled to match the slice, so don't bother
3115 * validating it and just reschedule.
3118 resched_curr(rq_of(cfs_rq
));
3122 * don't let the period tick interfere with the hrtick preemption
3124 if (!sched_feat(DOUBLE_TICK
) &&
3125 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
3129 if (cfs_rq
->nr_running
> 1)
3130 check_preempt_tick(cfs_rq
, curr
);
3134 /**************************************************
3135 * CFS bandwidth control machinery
3138 #ifdef CONFIG_CFS_BANDWIDTH
3140 #ifdef HAVE_JUMP_LABEL
3141 static struct static_key __cfs_bandwidth_used
;
3143 static inline bool cfs_bandwidth_used(void)
3145 return static_key_false(&__cfs_bandwidth_used
);
3148 void cfs_bandwidth_usage_inc(void)
3150 static_key_slow_inc(&__cfs_bandwidth_used
);
3153 void cfs_bandwidth_usage_dec(void)
3155 static_key_slow_dec(&__cfs_bandwidth_used
);
3157 #else /* HAVE_JUMP_LABEL */
3158 static bool cfs_bandwidth_used(void)
3163 void cfs_bandwidth_usage_inc(void) {}
3164 void cfs_bandwidth_usage_dec(void) {}
3165 #endif /* HAVE_JUMP_LABEL */
3168 * default period for cfs group bandwidth.
3169 * default: 0.1s, units: nanoseconds
3171 static inline u64
default_cfs_period(void)
3173 return 100000000ULL;
3176 static inline u64
sched_cfs_bandwidth_slice(void)
3178 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
3182 * Replenish runtime according to assigned quota and update expiration time.
3183 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3184 * additional synchronization around rq->lock.
3186 * requires cfs_b->lock
3188 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
3192 if (cfs_b
->quota
== RUNTIME_INF
)
3195 now
= sched_clock_cpu(smp_processor_id());
3196 cfs_b
->runtime
= cfs_b
->quota
;
3197 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
3200 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3202 return &tg
->cfs_bandwidth
;
3205 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3206 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3208 if (unlikely(cfs_rq
->throttle_count
))
3209 return cfs_rq
->throttled_clock_task
;
3211 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
3214 /* returns 0 on failure to allocate runtime */
3215 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3217 struct task_group
*tg
= cfs_rq
->tg
;
3218 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
3219 u64 amount
= 0, min_amount
, expires
;
3221 /* note: this is a positive sum as runtime_remaining <= 0 */
3222 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
3224 raw_spin_lock(&cfs_b
->lock
);
3225 if (cfs_b
->quota
== RUNTIME_INF
)
3226 amount
= min_amount
;
3229 * If the bandwidth pool has become inactive, then at least one
3230 * period must have elapsed since the last consumption.
3231 * Refresh the global state and ensure bandwidth timer becomes
3234 if (!cfs_b
->timer_active
) {
3235 __refill_cfs_bandwidth_runtime(cfs_b
);
3236 __start_cfs_bandwidth(cfs_b
, false);
3239 if (cfs_b
->runtime
> 0) {
3240 amount
= min(cfs_b
->runtime
, min_amount
);
3241 cfs_b
->runtime
-= amount
;
3245 expires
= cfs_b
->runtime_expires
;
3246 raw_spin_unlock(&cfs_b
->lock
);
3248 cfs_rq
->runtime_remaining
+= amount
;
3250 * we may have advanced our local expiration to account for allowed
3251 * spread between our sched_clock and the one on which runtime was
3254 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
3255 cfs_rq
->runtime_expires
= expires
;
3257 return cfs_rq
->runtime_remaining
> 0;
3261 * Note: This depends on the synchronization provided by sched_clock and the
3262 * fact that rq->clock snapshots this value.
3264 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3266 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3268 /* if the deadline is ahead of our clock, nothing to do */
3269 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
3272 if (cfs_rq
->runtime_remaining
< 0)
3276 * If the local deadline has passed we have to consider the
3277 * possibility that our sched_clock is 'fast' and the global deadline
3278 * has not truly expired.
3280 * Fortunately we can check determine whether this the case by checking
3281 * whether the global deadline has advanced. It is valid to compare
3282 * cfs_b->runtime_expires without any locks since we only care about
3283 * exact equality, so a partial write will still work.
3286 if (cfs_rq
->runtime_expires
!= cfs_b
->runtime_expires
) {
3287 /* extend local deadline, drift is bounded above by 2 ticks */
3288 cfs_rq
->runtime_expires
+= TICK_NSEC
;
3290 /* global deadline is ahead, expiration has passed */
3291 cfs_rq
->runtime_remaining
= 0;
3295 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3297 /* dock delta_exec before expiring quota (as it could span periods) */
3298 cfs_rq
->runtime_remaining
-= delta_exec
;
3299 expire_cfs_rq_runtime(cfs_rq
);
3301 if (likely(cfs_rq
->runtime_remaining
> 0))
3305 * if we're unable to extend our runtime we resched so that the active
3306 * hierarchy can be throttled
3308 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
3309 resched_curr(rq_of(cfs_rq
));
3312 static __always_inline
3313 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3315 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
3318 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
3321 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3323 return cfs_bandwidth_used() && cfs_rq
->throttled
;
3326 /* check whether cfs_rq, or any parent, is throttled */
3327 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3329 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
3333 * Ensure that neither of the group entities corresponding to src_cpu or
3334 * dest_cpu are members of a throttled hierarchy when performing group
3335 * load-balance operations.
3337 static inline int throttled_lb_pair(struct task_group
*tg
,
3338 int src_cpu
, int dest_cpu
)
3340 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
3342 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
3343 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
3345 return throttled_hierarchy(src_cfs_rq
) ||
3346 throttled_hierarchy(dest_cfs_rq
);
3349 /* updated child weight may affect parent so we have to do this bottom up */
3350 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
3352 struct rq
*rq
= data
;
3353 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3355 cfs_rq
->throttle_count
--;
3357 if (!cfs_rq
->throttle_count
) {
3358 /* adjust cfs_rq_clock_task() */
3359 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
3360 cfs_rq
->throttled_clock_task
;
3367 static int tg_throttle_down(struct task_group
*tg
, void *data
)
3369 struct rq
*rq
= data
;
3370 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3372 /* group is entering throttled state, stop time */
3373 if (!cfs_rq
->throttle_count
)
3374 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
3375 cfs_rq
->throttle_count
++;
3380 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
3382 struct rq
*rq
= rq_of(cfs_rq
);
3383 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3384 struct sched_entity
*se
;
3385 long task_delta
, dequeue
= 1;
3387 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
3389 /* freeze hierarchy runnable averages while throttled */
3391 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
3394 task_delta
= cfs_rq
->h_nr_running
;
3395 for_each_sched_entity(se
) {
3396 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
3397 /* throttled entity or throttle-on-deactivate */
3402 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
3403 qcfs_rq
->h_nr_running
-= task_delta
;
3405 if (qcfs_rq
->load
.weight
)
3410 sub_nr_running(rq
, task_delta
);
3412 cfs_rq
->throttled
= 1;
3413 cfs_rq
->throttled_clock
= rq_clock(rq
);
3414 raw_spin_lock(&cfs_b
->lock
);
3416 * Add to the _head_ of the list, so that an already-started
3417 * distribute_cfs_runtime will not see us
3419 list_add_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
3420 if (!cfs_b
->timer_active
)
3421 __start_cfs_bandwidth(cfs_b
, false);
3422 raw_spin_unlock(&cfs_b
->lock
);
3425 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
3427 struct rq
*rq
= rq_of(cfs_rq
);
3428 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3429 struct sched_entity
*se
;
3433 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
3435 cfs_rq
->throttled
= 0;
3437 update_rq_clock(rq
);
3439 raw_spin_lock(&cfs_b
->lock
);
3440 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
3441 list_del_rcu(&cfs_rq
->throttled_list
);
3442 raw_spin_unlock(&cfs_b
->lock
);
3444 /* update hierarchical throttle state */
3445 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
3447 if (!cfs_rq
->load
.weight
)
3450 task_delta
= cfs_rq
->h_nr_running
;
3451 for_each_sched_entity(se
) {
3455 cfs_rq
= cfs_rq_of(se
);
3457 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
3458 cfs_rq
->h_nr_running
+= task_delta
;
3460 if (cfs_rq_throttled(cfs_rq
))
3465 add_nr_running(rq
, task_delta
);
3467 /* determine whether we need to wake up potentially idle cpu */
3468 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
3472 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
3473 u64 remaining
, u64 expires
)
3475 struct cfs_rq
*cfs_rq
;
3477 u64 starting_runtime
= remaining
;
3480 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
3482 struct rq
*rq
= rq_of(cfs_rq
);
3484 raw_spin_lock(&rq
->lock
);
3485 if (!cfs_rq_throttled(cfs_rq
))
3488 runtime
= -cfs_rq
->runtime_remaining
+ 1;
3489 if (runtime
> remaining
)
3490 runtime
= remaining
;
3491 remaining
-= runtime
;
3493 cfs_rq
->runtime_remaining
+= runtime
;
3494 cfs_rq
->runtime_expires
= expires
;
3496 /* we check whether we're throttled above */
3497 if (cfs_rq
->runtime_remaining
> 0)
3498 unthrottle_cfs_rq(cfs_rq
);
3501 raw_spin_unlock(&rq
->lock
);
3508 return starting_runtime
- remaining
;
3512 * Responsible for refilling a task_group's bandwidth and unthrottling its
3513 * cfs_rqs as appropriate. If there has been no activity within the last
3514 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3515 * used to track this state.
3517 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
3519 u64 runtime
, runtime_expires
;
3522 /* no need to continue the timer with no bandwidth constraint */
3523 if (cfs_b
->quota
== RUNTIME_INF
)
3524 goto out_deactivate
;
3526 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3527 cfs_b
->nr_periods
+= overrun
;
3530 * idle depends on !throttled (for the case of a large deficit), and if
3531 * we're going inactive then everything else can be deferred
3533 if (cfs_b
->idle
&& !throttled
)
3534 goto out_deactivate
;
3537 * if we have relooped after returning idle once, we need to update our
3538 * status as actually running, so that other cpus doing
3539 * __start_cfs_bandwidth will stop trying to cancel us.
3541 cfs_b
->timer_active
= 1;
3543 __refill_cfs_bandwidth_runtime(cfs_b
);
3546 /* mark as potentially idle for the upcoming period */
3551 /* account preceding periods in which throttling occurred */
3552 cfs_b
->nr_throttled
+= overrun
;
3554 runtime_expires
= cfs_b
->runtime_expires
;
3557 * This check is repeated as we are holding onto the new bandwidth while
3558 * we unthrottle. This can potentially race with an unthrottled group
3559 * trying to acquire new bandwidth from the global pool. This can result
3560 * in us over-using our runtime if it is all used during this loop, but
3561 * only by limited amounts in that extreme case.
3563 while (throttled
&& cfs_b
->runtime
> 0) {
3564 runtime
= cfs_b
->runtime
;
3565 raw_spin_unlock(&cfs_b
->lock
);
3566 /* we can't nest cfs_b->lock while distributing bandwidth */
3567 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
3569 raw_spin_lock(&cfs_b
->lock
);
3571 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3573 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
3577 * While we are ensured activity in the period following an
3578 * unthrottle, this also covers the case in which the new bandwidth is
3579 * insufficient to cover the existing bandwidth deficit. (Forcing the
3580 * timer to remain active while there are any throttled entities.)
3587 cfs_b
->timer_active
= 0;
3591 /* a cfs_rq won't donate quota below this amount */
3592 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
3593 /* minimum remaining period time to redistribute slack quota */
3594 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
3595 /* how long we wait to gather additional slack before distributing */
3596 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
3599 * Are we near the end of the current quota period?
3601 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3602 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3603 * migrate_hrtimers, base is never cleared, so we are fine.
3605 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
3607 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
3610 /* if the call-back is running a quota refresh is already occurring */
3611 if (hrtimer_callback_running(refresh_timer
))
3614 /* is a quota refresh about to occur? */
3615 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
3616 if (remaining
< min_expire
)
3622 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
3624 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
3626 /* if there's a quota refresh soon don't bother with slack */
3627 if (runtime_refresh_within(cfs_b
, min_left
))
3630 start_bandwidth_timer(&cfs_b
->slack_timer
,
3631 ns_to_ktime(cfs_bandwidth_slack_period
));
3634 /* we know any runtime found here is valid as update_curr() precedes return */
3635 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3637 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3638 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
3640 if (slack_runtime
<= 0)
3643 raw_spin_lock(&cfs_b
->lock
);
3644 if (cfs_b
->quota
!= RUNTIME_INF
&&
3645 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
3646 cfs_b
->runtime
+= slack_runtime
;
3648 /* we are under rq->lock, defer unthrottling using a timer */
3649 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
3650 !list_empty(&cfs_b
->throttled_cfs_rq
))
3651 start_cfs_slack_bandwidth(cfs_b
);
3653 raw_spin_unlock(&cfs_b
->lock
);
3655 /* even if it's not valid for return we don't want to try again */
3656 cfs_rq
->runtime_remaining
-= slack_runtime
;
3659 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3661 if (!cfs_bandwidth_used())
3664 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
3667 __return_cfs_rq_runtime(cfs_rq
);
3671 * This is done with a timer (instead of inline with bandwidth return) since
3672 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3674 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
3676 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
3679 /* confirm we're still not at a refresh boundary */
3680 raw_spin_lock(&cfs_b
->lock
);
3681 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
3682 raw_spin_unlock(&cfs_b
->lock
);
3686 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
)
3687 runtime
= cfs_b
->runtime
;
3689 expires
= cfs_b
->runtime_expires
;
3690 raw_spin_unlock(&cfs_b
->lock
);
3695 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
3697 raw_spin_lock(&cfs_b
->lock
);
3698 if (expires
== cfs_b
->runtime_expires
)
3699 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
3700 raw_spin_unlock(&cfs_b
->lock
);
3704 * When a group wakes up we want to make sure that its quota is not already
3705 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3706 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3708 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
3710 if (!cfs_bandwidth_used())
3713 /* an active group must be handled by the update_curr()->put() path */
3714 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
3717 /* ensure the group is not already throttled */
3718 if (cfs_rq_throttled(cfs_rq
))
3721 /* update runtime allocation */
3722 account_cfs_rq_runtime(cfs_rq
, 0);
3723 if (cfs_rq
->runtime_remaining
<= 0)
3724 throttle_cfs_rq(cfs_rq
);
3727 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3728 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3730 if (!cfs_bandwidth_used())
3733 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
3737 * it's possible for a throttled entity to be forced into a running
3738 * state (e.g. set_curr_task), in this case we're finished.
