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
);
673 static inline void __update_task_entity_utilization(struct sched_entity
*se
);
675 /* Give new task start runnable values to heavy its load in infant time */
676 void init_task_runnable_average(struct task_struct
*p
)
680 slice
= sched_slice(task_cfs_rq(p
), &p
->se
) >> 10;
681 p
->se
.avg
.runnable_avg_sum
= p
->se
.avg
.running_avg_sum
= slice
;
682 p
->se
.avg
.avg_period
= slice
;
683 __update_task_entity_contrib(&p
->se
);
684 __update_task_entity_utilization(&p
->se
);
687 void init_task_runnable_average(struct task_struct
*p
)
693 * Update the current task's runtime statistics.
695 static void update_curr(struct cfs_rq
*cfs_rq
)
697 struct sched_entity
*curr
= cfs_rq
->curr
;
698 u64 now
= rq_clock_task(rq_of(cfs_rq
));
704 delta_exec
= now
- curr
->exec_start
;
705 if (unlikely((s64
)delta_exec
<= 0))
708 curr
->exec_start
= now
;
710 schedstat_set(curr
->statistics
.exec_max
,
711 max(delta_exec
, curr
->statistics
.exec_max
));
713 curr
->sum_exec_runtime
+= delta_exec
;
714 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
716 curr
->vruntime
+= calc_delta_fair(delta_exec
, curr
);
717 update_min_vruntime(cfs_rq
);
719 if (entity_is_task(curr
)) {
720 struct task_struct
*curtask
= task_of(curr
);
722 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
723 cpuacct_charge(curtask
, delta_exec
);
724 account_group_exec_runtime(curtask
, delta_exec
);
727 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
730 static void update_curr_fair(struct rq
*rq
)
732 update_curr(cfs_rq_of(&rq
->curr
->se
));
736 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
738 schedstat_set(se
->statistics
.wait_start
, rq_clock(rq_of(cfs_rq
)));
742 * Task is being enqueued - update stats:
744 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
747 * Are we enqueueing a waiting task? (for current tasks
748 * a dequeue/enqueue event is a NOP)
750 if (se
!= cfs_rq
->curr
)
751 update_stats_wait_start(cfs_rq
, se
);
755 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
757 schedstat_set(se
->statistics
.wait_max
, max(se
->statistics
.wait_max
,
758 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
));
759 schedstat_set(se
->statistics
.wait_count
, se
->statistics
.wait_count
+ 1);
760 schedstat_set(se
->statistics
.wait_sum
, se
->statistics
.wait_sum
+
761 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
762 #ifdef CONFIG_SCHEDSTATS
763 if (entity_is_task(se
)) {
764 trace_sched_stat_wait(task_of(se
),
765 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
768 schedstat_set(se
->statistics
.wait_start
, 0);
772 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
775 * Mark the end of the wait period if dequeueing a
778 if (se
!= cfs_rq
->curr
)
779 update_stats_wait_end(cfs_rq
, se
);
783 * We are picking a new current task - update its stats:
786 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
789 * We are starting a new run period:
791 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
794 /**************************************************
795 * Scheduling class queueing methods:
798 #ifdef CONFIG_NUMA_BALANCING
800 * Approximate time to scan a full NUMA task in ms. The task scan period is
801 * calculated based on the tasks virtual memory size and
802 * numa_balancing_scan_size.
804 unsigned int sysctl_numa_balancing_scan_period_min
= 1000;
805 unsigned int sysctl_numa_balancing_scan_period_max
= 60000;
807 /* Portion of address space to scan in MB */
808 unsigned int sysctl_numa_balancing_scan_size
= 256;
810 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
811 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
813 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
815 unsigned long rss
= 0;
816 unsigned long nr_scan_pages
;
819 * Calculations based on RSS as non-present and empty pages are skipped
820 * by the PTE scanner and NUMA hinting faults should be trapped based
823 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
824 rss
= get_mm_rss(p
->mm
);
828 rss
= round_up(rss
, nr_scan_pages
);
829 return rss
/ nr_scan_pages
;
832 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
833 #define MAX_SCAN_WINDOW 2560
835 static unsigned int task_scan_min(struct task_struct
*p
)
837 unsigned int scan_size
= ACCESS_ONCE(sysctl_numa_balancing_scan_size
);
838 unsigned int scan
, floor
;
839 unsigned int windows
= 1;
841 if (scan_size
< MAX_SCAN_WINDOW
)
842 windows
= MAX_SCAN_WINDOW
/ scan_size
;
843 floor
= 1000 / windows
;
845 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
846 return max_t(unsigned int, floor
, scan
);
849 static unsigned int task_scan_max(struct task_struct
*p
)
851 unsigned int smin
= task_scan_min(p
);
854 /* Watch for min being lower than max due to floor calculations */
855 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
856 return max(smin
, smax
);
859 static void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
861 rq
->nr_numa_running
+= (p
->numa_preferred_nid
!= -1);
862 rq
->nr_preferred_running
+= (p
->numa_preferred_nid
== task_node(p
));
865 static void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
867 rq
->nr_numa_running
-= (p
->numa_preferred_nid
!= -1);
868 rq
->nr_preferred_running
-= (p
->numa_preferred_nid
== task_node(p
));
874 spinlock_t lock
; /* nr_tasks, tasks */
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;
905 * The averaged statistics, shared & private, memory & cpu,
906 * occupy the first half of the array. The second half of the
907 * array is for current counters, which are averaged into the
908 * first set by task_numa_placement.
910 static inline int task_faults_idx(enum numa_faults_stats s
, int nid
, int priv
)
912 return NR_NUMA_HINT_FAULT_TYPES
* (s
* nr_node_ids
+ nid
) + priv
;
915 static inline unsigned long task_faults(struct task_struct
*p
, int nid
)
920 return p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
921 p
->numa_faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
924 static inline unsigned long group_faults(struct task_struct
*p
, int nid
)
929 return p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
930 p
->numa_group
->faults
[task_faults_idx(NUMA_MEM
, nid
, 1)];
933 static inline unsigned long group_faults_cpu(struct numa_group
*group
, int nid
)
935 return group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 0)] +
936 group
->faults_cpu
[task_faults_idx(NUMA_MEM
, nid
, 1)];
939 /* Handle placement on systems where not all nodes are directly connected. */
940 static unsigned long score_nearby_nodes(struct task_struct
*p
, int nid
,
941 int maxdist
, bool task
)
943 unsigned long score
= 0;
947 * All nodes are directly connected, and the same distance
948 * from each other. No need for fancy placement algorithms.
950 if (sched_numa_topology_type
== NUMA_DIRECT
)
954 * This code is called for each node, introducing N^2 complexity,
955 * which should be ok given the number of nodes rarely exceeds 8.
957 for_each_online_node(node
) {
958 unsigned long faults
;
959 int dist
= node_distance(nid
, node
);
962 * The furthest away nodes in the system are not interesting
963 * for placement; nid was already counted.
965 if (dist
== sched_max_numa_distance
|| node
== nid
)
969 * On systems with a backplane NUMA topology, compare groups
970 * of nodes, and move tasks towards the group with the most
971 * memory accesses. When comparing two nodes at distance
972 * "hoplimit", only nodes closer by than "hoplimit" are part
973 * of each group. Skip other nodes.
975 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
979 /* Add up the faults from nearby nodes. */
981 faults
= task_faults(p
, node
);
983 faults
= group_faults(p
, node
);
986 * On systems with a glueless mesh NUMA topology, there are
987 * no fixed "groups of nodes". Instead, nodes that are not
988 * directly connected bounce traffic through intermediate
989 * nodes; a numa_group can occupy any set of nodes.