3740 if (cfs_rq_throttled(cfs_rq
))
3743 throttle_cfs_rq(cfs_rq
);
3747 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
3749 struct cfs_bandwidth
*cfs_b
=
3750 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
3751 do_sched_cfs_slack_timer(cfs_b
);
3753 return HRTIMER_NORESTART
;
3756 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
3758 struct cfs_bandwidth
*cfs_b
=
3759 container_of(timer
, struct cfs_bandwidth
, period_timer
);
3764 raw_spin_lock(&cfs_b
->lock
);
3766 now
= hrtimer_cb_get_time(timer
);
3767 overrun
= hrtimer_forward(timer
, now
, cfs_b
->period
);
3772 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
3774 raw_spin_unlock(&cfs_b
->lock
);
3776 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
3779 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3781 raw_spin_lock_init(&cfs_b
->lock
);
3783 cfs_b
->quota
= RUNTIME_INF
;
3784 cfs_b
->period
= ns_to_ktime(default_cfs_period());
3786 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
3787 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3788 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
3789 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
3790 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
3793 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3795 cfs_rq
->runtime_enabled
= 0;
3796 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
3799 /* requires cfs_b->lock, may release to reprogram timer */
3800 void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
, bool force
)
3803 * The timer may be active because we're trying to set a new bandwidth
3804 * period or because we're racing with the tear-down path
3805 * (timer_active==0 becomes visible before the hrtimer call-back
3806 * terminates). In either case we ensure that it's re-programmed
3808 while (unlikely(hrtimer_active(&cfs_b
->period_timer
)) &&
3809 hrtimer_try_to_cancel(&cfs_b
->period_timer
) < 0) {
3810 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3811 raw_spin_unlock(&cfs_b
->lock
);
3813 raw_spin_lock(&cfs_b
->lock
);
3814 /* if someone else restarted the timer then we're done */
3815 if (!force
&& cfs_b
->timer_active
)
3819 cfs_b
->timer_active
= 1;
3820 start_bandwidth_timer(&cfs_b
->period_timer
, cfs_b
->period
);
3823 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
3825 hrtimer_cancel(&cfs_b
->period_timer
);
3826 hrtimer_cancel(&cfs_b
->slack_timer
);
3829 static void __maybe_unused
update_runtime_enabled(struct rq
*rq
)
3831 struct cfs_rq
*cfs_rq
;
3833 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
3834 struct cfs_bandwidth
*cfs_b
= &cfs_rq
->tg
->cfs_bandwidth
;
3836 raw_spin_lock(&cfs_b
->lock
);
3837 cfs_rq
->runtime_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
3838 raw_spin_unlock(&cfs_b
->lock
);
3842 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
3844 struct cfs_rq
*cfs_rq
;
3846 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
3847 if (!cfs_rq
->runtime_enabled
)
3851 * clock_task is not advancing so we just need to make sure
3852 * there's some valid quota amount
3854 cfs_rq
->runtime_remaining
= 1;
3856 * Offline rq is schedulable till cpu is completely disabled
3857 * in take_cpu_down(), so we prevent new cfs throttling here.
3859 cfs_rq
->runtime_enabled
= 0;
3861 if (cfs_rq_throttled(cfs_rq
))
3862 unthrottle_cfs_rq(cfs_rq
);
3866 #else /* CONFIG_CFS_BANDWIDTH */
3867 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3869 return rq_clock_task(rq_of(cfs_rq
));
3872 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
3873 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) { return false; }
3874 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
3875 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3877 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3882 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3887 static inline int throttled_lb_pair(struct task_group
*tg
,
3888 int src_cpu
, int dest_cpu
)
3893 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3895 #ifdef CONFIG_FAIR_GROUP_SCHED
3896 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
3899 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3903 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
3904 static inline void update_runtime_enabled(struct rq
*rq
) {}
3905 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
3907 #endif /* CONFIG_CFS_BANDWIDTH */
3909 /**************************************************
3910 * CFS operations on tasks:
3913 #ifdef CONFIG_SCHED_HRTICK
3914 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3916 struct sched_entity
*se
= &p
->se
;
3917 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3919 WARN_ON(task_rq(p
) != rq
);
3921 if (cfs_rq
->nr_running
> 1) {
3922 u64 slice
= sched_slice(cfs_rq
, se
);
3923 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
3924 s64 delta
= slice
- ran
;
3931 hrtick_start(rq
, delta
);
3936 * called from enqueue/dequeue and updates the hrtick when the
3937 * current task is from our class and nr_running is low enough
3940 static void hrtick_update(struct rq
*rq
)
3942 struct task_struct
*curr
= rq
->curr
;
3944 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
3947 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
3948 hrtick_start_fair(rq
, curr
);
3950 #else /* !CONFIG_SCHED_HRTICK */
3952 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
3956 static inline void hrtick_update(struct rq
*rq
)
3962 * The enqueue_task method is called before nr_running is
3963 * increased. Here we update the fair scheduling stats and
3964 * then put the task into the rbtree:
3967 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
3969 struct cfs_rq
*cfs_rq
;
3970 struct sched_entity
*se
= &p
->se
;
3972 for_each_sched_entity(se
) {
3975 cfs_rq
= cfs_rq_of(se
);
3976 enqueue_entity(cfs_rq
, se
, flags
);
3979 * end evaluation on encountering a throttled cfs_rq
3981 * note: in the case of encountering a throttled cfs_rq we will
3982 * post the final h_nr_running increment below.
3984 if (cfs_rq_throttled(cfs_rq
))
3986 cfs_rq
->h_nr_running
++;
3988 flags
= ENQUEUE_WAKEUP
;
3991 for_each_sched_entity(se
) {
3992 cfs_rq
= cfs_rq_of(se
);
3993 cfs_rq
->h_nr_running
++;
3995 if (cfs_rq_throttled(cfs_rq
))
3998 update_cfs_shares(cfs_rq
);
3999 update_entity_load_avg(se
, 1);
4003 update_rq_runnable_avg(rq
, rq
->nr_running
);
4004 add_nr_running(rq
, 1);
4009 static void set_next_buddy(struct sched_entity
*se
);
4012 * The dequeue_task method is called before nr_running is
4013 * decreased. We remove the task from the rbtree and
4014 * update the fair scheduling stats:
4016 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
4018 struct cfs_rq
*cfs_rq
;
4019 struct sched_entity
*se
= &p
->se
;
4020 int task_sleep
= flags
& DEQUEUE_SLEEP
;
4022 for_each_sched_entity(se
) {
4023 cfs_rq
= cfs_rq_of(se
);
4024 dequeue_entity(cfs_rq
, se
, flags
);
4027 * end evaluation on encountering a throttled cfs_rq
4029 * note: in the case of encountering a throttled cfs_rq we will
4030 * post the final h_nr_running decrement below.
4032 if (cfs_rq_throttled(cfs_rq
))
4034 cfs_rq
->h_nr_running
--;
4036 /* Don't dequeue parent if it has other entities besides us */
4037 if (cfs_rq
->load
.weight
) {
4039 * Bias pick_next to pick a task from this cfs_rq, as
4040 * p is sleeping when it is within its sched_slice.
4042 if (task_sleep
&& parent_entity(se
))
4043 set_next_buddy(parent_entity(se
));
4045 /* avoid re-evaluating load for this entity */
4046 se
= parent_entity(se
);
4049 flags
|= DEQUEUE_SLEEP
;
4052 for_each_sched_entity(se
) {
4053 cfs_rq
= cfs_rq_of(se
);
4054 cfs_rq
->h_nr_running
--;
4056 if (cfs_rq_throttled(cfs_rq
))
4059 update_cfs_shares(cfs_rq
);
4060 update_entity_load_avg(se
, 1);
4064 sub_nr_running(rq
, 1);
4065 update_rq_runnable_avg(rq
, 1);
4071 /* Used instead of source_load when we know the type == 0 */
4072 static unsigned long weighted_cpuload(const int cpu
)
4074 return cpu_rq(cpu
)->cfs
.runnable_load_avg
;
4078 * Return a low guess at the load of a migration-source cpu weighted
4079 * according to the scheduling class and "nice" value.
4081 * We want to under-estimate the load of migration sources, to
4082 * balance conservatively.
4084 static unsigned long source_load(int cpu
, int type
)
4086 struct rq
*rq
= cpu_rq(cpu
);
4087 unsigned long total
= weighted_cpuload(cpu
);
4089 if (type
== 0 || !sched_feat(LB_BIAS
))
4092 return min(rq
->cpu_load
[type
-1], total
);
4096 * Return a high guess at the load of a migration-target cpu weighted
4097 * according to the scheduling class and "nice" value.
4099 static unsigned long target_load(int cpu
, int type
)
4101 struct rq
*rq
= cpu_rq(cpu
);
4102 unsigned long total
= weighted_cpuload(cpu
);
4104 if (type
== 0 || !sched_feat(LB_BIAS
))
4107 return max(rq
->cpu_load
[type
-1], total
);
4110 static unsigned long capacity_of(int cpu
)
4112 return cpu_rq(cpu
)->cpu_capacity
;
4115 static unsigned long cpu_avg_load_per_task(int cpu
)
4117 struct rq
*rq
= cpu_rq(cpu
);
4118 unsigned long nr_running
= ACCESS_ONCE(rq
->cfs
.h_nr_running
);
4119 unsigned long load_avg
= rq
->cfs
.runnable_load_avg
;
4122 return load_avg
/ nr_running
;
4127 static void record_wakee(struct task_struct
*p
)
4130 * Rough decay (wiping) for cost saving, don't worry
4131 * about the boundary, really active task won't care
4134 if (time_after(jiffies
, current
->wakee_flip_decay_ts
+ HZ
)) {
4135 current
->wakee_flips
>>= 1;
4136 current
->wakee_flip_decay_ts
= jiffies
;
4139 if (current
->last_wakee
!= p
) {
4140 current
->last_wakee
= p
;
4141 current
->wakee_flips
++;
4145 static void task_waking_fair(struct task_struct
*p
)
4147 struct sched_entity
*se
= &p
->se
;
4148 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4151 #ifndef CONFIG_64BIT
4152 u64 min_vruntime_copy
;
4155 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
4157 min_vruntime
= cfs_rq
->min_vruntime
;
4158 } while (min_vruntime
!= min_vruntime_copy
);
4160 min_vruntime
= cfs_rq
->min_vruntime
;
4163 se
->vruntime
-= min_vruntime
;
4167 #ifdef CONFIG_FAIR_GROUP_SCHED
4169 * effective_load() calculates the load change as seen from the root_task_group
4171 * Adding load to a group doesn't make a group heavier, but can cause movement
4172 * of group shares between cpus. Assuming the shares were perfectly aligned one
4173 * can calculate the shift in shares.
4175 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4176 * on this @cpu and results in a total addition (subtraction) of @wg to the
4177 * total group weight.
4179 * Given a runqueue weight distribution (rw_i) we can compute a shares
4180 * distribution (s_i) using:
4182 * s_i = rw_i / \Sum rw_j (1)
4184 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4185 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4186 * shares distribution (s_i):
4188 * rw_i = { 2, 4, 1, 0 }
4189 * s_i = { 2/7, 4/7, 1/7, 0 }
4191 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4192 * task used to run on and the CPU the waker is running on), we need to
4193 * compute the effect of waking a task on either CPU and, in case of a sync
4194 * wakeup, compute the effect of the current task going to sleep.
4196 * So for a change of @wl to the local @cpu with an overall group weight change
4197 * of @wl we can compute the new shares distribution (s'_i) using:
4199 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4201 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4202 * differences in waking a task to CPU 0. The additional task changes the
4203 * weight and shares distributions like:
4205 * rw'_i = { 3, 4, 1, 0 }
4206 * s'_i = { 3/8, 4/8, 1/8, 0 }
4208 * We can then compute the difference in effective weight by using:
4210 * dw_i = S * (s'_i - s_i) (3)
4212 * Where 'S' is the group weight as seen by its parent.
4214 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4215 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4216 * 4/7) times the weight of the group.
4218 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4220 struct sched_entity
*se
= tg
->se
[cpu
];
4222 if (!tg
->parent
) /* the trivial, non-cgroup case */
4225 for_each_sched_entity(se
) {
4231 * W = @wg + \Sum rw_j
4233 W
= wg
+ calc_tg_weight(tg
, se
->my_q
);
4238 w
= se
->my_q
->load
.weight
+ wl
;
4241 * wl = S * s'_i; see (2)
4244 wl
= (w
* tg
->shares
) / W
;
4249 * Per the above, wl is the new se->load.weight value; since
4250 * those are clipped to [MIN_SHARES, ...) do so now. See
4251 * calc_cfs_shares().
4253 if (wl
< MIN_SHARES
)
4257 * wl = dw_i = S * (s'_i - s_i); see (3)
4259 wl
-= se
->load
.weight
;
4262 * Recursively apply this logic to all parent groups to compute
4263 * the final effective load change on the root group. Since
4264 * only the @tg group gets extra weight, all parent groups can
4265 * only redistribute existing shares. @wl is the shift in shares
4266 * resulting from this level per the above.
4275 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4282 static int wake_wide(struct task_struct
*p
)
4284 int factor
= this_cpu_read(sd_llc_size
);
4287 * Yeah, it's the switching-frequency, could means many wakee or
4288 * rapidly switch, use factor here will just help to automatically
4289 * adjust the loose-degree, so bigger node will lead to more pull.
4291 if (p
->wakee_flips
> factor
) {
4293 * wakee is somewhat hot, it needs certain amount of cpu
4294 * resource, so if waker is far more hot, prefer to leave
4297 if (current
->wakee_flips
> (factor
* p
->wakee_flips
))
4304 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
4306 s64 this_load
, load
;
4307 s64 this_eff_load
, prev_eff_load
;
4308 int idx
, this_cpu
, prev_cpu
;
4309 struct task_group
*tg
;
4310 unsigned long weight
;
4314 * If we wake multiple tasks be careful to not bounce
4315 * ourselves around too much.
4321 this_cpu
= smp_processor_id();
4322 prev_cpu
= task_cpu(p
);
4323 load
= source_load(prev_cpu
, idx
);
4324 this_load
= target_load(this_cpu
, idx
);
4327 * If sync wakeup then subtract the (maximum possible)
4328 * effect of the currently running task from the load
4329 * of the current CPU:
4332 tg
= task_group(current
);
4333 weight
= current
->se
.load
.weight
;
4335 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
4336 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
4340 weight
= p
->se
.load
.weight
;
4343 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4344 * due to the sync cause above having dropped this_load to 0, we'll
4345 * always have an imbalance, but there's really nothing you can do
4346 * about that, so that's good too.
4348 * Otherwise check if either cpus are near enough in load to allow this
4349 * task to be woken on this_cpu.