990 * The further away a node is, the less the faults count.
991 * This seems to result in good task placement.
993 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
994 faults
*= (sched_max_numa_distance
- dist
);
995 faults
/= (sched_max_numa_distance
- LOCAL_DISTANCE
);
1005 * These return the fraction of accesses done by a particular task, or
1006 * task group, on a particular numa node. The group weight is given a
1007 * larger multiplier, in order to group tasks together that are almost
1008 * evenly spread out between numa nodes.
1010 static inline unsigned long task_weight(struct task_struct
*p
, int nid
,
1013 unsigned long faults
, total_faults
;
1015 if (!p
->numa_faults
)
1018 total_faults
= p
->total_numa_faults
;
1023 faults
= task_faults(p
, nid
);
1024 faults
+= score_nearby_nodes(p
, nid
, dist
, true);
1026 return 1000 * faults
/ total_faults
;
1029 static inline unsigned long group_weight(struct task_struct
*p
, int nid
,
1032 unsigned long faults
, total_faults
;
1037 total_faults
= p
->numa_group
->total_faults
;
1042 faults
= group_faults(p
, nid
);
1043 faults
+= score_nearby_nodes(p
, nid
, dist
, false);
1045 return 1000 * faults
/ total_faults
;
1048 bool should_numa_migrate_memory(struct task_struct
*p
, struct page
* page
,
1049 int src_nid
, int dst_cpu
)
1051 struct numa_group
*ng
= p
->numa_group
;
1052 int dst_nid
= cpu_to_node(dst_cpu
);
1053 int last_cpupid
, this_cpupid
;
1055 this_cpupid
= cpu_pid_to_cpupid(dst_cpu
, current
->pid
);
1058 * Multi-stage node selection is used in conjunction with a periodic
1059 * migration fault to build a temporal task<->page relation. By using
1060 * a two-stage filter we remove short/unlikely relations.
1062 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1063 * a task's usage of a particular page (n_p) per total usage of this
1064 * page (n_t) (in a given time-span) to a probability.
1066 * Our periodic faults will sample this probability and getting the
1067 * same result twice in a row, given these samples are fully
1068 * independent, is then given by P(n)^2, provided our sample period
1069 * is sufficiently short compared to the usage pattern.
1071 * This quadric squishes small probabilities, making it less likely we
1072 * act on an unlikely task<->page relation.
1074 last_cpupid
= page_cpupid_xchg_last(page
, this_cpupid
);
1075 if (!cpupid_pid_unset(last_cpupid
) &&
1076 cpupid_to_nid(last_cpupid
) != dst_nid
)
1079 /* Always allow migrate on private faults */
1080 if (cpupid_match_pid(p
, last_cpupid
))
1083 /* A shared fault, but p->numa_group has not been set up yet. */
1088 * Do not migrate if the destination is not a node that
1089 * is actively used by this numa group.
1091 if (!node_isset(dst_nid
, ng
->active_nodes
))
1095 * Source is a node that is not actively used by this
1096 * numa group, while the destination is. Migrate.
1098 if (!node_isset(src_nid
, ng
->active_nodes
))
1102 * Both source and destination are nodes in active
1103 * use by this numa group. Maximize memory bandwidth
1104 * by migrating from more heavily used groups, to less
1105 * heavily used ones, spreading the load around.
1106 * Use a 1/4 hysteresis to avoid spurious page movement.
1108 return group_faults(p
, dst_nid
) < (group_faults(p
, src_nid
) * 3 / 4);
1111 static unsigned long weighted_cpuload(const int cpu
);
1112 static unsigned long source_load(int cpu
, int type
);
1113 static unsigned long target_load(int cpu
, int type
);
1114 static unsigned long capacity_of(int cpu
);
1115 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
);
1117 /* Cached statistics for all CPUs within a node */
1119 unsigned long nr_running
;
1122 /* Total compute capacity of CPUs on a node */
1123 unsigned long compute_capacity
;
1125 /* Approximate capacity in terms of runnable tasks on a node */
1126 unsigned long task_capacity
;
1127 int has_free_capacity
;
1131 * XXX borrowed from update_sg_lb_stats
1133 static void update_numa_stats(struct numa_stats
*ns
, int nid
)
1135 int smt
, cpu
, cpus
= 0;
1136 unsigned long capacity
;
1138 memset(ns
, 0, sizeof(*ns
));
1139 for_each_cpu(cpu
, cpumask_of_node(nid
)) {
1140 struct rq
*rq
= cpu_rq(cpu
);
1142 ns
->nr_running
+= rq
->nr_running
;
1143 ns
->load
+= weighted_cpuload(cpu
);
1144 ns
->compute_capacity
+= capacity_of(cpu
);
1150 * If we raced with hotplug and there are no CPUs left in our mask
1151 * the @ns structure is NULL'ed and task_numa_compare() will
1152 * not find this node attractive.
1154 * We'll either bail at !has_free_capacity, or we'll detect a huge
1155 * imbalance and bail there.
1160 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1161 smt
= DIV_ROUND_UP(SCHED_CAPACITY_SCALE
* cpus
, ns
->compute_capacity
);
1162 capacity
= cpus
/ smt
; /* cores */
1164 ns
->task_capacity
= min_t(unsigned, capacity
,
1165 DIV_ROUND_CLOSEST(ns
->compute_capacity
, SCHED_CAPACITY_SCALE
));
1166 ns
->has_free_capacity
= (ns
->nr_running
< ns
->task_capacity
);
1169 struct task_numa_env
{
1170 struct task_struct
*p
;
1172 int src_cpu
, src_nid
;
1173 int dst_cpu
, dst_nid
;
1175 struct numa_stats src_stats
, dst_stats
;
1180 struct task_struct
*best_task
;
1185 static void task_numa_assign(struct task_numa_env
*env
,
1186 struct task_struct
*p
, long imp
)
1189 put_task_struct(env
->best_task
);
1194 env
->best_imp
= imp
;
1195 env
->best_cpu
= env
->dst_cpu
;
1198 static bool load_too_imbalanced(long src_load
, long dst_load
,
1199 struct task_numa_env
*env
)
1201 long src_capacity
, dst_capacity
;
1203 long load_a
, load_b
;
1208 * The load is corrected for the CPU capacity available on each node.
1211 * ------------ vs ---------
1212 * src_capacity dst_capacity
1214 src_capacity
= env
->src_stats
.compute_capacity
;
1215 dst_capacity
= env
->dst_stats
.compute_capacity
;
1217 /* We care about the slope of the imbalance, not the direction. */
1220 if (load_a
< load_b
)
1221 swap(load_a
, load_b
);
1223 /* Is the difference below the threshold? */
1224 imb
= load_a
* src_capacity
* 100 -
1225 load_b
* dst_capacity
* env
->imbalance_pct
;
1230 * The imbalance is above the allowed threshold.
1231 * Allow a move that brings us closer to a balanced situation,
1232 * without moving things past the point of balance.
1234 orig_src_load
= env
->src_stats
.load
;
1237 * In a task swap, there will be one load moving from src to dst,
1238 * and another moving back. This is the net sum of both moves.
1239 * A simple task move will always have a positive value.
1240 * Allow the move if it brings the system closer to a balanced
1241 * situation, without crossing over the balance point.
1243 moved_load
= orig_src_load
- src_load
;
1246 /* Moving src -> dst. Did we overshoot balance? */
1247 return src_load
* dst_capacity
< dst_load
* src_capacity
;
1249 /* Moving dst -> src. Did we overshoot balance? */
1250 return dst_load
* src_capacity
< src_load
* dst_capacity
;
1254 * This checks if the overall compute and NUMA accesses of the system would
1255 * be improved if the source tasks was migrated to the target dst_cpu taking
1256 * into account that it might be best if task running on the dst_cpu should
1257 * be exchanged with the source task
1259 static void task_numa_compare(struct task_numa_env
*env
,
1260 long taskimp
, long groupimp
)
1262 struct rq
*src_rq
= cpu_rq(env
->src_cpu
);
1263 struct rq
*dst_rq
= cpu_rq(env
->dst_cpu
);
1264 struct task_struct
*cur
;
1265 long src_load
, dst_load
;
1267 long imp
= env
->p
->numa_group
? groupimp
: taskimp
;
1269 int dist
= env
->dist
;
1273 raw_spin_lock_irq(&dst_rq
->lock
);
1276 * No need to move the exiting task, and this ensures that ->curr
1277 * wasn't reaped and thus get_task_struct() in task_numa_assign()
1278 * is safe under RCU read lock.
1279 * Note that rcu_read_lock() itself can't protect from the final
1280 * put_task_struct() after the last schedule().
1282 if ((cur
->flags
& PF_EXITING
) || is_idle_task(cur
))
1284 raw_spin_unlock_irq(&dst_rq
->lock
);
1287 * Because we have preemption enabled we can get migrated around and
1288 * end try selecting ourselves (current == env->p) as a swap candidate.
1294 * "imp" is the fault differential for the source task between the
1295 * source and destination node. Calculate the total differential for
1296 * the source task and potential destination task. The more negative
1297 * the value is, the more rmeote accesses that would be expected to
1298 * be incurred if the tasks were swapped.
1301 /* Skip this swap candidate if cannot move to the source cpu */
1302 if (!cpumask_test_cpu(env
->src_cpu
, tsk_cpus_allowed(cur
)))
1306 * If dst and source tasks are in the same NUMA group, or not
1307 * in any group then look only at task weights.
1309 if (cur
->numa_group
== env
->p
->numa_group
) {
1310 imp
= taskimp
+ task_weight(cur
, env
->src_nid
, dist
) -
1311 task_weight(cur
, env
->dst_nid
, dist
);
1313 * Add some hysteresis to prevent swapping the
1314 * tasks within a group over tiny differences.
1316 if (cur
->numa_group
)
1320 * Compare the group weights. If a task is all by
1321 * itself (not part of a group), use the task weight
1324 if (cur
->numa_group
)
1325 imp
+= group_weight(cur
, env
->src_nid
, dist
) -
1326 group_weight(cur
, env
->dst_nid
, dist
);
1328 imp
+= task_weight(cur
, env
->src_nid
, dist
) -
1329 task_weight(cur
, env
->dst_nid
, dist
);
1333 if (imp
<= env
->best_imp
&& moveimp
<= env
->best_imp
)
1337 /* Is there capacity at our destination? */
1338 if (env
->src_stats
.nr_running
<= env
->src_stats
.task_capacity
&&
1339 !env
->dst_stats
.has_free_capacity
)
1345 /* Balance doesn't matter much if we're running a task per cpu */
1346 if (imp
> env
->best_imp
&& src_rq
->nr_running
== 1 &&
1347 dst_rq
->nr_running
== 1)
1351 * In the overloaded case, try and keep the load balanced.
1354 load
= task_h_load(env
->p
);
1355 dst_load
= env
->dst_stats
.load
+ load
;
1356 src_load
= env
->src_stats
.load
- load
;
1358 if (moveimp
> imp
&& moveimp
> env
->best_imp
) {
1360 * If the improvement from just moving env->p direction is
1361 * better than swapping tasks around, check if a move is
1362 * possible. Store a slightly smaller score than moveimp,
1363 * so an actually idle CPU will win.
1365 if (!load_too_imbalanced(src_load
, dst_load
, env
)) {
1372 if (imp
<= env
->best_imp
)
1376 load
= task_h_load(cur
);
1381 if (load_too_imbalanced(src_load
, dst_load
, env
))
1385 * One idle CPU per node is evaluated for a task numa move.
1386 * Call select_idle_sibling to maybe find a better one.
1389 env
->dst_cpu
= select_idle_sibling(env
->p
, env
->dst_cpu
);
1392 task_numa_assign(env
, cur
, imp
);
1397 static void task_numa_find_cpu(struct task_numa_env
*env
,
1398 long taskimp
, long groupimp
)
1402 for_each_cpu(cpu
, cpumask_of_node(env
->dst_nid
)) {
1403 /* Skip this CPU if the source task cannot migrate */
1404 if (!cpumask_test_cpu(cpu
, tsk_cpus_allowed(env
->p
)))
1408 task_numa_compare(env
, taskimp
, groupimp
);
1412 static int task_numa_migrate(struct task_struct
*p
)
1414 struct task_numa_env env
= {
1417 .src_cpu
= task_cpu(p
),
1418 .src_nid
= task_node(p
),
1420 .imbalance_pct
= 112,
1426 struct sched_domain
*sd
;
1427 unsigned long taskweight
, groupweight
;
1429 long taskimp
, groupimp
;
1432 * Pick the lowest SD_NUMA domain, as that would have the smallest
1433 * imbalance and would be the first to start moving tasks about.
1435 * And we want to avoid any moving of tasks about, as that would create
1436 * random movement of tasks -- counter the numa conditions we're trying
1440 sd
= rcu_dereference(per_cpu(sd_numa
, env
.src_cpu
));
1442 env
.imbalance_pct
= 100 + (sd
->imbalance_pct
- 100) / 2;
1446 * Cpusets can break the scheduler domain tree into smaller
1447 * balance domains, some of which do not cross NUMA boundaries.
1448 * Tasks that are "trapped" in such domains cannot be migrated
1449 * elsewhere, so there is no point in (re)trying.
1451 if (unlikely(!sd
)) {
1452 p
->numa_preferred_nid
= task_node(p
);
1456 env
.dst_nid
= p
->numa_preferred_nid
;
1457 dist
= env
.dist
= node_distance(env
.src_nid
, env
.dst_nid
);
1458 taskweight
= task_weight(p
, env
.src_nid
, dist
);
1459 groupweight
= group_weight(p
, env
.src_nid
, dist
);
1460 update_numa_stats(&env
.src_stats
, env
.src_nid
);
1461 taskimp
= task_weight(p
, env
.dst_nid
, dist
) - taskweight
;
1462 groupimp
= group_weight(p
, env
.dst_nid
, dist
) - groupweight
;
1463 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1465 /* Try to find a spot on the preferred nid. */
1466 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1469 * Look at other nodes in these cases:
1470 * - there is no space available on the preferred_nid
1471 * - the task is part of a numa_group that is interleaved across
1472 * multiple NUMA nodes; in order to better consolidate the group,
1473 * we need to check other locations.