4351 this_eff_load
= 100;
4352 this_eff_load
*= capacity_of(prev_cpu
);
4354 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
4355 prev_eff_load
*= capacity_of(this_cpu
);
4357 if (this_load
> 0) {
4358 this_eff_load
*= this_load
+
4359 effective_load(tg
, this_cpu
, weight
, weight
);
4361 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
4364 balanced
= this_eff_load
<= prev_eff_load
;
4366 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
4371 schedstat_inc(sd
, ttwu_move_affine
);
4372 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
4378 * find_idlest_group finds and returns the least busy CPU group within the
4381 static struct sched_group
*
4382 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
4383 int this_cpu
, int sd_flag
)
4385 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
4386 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
4387 int load_idx
= sd
->forkexec_idx
;
4388 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
4390 if (sd_flag
& SD_BALANCE_WAKE
)
4391 load_idx
= sd
->wake_idx
;
4394 unsigned long load
, avg_load
;
4398 /* Skip over this group if it has no CPUs allowed */
4399 if (!cpumask_intersects(sched_group_cpus(group
),
4400 tsk_cpus_allowed(p
)))
4403 local_group
= cpumask_test_cpu(this_cpu
,
4404 sched_group_cpus(group
));
4406 /* Tally up the load of all CPUs in the group */
4409 for_each_cpu(i
, sched_group_cpus(group
)) {
4410 /* Bias balancing toward cpus of our domain */
4412 load
= source_load(i
, load_idx
);
4414 load
= target_load(i
, load_idx
);
4419 /* Adjust by relative CPU capacity of the group */
4420 avg_load
= (avg_load
* SCHED_CAPACITY_SCALE
) / group
->sgc
->capacity
;
4423 this_load
= avg_load
;
4424 } else if (avg_load
< min_load
) {
4425 min_load
= avg_load
;
4428 } while (group
= group
->next
, group
!= sd
->groups
);
4430 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
4436 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4439 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
4441 unsigned long load
, min_load
= ULONG_MAX
;
4442 unsigned int min_exit_latency
= UINT_MAX
;
4443 u64 latest_idle_timestamp
= 0;
4444 int least_loaded_cpu
= this_cpu
;
4445 int shallowest_idle_cpu
= -1;
4448 /* Traverse only the allowed CPUs */
4449 for_each_cpu_and(i
, sched_group_cpus(group
), tsk_cpus_allowed(p
)) {
4451 struct rq
*rq
= cpu_rq(i
);
4452 struct cpuidle_state
*idle
= idle_get_state(rq
);
4453 if (idle
&& idle
->exit_latency
< min_exit_latency
) {
4455 * We give priority to a CPU whose idle state
4456 * has the smallest exit latency irrespective
4457 * of any idle timestamp.
4459 min_exit_latency
= idle
->exit_latency
;
4460 latest_idle_timestamp
= rq
->idle_stamp
;
4461 shallowest_idle_cpu
= i
;
4462 } else if ((!idle
|| idle
->exit_latency
== min_exit_latency
) &&
4463 rq
->idle_stamp
> latest_idle_timestamp
) {
4465 * If equal or no active idle state, then
4466 * the most recently idled CPU might have
4469 latest_idle_timestamp
= rq
->idle_stamp
;
4470 shallowest_idle_cpu
= i
;
4473 load
= weighted_cpuload(i
);
4474 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
4476 least_loaded_cpu
= i
;
4481 return shallowest_idle_cpu
!= -1 ? shallowest_idle_cpu
: least_loaded_cpu
;
4485 * Try and locate an idle CPU in the sched_domain.
4487 static int select_idle_sibling(struct task_struct
*p
, int target
)
4489 struct sched_domain
*sd
;
4490 struct sched_group
*sg
;
4491 int i
= task_cpu(p
);
4493 if (idle_cpu(target
))
4497 * If the prevous cpu is cache affine and idle, don't be stupid.
4499 if (i
!= target
&& cpus_share_cache(i
, target
) && idle_cpu(i
))
4503 * Otherwise, iterate the domains and find an elegible idle cpu.
4505 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
4506 for_each_lower_domain(sd
) {
4509 if (!cpumask_intersects(sched_group_cpus(sg
),
4510 tsk_cpus_allowed(p
)))
4513 for_each_cpu(i
, sched_group_cpus(sg
)) {
4514 if (i
== target
|| !idle_cpu(i
))
4518 target
= cpumask_first_and(sched_group_cpus(sg
),
4519 tsk_cpus_allowed(p
));
4523 } while (sg
!= sd
->groups
);
4530 * select_task_rq_fair: Select target runqueue for the waking task in domains
4531 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4532 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4534 * Balances load by selecting the idlest cpu in the idlest group, or under
4535 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4537 * Returns the target cpu number.
4539 * preempt must be disabled.
4542 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
4544 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
4545 int cpu
= smp_processor_id();
4547 int want_affine
= 0;
4548 int sync
= wake_flags
& WF_SYNC
;
4550 if (p
->nr_cpus_allowed
== 1)
4553 if (sd_flag
& SD_BALANCE_WAKE
)
4554 want_affine
= cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
));
4557 for_each_domain(cpu
, tmp
) {
4558 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
4562 * If both cpu and prev_cpu are part of this domain,
4563 * cpu is a valid SD_WAKE_AFFINE target.
4565 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
4566 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
4571 if (tmp
->flags
& sd_flag
)
4575 if (affine_sd
&& cpu
!= prev_cpu
&& wake_affine(affine_sd
, p
, sync
))
4578 if (sd_flag
& SD_BALANCE_WAKE
) {
4579 new_cpu
= select_idle_sibling(p
, prev_cpu
);
4584 struct sched_group
*group
;
4587 if (!(sd
->flags
& sd_flag
)) {
4592 group
= find_idlest_group(sd
, p
, cpu
, sd_flag
);
4598 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
4599 if (new_cpu
== -1 || new_cpu
== cpu
) {
4600 /* Now try balancing at a lower domain level of cpu */
4605 /* Now try balancing at a lower domain level of new_cpu */
4607 weight
= sd
->span_weight
;
4609 for_each_domain(cpu
, tmp
) {
4610 if (weight
<= tmp
->span_weight
)
4612 if (tmp
->flags
& sd_flag
)
4615 /* while loop will break here if sd == NULL */
4624 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4625 * cfs_rq_of(p) references at time of call are still valid and identify the
4626 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4627 * other assumptions, including the state of rq->lock, should be made.
4630 migrate_task_rq_fair(struct task_struct
*p
, int next_cpu
)
4632 struct sched_entity
*se
= &p
->se
;
4633 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4636 * Load tracking: accumulate removed load so that it can be processed
4637 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4638 * to blocked load iff they have a positive decay-count. It can never
4639 * be negative here since on-rq tasks have decay-count == 0.
4641 if (se
->avg
.decay_count
) {
4642 se
->avg
.decay_count
= -__synchronize_entity_decay(se
);
4643 atomic_long_add(se
->avg
.load_avg_contrib
,
4644 &cfs_rq
->removed_load
);
4647 /* We have migrated, no longer consider this task hot */
4650 #endif /* CONFIG_SMP */
4652 static unsigned long
4653 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
4655 unsigned long gran
= sysctl_sched_wakeup_granularity
;
4658 * Since its curr running now, convert the gran from real-time
4659 * to virtual-time in his units.
4661 * By using 'se' instead of 'curr' we penalize light tasks, so
4662 * they get preempted easier. That is, if 'se' < 'curr' then
4663 * the resulting gran will be larger, therefore penalizing the
4664 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4665 * be smaller, again penalizing the lighter task.
4667 * This is especially important for buddies when the leftmost
4668 * task is higher priority than the buddy.
4670 return calc_delta_fair(gran
, se
);
4674 * Should 'se' preempt 'curr'.
4688 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
4690 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
4695 gran
= wakeup_gran(curr
, se
);
4702 static void set_last_buddy(struct sched_entity
*se
)
4704 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4707 for_each_sched_entity(se
)
4708 cfs_rq_of(se
)->last
= se
;
4711 static void set_next_buddy(struct sched_entity
*se
)
4713 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4716 for_each_sched_entity(se
)
4717 cfs_rq_of(se
)->next
= se
;
4720 static void set_skip_buddy(struct sched_entity
*se
)
4722 for_each_sched_entity(se
)
4723 cfs_rq_of(se
)->skip
= se
;
4727 * Preempt the current task with a newly woken task if needed:
4729 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
4731 struct task_struct
*curr
= rq
->curr
;
4732 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
4733 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4734 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
4735 int next_buddy_marked
= 0;
4737 if (unlikely(se
== pse
))
4741 * This is possible from callers such as attach_tasks(), in which we
4742 * unconditionally check_prempt_curr() after an enqueue (which may have
4743 * lead to a throttle). This both saves work and prevents false
4744 * next-buddy nomination below.
4746 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
4749 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
4750 set_next_buddy(pse
);
4751 next_buddy_marked
= 1;
4755 * We can come here with TIF_NEED_RESCHED already set from new task
4758 * Note: this also catches the edge-case of curr being in a throttled
4759 * group (e.g. via set_curr_task), since update_curr() (in the
4760 * enqueue of curr) will have resulted in resched being set. This
4761 * prevents us from potentially nominating it as a false LAST_BUDDY
4764 if (test_tsk_need_resched(curr
))
4767 /* Idle tasks are by definition preempted by non-idle tasks. */
4768 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
4769 likely(p
->policy
!= SCHED_IDLE
))
4773 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4774 * is driven by the tick):
4776 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
4779 find_matching_se(&se
, &pse
);
4780 update_curr(cfs_rq_of(se
));
4782 if (wakeup_preempt_entity(se
, pse
) == 1) {
4784 * Bias pick_next to pick the sched entity that is
4785 * triggering this preemption.
4787 if (!next_buddy_marked
)
4788 set_next_buddy(pse
);
4797 * Only set the backward buddy when the current task is still
4798 * on the rq. This can happen when a wakeup gets interleaved
4799 * with schedule on the ->pre_schedule() or idle_balance()
4800 * point, either of which can * drop the rq lock.
4802 * Also, during early boot the idle thread is in the fair class,
4803 * for obvious reasons its a bad idea to schedule back to it.
4805 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
4808 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
4812 static struct task_struct
*
4813 pick_next_task_fair(struct rq
*rq
, struct task_struct
*prev
)
4815 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
4816 struct sched_entity
*se
;
4817 struct task_struct
*p
;
4821 #ifdef CONFIG_FAIR_GROUP_SCHED
4822 if (!cfs_rq
->nr_running
)
4825 if (prev
->sched_class
!= &fair_sched_class
)
4829 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
4830 * likely that a next task is from the same cgroup as the current.
4832 * Therefore attempt to avoid putting and setting the entire cgroup
4833 * hierarchy, only change the part that actually changes.
4837 struct sched_entity
*curr
= cfs_rq
->curr
;
4840 * Since we got here without doing put_prev_entity() we also
4841 * have to consider cfs_rq->curr. If it is still a runnable
4842 * entity, update_curr() will update its vruntime, otherwise
4843 * forget we've ever seen it.
4845 if (curr
&& curr
->on_rq
)
4846 update_curr(cfs_rq
);
4851 * This call to check_cfs_rq_runtime() will do the throttle and
4852 * dequeue its entity in the parent(s). Therefore the 'simple'
4853 * nr_running test will indeed be correct.
4855 if (unlikely(check_cfs_rq_runtime(cfs_rq
)))
4858 se
= pick_next_entity(cfs_rq
, curr
);
4859 cfs_rq
= group_cfs_rq(se
);
4865 * Since we haven't yet done put_prev_entity and if the selected task
4866 * is a different task than we started out with, try and touch the
4867 * least amount of cfs_rqs.
4870 struct sched_entity
*pse
= &prev
->se
;
4872 while (!(cfs_rq
= is_same_group(se
, pse
))) {
4873 int se_depth
= se
->depth
;
4874 int pse_depth
= pse
->depth
;
4876 if (se_depth
<= pse_depth
) {
4877 put_prev_entity(cfs_rq_of(pse
), pse
);
4878 pse
= parent_entity(pse
);
4880 if (se_depth
>= pse_depth
) {
4881 set_next_entity(cfs_rq_of(se
), se
);
4882 se
= parent_entity(se
);
4886 put_prev_entity(cfs_rq
, pse
);
4887 set_next_entity(cfs_rq
, se
);
4890 if (hrtick_enabled(rq
))
4891 hrtick_start_fair(rq
, p
);
4898 if (!cfs_rq
->nr_running
)
4901 put_prev_task(rq
, prev
);
4904 se
= pick_next_entity(cfs_rq
, NULL
);
4905 set_next_entity(cfs_rq
, se
);
4906 cfs_rq
= group_cfs_rq(se
);
4911 if (hrtick_enabled(rq
))
4912 hrtick_start_fair(rq
, p
);
4917 new_tasks
= idle_balance(rq
);
4919 * Because idle_balance() releases (and re-acquires) rq->lock, it is
4920 * possible for any higher priority task to appear. In that case we
4921 * must re-start the pick_next_entity() loop.
4933 * Account for a descheduled task:
4935 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
4937 struct sched_entity
*se
= &prev
->se
;
4938 struct cfs_rq
*cfs_rq
;
4940 for_each_sched_entity(se
) {
4941 cfs_rq
= cfs_rq_of(se
);
4942 put_prev_entity(cfs_rq
, se
);
4947 * sched_yield() is very simple
4949 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4951 static void yield_task_fair(struct rq
*rq
)
4953 struct task_struct
*curr
= rq
->curr
;
4954 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
4955 struct sched_entity
*se
= &curr
->se
;
4958 * Are we the only task in the tree?
4960 if (unlikely(rq
->nr_running
== 1))
4963 clear_buddies(cfs_rq
, se
);
4965 if (curr
->policy
!= SCHED_BATCH
) {
4966 update_rq_clock(rq
);
4968 * Update run-time statistics of the 'current'.
4970 update_curr(cfs_rq
);
4972 * Tell update_rq_clock() that we've just updated,
4973 * so we don't do microscopic update in schedule()
4974 * and double the fastpath cost.
4976 rq
->skip_clock_update
= 1;
4982 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
4984 struct sched_entity
*se
= &p
->se
;
4986 /* throttled hierarchies are not runnable */
4987 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
4990 /* Tell the scheduler that we'd really like pse to run next. */
4993 yield_task_fair(rq
);
4999 /**************************************************
5000 * Fair scheduling class load-balancing methods.
5004 * The purpose of load-balancing is to achieve the same basic fairness the
5005 * per-cpu scheduler provides, namely provide a proportional amount of compute
5006 * time to each task. This is expressed in the following equation:
5008 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5010 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5011 * W_i,0 is defined as:
5013 * W_i,0 = \Sum_j w_i,j (2)
5015 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5016 * is derived from the nice value as per prio_to_weight[].
5018 * The weight average is an exponential decay average of the instantaneous
5021 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5023 * C_i is the compute capacity of cpu i, typically it is the
5024 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5025 * can also include other factors [XXX].
5027 * To achieve this balance we define a measure of imbalance which follows
5028 * directly from (1):
5030 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5032 * We them move tasks around to minimize the imbalance. In the continuous
5033 * function space it is obvious this converges, in the discrete case we get
5034 * a few fun cases generally called infeasible weight scenarios.
5037 * - infeasible weights;
5038 * - local vs global optima in the discrete case. ]
5043 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5044 * for all i,j solution, we create a tree of cpus that follows the hardware
5045 * topology where each level pairs two lower groups (or better). This results
5046 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5047 * tree to only the first of the previous level and we decrease the frequency
5048 * of load-balance at each level inv. proportional to the number of cpus in
5054 * \Sum { --- * --- * 2^i } = O(n) (5)
5056 * `- size of each group
5057 * | | `- number of cpus doing load-balance
5059 * `- sum over all levels
5061 * Coupled with a limit on how many tasks we can migrate every balance pass,
5062 * this makes (5) the runtime complexity of the balancer.