1475 if (env
.best_cpu
== -1 || (p
->numa_group
&&
1476 nodes_weight(p
->numa_group
->active_nodes
) > 1)) {
1477 for_each_online_node(nid
) {
1478 if (nid
== env
.src_nid
|| nid
== p
->numa_preferred_nid
)
1481 dist
= node_distance(env
.src_nid
, env
.dst_nid
);
1482 if (sched_numa_topology_type
== NUMA_BACKPLANE
&&
1484 taskweight
= task_weight(p
, env
.src_nid
, dist
);
1485 groupweight
= group_weight(p
, env
.src_nid
, dist
);
1488 /* Only consider nodes where both task and groups benefit */
1489 taskimp
= task_weight(p
, nid
, dist
) - taskweight
;
1490 groupimp
= group_weight(p
, nid
, dist
) - groupweight
;
1491 if (taskimp
< 0 && groupimp
< 0)
1496 update_numa_stats(&env
.dst_stats
, env
.dst_nid
);
1497 task_numa_find_cpu(&env
, taskimp
, groupimp
);
1502 * If the task is part of a workload that spans multiple NUMA nodes,
1503 * and is migrating into one of the workload's active nodes, remember
1504 * this node as the task's preferred numa node, so the workload can
1506 * A task that migrated to a second choice node will be better off
1507 * trying for a better one later. Do not set the preferred node here.
1509 if (p
->numa_group
) {
1510 if (env
.best_cpu
== -1)
1515 if (node_isset(nid
, p
->numa_group
->active_nodes
))
1516 sched_setnuma(p
, env
.dst_nid
);
1519 /* No better CPU than the current one was found. */
1520 if (env
.best_cpu
== -1)
1524 * Reset the scan period if the task is being rescheduled on an
1525 * alternative node to recheck if the tasks is now properly placed.
1527 p
->numa_scan_period
= task_scan_min(p
);
1529 if (env
.best_task
== NULL
) {
1530 ret
= migrate_task_to(p
, env
.best_cpu
);
1532 trace_sched_stick_numa(p
, env
.src_cpu
, env
.best_cpu
);
1536 ret
= migrate_swap(p
, env
.best_task
);
1538 trace_sched_stick_numa(p
, env
.src_cpu
, task_cpu(env
.best_task
));
1539 put_task_struct(env
.best_task
);
1543 /* Attempt to migrate a task to a CPU on the preferred node. */
1544 static void numa_migrate_preferred(struct task_struct
*p
)
1546 unsigned long interval
= HZ
;
1548 /* This task has no NUMA fault statistics yet */
1549 if (unlikely(p
->numa_preferred_nid
== -1 || !p
->numa_faults
))
1552 /* Periodically retry migrating the task to the preferred node */
1553 interval
= min(interval
, msecs_to_jiffies(p
->numa_scan_period
) / 16);
1554 p
->numa_migrate_retry
= jiffies
+ interval
;
1556 /* Success if task is already running on preferred CPU */
1557 if (task_node(p
) == p
->numa_preferred_nid
)
1560 /* Otherwise, try migrate to a CPU on the preferred node */
1561 task_numa_migrate(p
);
1565 * Find the nodes on which the workload is actively running. We do this by
1566 * tracking the nodes from which NUMA hinting faults are triggered. This can
1567 * be different from the set of nodes where the workload's memory is currently
1570 * The bitmask is used to make smarter decisions on when to do NUMA page
1571 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1572 * are added when they cause over 6/16 of the maximum number of faults, but
1573 * only removed when they drop below 3/16.
1575 static void update_numa_active_node_mask(struct numa_group
*numa_group
)
1577 unsigned long faults
, max_faults
= 0;
1580 for_each_online_node(nid
) {
1581 faults
= group_faults_cpu(numa_group
, nid
);
1582 if (faults
> max_faults
)
1583 max_faults
= faults
;
1586 for_each_online_node(nid
) {
1587 faults
= group_faults_cpu(numa_group
, nid
);
1588 if (!node_isset(nid
, numa_group
->active_nodes
)) {
1589 if (faults
> max_faults
* 6 / 16)
1590 node_set(nid
, numa_group
->active_nodes
);
1591 } else if (faults
< max_faults
* 3 / 16)
1592 node_clear(nid
, numa_group
->active_nodes
);
1597 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1598 * increments. The more local the fault statistics are, the higher the scan
1599 * period will be for the next scan window. If local/(local+remote) ratio is
1600 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1601 * the scan period will decrease. Aim for 70% local accesses.
1603 #define NUMA_PERIOD_SLOTS 10
1604 #define NUMA_PERIOD_THRESHOLD 7
1607 * Increase the scan period (slow down scanning) if the majority of
1608 * our memory is already on our local node, or if the majority of
1609 * the page accesses are shared with other processes.
1610 * Otherwise, decrease the scan period.
1612 static void update_task_scan_period(struct task_struct
*p
,
1613 unsigned long shared
, unsigned long private)
1615 unsigned int period_slot
;
1619 unsigned long remote
= p
->numa_faults_locality
[0];
1620 unsigned long local
= p
->numa_faults_locality
[1];
1623 * If there were no record hinting faults then either the task is
1624 * completely idle or all activity is areas that are not of interest
1625 * to automatic numa balancing. Scan slower
1627 if (local
+ shared
== 0) {
1628 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1629 p
->numa_scan_period
<< 1);
1631 p
->mm
->numa_next_scan
= jiffies
+
1632 msecs_to_jiffies(p
->numa_scan_period
);
1638 * Prepare to scale scan period relative to the current period.
1639 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1640 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1641 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1643 period_slot
= DIV_ROUND_UP(p
->numa_scan_period
, NUMA_PERIOD_SLOTS
);
1644 ratio
= (local
* NUMA_PERIOD_SLOTS
) / (local
+ remote
);
1645 if (ratio
>= NUMA_PERIOD_THRESHOLD
) {
1646 int slot
= ratio
- NUMA_PERIOD_THRESHOLD
;
1649 diff
= slot
* period_slot
;
1651 diff
= -(NUMA_PERIOD_THRESHOLD
- ratio
) * period_slot
;
1654 * Scale scan rate increases based on sharing. There is an
1655 * inverse relationship between the degree of sharing and
1656 * the adjustment made to the scanning period. Broadly
1657 * speaking the intent is that there is little point
1658 * scanning faster if shared accesses dominate as it may
1659 * simply bounce migrations uselessly
1661 ratio
= DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS
, (private + shared
+ 1));
1662 diff
= (diff
* ratio
) / NUMA_PERIOD_SLOTS
;
1665 p
->numa_scan_period
= clamp(p
->numa_scan_period
+ diff
,
1666 task_scan_min(p
), task_scan_max(p
));
1667 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
1671 * Get the fraction of time the task has been running since the last
1672 * NUMA placement cycle. The scheduler keeps similar statistics, but
1673 * decays those on a 32ms period, which is orders of magnitude off
1674 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1675 * stats only if the task is so new there are no NUMA statistics yet.
1677 static u64
numa_get_avg_runtime(struct task_struct
*p
, u64
*period
)
1679 u64 runtime
, delta
, now
;
1680 /* Use the start of this time slice to avoid calculations. */
1681 now
= p
->se
.exec_start
;
1682 runtime
= p
->se
.sum_exec_runtime
;
1684 if (p
->last_task_numa_placement
) {
1685 delta
= runtime
- p
->last_sum_exec_runtime
;
1686 *period
= now
- p
->last_task_numa_placement
;
1688 delta
= p
->se
.avg
.runnable_avg_sum
;
1689 *period
= p
->se
.avg
.avg_period
;
1692 p
->last_sum_exec_runtime
= runtime
;
1693 p
->last_task_numa_placement
= now
;
1699 * Determine the preferred nid for a task in a numa_group. This needs to
1700 * be done in a way that produces consistent results with group_weight,
1701 * otherwise workloads might not converge.
1703 static int preferred_group_nid(struct task_struct
*p
, int nid
)
1708 /* Direct connections between all NUMA nodes. */
1709 if (sched_numa_topology_type
== NUMA_DIRECT
)
1713 * On a system with glueless mesh NUMA topology, group_weight
1714 * scores nodes according to the number of NUMA hinting faults on
1715 * both the node itself, and on nearby nodes.
1717 if (sched_numa_topology_type
== NUMA_GLUELESS_MESH
) {
1718 unsigned long score
, max_score
= 0;
1719 int node
, max_node
= nid
;
1721 dist
= sched_max_numa_distance
;
1723 for_each_online_node(node
) {
1724 score
= group_weight(p
, node
, dist
);
1725 if (score
> max_score
) {
1734 * Finding the preferred nid in a system with NUMA backplane
1735 * interconnect topology is more involved. The goal is to locate
1736 * tasks from numa_groups near each other in the system, and
1737 * untangle workloads from different sides of the system. This requires
1738 * searching down the hierarchy of node groups, recursively searching
1739 * inside the highest scoring group of nodes. The nodemask tricks
1740 * keep the complexity of the search down.
1742 nodes
= node_online_map
;
1743 for (dist
= sched_max_numa_distance
; dist
> LOCAL_DISTANCE
; dist
--) {
1744 unsigned long max_faults
= 0;
1745 nodemask_t max_group
= NODE_MASK_NONE
;
1748 /* Are there nodes at this distance from each other? */
1749 if (!find_numa_distance(dist
))
1752 for_each_node_mask(a
, nodes
) {
1753 unsigned long faults
= 0;
1754 nodemask_t this_group
;
1755 nodes_clear(this_group
);
1757 /* Sum group's NUMA faults; includes a==b case. */
1758 for_each_node_mask(b
, nodes
) {
1759 if (node_distance(a
, b
) < dist
) {
1760 faults
+= group_faults(p
, b
);
1761 node_set(b
, this_group
);
1762 node_clear(b
, nodes
);
1766 /* Remember the top group. */
1767 if (faults
> max_faults
) {
1768 max_faults
= faults
;
1769 max_group
= this_group
;
1771 * subtle: at the smallest distance there is
1772 * just one node left in each "group", the
1773 * winner is the preferred nid.
1778 /* Next round, evaluate the nodes within max_group. */
1786 static void task_numa_placement(struct task_struct
*p
)
1788 int seq
, nid
, max_nid
= -1, max_group_nid
= -1;
1789 unsigned long max_faults
= 0, max_group_faults
= 0;
1790 unsigned long fault_types
[2] = { 0, 0 };
1791 unsigned long total_faults
;
1792 u64 runtime
, period
;
1793 spinlock_t
*group_lock
= NULL
;
1795 seq
= ACCESS_ONCE(p
->mm
->numa_scan_seq
);
1796 if (p
->numa_scan_seq
== seq
)
1798 p
->numa_scan_seq
= seq
;
1799 p
->numa_scan_period_max
= task_scan_max(p
);
1801 total_faults
= p
->numa_faults_locality
[0] +
1802 p
->numa_faults_locality
[1];
1803 runtime
= numa_get_avg_runtime(p
, &period
);
1805 /* If the task is part of a group prevent parallel updates to group stats */
1806 if (p
->numa_group
) {
1807 group_lock
= &p
->numa_group
->lock
;
1808 spin_lock_irq(group_lock
);
1811 /* Find the node with the highest number of faults */
1812 for_each_online_node(nid
) {
1813 /* Keep track of the offsets in numa_faults array */
1814 int mem_idx
, membuf_idx
, cpu_idx
, cpubuf_idx
;
1815 unsigned long faults
= 0, group_faults
= 0;
1818 for (priv
= 0; priv
< NR_NUMA_HINT_FAULT_TYPES
; priv
++) {
1819 long diff
, f_diff
, f_weight
;
1821 mem_idx
= task_faults_idx(NUMA_MEM
, nid
, priv
);
1822 membuf_idx
= task_faults_idx(NUMA_MEMBUF
, nid
, priv
);
1823 cpu_idx
= task_faults_idx(NUMA_CPU
, nid
, priv
);
1824 cpubuf_idx
= task_faults_idx(NUMA_CPUBUF
, nid
, priv
);
1826 /* Decay existing window, copy faults since last scan */
1827 diff
= p
->numa_faults
[membuf_idx
] - p
->numa_faults
[mem_idx
] / 2;
1828 fault_types
[priv
] += p
->numa_faults
[membuf_idx
];
1829 p
->numa_faults
[membuf_idx
] = 0;
1832 * Normalize the faults_from, so all tasks in a group
1833 * count according to CPU use, instead of by the raw
1834 * number of faults. Tasks with little runtime have
1835 * little over-all impact on throughput, and thus their
1836 * faults are less important.
1838 f_weight
= div64_u64(runtime
<< 16, period
+ 1);
1839 f_weight
= (f_weight
* p
->numa_faults
[cpubuf_idx
]) /
1841 f_diff
= f_weight
- p
->numa_faults
[cpu_idx
] / 2;
1842 p
->numa_faults
[cpubuf_idx
] = 0;
1844 p
->numa_faults
[mem_idx
] += diff
;
1845 p
->numa_faults
[cpu_idx
] += f_diff
;
1846 faults
+= p
->numa_faults
[mem_idx
];
1847 p
->total_numa_faults
+= diff
;
1848 if (p
->numa_group
) {
1850 * safe because we can only change our own group
1852 * mem_idx represents the offset for a given
1853 * nid and priv in a specific region because it
1854 * is at the beginning of the numa_faults array.
1856 p
->numa_group
->faults
[mem_idx
] += diff
;
1857 p
->numa_group
->faults_cpu
[mem_idx
] += f_diff
;
1858 p
->numa_group
->total_faults
+= diff
;
1859 group_faults
+= p
->numa_group
->faults
[mem_idx
];
1863 if (faults
> max_faults
) {
1864 max_faults
= faults
;
1868 if (group_faults
> max_group_faults
) {
1869 max_group_faults
= group_faults
;
1870 max_group_nid
= nid
;
1874 update_task_scan_period(p
, fault_types
[0], fault_types
[1]);
1876 if (p
->numa_group
) {
1877 update_numa_active_node_mask(p
->numa_group
);
1878 spin_unlock_irq(group_lock
);
1879 max_nid
= preferred_group_nid(p
, max_group_nid
);
1883 /* Set the new preferred node */
1884 if (max_nid
!= p
->numa_preferred_nid
)
1885 sched_setnuma(p
, max_nid
);
1887 if (task_node(p
) != p
->numa_preferred_nid
)
1888 numa_migrate_preferred(p
);
1892 static inline int get_numa_group(struct numa_group
*grp
)
1894 return atomic_inc_not_zero(&grp
->refcount
);
1897 static inline void put_numa_group(struct numa_group
*grp
)
1899 if (atomic_dec_and_test(&grp
->refcount
))
1900 kfree_rcu(grp
, rcu
);
1903 static void task_numa_group(struct task_struct
*p
, int cpupid
, int flags
,
1906 struct numa_group
*grp
, *my_grp
;
1907 struct task_struct
*tsk
;
1909 int cpu
= cpupid_to_cpu(cpupid
);
1912 if (unlikely(!p
->numa_group
)) {
1913 unsigned int size
= sizeof(struct numa_group
) +
1914 4*nr_node_ids
*sizeof(unsigned long);
1916 grp
= kzalloc(size
, GFP_KERNEL
| __GFP_NOWARN
);
1920 atomic_set(&grp
->refcount
, 1);
1921 spin_lock_init(&grp
->lock
);
1923 /* Second half of the array tracks nids where faults happen */
1924 grp
->faults_cpu
= grp
->faults
+ NR_NUMA_HINT_FAULT_TYPES
*
1927 node_set(task_node(current
), grp
->active_nodes
);
1929 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
1930 grp
->faults
[i
] = p
->numa_faults
[i
];
1932 grp
->total_faults
= p
->total_numa_faults
;
1935 rcu_assign_pointer(p
->numa_group
, grp
);
1939 tsk
= ACCESS_ONCE(cpu_rq(cpu
)->curr
);
1941 if (!cpupid_match_pid(tsk
, cpupid
))
1944 grp
= rcu_dereference(tsk
->numa_group
);
1948 my_grp
= p
->numa_group
;
1953 * Only join the other group if its bigger; if we're the bigger group,
1954 * the other task will join us.