5064 * An important property here is that each CPU is still (indirectly) connected
5065 * to every other cpu in at most O(log n) steps:
5067 * The adjacency matrix of the resulting graph is given by:
5070 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5073 * And you'll find that:
5075 * A^(log_2 n)_i,j != 0 for all i,j (7)
5077 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5078 * The task movement gives a factor of O(m), giving a convergence complexity
5081 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5086 * In order to avoid CPUs going idle while there's still work to do, new idle
5087 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5088 * tree itself instead of relying on other CPUs to bring it work.
5090 * This adds some complexity to both (5) and (8) but it reduces the total idle
5098 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5101 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5106 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5108 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5110 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5113 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5114 * rewrite all of this once again.]
5117 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
5119 enum fbq_type
{ regular
, remote
, all
};
5121 #define LBF_ALL_PINNED 0x01
5122 #define LBF_NEED_BREAK 0x02
5123 #define LBF_DST_PINNED 0x04
5124 #define LBF_SOME_PINNED 0x08
5127 struct sched_domain
*sd
;
5135 struct cpumask
*dst_grpmask
;
5137 enum cpu_idle_type idle
;
5139 /* The set of CPUs under consideration for load-balancing */
5140 struct cpumask
*cpus
;
5145 unsigned int loop_break
;
5146 unsigned int loop_max
;
5148 enum fbq_type fbq_type
;
5149 struct list_head tasks
;
5153 * Is this task likely cache-hot:
5155 static int task_hot(struct task_struct
*p
, struct lb_env
*env
)
5159 lockdep_assert_held(&env
->src_rq
->lock
);
5161 if (p
->sched_class
!= &fair_sched_class
)
5164 if (unlikely(p
->policy
== SCHED_IDLE
))
5168 * Buddy candidates are cache hot:
5170 if (sched_feat(CACHE_HOT_BUDDY
) && env
->dst_rq
->nr_running
&&
5171 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
5172 &p
->se
== cfs_rq_of(&p
->se
)->last
))
5175 if (sysctl_sched_migration_cost
== -1)
5177 if (sysctl_sched_migration_cost
== 0)
5180 delta
= rq_clock_task(env
->src_rq
) - p
->se
.exec_start
;
5182 return delta
< (s64
)sysctl_sched_migration_cost
;
5185 #ifdef CONFIG_NUMA_BALANCING
5186 /* Returns true if the destination node has incurred more faults */
5187 static bool migrate_improves_locality(struct task_struct
*p
, struct lb_env
*env
)
5189 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
5190 int src_nid
, dst_nid
;
5192 if (!sched_feat(NUMA_FAVOUR_HIGHER
) || !p
->numa_faults_memory
||
5193 !(env
->sd
->flags
& SD_NUMA
)) {
5197 src_nid
= cpu_to_node(env
->src_cpu
);
5198 dst_nid
= cpu_to_node(env
->dst_cpu
);
5200 if (src_nid
== dst_nid
)
5204 /* Task is already in the group's interleave set. */
5205 if (node_isset(src_nid
, numa_group
->active_nodes
))
5208 /* Task is moving into the group's interleave set. */
5209 if (node_isset(dst_nid
, numa_group
->active_nodes
))
5212 return group_faults(p
, dst_nid
) > group_faults(p
, src_nid
);
5215 /* Encourage migration to the preferred node. */
5216 if (dst_nid
== p
->numa_preferred_nid
)
5219 return task_faults(p
, dst_nid
) > task_faults(p
, src_nid
);
5223 static bool migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
5225 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
5226 int src_nid
, dst_nid
;
5228 if (!sched_feat(NUMA
) || !sched_feat(NUMA_RESIST_LOWER
))
5231 if (!p
->numa_faults_memory
|| !(env
->sd
->flags
& SD_NUMA
))
5234 src_nid
= cpu_to_node(env
->src_cpu
);
5235 dst_nid
= cpu_to_node(env
->dst_cpu
);
5237 if (src_nid
== dst_nid
)
5241 /* Task is moving within/into the group's interleave set. */
5242 if (node_isset(dst_nid
, numa_group
->active_nodes
))
5245 /* Task is moving out of the group's interleave set. */
5246 if (node_isset(src_nid
, numa_group
->active_nodes
))
5249 return group_faults(p
, dst_nid
) < group_faults(p
, src_nid
);
5252 /* Migrating away from the preferred node is always bad. */
5253 if (src_nid
== p
->numa_preferred_nid
)
5256 return task_faults(p
, dst_nid
) < task_faults(p
, src_nid
);
5260 static inline bool migrate_improves_locality(struct task_struct
*p
,
5266 static inline bool migrate_degrades_locality(struct task_struct
*p
,
5274 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5277 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
5279 int tsk_cache_hot
= 0;
5281 lockdep_assert_held(&env
->src_rq
->lock
);
5284 * We do not migrate tasks that are:
5285 * 1) throttled_lb_pair, or
5286 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5287 * 3) running (obviously), or
5288 * 4) are cache-hot on their current CPU.
5290 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
5293 if (!cpumask_test_cpu(env
->dst_cpu
, tsk_cpus_allowed(p
))) {
5296 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
5298 env
->flags
|= LBF_SOME_PINNED
;
5301 * Remember if this task can be migrated to any other cpu in
5302 * our sched_group. We may want to revisit it if we couldn't
5303 * meet load balance goals by pulling other tasks on src_cpu.
5305 * Also avoid computing new_dst_cpu if we have already computed
5306 * one in current iteration.
5308 if (!env
->dst_grpmask
|| (env
->flags
& LBF_DST_PINNED
))
5311 /* Prevent to re-select dst_cpu via env's cpus */
5312 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
5313 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
))) {
5314 env
->flags
|= LBF_DST_PINNED
;
5315 env
->new_dst_cpu
= cpu
;
5323 /* Record that we found atleast one task that could run on dst_cpu */
5324 env
->flags
&= ~LBF_ALL_PINNED
;
5326 if (task_running(env
->src_rq
, p
)) {
5327 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
5332 * Aggressive migration if:
5333 * 1) destination numa is preferred
5334 * 2) task is cache cold, or
5335 * 3) too many balance attempts have failed.
5337 tsk_cache_hot
= task_hot(p
, env
);
5339 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
5341 if (migrate_improves_locality(p
, env
) || !tsk_cache_hot
||
5342 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
5343 if (tsk_cache_hot
) {
5344 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
5345 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
5350 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
5355 * detach_task() -- detach the task for the migration specified in env
5357 static void detach_task(struct task_struct
*p
, struct lb_env
*env
)
5359 lockdep_assert_held(&env
->src_rq
->lock
);
5361 deactivate_task(env
->src_rq
, p
, 0);
5362 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
5363 set_task_cpu(p
, env
->dst_cpu
);
5367 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5368 * part of active balancing operations within "domain".
5370 * Returns a task if successful and NULL otherwise.
5372 static struct task_struct
*detach_one_task(struct lb_env
*env
)
5374 struct task_struct
*p
, *n
;
5376 lockdep_assert_held(&env
->src_rq
->lock
);
5378 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
5379 if (!can_migrate_task(p
, env
))
5382 detach_task(p
, env
);
5385 * Right now, this is only the second place where
5386 * lb_gained[env->idle] is updated (other is detach_tasks)
5387 * so we can safely collect stats here rather than
5388 * inside detach_tasks().
5390 schedstat_inc(env
->sd
, lb_gained
[env
->idle
]);
5396 static const unsigned int sched_nr_migrate_break
= 32;
5399 * detach_tasks() -- tries to detach up to imbalance weighted load from
5400 * busiest_rq, as part of a balancing operation within domain "sd".
5402 * Returns number of detached tasks if successful and 0 otherwise.
5404 static int detach_tasks(struct lb_env
*env
)
5406 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
5407 struct task_struct
*p
;
5411 lockdep_assert_held(&env
->src_rq
->lock
);
5413 if (env
->imbalance
<= 0)
5416 while (!list_empty(tasks
)) {
5417 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
5420 /* We've more or less seen every task there is, call it quits */
5421 if (env
->loop
> env
->loop_max
)
5424 /* take a breather every nr_migrate tasks */
5425 if (env
->loop
> env
->loop_break
) {
5426 env
->loop_break
+= sched_nr_migrate_break
;
5427 env
->flags
|= LBF_NEED_BREAK
;
5431 if (!can_migrate_task(p
, env
))
5434 load
= task_h_load(p
);
5436 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
5439 if ((load
/ 2) > env
->imbalance
)
5442 detach_task(p
, env
);
5443 list_add(&p
->se
.group_node
, &env
->tasks
);
5446 env
->imbalance
-= load
;
5448 #ifdef CONFIG_PREEMPT
5450 * NEWIDLE balancing is a source of latency, so preemptible
5451 * kernels will stop after the first task is detached to minimize
5452 * the critical section.
5454 if (env
->idle
== CPU_NEWLY_IDLE
)
5459 * We only want to steal up to the prescribed amount of
5462 if (env
->imbalance
<= 0)
5467 list_move_tail(&p
->se
.group_node
, tasks
);
5471 * Right now, this is one of only two places we collect this stat
5472 * so we can safely collect detach_one_task() stats here rather
5473 * than inside detach_one_task().
5475 schedstat_add(env
->sd
, lb_gained
[env
->idle
], detached
);
5481 * attach_task() -- attach the task detached by detach_task() to its new rq.
5483 static void attach_task(struct rq
*rq
, struct task_struct
*p
)
5485 lockdep_assert_held(&rq
->lock
);
5487 BUG_ON(task_rq(p
) != rq
);
5488 p
->on_rq
= TASK_ON_RQ_QUEUED
;
5489 activate_task(rq
, p
, 0);
5490 check_preempt_curr(rq
, p
, 0);
5494 * attach_one_task() -- attaches the task returned from detach_one_task() to
5497 static void attach_one_task(struct rq
*rq
, struct task_struct
*p
)
5499 raw_spin_lock(&rq
->lock
);
5501 raw_spin_unlock(&rq
->lock
);
5505 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5508 static void attach_tasks(struct lb_env
*env
)
5510 struct list_head
*tasks
= &env
->tasks
;
5511 struct task_struct
*p
;
5513 raw_spin_lock(&env
->dst_rq
->lock
);
5515 while (!list_empty(tasks
)) {
5516 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
5517 list_del_init(&p
->se
.group_node
);
5519 attach_task(env
->dst_rq
, p
);
5522 raw_spin_unlock(&env
->dst_rq
->lock
);
5525 #ifdef CONFIG_FAIR_GROUP_SCHED
5527 * update tg->load_weight by folding this cpu's load_avg
5529 static void __update_blocked_averages_cpu(struct task_group
*tg
, int cpu
)
5531 struct sched_entity
*se
= tg
->se
[cpu
];
5532 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu
];
5534 /* throttled entities do not contribute to load */
5535 if (throttled_hierarchy(cfs_rq
))
5538 update_cfs_rq_blocked_load(cfs_rq
, 1);
5541 update_entity_load_avg(se
, 1);
5543 * We pivot on our runnable average having decayed to zero for
5544 * list removal. This generally implies that all our children
5545 * have also been removed (modulo rounding error or bandwidth
5546 * control); however, such cases are rare and we can fix these
5549 * TODO: fix up out-of-order children on enqueue.
5551 if (!se
->avg
.runnable_avg_sum
&& !cfs_rq
->nr_running
)
5552 list_del_leaf_cfs_rq(cfs_rq
);
5554 struct rq
*rq
= rq_of(cfs_rq
);
5555 update_rq_runnable_avg(rq
, rq
->nr_running
);
5559 static void update_blocked_averages(int cpu
)
5561 struct rq
*rq
= cpu_rq(cpu
);
5562 struct cfs_rq
*cfs_rq
;
5563 unsigned long flags
;
5565 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5566 update_rq_clock(rq
);
5568 * Iterates the task_group tree in a bottom up fashion, see
5569 * list_add_leaf_cfs_rq() for details.
5571 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
5573 * Note: We may want to consider periodically releasing
5574 * rq->lock about these updates so that creating many task
5575 * groups does not result in continually extending hold time.
5577 __update_blocked_averages_cpu(cfs_rq
->tg
, rq
->cpu
);
5580 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5584 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5585 * This needs to be done in a top-down fashion because the load of a child
5586 * group is a fraction of its parents load.
5588 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
5590 struct rq
*rq
= rq_of(cfs_rq
);
5591 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
5592 unsigned long now
= jiffies
;
5595 if (cfs_rq
->last_h_load_update
== now
)
5598 cfs_rq
->h_load_next
= NULL
;
5599 for_each_sched_entity(se
) {
5600 cfs_rq
= cfs_rq_of(se
);
5601 cfs_rq
->h_load_next
= se
;
5602 if (cfs_rq
->last_h_load_update
== now
)
5607 cfs_rq
->h_load
= cfs_rq
->runnable_load_avg
;
5608 cfs_rq
->last_h_load_update
= now
;
5611 while ((se
= cfs_rq
->h_load_next
) != NULL
) {
5612 load
= cfs_rq
->h_load
;
5613 load
= div64_ul(load
* se
->avg
.load_avg_contrib
,
5614 cfs_rq
->runnable_load_avg
+ 1);
5615 cfs_rq
= group_cfs_rq(se
);
5616 cfs_rq
->h_load
= load
;
5617 cfs_rq
->last_h_load_update
= now
;
5621 static unsigned long task_h_load(struct task_struct
*p
)
5623 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
5625 update_cfs_rq_h_load(cfs_rq
);
5626 return div64_ul(p
->se
.avg
.load_avg_contrib
* cfs_rq
->h_load
,
5627 cfs_rq
->runnable_load_avg
+ 1);
5630 static inline void update_blocked_averages(int cpu
)
5634 static unsigned long task_h_load(struct task_struct
*p
)
5636 return p
->se
.avg
.load_avg_contrib
;
5640 /********** Helpers for find_busiest_group ************************/
5649 * sg_lb_stats - stats of a sched_group required for load_balancing
5651 struct sg_lb_stats
{
5652 unsigned long avg_load
; /*Avg load across the CPUs of the group */
5653 unsigned long group_load
; /* Total load over the CPUs of the group */
5654 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
5655 unsigned long load_per_task
;
5656 unsigned long group_capacity
;
5657 unsigned int sum_nr_running
; /* Nr tasks running in the group */
5658 unsigned int group_capacity_factor
;
5659 unsigned int idle_cpus
;
5660 unsigned int group_weight
;
5661 enum group_type group_type
;
5662 int group_has_free_capacity
;
5663 #ifdef CONFIG_NUMA_BALANCING
5664 unsigned int nr_numa_running
;
5665 unsigned int nr_preferred_running
;
5670 * sd_lb_stats - Structure to store the statistics of a sched_domain
5671 * during load balancing.
5673 struct sd_lb_stats
{
5674 struct sched_group
*busiest
; /* Busiest group in this sd */
5675 struct sched_group
*local
; /* Local group in this sd */
5676 unsigned long total_load
; /* Total load of all groups in sd */
5677 unsigned long total_capacity
; /* Total capacity of all groups in sd */
5678 unsigned long avg_load
; /* Average load across all groups in sd */
5680 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
5681 struct sg_lb_stats local_stat
; /* Statistics of the local group */
5684 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
5687 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5688 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5689 * We must however clear busiest_stat::avg_load because
5690 * update_sd_pick_busiest() reads this before assignment.