1956 if (my_grp
->nr_tasks
> grp
->nr_tasks
)
1960 * Tie-break on the grp address.
1962 if (my_grp
->nr_tasks
== grp
->nr_tasks
&& my_grp
> grp
)
1965 /* Always join threads in the same process. */
1966 if (tsk
->mm
== current
->mm
)
1969 /* Simple filter to avoid false positives due to PID collisions */
1970 if (flags
& TNF_SHARED
)
1973 /* Update priv based on whether false sharing was detected */
1976 if (join
&& !get_numa_group(grp
))
1984 BUG_ON(irqs_disabled());
1985 double_lock_irq(&my_grp
->lock
, &grp
->lock
);
1987 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++) {
1988 my_grp
->faults
[i
] -= p
->numa_faults
[i
];
1989 grp
->faults
[i
] += p
->numa_faults
[i
];
1991 my_grp
->total_faults
-= p
->total_numa_faults
;
1992 grp
->total_faults
+= p
->total_numa_faults
;
1997 spin_unlock(&my_grp
->lock
);
1998 spin_unlock_irq(&grp
->lock
);
2000 rcu_assign_pointer(p
->numa_group
, grp
);
2002 put_numa_group(my_grp
);
2010 void task_numa_free(struct task_struct
*p
)
2012 struct numa_group
*grp
= p
->numa_group
;
2013 void *numa_faults
= p
->numa_faults
;
2014 unsigned long flags
;
2018 spin_lock_irqsave(&grp
->lock
, flags
);
2019 for (i
= 0; i
< NR_NUMA_HINT_FAULT_STATS
* nr_node_ids
; i
++)
2020 grp
->faults
[i
] -= p
->numa_faults
[i
];
2021 grp
->total_faults
-= p
->total_numa_faults
;
2024 spin_unlock_irqrestore(&grp
->lock
, flags
);
2025 RCU_INIT_POINTER(p
->numa_group
, NULL
);
2026 put_numa_group(grp
);
2029 p
->numa_faults
= NULL
;
2034 * Got a PROT_NONE fault for a page on @node.
2036 void task_numa_fault(int last_cpupid
, int mem_node
, int pages
, int flags
)
2038 struct task_struct
*p
= current
;
2039 bool migrated
= flags
& TNF_MIGRATED
;
2040 int cpu_node
= task_node(current
);
2041 int local
= !!(flags
& TNF_FAULT_LOCAL
);
2044 if (!numabalancing_enabled
)
2047 /* for example, ksmd faulting in a user's mm */
2051 /* Allocate buffer to track faults on a per-node basis */
2052 if (unlikely(!p
->numa_faults
)) {
2053 int size
= sizeof(*p
->numa_faults
) *
2054 NR_NUMA_HINT_FAULT_BUCKETS
* nr_node_ids
;
2056 p
->numa_faults
= kzalloc(size
, GFP_KERNEL
|__GFP_NOWARN
);
2057 if (!p
->numa_faults
)
2060 p
->total_numa_faults
= 0;
2061 memset(p
->numa_faults_locality
, 0, sizeof(p
->numa_faults_locality
));
2065 * First accesses are treated as private, otherwise consider accesses
2066 * to be private if the accessing pid has not changed
2068 if (unlikely(last_cpupid
== (-1 & LAST_CPUPID_MASK
))) {
2071 priv
= cpupid_match_pid(p
, last_cpupid
);
2072 if (!priv
&& !(flags
& TNF_NO_GROUP
))
2073 task_numa_group(p
, last_cpupid
, flags
, &priv
);
2077 * If a workload spans multiple NUMA nodes, a shared fault that
2078 * occurs wholly within the set of nodes that the workload is
2079 * actively using should be counted as local. This allows the
2080 * scan rate to slow down when a workload has settled down.
2082 if (!priv
&& !local
&& p
->numa_group
&&
2083 node_isset(cpu_node
, p
->numa_group
->active_nodes
) &&
2084 node_isset(mem_node
, p
->numa_group
->active_nodes
))
2087 task_numa_placement(p
);
2090 * Retry task to preferred node migration periodically, in case it
2091 * case it previously failed, or the scheduler moved us.
2093 if (time_after(jiffies
, p
->numa_migrate_retry
))
2094 numa_migrate_preferred(p
);
2097 p
->numa_pages_migrated
+= pages
;
2099 p
->numa_faults
[task_faults_idx(NUMA_MEMBUF
, mem_node
, priv
)] += pages
;
2100 p
->numa_faults
[task_faults_idx(NUMA_CPUBUF
, cpu_node
, priv
)] += pages
;
2101 p
->numa_faults_locality
[local
] += pages
;
2104 static void reset_ptenuma_scan(struct task_struct
*p
)
2106 ACCESS_ONCE(p
->mm
->numa_scan_seq
)++;
2107 p
->mm
->numa_scan_offset
= 0;
2111 * The expensive part of numa migration is done from task_work context.
2112 * Triggered from task_tick_numa().
2114 void task_numa_work(struct callback_head
*work
)
2116 unsigned long migrate
, next_scan
, now
= jiffies
;
2117 struct task_struct
*p
= current
;
2118 struct mm_struct
*mm
= p
->mm
;
2119 struct vm_area_struct
*vma
;
2120 unsigned long start
, end
;
2121 unsigned long nr_pte_updates
= 0;
2124 WARN_ON_ONCE(p
!= container_of(work
, struct task_struct
, numa_work
));
2126 work
->next
= work
; /* protect against double add */
2128 * Who cares about NUMA placement when they're dying.
2130 * NOTE: make sure not to dereference p->mm before this check,
2131 * exit_task_work() happens _after_ exit_mm() so we could be called
2132 * without p->mm even though we still had it when we enqueued this
2135 if (p
->flags
& PF_EXITING
)
2138 if (!mm
->numa_next_scan
) {
2139 mm
->numa_next_scan
= now
+
2140 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2144 * Enforce maximal scan/migration frequency..
2146 migrate
= mm
->numa_next_scan
;
2147 if (time_before(now
, migrate
))
2150 if (p
->numa_scan_period
== 0) {
2151 p
->numa_scan_period_max
= task_scan_max(p
);
2152 p
->numa_scan_period
= task_scan_min(p
);
2155 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
2156 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
2160 * Delay this task enough that another task of this mm will likely win
2161 * the next time around.
2163 p
->node_stamp
+= 2 * TICK_NSEC
;
2165 start
= mm
->numa_scan_offset
;
2166 pages
= sysctl_numa_balancing_scan_size
;
2167 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
2171 down_read(&mm
->mmap_sem
);
2172 vma
= find_vma(mm
, start
);
2174 reset_ptenuma_scan(p
);
2178 for (; vma
; vma
= vma
->vm_next
) {
2179 if (!vma_migratable(vma
) || !vma_policy_mof(vma
))
2183 * Shared library pages mapped by multiple processes are not
2184 * migrated as it is expected they are cache replicated. Avoid
2185 * hinting faults in read-only file-backed mappings or the vdso
2186 * as migrating the pages will be of marginal benefit.
2189 (vma
->vm_file
&& (vma
->vm_flags
& (VM_READ
|VM_WRITE
)) == (VM_READ
)))
2193 * Skip inaccessible VMAs to avoid any confusion between
2194 * PROT_NONE and NUMA hinting ptes
2196 if (!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)))
2200 start
= max(start
, vma
->vm_start
);
2201 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
2202 end
= min(end
, vma
->vm_end
);
2203 nr_pte_updates
+= change_prot_numa(vma
, start
, end
);
2206 * Scan sysctl_numa_balancing_scan_size but ensure that
2207 * at least one PTE is updated so that unused virtual
2208 * address space is quickly skipped.
2211 pages
-= (end
- start
) >> PAGE_SHIFT
;
2218 } while (end
!= vma
->vm_end
);
2223 * It is possible to reach the end of the VMA list but the last few
2224 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2225 * would find the !migratable VMA on the next scan but not reset the
2226 * scanner to the start so check it now.
2229 mm
->numa_scan_offset
= start
;
2231 reset_ptenuma_scan(p
);
2232 up_read(&mm
->mmap_sem
);
2236 * Drive the periodic memory faults..
2238 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2240 struct callback_head
*work
= &curr
->numa_work
;
2244 * We don't care about NUMA placement if we don't have memory.
2246 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
2250 * Using runtime rather than walltime has the dual advantage that
2251 * we (mostly) drive the selection from busy threads and that the
2252 * task needs to have done some actual work before we bother with
2255 now
= curr
->se
.sum_exec_runtime
;
2256 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
2258 if (now
- curr
->node_stamp
> period
) {
2259 if (!curr
->node_stamp
)
2260 curr
->numa_scan_period
= task_scan_min(curr
);
2261 curr
->node_stamp
+= period
;
2263 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
2264 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
2265 task_work_add(curr
, work
, true);
2270 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
2274 static inline void account_numa_enqueue(struct rq
*rq
, struct task_struct
*p
)
2278 static inline void account_numa_dequeue(struct rq
*rq
, struct task_struct
*p
)
2281 #endif /* CONFIG_NUMA_BALANCING */
2284 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2286 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
2287 if (!parent_entity(se
))
2288 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2290 if (entity_is_task(se
)) {
2291 struct rq
*rq
= rq_of(cfs_rq
);
2293 account_numa_enqueue(rq
, task_of(se
));
2294 list_add(&se
->group_node
, &rq
->cfs_tasks
);
2297 cfs_rq
->nr_running
++;
2301 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2303 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
2304 if (!parent_entity(se
))
2305 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
2306 if (entity_is_task(se
)) {
2307 account_numa_dequeue(rq_of(cfs_rq
), task_of(se
));
2308 list_del_init(&se
->group_node
);
2310 cfs_rq
->nr_running
--;
2313 #ifdef CONFIG_FAIR_GROUP_SCHED
2315 static inline long calc_tg_weight(struct task_group
*tg
, struct cfs_rq
*cfs_rq
)
2320 * Use this CPU's actual weight instead of the last load_contribution
2321 * to gain a more accurate current total weight. See
2322 * update_cfs_rq_load_contribution().
2324 tg_weight
= atomic_long_read(&tg
->load_avg
);
2325 tg_weight
-= cfs_rq
->tg_load_contrib
;
2326 tg_weight
+= cfs_rq
->load
.weight
;
2331 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2333 long tg_weight
, load
, shares
;
2335 tg_weight
= calc_tg_weight(tg
, cfs_rq
);
2336 load
= cfs_rq
->load
.weight
;
2338 shares
= (tg
->shares
* load
);
2340 shares
/= tg_weight
;
2342 if (shares
< MIN_SHARES
)
2343 shares
= MIN_SHARES
;
2344 if (shares
> tg
->shares
)
2345 shares
= tg
->shares
;
2349 # else /* CONFIG_SMP */
2350 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
2354 # endif /* CONFIG_SMP */
2355 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
2356 unsigned long weight
)
2359 /* commit outstanding execution time */
2360 if (cfs_rq
->curr
== se
)
2361 update_curr(cfs_rq
);
2362 account_entity_dequeue(cfs_rq
, se
);
2365 update_load_set(&se
->load
, weight
);
2368 account_entity_enqueue(cfs_rq
, se
);
2371 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
2373 static void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2375 struct task_group
*tg
;
2376 struct sched_entity
*se
;
2380 se
= tg
->se
[cpu_of(rq_of(cfs_rq
))];
2381 if (!se
|| throttled_hierarchy(cfs_rq
))
2384 if (likely(se
->load
.weight
== tg
->shares
))
2387 shares
= calc_cfs_shares(cfs_rq
, tg
);
2389 reweight_entity(cfs_rq_of(se
), se
, shares
);
2391 #else /* CONFIG_FAIR_GROUP_SCHED */
2392 static inline void update_cfs_shares(struct cfs_rq
*cfs_rq
)
2395 #endif /* CONFIG_FAIR_GROUP_SCHED */
2399 * We choose a half-life close to 1 scheduling period.
2400 * Note: The tables below are dependent on this value.
2402 #define LOAD_AVG_PERIOD 32
2403 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2404 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2406 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2407 static const u32 runnable_avg_yN_inv
[] = {
2408 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2409 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2410 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2411 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2412 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2413 0x85aac367, 0x82cd8698,
2417 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2418 * over-estimates when re-combining.
2420 static const u32 runnable_avg_yN_sum
[] = {
2421 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2422 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2423 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2428 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2430 static __always_inline u64
decay_load(u64 val
, u64 n
)
2432 unsigned int local_n
;
2436 else if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
2439 /* after bounds checking we can collapse to 32-bit */
2443 * As y^PERIOD = 1/2, we can combine
2444 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2445 * With a look-up table which covers y^n (n<PERIOD)
2447 * To achieve constant time decay_load.
2449 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
2450 val
>>= local_n
/ LOAD_AVG_PERIOD
;
2451 local_n
%= LOAD_AVG_PERIOD
;
2454 val
*= runnable_avg_yN_inv
[local_n
];
2455 /* We don't use SRR here since we always want to round down. */
2460 * For updates fully spanning n periods, the contribution to runnable
2461 * average will be: \Sum 1024*y^n
2463 * We can compute this reasonably efficiently by combining:
2464 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2466 static u32
__compute_runnable_contrib(u64 n
)
2470 if (likely(n
<= LOAD_AVG_PERIOD
))
2471 return runnable_avg_yN_sum
[n
];
2472 else if (unlikely(n
>= LOAD_AVG_MAX_N
))
2473 return LOAD_AVG_MAX
;
2475 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2477 contrib
/= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2478 contrib
+= runnable_avg_yN_sum
[LOAD_AVG_PERIOD
];
2480 n
-= LOAD_AVG_PERIOD
;
2481 } while (n
> LOAD_AVG_PERIOD
);
2483 contrib
= decay_load(contrib
, n
);
2484 return contrib
+ runnable_avg_yN_sum
[n
];
2487 unsigned long __weak
arch_scale_freq_capacity(struct sched_domain
*sd
, int cpu
);
2490 * We can represent the historical contribution to runnable average as the
2491 * coefficients of a geometric series. To do this we sub-divide our runnable
2492 * history into segments of approximately 1ms (1024us); label the segment that
2493 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2495 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2497 * (now) (~1ms ago) (~2ms ago)
2499 * Let u_i denote the fraction of p_i that the entity was runnable.