5692 *sds
= (struct sd_lb_stats
){
5696 .total_capacity
= 0UL,
5699 .sum_nr_running
= 0,
5700 .group_type
= group_other
,
5706 * get_sd_load_idx - Obtain the load index for a given sched domain.
5707 * @sd: The sched_domain whose load_idx is to be obtained.
5708 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5710 * Return: The load index.
5712 static inline int get_sd_load_idx(struct sched_domain
*sd
,
5713 enum cpu_idle_type idle
)
5719 load_idx
= sd
->busy_idx
;
5722 case CPU_NEWLY_IDLE
:
5723 load_idx
= sd
->newidle_idx
;
5726 load_idx
= sd
->idle_idx
;
5733 static unsigned long default_scale_capacity(struct sched_domain
*sd
, int cpu
)
5735 return SCHED_CAPACITY_SCALE
;
5738 unsigned long __weak
arch_scale_freq_capacity(struct sched_domain
*sd
, int cpu
)
5740 return default_scale_capacity(sd
, cpu
);
5743 static unsigned long default_scale_cpu_capacity(struct sched_domain
*sd
, int cpu
)
5745 if ((sd
->flags
& SD_SHARE_CPUCAPACITY
) && (sd
->span_weight
> 1))
5746 return sd
->smt_gain
/ sd
->span_weight
;
5748 return SCHED_CAPACITY_SCALE
;
5751 unsigned long __weak
arch_scale_cpu_capacity(struct sched_domain
*sd
, int cpu
)
5753 return default_scale_cpu_capacity(sd
, cpu
);
5756 static unsigned long scale_rt_capacity(int cpu
)
5758 struct rq
*rq
= cpu_rq(cpu
);
5759 u64 total
, available
, age_stamp
, avg
;
5763 * Since we're reading these variables without serialization make sure
5764 * we read them once before doing sanity checks on them.
5766 age_stamp
= ACCESS_ONCE(rq
->age_stamp
);
5767 avg
= ACCESS_ONCE(rq
->rt_avg
);
5769 delta
= rq_clock(rq
) - age_stamp
;
5770 if (unlikely(delta
< 0))
5773 total
= sched_avg_period() + delta
;
5775 if (unlikely(total
< avg
)) {
5776 /* Ensures that capacity won't end up being negative */
5779 available
= total
- avg
;
5782 if (unlikely((s64
)total
< SCHED_CAPACITY_SCALE
))
5783 total
= SCHED_CAPACITY_SCALE
;
5785 total
>>= SCHED_CAPACITY_SHIFT
;
5787 return div_u64(available
, total
);
5790 static void update_cpu_capacity(struct sched_domain
*sd
, int cpu
)
5792 unsigned long capacity
= SCHED_CAPACITY_SCALE
;
5793 struct sched_group
*sdg
= sd
->groups
;
5795 if (sched_feat(ARCH_CAPACITY
))
5796 capacity
*= arch_scale_cpu_capacity(sd
, cpu
);
5798 capacity
*= default_scale_cpu_capacity(sd
, cpu
);
5800 capacity
>>= SCHED_CAPACITY_SHIFT
;
5802 sdg
->sgc
->capacity_orig
= capacity
;
5804 if (sched_feat(ARCH_CAPACITY
))
5805 capacity
*= arch_scale_freq_capacity(sd
, cpu
);
5807 capacity
*= default_scale_capacity(sd
, cpu
);
5809 capacity
>>= SCHED_CAPACITY_SHIFT
;
5811 capacity
*= scale_rt_capacity(cpu
);
5812 capacity
>>= SCHED_CAPACITY_SHIFT
;
5817 cpu_rq(cpu
)->cpu_capacity
= capacity
;
5818 sdg
->sgc
->capacity
= capacity
;
5821 void update_group_capacity(struct sched_domain
*sd
, int cpu
)
5823 struct sched_domain
*child
= sd
->child
;
5824 struct sched_group
*group
, *sdg
= sd
->groups
;
5825 unsigned long capacity
, capacity_orig
;
5826 unsigned long interval
;
5828 interval
= msecs_to_jiffies(sd
->balance_interval
);
5829 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
5830 sdg
->sgc
->next_update
= jiffies
+ interval
;
5833 update_cpu_capacity(sd
, cpu
);
5837 capacity_orig
= capacity
= 0;
5839 if (child
->flags
& SD_OVERLAP
) {
5841 * SD_OVERLAP domains cannot assume that child groups
5842 * span the current group.
5845 for_each_cpu(cpu
, sched_group_cpus(sdg
)) {
5846 struct sched_group_capacity
*sgc
;
5847 struct rq
*rq
= cpu_rq(cpu
);
5850 * build_sched_domains() -> init_sched_groups_capacity()
5851 * gets here before we've attached the domains to the
5854 * Use capacity_of(), which is set irrespective of domains
5855 * in update_cpu_capacity().
5857 * This avoids capacity/capacity_orig from being 0 and
5858 * causing divide-by-zero issues on boot.
5860 * Runtime updates will correct capacity_orig.
5862 if (unlikely(!rq
->sd
)) {
5863 capacity_orig
+= capacity_of(cpu
);
5864 capacity
+= capacity_of(cpu
);
5868 sgc
= rq
->sd
->groups
->sgc
;
5869 capacity_orig
+= sgc
->capacity_orig
;
5870 capacity
+= sgc
->capacity
;
5874 * !SD_OVERLAP domains can assume that child groups
5875 * span the current group.
5878 group
= child
->groups
;
5880 capacity_orig
+= group
->sgc
->capacity_orig
;
5881 capacity
+= group
->sgc
->capacity
;
5882 group
= group
->next
;
5883 } while (group
!= child
->groups
);
5886 sdg
->sgc
->capacity_orig
= capacity_orig
;
5887 sdg
->sgc
->capacity
= capacity
;
5891 * Try and fix up capacity for tiny siblings, this is needed when
5892 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5893 * which on its own isn't powerful enough.
5895 * See update_sd_pick_busiest() and check_asym_packing().
5898 fix_small_capacity(struct sched_domain
*sd
, struct sched_group
*group
)
5901 * Only siblings can have significantly less than SCHED_CAPACITY_SCALE
5903 if (!(sd
->flags
& SD_SHARE_CPUCAPACITY
))
5907 * If ~90% of the cpu_capacity is still there, we're good.
5909 if (group
->sgc
->capacity
* 32 > group
->sgc
->capacity_orig
* 29)
5916 * Group imbalance indicates (and tries to solve) the problem where balancing
5917 * groups is inadequate due to tsk_cpus_allowed() constraints.
5919 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5920 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5923 * { 0 1 2 3 } { 4 5 6 7 }
5926 * If we were to balance group-wise we'd place two tasks in the first group and
5927 * two tasks in the second group. Clearly this is undesired as it will overload
5928 * cpu 3 and leave one of the cpus in the second group unused.
5930 * The current solution to this issue is detecting the skew in the first group
5931 * by noticing the lower domain failed to reach balance and had difficulty
5932 * moving tasks due to affinity constraints.
5934 * When this is so detected; this group becomes a candidate for busiest; see
5935 * update_sd_pick_busiest(). And calculate_imbalance() and
5936 * find_busiest_group() avoid some of the usual balance conditions to allow it
5937 * to create an effective group imbalance.
5939 * This is a somewhat tricky proposition since the next run might not find the
5940 * group imbalance and decide the groups need to be balanced again. A most
5941 * subtle and fragile situation.
5944 static inline int sg_imbalanced(struct sched_group
*group
)
5946 return group
->sgc
->imbalance
;
5950 * Compute the group capacity factor.
5952 * Avoid the issue where N*frac(smt_capacity) >= 1 creates 'phantom' cores by
5953 * first dividing out the smt factor and computing the actual number of cores
5954 * and limit unit capacity with that.
5956 static inline int sg_capacity_factor(struct lb_env
*env
, struct sched_group
*group
)
5958 unsigned int capacity_factor
, smt
, cpus
;
5959 unsigned int capacity
, capacity_orig
;
5961 capacity
= group
->sgc
->capacity
;
5962 capacity_orig
= group
->sgc
->capacity_orig
;
5963 cpus
= group
->group_weight
;
5965 /* smt := ceil(cpus / capacity), assumes: 1 < smt_capacity < 2 */
5966 smt
= DIV_ROUND_UP(SCHED_CAPACITY_SCALE
* cpus
, capacity_orig
);
5967 capacity_factor
= cpus
/ smt
; /* cores */
5969 capacity_factor
= min_t(unsigned,
5970 capacity_factor
, DIV_ROUND_CLOSEST(capacity
, SCHED_CAPACITY_SCALE
));
5971 if (!capacity_factor
)
5972 capacity_factor
= fix_small_capacity(env
->sd
, group
);
5974 return capacity_factor
;
5977 static enum group_type
5978 group_classify(struct sched_group
*group
, struct sg_lb_stats
*sgs
)
5980 if (sgs
->sum_nr_running
> sgs
->group_capacity_factor
)
5981 return group_overloaded
;
5983 if (sg_imbalanced(group
))
5984 return group_imbalanced
;
5990 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5991 * @env: The load balancing environment.
5992 * @group: sched_group whose statistics are to be updated.
5993 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5994 * @local_group: Does group contain this_cpu.
5995 * @sgs: variable to hold the statistics for this group.
5996 * @overload: Indicate more than one runnable task for any CPU.
5998 static inline void update_sg_lb_stats(struct lb_env
*env
,
5999 struct sched_group
*group
, int load_idx
,
6000 int local_group
, struct sg_lb_stats
*sgs
,
6006 memset(sgs
, 0, sizeof(*sgs
));
6008 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
6009 struct rq
*rq
= cpu_rq(i
);
6011 /* Bias balancing toward cpus of our domain */
6013 load
= target_load(i
, load_idx
);
6015 load
= source_load(i
, load_idx
);
6017 sgs
->group_load
+= load
;
6018 sgs
->sum_nr_running
+= rq
->cfs
.h_nr_running
;
6020 if (rq
->nr_running
> 1)
6023 #ifdef CONFIG_NUMA_BALANCING
6024 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
6025 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
6027 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
6032 /* Adjust by relative CPU capacity of the group */
6033 sgs
->group_capacity
= group
->sgc
->capacity
;
6034 sgs
->avg_load
= (sgs
->group_load
*SCHED_CAPACITY_SCALE
) / sgs
->group_capacity
;
6036 if (sgs
->sum_nr_running
)
6037 sgs
->load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
6039 sgs
->group_weight
= group
->group_weight
;
6040 sgs
->group_capacity_factor
= sg_capacity_factor(env
, group
);
6041 sgs
->group_type
= group_classify(group
, sgs
);
6043 if (sgs
->group_capacity_factor
> sgs
->sum_nr_running
)
6044 sgs
->group_has_free_capacity
= 1;
6048 * update_sd_pick_busiest - return 1 on busiest group
6049 * @env: The load balancing environment.
6050 * @sds: sched_domain statistics
6051 * @sg: sched_group candidate to be checked for being the busiest
6052 * @sgs: sched_group statistics
6054 * Determine if @sg is a busier group than the previously selected
6057 * Return: %true if @sg is a busier group than the previously selected
6058 * busiest group. %false otherwise.
6060 static bool update_sd_pick_busiest(struct lb_env
*env
,
6061 struct sd_lb_stats
*sds
,
6062 struct sched_group
*sg
,
6063 struct sg_lb_stats
*sgs
)
6065 struct sg_lb_stats
*busiest
= &sds
->busiest_stat
;
6067 if (sgs
->group_type
> busiest
->group_type
)
6070 if (sgs
->group_type
< busiest
->group_type
)
6073 if (sgs
->avg_load
<= busiest
->avg_load
)
6076 /* This is the busiest node in its class. */
6077 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
6081 * ASYM_PACKING needs to move all the work to the lowest
6082 * numbered CPUs in the group, therefore mark all groups
6083 * higher than ourself as busy.
6085 if (sgs
->sum_nr_running
&& env
->dst_cpu
< group_first_cpu(sg
)) {
6089 if (group_first_cpu(sds
->busiest
) > group_first_cpu(sg
))
6096 #ifdef CONFIG_NUMA_BALANCING
6097 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
6099 if (sgs
->sum_nr_running
> sgs
->nr_numa_running
)
6101 if (sgs
->sum_nr_running
> sgs
->nr_preferred_running
)
6106 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
6108 if (rq
->nr_running
> rq
->nr_numa_running
)
6110 if (rq
->nr_running
> rq
->nr_preferred_running
)
6115 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
6120 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
6124 #endif /* CONFIG_NUMA_BALANCING */
6127 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6128 * @env: The load balancing environment.
6129 * @sds: variable to hold the statistics for this sched_domain.
6131 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6133 struct sched_domain
*child
= env
->sd
->child
;
6134 struct sched_group
*sg
= env
->sd
->groups
;
6135 struct sg_lb_stats tmp_sgs
;
6136 int load_idx
, prefer_sibling
= 0;
6137 bool overload
= false;
6139 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
6142 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
6145 struct sg_lb_stats
*sgs
= &tmp_sgs
;
6148 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_cpus(sg
));
6151 sgs
= &sds
->local_stat
;
6153 if (env
->idle
!= CPU_NEWLY_IDLE
||
6154 time_after_eq(jiffies
, sg
->sgc
->next_update
))
6155 update_group_capacity(env
->sd
, env
->dst_cpu
);
6158 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, sgs
,
6165 * In case the child domain prefers tasks go to siblings
6166 * first, lower the sg capacity factor to one so that we'll try
6167 * and move all the excess tasks away. We lower the capacity
6168 * of a group only if the local group has the capacity to fit
6169 * these excess tasks, i.e. nr_running < group_capacity_factor. The
6170 * extra check prevents the case where you always pull from the
6171 * heaviest group when it is already under-utilized (possible
6172 * with a large weight task outweighs the tasks on the system).
6174 if (prefer_sibling
&& sds
->local
&&
6175 sds
->local_stat
.group_has_free_capacity
)
6176 sgs
->group_capacity_factor
= min(sgs
->group_capacity_factor
, 1U);
6178 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
6180 sds
->busiest_stat
= *sgs
;
6184 /* Now, start updating sd_lb_stats */
6185 sds
->total_load
+= sgs
->group_load
;
6186 sds
->total_capacity
+= sgs
->group_capacity
;
6189 } while (sg
!= env
->sd
->groups
);
6191 if (env
->sd
->flags
& SD_NUMA
)
6192 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
6194 if (!env
->sd
->parent
) {
6195 /* update overload indicator if we are at root domain */
6196 if (env
->dst_rq
->rd
->overload
!= overload
)
6197 env
->dst_rq
->rd
->overload
= overload
;
6203 * check_asym_packing - Check to see if the group is packed into the
6206 * This is primarily intended to used at the sibling level. Some
6207 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6208 * case of POWER7, it can move to lower SMT modes only when higher
6209 * threads are idle. When in lower SMT modes, the threads will
6210 * perform better since they share less core resources. Hence when we
6211 * have idle threads, we want them to be the higher ones.
6213 * This packing function is run on idle threads. It checks to see if
6214 * the busiest CPU in this domain (core in the P7 case) has a higher
6215 * CPU number than the packing function is being run on. Here we are
6216 * assuming lower CPU number will be equivalent to lower a SMT thread
6219 * Return: 1 when packing is required and a task should be moved to
6220 * this CPU. The amount of the imbalance is returned in *imbalance.
6222 * @env: The load balancing environment.
6223 * @sds: Statistics of the sched_domain which is to be packed
6225 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6229 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
6235 busiest_cpu
= group_first_cpu(sds
->busiest
);
6236 if (env
->dst_cpu
> busiest_cpu
)
6239 env
->imbalance
= DIV_ROUND_CLOSEST(
6240 sds
->busiest_stat
.avg_load
* sds
->busiest_stat
.group_capacity
,
6241 SCHED_CAPACITY_SCALE
);
6247 * fix_small_imbalance - Calculate the minor imbalance that exists
6248 * amongst the groups of a sched_domain, during
6250 * @env: The load balancing environment.
6251 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6254 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6256 unsigned long tmp
, capa_now
= 0, capa_move
= 0;
6257 unsigned int imbn
= 2;
6258 unsigned long scaled_busy_load_per_task
;
6259 struct sg_lb_stats
*local
, *busiest
;
6261 local
= &sds
->local_stat
;
6262 busiest
= &sds
->busiest_stat
;
6264 if (!local
->sum_nr_running
)
6265 local
->load_per_task
= cpu_avg_load_per_task(env
->dst_cpu
);
6266 else if (busiest
->load_per_task
> local
->load_per_task
)
6269 scaled_busy_load_per_task
=
6270 (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
6271 busiest
->group_capacity
;
6273 if (busiest
->avg_load
+ scaled_busy_load_per_task
>=
6274 local
->avg_load
+ (scaled_busy_load_per_task
* imbn
)) {
6275 env
->imbalance
= busiest
->load_per_task
;
6280 * OK, we don't have enough imbalance to justify moving tasks,
6281 * however we may be able to increase total CPU capacity used by
6285 capa_now
+= busiest
->group_capacity
*
6286 min(busiest
->load_per_task
, busiest
->avg_load
);
6287 capa_now
+= local
->group_capacity
*
6288 min(local
->load_per_task
, local
->avg_load
);
6289 capa_now
/= SCHED_CAPACITY_SCALE
;
6291 /* Amount of load we'd subtract */
6292 if (busiest
->avg_load
> scaled_busy_load_per_task
) {
6293 capa_move
+= busiest
->group_capacity
*
6294 min(busiest
->load_per_task
,
6295 busiest
->avg_load
- scaled_busy_load_per_task
);
6298 /* Amount of load we'd add */
6299 if (busiest
->avg_load
* busiest
->group_capacity
<
6300 busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) {
6301 tmp
= (busiest
->avg_load
* busiest
->group_capacity
) /
6302 local
->group_capacity
;
6304 tmp
= (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
6305 local
->group_capacity
;
6307 capa_move
+= local
->group_capacity
*
6308 min(local
->load_per_task
, local
->avg_load
+ tmp
);
6309 capa_move
/= SCHED_CAPACITY_SCALE
;
6311 /* Move if we gain throughput */
6312 if (capa_move
> capa_now
)
6313 env
->imbalance
= busiest
->load_per_task
;
6317 * calculate_imbalance - Calculate the amount of imbalance present within the
6318 * groups of a given sched_domain during load balance.
6319 * @env: load balance environment
6320 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6322 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6324 unsigned long max_pull
, load_above_capacity
= ~0UL;
6325 struct sg_lb_stats
*local
, *busiest
;
6327 local
= &sds
->local_stat
;
6328 busiest
= &sds
->busiest_stat
;
6330 if (busiest
->group_type
== group_imbalanced
) {
6332 * In the group_imb case we cannot rely on group-wide averages
6333 * to ensure cpu-load equilibrium, look at wider averages. XXX
6335 busiest
->load_per_task
=
6336 min(busiest
->load_per_task
, sds
->avg_load
);
6340 * In the presence of smp nice balancing, certain scenarios can have
6341 * max load less than avg load(as we skip the groups at or below
6342 * its cpu_capacity, while calculating max_load..)
6344 if (busiest
->avg_load
<= sds
->avg_load
||
6345 local
->avg_load
>= sds
->avg_load
) {
6347 return fix_small_imbalance(env
, sds
);
6351 * If there aren't any idle cpus, avoid creating some.
6353 if (busiest
->group_type
== group_overloaded
&&
6354 local
->group_type
== group_overloaded
) {
6355 load_above_capacity
=
6356 (busiest
->sum_nr_running
- busiest
->group_capacity_factor
);
6358 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_CAPACITY_SCALE
);
6359 load_above_capacity
/= busiest
->group_capacity
;
6363 * We're trying to get all the cpus to the average_load, so we don't
6364 * want to push ourselves above the average load, nor do we wish to
6365 * reduce the max loaded cpu below the average load. At the same time,
6366 * we also don't want to reduce the group load below the group capacity
6367 * (so that we can implement power-savings policies etc). Thus we look
6368 * for the minimum possible imbalance.
6370 max_pull
= min(busiest
->avg_load
- sds
->avg_load
, load_above_capacity
);
6372 /* How much load to actually move to equalise the imbalance */
6373 env
->imbalance
= min(
6374 max_pull
* busiest
->group_capacity
,
6375 (sds
->avg_load
- local
->avg_load
) * local
->group_capacity
6376 ) / SCHED_CAPACITY_SCALE
;
6379 * if *imbalance is less than the average load per runnable task
6380 * there is no guarantee that any tasks will be moved so we'll have
6381 * a think about bumping its value to force at least one task to be
6384 if (env
->imbalance
< busiest
->load_per_task
)
6385 return fix_small_imbalance(env
, sds
);
6388 /******* find_busiest_group() helpers end here *********************/
6391 * find_busiest_group - Returns the busiest group within the sched_domain
6392 * if there is an imbalance. If there isn't an imbalance, and
6393 * the user has opted for power-savings, it returns a group whose
6394 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6395 * such a group exists.
6397 * Also calculates the amount of weighted load which should be moved
6398 * to restore balance.
6400 * @env: The load balancing environment.
6402 * Return: - The busiest group if imbalance exists.
6403 * - If no imbalance and user has opted for power-savings balance,
6404 * return the least loaded group whose CPUs can be
6405 * put to idle by rebalancing its tasks onto our group.
6407 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
6409 struct sg_lb_stats
*local
, *busiest
;
6410 struct sd_lb_stats sds
;
6412 init_sd_lb_stats(&sds
);
6415 * Compute the various statistics relavent for load balancing at
6418 update_sd_lb_stats(env
, &sds
);
6419 local
= &sds
.local_stat
;
6420 busiest
= &sds
.busiest_stat
;
6422 if ((env
->idle
== CPU_IDLE
|| env
->idle
== CPU_NEWLY_IDLE
) &&
6423 check_asym_packing(env
, &sds
))
6426 /* There is no busy sibling group to pull tasks from */
6427 if (!sds
.busiest
|| busiest
->sum_nr_running
== 0)
6430 sds
.avg_load
= (SCHED_CAPACITY_SCALE
* sds
.total_load
)
6431 / sds
.total_capacity
;
6434 * If the busiest group is imbalanced the below checks don't
6435 * work because they assume all things are equal, which typically
6436 * isn't true due to cpus_allowed constraints and the like.
6438 if (busiest
->group_type
== group_imbalanced
)
6441 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6442 if (env
->idle
== CPU_NEWLY_IDLE
&& local
->group_has_free_capacity
&&
6443 !busiest
->group_has_free_capacity
)
6447 * If the local group is busier than the selected busiest group
6448 * don't try and pull any tasks.
6450 if (local
->avg_load
>= busiest
->avg_load
)
6454 * Don't pull any tasks if this group is already above the domain
6457 if (local
->avg_load
>= sds
.avg_load
)
6460 if (env
->idle
== CPU_IDLE
) {
6462 * This cpu is idle. If the busiest group is not overloaded
6463 * and there is no imbalance between this and busiest group
6464 * wrt idle cpus, it is balanced. The imbalance becomes
6465 * significant if the diff is greater than 1 otherwise we
6466 * might end up to just move the imbalance on another group
6468 if ((busiest
->group_type
!= group_overloaded
) &&
6469 (local
->idle_cpus
<= (busiest
->idle_cpus
+ 1)))
6473 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6474 * imbalance_pct to be conservative.
6476 if (100 * busiest
->avg_load
<=
6477 env
->sd
->imbalance_pct
* local
->avg_load
)
6482 /* Looks like there is an imbalance. Compute it */
6483 calculate_imbalance(env
, &sds
);
6492 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6494 static struct rq
*find_busiest_queue(struct lb_env
*env
,
6495 struct sched_group
*group
)
6497 struct rq
*busiest
= NULL
, *rq
;
6498 unsigned long busiest_load
= 0, busiest_capacity
= 1;
6501 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
6502 unsigned long capacity
, capacity_factor
, wl
;
6506 rt
= fbq_classify_rq(rq
);
6509 * We classify groups/runqueues into three groups:
6510 * - regular: there are !numa tasks
6511 * - remote: there are numa tasks that run on the 'wrong' node
6512 * - all: there is no distinction
6514 * In order to avoid migrating ideally placed numa tasks,
6515 * ignore those when there's better options.
6517 * If we ignore the actual busiest queue to migrate another
6518 * task, the next balance pass can still reduce the busiest
6519 * queue by moving tasks around inside the node.
6521 * If we cannot move enough load due to this classification
6522 * the next pass will adjust the group classification and
6523 * allow migration of more tasks.
6525 * Both cases only affect the total convergence complexity.
6527 if (rt
> env
->fbq_type
)
6530 capacity
= capacity_of(i
);
6531 capacity_factor
= DIV_ROUND_CLOSEST(capacity
, SCHED_CAPACITY_SCALE
);
6532 if (!capacity_factor
)
6533 capacity_factor
= fix_small_capacity(env
->sd
, group
);
6535 wl
= weighted_cpuload(i
);
6538 * When comparing with imbalance, use weighted_cpuload()
6539 * which is not scaled with the cpu capacity.
6541 if (capacity_factor
&& rq
->nr_running
== 1 && wl
> env
->imbalance
)
6545 * For the load comparisons with the other cpu's, consider
6546 * the weighted_cpuload() scaled with the cpu capacity, so
6547 * that the load can be moved away from the cpu that is
6548 * potentially running at a lower capacity.
6550 * Thus we're looking for max(wl_i / capacity_i), crosswise
6551 * multiplication to rid ourselves of the division works out
6552 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6553 * our previous maximum.
6555 if (wl
* busiest_capacity
> busiest_load
* capacity
) {
6557 busiest_capacity
= capacity
;
6566 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6567 * so long as it is large enough.
6569 #define MAX_PINNED_INTERVAL 512
6571 /* Working cpumask for load_balance and load_balance_newidle. */
6572 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6574 static int need_active_balance(struct lb_env
*env
)
6576 struct sched_domain
*sd
= env
->sd
;
6578 if (env
->idle
== CPU_NEWLY_IDLE
) {
6581 * ASYM_PACKING needs to force migrate tasks from busy but
6582 * higher numbered CPUs in order to pack all tasks in the
6583 * lowest numbered CPUs.
6585 if ((sd
->flags
& SD_ASYM_PACKING
) && env
->src_cpu
> env
->dst_cpu
)
6589 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
6592 static int active_load_balance_cpu_stop(void *data
);
6594 static int should_we_balance(struct lb_env
*env
)
6596 struct sched_group
*sg
= env
->sd
->groups
;
6597 struct cpumask
*sg_cpus
, *sg_mask
;
6598 int cpu
, balance_cpu
= -1;
6601 * In the newly idle case, we will allow all the cpu's
6602 * to do the newly idle load balance.
6604 if (env
->idle
== CPU_NEWLY_IDLE
)
6607 sg_cpus
= sched_group_cpus(sg
);
6608 sg_mask
= sched_group_mask(sg
);
6609 /* Try to find first idle cpu */
6610 for_each_cpu_and(cpu
, sg_cpus
, env
->cpus
) {
6611 if (!cpumask_test_cpu(cpu
, sg_mask
) || !idle_cpu(cpu
))
6618 if (balance_cpu
== -1)
6619 balance_cpu
= group_balance_cpu(sg
);
6622 * First idle cpu or the first cpu(busiest) in this sched group
6623 * is eligible for doing load balancing at this and above domains.
6625 return balance_cpu
== env
->dst_cpu
;
6629 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6630 * tasks if there is an imbalance.
6632 static int load_balance(int this_cpu
, struct rq
*this_rq
,
6633 struct sched_domain
*sd
, enum cpu_idle_type idle
,
6634 int *continue_balancing
)
6636 int ld_moved
, cur_ld_moved
, active_balance
= 0;
6637 struct sched_domain
*sd_parent
= sd
->parent
;
6638 struct sched_group
*group
;
6640 unsigned long flags
;
6641 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(load_balance_mask
);
6643 struct lb_env env
= {
6645 .dst_cpu
= this_cpu
,
6647 .dst_grpmask
= sched_group_cpus(sd
->groups
),
6649 .loop_break
= sched_nr_migrate_break
,
6652 .tasks
= LIST_HEAD_INIT(env
.tasks
),
6656 * For NEWLY_IDLE load_balancing, we don't need to consider
6657 * other cpus in our group
6659 if (idle
== CPU_NEWLY_IDLE
)
6660 env
.dst_grpmask
= NULL
;
6662 cpumask_copy(cpus
, cpu_active_mask
);
6664 schedstat_inc(sd
, lb_count
[idle
]);
6667 if (!should_we_balance(&env
)) {
6668 *continue_balancing
= 0;
6672 group
= find_busiest_group(&env
);
6674 schedstat_inc(sd
, lb_nobusyg
[idle
]);
6678 busiest
= find_busiest_queue(&env
, group
);
6680 schedstat_inc(sd
, lb_nobusyq
[idle
]);
6684 BUG_ON(busiest
== env
.dst_rq
);
6686 schedstat_add(sd
, lb_imbalance
[idle
], env
.imbalance
);
6689 if (busiest
->nr_running
> 1) {
6691 * Attempt to move tasks. If find_busiest_group has found
6692 * an imbalance but busiest->nr_running <= 1, the group is
6693 * still unbalanced. ld_moved simply stays zero, so it is
6694 * correctly treated as an imbalance.
6696 env
.flags
|= LBF_ALL_PINNED
;
6697 env
.src_cpu
= busiest
->cpu
;
6698 env
.src_rq
= busiest
;
6699 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
6702 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
6705 * cur_ld_moved - load moved in current iteration
6706 * ld_moved - cumulative load moved across iterations
6708 cur_ld_moved
= detach_tasks(&env
);
6711 * We've detached some tasks from busiest_rq. Every
6712 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
6713 * unlock busiest->lock, and we are able to be sure
6714 * that nobody can manipulate the tasks in parallel.
6715 * See task_rq_lock() family for the details.
6718 raw_spin_unlock(&busiest
->lock
);
6722 ld_moved
+= cur_ld_moved
;
6725 local_irq_restore(flags
);
6727 if (env
.flags
& LBF_NEED_BREAK
) {
6728 env
.flags
&= ~LBF_NEED_BREAK
;
6733 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6734 * us and move them to an alternate dst_cpu in our sched_group
6735 * where they can run. The upper limit on how many times we
6736 * iterate on same src_cpu is dependent on number of cpus in our
6739 * This changes load balance semantics a bit on who can move
6740 * load to a given_cpu. In addition to the given_cpu itself
6741 * (or a ilb_cpu acting on its behalf where given_cpu is
6742 * nohz-idle), we now have balance_cpu in a position to move
6743 * load to given_cpu. In rare situations, this may cause
6744 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6745 * _independently_ and at _same_ time to move some load to
6746 * given_cpu) causing exceess load to be moved to given_cpu.
6747 * This however should not happen so much in practice and
6748 * moreover subsequent load balance cycles should correct the
6749 * excess load moved.
6751 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
6753 /* Prevent to re-select dst_cpu via env's cpus */
6754 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
6756 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
6757 env
.dst_cpu
= env
.new_dst_cpu
;
6758 env
.flags
&= ~LBF_DST_PINNED
;
6760 env
.loop_break
= sched_nr_migrate_break
;
6763 * Go back to "more_balance" rather than "redo" since we
6764 * need to continue with same src_cpu.
6770 * We failed to reach balance because of affinity.
6773 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
6775 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0)
6776 *group_imbalance
= 1;
6779 /* All tasks on this runqueue were pinned by CPU affinity */
6780 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
6781 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
6782 if (!cpumask_empty(cpus
)) {
6784 env
.loop_break
= sched_nr_migrate_break
;
6787 goto out_all_pinned
;
6792 schedstat_inc(sd
, lb_failed
[idle
]);
6794 * Increment the failure counter only on periodic balance.
6795 * We do not want newidle balance, which can be very
6796 * frequent, pollute the failure counter causing
6797 * excessive cache_hot migrations and active balances.
6799 if (idle
!= CPU_NEWLY_IDLE
)
6800 sd
->nr_balance_failed
++;
6802 if (need_active_balance(&env
)) {
6803 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
6805 /* don't kick the active_load_balance_cpu_stop,
6806 * if the curr task on busiest cpu can't be
6809 if (!cpumask_test_cpu(this_cpu
,
6810 tsk_cpus_allowed(busiest
->curr
))) {
6811 raw_spin_unlock_irqrestore(&busiest
->lock
,
6813 env
.flags
|= LBF_ALL_PINNED
;
6814 goto out_one_pinned
;
6818 * ->active_balance synchronizes accesses to
6819 * ->active_balance_work. Once set, it's cleared
6820 * only after active load balance is finished.
6822 if (!busiest
->active_balance
) {
6823 busiest
->active_balance
= 1;
6824 busiest
->push_cpu
= this_cpu
;
6827 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
6829 if (active_balance
) {
6830 stop_one_cpu_nowait(cpu_of(busiest
),
6831 active_load_balance_cpu_stop
, busiest
,
6832 &busiest
->active_balance_work
);
6836 * We've kicked active balancing, reset the failure
6839 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
6842 sd
->nr_balance_failed
= 0;
6844 if (likely(!active_balance
)) {
6845 /* We were unbalanced, so reset the balancing interval */
6846 sd
->balance_interval
= sd
->min_interval
;
6849 * If we've begun active balancing, start to back off. This
6850 * case may not be covered by the all_pinned logic if there
6851 * is only 1 task on the busy runqueue (because we don't call
6854 if (sd
->balance_interval
< sd
->max_interval
)
6855 sd
->balance_interval
*= 2;
6862 * We reach balance although we may have faced some affinity
6863 * constraints. Clear the imbalance flag if it was set.
6866 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
6868 if (*group_imbalance
)
6869 *group_imbalance
= 0;
6874 * We reach balance because all tasks are pinned at this level so
6875 * we can't migrate them. Let the imbalance flag set so parent level
6876 * can try to migrate them.
6878 schedstat_inc(sd
, lb_balanced
[idle
]);
6880 sd
->nr_balance_failed
= 0;
6883 /* tune up the balancing interval */
6884 if (((env
.flags
& LBF_ALL_PINNED
) &&
6885 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
6886 (sd
->balance_interval
< sd
->max_interval
))
6887 sd
->balance_interval
*= 2;
6894 static inline unsigned long
6895 get_sd_balance_interval(struct sched_domain
*sd
, int cpu_busy
)
6897 unsigned long interval
= sd
->balance_interval
;
6900 interval
*= sd
->busy_factor
;
6902 /* scale ms to jiffies */
6903 interval
= msecs_to_jiffies(interval
);
6904 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
6910 update_next_balance(struct sched_domain
*sd
, int cpu_busy
, unsigned long *next_balance
)
6912 unsigned long interval
, next
;
6914 interval
= get_sd_balance_interval(sd
, cpu_busy
);
6915 next
= sd
->last_balance
+ interval
;
6917 if (time_after(*next_balance
, next
))
6918 *next_balance
= next
;
6922 * idle_balance is called by schedule() if this_cpu is about to become
6923 * idle. Attempts to pull tasks from other CPUs.
6925 static int idle_balance(struct rq
*this_rq
)
6927 unsigned long next_balance
= jiffies
+ HZ
;
6928 int this_cpu
= this_rq
->cpu
;
6929 struct sched_domain
*sd
;
6930 int pulled_task
= 0;
6933 idle_enter_fair(this_rq
);
6936 * We must set idle_stamp _before_ calling idle_balance(), such that we
6937 * measure the duration of idle_balance() as idle time.
6939 this_rq
->idle_stamp
= rq_clock(this_rq
);
6941 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
||
6942 !this_rq
->rd
->overload
) {
6944 sd
= rcu_dereference_check_sched_domain(this_rq
->sd
);
6946 update_next_balance(sd
, 0, &next_balance
);
6953 * Drop the rq->lock, but keep IRQ/preempt disabled.
6955 raw_spin_unlock(&this_rq
->lock
);
6957 update_blocked_averages(this_cpu
);
6959 for_each_domain(this_cpu
, sd
) {
6960 int continue_balancing
= 1;
6961 u64 t0
, domain_cost
;
6963 if (!(sd
->flags
& SD_LOAD_BALANCE
))
6966 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
) {
6967 update_next_balance(sd
, 0, &next_balance
);
6971 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
6972 t0
= sched_clock_cpu(this_cpu
);
6974 pulled_task
= load_balance(this_cpu
, this_rq
,
6976 &continue_balancing
);
6978 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
6979 if (domain_cost
> sd
->max_newidle_lb_cost
)
6980 sd
->max_newidle_lb_cost
= domain_cost
;
6982 curr_cost
+= domain_cost
;
6985 update_next_balance(sd
, 0, &next_balance
);
6988 * Stop searching for tasks to pull if there are
6989 * now runnable tasks on this rq.
6991 if (pulled_task
|| this_rq
->nr_running
> 0)
6996 raw_spin_lock(&this_rq
->lock
);
6998 if (curr_cost
> this_rq
->max_idle_balance_cost
)
6999 this_rq
->max_idle_balance_cost
= curr_cost
;
7002 * While browsing the domains, we released the rq lock, a task could
7003 * have been enqueued in the meantime. Since we're not going idle,
7004 * pretend we pulled a task.
7006 if (this_rq
->cfs
.h_nr_running
&& !pulled_task
)
7010 /* Move the next balance forward */
7011 if (time_after(this_rq
->next_balance
, next_balance
))
7012 this_rq
->next_balance
= next_balance
;
7014 /* Is there a task of a high priority class? */
7015 if (this_rq
->nr_running
!= this_rq
->cfs
.h_nr_running
)
7019 idle_exit_fair(this_rq
);
7020 this_rq
->idle_stamp
= 0;
7027 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7028 * running tasks off the busiest CPU onto idle CPUs. It requires at
7029 * least 1 task to be running on each physical CPU where possible, and
7030 * avoids physical / logical imbalances.
7032 static int active_load_balance_cpu_stop(void *data
)
7034 struct rq
*busiest_rq
= data
;
7035 int busiest_cpu
= cpu_of(busiest_rq
);
7036 int target_cpu
= busiest_rq
->push_cpu
;
7037 struct rq
*target_rq
= cpu_rq(target_cpu
);
7038 struct sched_domain
*sd
;
7039 struct task_struct
*p
= NULL
;
7041 raw_spin_lock_irq(&busiest_rq
->lock
);
7043 /* make sure the requested cpu hasn't gone down in the meantime */
7044 if (unlikely(busiest_cpu
!= smp_processor_id() ||
7045 !busiest_rq
->active_balance
))
7048 /* Is there any task to move? */
7049 if (busiest_rq
->nr_running
<= 1)
7053 * This condition is "impossible", if it occurs
7054 * we need to fix it. Originally reported by
7055 * Bjorn Helgaas on a 128-cpu setup.
7057 BUG_ON(busiest_rq
== target_rq
);
7059 /* Search for an sd spanning us and the target CPU. */
7061 for_each_domain(target_cpu
, sd
) {
7062 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
7063 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
7068 struct lb_env env
= {
7070 .dst_cpu
= target_cpu
,
7071 .dst_rq
= target_rq
,
7072 .src_cpu
= busiest_rq
->cpu
,
7073 .src_rq
= busiest_rq
,
7077 schedstat_inc(sd
, alb_count
);
7079 p
= detach_one_task(&env
);
7081 schedstat_inc(sd
, alb_pushed
);
7083 schedstat_inc(sd
, alb_failed
);
7087 busiest_rq
->active_balance
= 0;
7088 raw_spin_unlock(&busiest_rq
->lock
);
7091 attach_one_task(target_rq
, p
);
7098 static inline int on_null_domain(struct rq
*rq
)
7100 return unlikely(!rcu_dereference_sched(rq
->sd
));
7103 #ifdef CONFIG_NO_HZ_COMMON
7105 * idle load balancing details
7106 * - When one of the busy CPUs notice that there may be an idle rebalancing
7107 * needed, they will kick the idle load balancer, which then does idle
7108 * load balancing for all the idle CPUs.
7111 cpumask_var_t idle_cpus_mask
;
7113 unsigned long next_balance
; /* in jiffy units */
7114 } nohz ____cacheline_aligned
;
7116 static inline int find_new_ilb(void)
7118 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
7120 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
7127 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7128 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7129 * CPU (if there is one).
7131 static void nohz_balancer_kick(void)
7135 nohz
.next_balance
++;
7137 ilb_cpu
= find_new_ilb();
7139 if (ilb_cpu
>= nr_cpu_ids
)
7142 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
7145 * Use smp_send_reschedule() instead of resched_cpu().
7146 * This way we generate a sched IPI on the target cpu which
7147 * is idle. And the softirq performing nohz idle load balance
7148 * will be run before returning from the IPI.
7150 smp_send_reschedule(ilb_cpu
);
7154 static inline void nohz_balance_exit_idle(int cpu
)
7156 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
7158 * Completely isolated CPUs don't ever set, so we must test.
7160 if (likely(cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))) {
7161 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
7162 atomic_dec(&nohz
.nr_cpus
);
7164 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
7168 static inline void set_cpu_sd_state_busy(void)
7170 struct sched_domain
*sd
;
7171 int cpu
= smp_processor_id();
7174 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7176 if (!sd
|| !sd
->nohz_idle
)
7180 atomic_inc(&sd
->groups
->sgc
->nr_busy_cpus
);
7185 void set_cpu_sd_state_idle(void)
7187 struct sched_domain
*sd
;
7188 int cpu
= smp_processor_id();
7191 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7193 if (!sd
|| sd
->nohz_idle
)
7197 atomic_dec(&sd
->groups
->sgc
->nr_busy_cpus
);
7203 * This routine will record that the cpu is going idle with tick stopped.
7204 * This info will be used in performing idle load balancing in the future.
7206 void nohz_balance_enter_idle(int cpu
)
7209 * If this cpu is going down, then nothing needs to be done.
7211 if (!cpu_active(cpu
))
7214 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
7218 * If we're a completely isolated CPU, we don't play.
7220 if (on_null_domain(cpu_rq(cpu
)))
7223 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
7224 atomic_inc(&nohz
.nr_cpus
);
7225 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
7228 static int sched_ilb_notifier(struct notifier_block
*nfb
,
7229 unsigned long action
, void *hcpu
)
7231 switch (action
& ~CPU_TASKS_FROZEN
) {
7233 nohz_balance_exit_idle(smp_processor_id());
7241 static DEFINE_SPINLOCK(balancing
);
7244 * Scale the max load_balance interval with the number of CPUs in the system.
7245 * This trades load-balance latency on larger machines for less cross talk.
7247 void update_max_interval(void)
7249 max_load_balance_interval
= HZ
*num_online_cpus()/10;
7253 * It checks each scheduling domain to see if it is due to be balanced,
7254 * and initiates a balancing operation if so.
7256 * Balancing parameters are set up in init_sched_domains.
7258 static void rebalance_domains(struct rq
*rq
, enum cpu_idle_type idle
)
7260 int continue_balancing
= 1;
7262 unsigned long interval
;
7263 struct sched_domain
*sd
;
7264 /* Earliest time when we have to do rebalance again */
7265 unsigned long next_balance
= jiffies
+ 60*HZ
;
7266 int update_next_balance
= 0;
7267 int need_serialize
, need_decay
= 0;
7270 update_blocked_averages(cpu
);
7273 for_each_domain(cpu
, sd
) {
7275 * Decay the newidle max times here because this is a regular
7276 * visit to all the domains. Decay ~1% per second.
7278 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
7279 sd
->max_newidle_lb_cost
=
7280 (sd
->max_newidle_lb_cost
* 253) / 256;
7281 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
7284 max_cost
+= sd
->max_newidle_lb_cost
;
7286 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7290 * Stop the load balance at this level. There is another
7291 * CPU in our sched group which is doing load balancing more
7294 if (!continue_balancing
) {
7300 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7302 need_serialize
= sd
->flags
& SD_SERIALIZE
;
7303 if (need_serialize
) {
7304 if (!spin_trylock(&balancing
))
7308 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
7309 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
7311 * The LBF_DST_PINNED logic could have changed
7312 * env->dst_cpu, so we can't know our idle
7313 * state even if we migrated tasks. Update it.
7315 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
7317 sd
->last_balance
= jiffies
;
7318 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7321 spin_unlock(&balancing
);
7323 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
7324 next_balance
= sd
->last_balance
+ interval
;
7325 update_next_balance
= 1;
7330 * Ensure the rq-wide value also decays but keep it at a
7331 * reasonable floor to avoid funnies with rq->avg_idle.
7333 rq
->max_idle_balance_cost
=
7334 max((u64
)sysctl_sched_migration_cost
, max_cost
);
7339 * next_balance will be updated only when there is a need.
7340 * When the cpu is attached to null domain for ex, it will not be
7343 if (likely(update_next_balance
))
7344 rq
->next_balance
= next_balance
;
7347 #ifdef CONFIG_NO_HZ_COMMON
7349 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7350 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7352 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
7354 int this_cpu
= this_rq
->cpu
;
7358 if (idle
!= CPU_IDLE
||
7359 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
7362 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
7363 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
7367 * If this cpu gets work to do, stop the load balancing
7368 * work being done for other cpus. Next load
7369 * balancing owner will pick it up.
7374 rq
= cpu_rq(balance_cpu
);
7377 * If time for next balance is due,
7380 if (time_after_eq(jiffies
, rq
->next_balance
)) {
7381 raw_spin_lock_irq(&rq
->lock
);
7382 update_rq_clock(rq
);
7383 update_idle_cpu_load(rq
);
7384 raw_spin_unlock_irq(&rq
->lock
);
7385 rebalance_domains(rq
, CPU_IDLE
);
7388 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
7389 this_rq
->next_balance
= rq
->next_balance
;
7391 nohz
.next_balance
= this_rq
->next_balance
;
7393 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
7397 * Current heuristic for kicking the idle load balancer in the presence
7398 * of an idle cpu is the system.
7399 * - This rq has more than one task.
7400 * - At any scheduler domain level, this cpu's scheduler group has multiple
7401 * busy cpu's exceeding the group's capacity.
7402 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7403 * domain span are idle.
7405 static inline int nohz_kick_needed(struct rq
*rq
)
7407 unsigned long now
= jiffies
;
7408 struct sched_domain
*sd
;
7409 struct sched_group_capacity
*sgc
;
7410 int nr_busy
, cpu
= rq
->cpu
;
7412 if (unlikely(rq
->idle_balance
))
7416 * We may be recently in ticked or tickless idle mode. At the first
7417 * busy tick after returning from idle, we will update the busy stats.
7419 set_cpu_sd_state_busy();
7420 nohz_balance_exit_idle(cpu
);
7423 * None are in tickless mode and hence no need for NOHZ idle load
7426 if (likely(!atomic_read(&nohz
.nr_cpus
)))
7429 if (time_before(now
, nohz
.next_balance
))
7432 if (rq
->nr_running
>= 2)
7436 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7439 sgc
= sd
->groups
->sgc
;
7440 nr_busy
= atomic_read(&sgc
->nr_busy_cpus
);
7443 goto need_kick_unlock
;
7446 sd
= rcu_dereference(per_cpu(sd_asym
, cpu
));
7448 if (sd
&& (cpumask_first_and(nohz
.idle_cpus_mask
,
7449 sched_domain_span(sd
)) < cpu
))
7450 goto need_kick_unlock
;
7461 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
) { }
7465 * run_rebalance_domains is triggered when needed from the scheduler tick.
7466 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7468 static void run_rebalance_domains(struct softirq_action
*h
)
7470 struct rq
*this_rq
= this_rq();
7471 enum cpu_idle_type idle
= this_rq
->idle_balance
?
7472 CPU_IDLE
: CPU_NOT_IDLE
;
7474 rebalance_domains(this_rq
, idle
);
7477 * If this cpu has a pending nohz_balance_kick, then do the
7478 * balancing on behalf of the other idle cpus whose ticks are
7481 nohz_idle_balance(this_rq
, idle
);
7485 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7487 void trigger_load_balance(struct rq
*rq
)
7489 /* Don't need to rebalance while attached to NULL domain */
7490 if (unlikely(on_null_domain(rq
)))
7493 if (time_after_eq(jiffies
, rq
->next_balance
))
7494 raise_softirq(SCHED_SOFTIRQ
);
7495 #ifdef CONFIG_NO_HZ_COMMON
7496 if (nohz_kick_needed(rq
))
7497 nohz_balancer_kick();
7501 static void rq_online_fair(struct rq
*rq
)
7505 update_runtime_enabled(rq
);
7508 static void rq_offline_fair(struct rq
*rq
)
7512 /* Ensure any throttled groups are reachable by pick_next_task */
7513 unthrottle_offline_cfs_rqs(rq
);
7516 #endif /* CONFIG_SMP */
7519 * scheduler tick hitting a task of our scheduling class:
7521 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
7523 struct cfs_rq
*cfs_rq
;
7524 struct sched_entity
*se
= &curr
->se
;
7526 for_each_sched_entity(se
) {
7527 cfs_rq
= cfs_rq_of(se
);
7528 entity_tick(cfs_rq
, se
, queued
);
7531 if (numabalancing_enabled
)
7532 task_tick_numa(rq
, curr
);
7534 update_rq_runnable_avg(rq
, 1);
7538 * called on fork with the child task as argument from the parent's context
7539 * - child not yet on the tasklist
7540 * - preemption disabled
7542 static void task_fork_fair(struct task_struct
*p
)
7544 struct cfs_rq
*cfs_rq
;
7545 struct sched_entity
*se
= &p
->se
, *curr
;
7546 int this_cpu
= smp_processor_id();
7547 struct rq
*rq
= this_rq();
7548 unsigned long flags
;
7550 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7552 update_rq_clock(rq
);
7554 cfs_rq
= task_cfs_rq(current
);
7555 curr
= cfs_rq
->curr
;
7558 * Not only the cpu but also the task_group of the parent might have
7559 * been changed after parent->se.parent,cfs_rq were copied to
7560 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7561 * of child point to valid ones.
7564 __set_task_cpu(p
, this_cpu
);
7567 update_curr(cfs_rq
);
7570 se
->vruntime
= curr
->vruntime
;
7571 place_entity(cfs_rq
, se
, 1);
7573 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
7575 * Upon rescheduling, sched_class::put_prev_task() will place
7576 * 'current' within the tree based on its new key value.
7578 swap(curr
->vruntime
, se
->vruntime
);
7582 se
->vruntime
-= cfs_rq
->min_vruntime
;
7584 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7588 * Priority of the task has changed. Check to see if we preempt
7592 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
7594 if (!task_on_rq_queued(p
))
7598 * Reschedule if we are currently running on this runqueue and
7599 * our priority decreased, or if we are not currently running on
7600 * this runqueue and our priority is higher than the current's
7602 if (rq
->curr
== p
) {
7603 if (p
->prio
> oldprio
)
7606 check_preempt_curr(rq
, p
, 0);
7609 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
7611 struct sched_entity
*se
= &p
->se
;
7612 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7615 * Ensure the task's vruntime is normalized, so that when it's
7616 * switched back to the fair class the enqueue_entity(.flags=0) will
7617 * do the right thing.
7619 * If it's queued, then the dequeue_entity(.flags=0) will already
7620 * have normalized the vruntime, if it's !queued, then only when
7621 * the task is sleeping will it still have non-normalized vruntime.
7623 if (!task_on_rq_queued(p
) && p
->state
!= TASK_RUNNING
) {
7625 * Fix up our vruntime so that the current sleep doesn't
7626 * cause 'unlimited' sleep bonus.
7628 place_entity(cfs_rq
, se
, 0);
7629 se
->vruntime
-= cfs_rq
->min_vruntime
;
7634 * Remove our load from contribution when we leave sched_fair
7635 * and ensure we don't carry in an old decay_count if we
7638 if (se
->avg
.decay_count
) {
7639 __synchronize_entity_decay(se
);
7640 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
7646 * We switched to the sched_fair class.
7648 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
7650 #ifdef CONFIG_FAIR_GROUP_SCHED
7651 struct sched_entity
*se
= &p
->se
;
7653 * Since the real-depth could have been changed (only FAIR
7654 * class maintain depth value), reset depth properly.
7656 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
7658 if (!task_on_rq_queued(p
))
7662 * We were most likely switched from sched_rt, so
7663 * kick off the schedule if running, otherwise just see
7664 * if we can still preempt the current task.
7669 check_preempt_curr(rq
, p
, 0);
7672 /* Account for a task changing its policy or group.
7674 * This routine is mostly called to set cfs_rq->curr field when a task
7675 * migrates between groups/classes.
7677 static void set_curr_task_fair(struct rq
*rq
)
7679 struct sched_entity
*se
= &rq
->curr
->se
;
7681 for_each_sched_entity(se
) {
7682 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7684 set_next_entity(cfs_rq
, se
);
7685 /* ensure bandwidth has been allocated on our new cfs_rq */
7686 account_cfs_rq_runtime(cfs_rq
, 0);
7690 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
7692 cfs_rq
->tasks_timeline
= RB_ROOT
;
7693 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7694 #ifndef CONFIG_64BIT
7695 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
7698 atomic64_set(&cfs_rq
->decay_counter
, 1);
7699 atomic_long_set(&cfs_rq
->removed_load
, 0);
7703 #ifdef CONFIG_FAIR_GROUP_SCHED
7704 static void task_move_group_fair(struct task_struct
*p
, int queued
)
7706 struct sched_entity
*se
= &p
->se
;
7707 struct cfs_rq
*cfs_rq
;
7710 * If the task was not on the rq at the time of this cgroup movement
7711 * it must have been asleep, sleeping tasks keep their ->vruntime
7712 * absolute on their old rq until wakeup (needed for the fair sleeper
7713 * bonus in place_entity()).
7715 * If it was on the rq, we've just 'preempted' it, which does convert
7716 * ->vruntime to a relative base.
7718 * Make sure both cases convert their relative position when migrating
7719 * to another cgroup's rq. This does somewhat interfere with the
7720 * fair sleeper stuff for the first placement, but who cares.
7723 * When !queued, vruntime of the task has usually NOT been normalized.
7724 * But there are some cases where it has already been normalized:
7726 * - Moving a forked child which is waiting for being woken up by
7727 * wake_up_new_task().
7728 * - Moving a task which has been woken up by try_to_wake_up() and
7729 * waiting for actually being woken up by sched_ttwu_pending().
7731 * To prevent boost or penalty in the new cfs_rq caused by delta
7732 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7734 if (!queued
&& (!se
->sum_exec_runtime
|| p
->state
== TASK_WAKING
))
7738 se
->vruntime
-= cfs_rq_of(se
)->min_vruntime
;
7739 set_task_rq(p
, task_cpu(p
));
7740 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
7742 cfs_rq
= cfs_rq_of(se
);
7743 se
->vruntime
+= cfs_rq
->min_vruntime
;
7746 * migrate_task_rq_fair() will have removed our previous
7747 * contribution, but we must synchronize for ongoing future
7750 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
7751 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
7756 void free_fair_sched_group(struct task_group
*tg
)
7760 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7762 for_each_possible_cpu(i
) {
7764 kfree(tg
->cfs_rq
[i
]);
7773 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7775 struct cfs_rq
*cfs_rq
;
7776 struct sched_entity
*se
;
7779 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
7782 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
7786 tg
->shares
= NICE_0_LOAD
;
7788 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
7790 for_each_possible_cpu(i
) {
7791 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
7792 GFP_KERNEL
, cpu_to_node(i
));
7796 se
= kzalloc_node(sizeof(struct sched_entity
),
7797 GFP_KERNEL
, cpu_to_node(i
));
7801 init_cfs_rq(cfs_rq
);
7802 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
7813 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
7815 struct rq
*rq
= cpu_rq(cpu
);
7816 unsigned long flags
;
7819 * Only empty task groups can be destroyed; so we can speculatively
7820 * check on_list without danger of it being re-added.
7822 if (!tg
->cfs_rq
[cpu
]->on_list
)
7825 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7826 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
7827 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7830 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7831 struct sched_entity
*se
, int cpu
,
7832 struct sched_entity
*parent
)
7834 struct rq
*rq
= cpu_rq(cpu
);
7838 init_cfs_rq_runtime(cfs_rq
);
7840 tg
->cfs_rq
[cpu
] = cfs_rq
;
7843 /* se could be NULL for root_task_group */
7848 se
->cfs_rq
= &rq
->cfs
;
7851 se
->cfs_rq
= parent
->my_q
;
7852 se
->depth
= parent
->depth
+ 1;
7856 /* guarantee group entities always have weight */
7857 update_load_set(&se
->load
, NICE_0_LOAD
);
7858 se
->parent
= parent
;
7861 static DEFINE_MUTEX(shares_mutex
);
7863 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
7866 unsigned long flags
;
7869 * We can't change the weight of the root cgroup.
7874 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
7876 mutex_lock(&shares_mutex
);
7877 if (tg
->shares
== shares
)
7880 tg
->shares
= shares
;
7881 for_each_possible_cpu(i
) {
7882 struct rq
*rq
= cpu_rq(i
);
7883 struct sched_entity
*se
;
7886 /* Propagate contribution to hierarchy */
7887 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7889 /* Possible calls to update_curr() need rq clock */
7890 update_rq_clock(rq
);
7891 for_each_sched_entity(se
)
7892 update_cfs_shares(group_cfs_rq(se
));
7893 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7897 mutex_unlock(&shares_mutex
);
7900 #else /* CONFIG_FAIR_GROUP_SCHED */
7902 void free_fair_sched_group(struct task_group
*tg
) { }
7904 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7909 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
) { }
7911 #endif /* CONFIG_FAIR_GROUP_SCHED */
7914 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
7916 struct sched_entity
*se
= &task
->se
;
7917 unsigned int rr_interval
= 0;
7920 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7923 if (rq
->cfs
.load
.weight
)
7924 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
7930 * All the scheduling class methods:
7932 const struct sched_class fair_sched_class
= {
7933 .next
= &idle_sched_class
,
7934 .enqueue_task
= enqueue_task_fair
,
7935 .dequeue_task
= dequeue_task_fair
,
7936 .yield_task
= yield_task_fair
,
7937 .yield_to_task
= yield_to_task_fair
,
7939 .check_preempt_curr
= check_preempt_wakeup
,
7941 .pick_next_task
= pick_next_task_fair
,
7942 .put_prev_task
= put_prev_task_fair
,
7945 .select_task_rq
= select_task_rq_fair
,
7946 .migrate_task_rq
= migrate_task_rq_fair
,
7948 .rq_online
= rq_online_fair
,
7949 .rq_offline
= rq_offline_fair
,
7951 .task_waking
= task_waking_fair
,
7954 .set_curr_task
= set_curr_task_fair
,
7955 .task_tick
= task_tick_fair
,
7956 .task_fork
= task_fork_fair
,
7958 .prio_changed
= prio_changed_fair
,
7959 .switched_from
= switched_from_fair
,
7960 .switched_to
= switched_to_fair
,
7962 .get_rr_interval
= get_rr_interval_fair
,
7964 .update_curr
= update_curr_fair
,
7966 #ifdef CONFIG_FAIR_GROUP_SCHED
7967 .task_move_group
= task_move_group_fair
,
7971 #ifdef CONFIG_SCHED_DEBUG
7972 void print_cfs_stats(struct seq_file
*m
, int cpu
)
7974 struct cfs_rq
*cfs_rq
;
7977 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
7978 print_cfs_rq(m
, cpu
, cfs_rq
);
7983 __init
void init_sched_fair_class(void)
7986 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7988 #ifdef CONFIG_NO_HZ_COMMON
7989 nohz
.next_balance
= jiffies
;
7990 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
7991 cpu_notifier(sched_ilb_notifier
, 0);