2501 * We then designate the fractions u_i as our co-efficients, yielding the
2502 * following representation of historical load:
2503 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2505 * We choose y based on the with of a reasonably scheduling period, fixing:
2508 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2509 * approximately half as much as the contribution to load within the last ms
2512 * When a period "rolls over" and we have new u_0`, multiplying the previous
2513 * sum again by y is sufficient to update:
2514 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2515 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2517 static __always_inline
int __update_entity_runnable_avg(u64 now
, int cpu
,
2518 struct sched_avg
*sa
,
2523 u32 runnable_contrib
;
2524 int delta_w
, decayed
= 0;
2525 unsigned long scale_freq
= arch_scale_freq_capacity(NULL
, cpu
);
2527 delta
= now
- sa
->last_runnable_update
;
2529 * This should only happen when time goes backwards, which it
2530 * unfortunately does during sched clock init when we swap over to TSC.
2532 if ((s64
)delta
< 0) {
2533 sa
->last_runnable_update
= now
;
2538 * Use 1024ns as the unit of measurement since it's a reasonable
2539 * approximation of 1us and fast to compute.
2544 sa
->last_runnable_update
= now
;
2546 /* delta_w is the amount already accumulated against our next period */
2547 delta_w
= sa
->avg_period
% 1024;
2548 if (delta
+ delta_w
>= 1024) {
2549 /* period roll-over */
2553 * Now that we know we're crossing a period boundary, figure
2554 * out how much from delta we need to complete the current
2555 * period and accrue it.
2557 delta_w
= 1024 - delta_w
;
2559 sa
->runnable_avg_sum
+= delta_w
;
2561 sa
->running_avg_sum
+= delta_w
* scale_freq
2562 >> SCHED_CAPACITY_SHIFT
;
2563 sa
->avg_period
+= delta_w
;
2567 /* Figure out how many additional periods this update spans */
2568 periods
= delta
/ 1024;
2571 sa
->runnable_avg_sum
= decay_load(sa
->runnable_avg_sum
,
2573 sa
->running_avg_sum
= decay_load(sa
->running_avg_sum
,
2575 sa
->avg_period
= decay_load(sa
->avg_period
,
2578 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2579 runnable_contrib
= __compute_runnable_contrib(periods
);
2581 sa
->runnable_avg_sum
+= runnable_contrib
;
2583 sa
->running_avg_sum
+= runnable_contrib
* scale_freq
2584 >> SCHED_CAPACITY_SHIFT
;
2585 sa
->avg_period
+= runnable_contrib
;
2588 /* Remainder of delta accrued against u_0` */
2590 sa
->runnable_avg_sum
+= delta
;
2592 sa
->running_avg_sum
+= delta
* scale_freq
2593 >> SCHED_CAPACITY_SHIFT
;
2594 sa
->avg_period
+= delta
;
2599 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2600 static inline u64
__synchronize_entity_decay(struct sched_entity
*se
)
2602 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2603 u64 decays
= atomic64_read(&cfs_rq
->decay_counter
);
2605 decays
-= se
->avg
.decay_count
;
2606 se
->avg
.decay_count
= 0;
2610 se
->avg
.load_avg_contrib
= decay_load(se
->avg
.load_avg_contrib
, decays
);
2611 se
->avg
.utilization_avg_contrib
=
2612 decay_load(se
->avg
.utilization_avg_contrib
, decays
);
2617 #ifdef CONFIG_FAIR_GROUP_SCHED
2618 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2621 struct task_group
*tg
= cfs_rq
->tg
;
2624 tg_contrib
= cfs_rq
->runnable_load_avg
+ cfs_rq
->blocked_load_avg
;
2625 tg_contrib
-= cfs_rq
->tg_load_contrib
;
2630 if (force_update
|| abs(tg_contrib
) > cfs_rq
->tg_load_contrib
/ 8) {
2631 atomic_long_add(tg_contrib
, &tg
->load_avg
);
2632 cfs_rq
->tg_load_contrib
+= tg_contrib
;
2637 * Aggregate cfs_rq runnable averages into an equivalent task_group
2638 * representation for computing load contributions.
2640 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2641 struct cfs_rq
*cfs_rq
)
2643 struct task_group
*tg
= cfs_rq
->tg
;
2646 /* The fraction of a cpu used by this cfs_rq */
2647 contrib
= div_u64((u64
)sa
->runnable_avg_sum
<< NICE_0_SHIFT
,
2648 sa
->avg_period
+ 1);
2649 contrib
-= cfs_rq
->tg_runnable_contrib
;
2651 if (abs(contrib
) > cfs_rq
->tg_runnable_contrib
/ 64) {
2652 atomic_add(contrib
, &tg
->runnable_avg
);
2653 cfs_rq
->tg_runnable_contrib
+= contrib
;
2657 static inline void __update_group_entity_contrib(struct sched_entity
*se
)
2659 struct cfs_rq
*cfs_rq
= group_cfs_rq(se
);
2660 struct task_group
*tg
= cfs_rq
->tg
;
2665 contrib
= cfs_rq
->tg_load_contrib
* tg
->shares
;
2666 se
->avg
.load_avg_contrib
= div_u64(contrib
,
2667 atomic_long_read(&tg
->load_avg
) + 1);
2670 * For group entities we need to compute a correction term in the case
2671 * that they are consuming <1 cpu so that we would contribute the same
2672 * load as a task of equal weight.
2674 * Explicitly co-ordinating this measurement would be expensive, but
2675 * fortunately the sum of each cpus contribution forms a usable
2676 * lower-bound on the true value.
2678 * Consider the aggregate of 2 contributions. Either they are disjoint
2679 * (and the sum represents true value) or they are disjoint and we are
2680 * understating by the aggregate of their overlap.
2682 * Extending this to N cpus, for a given overlap, the maximum amount we
2683 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2684 * cpus that overlap for this interval and w_i is the interval width.
2686 * On a small machine; the first term is well-bounded which bounds the
2687 * total error since w_i is a subset of the period. Whereas on a
2688 * larger machine, while this first term can be larger, if w_i is the
2689 * of consequential size guaranteed to see n_i*w_i quickly converge to
2690 * our upper bound of 1-cpu.
2692 runnable_avg
= atomic_read(&tg
->runnable_avg
);
2693 if (runnable_avg
< NICE_0_LOAD
) {
2694 se
->avg
.load_avg_contrib
*= runnable_avg
;
2695 se
->avg
.load_avg_contrib
>>= NICE_0_SHIFT
;
2699 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
)
2701 __update_entity_runnable_avg(rq_clock_task(rq
), cpu_of(rq
), &rq
->avg
,
2702 runnable
, runnable
);
2703 __update_tg_runnable_avg(&rq
->avg
, &rq
->cfs
);
2705 #else /* CONFIG_FAIR_GROUP_SCHED */
2706 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
2707 int force_update
) {}
2708 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
2709 struct cfs_rq
*cfs_rq
) {}
2710 static inline void __update_group_entity_contrib(struct sched_entity
*se
) {}
2711 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2712 #endif /* CONFIG_FAIR_GROUP_SCHED */
2714 static inline void __update_task_entity_contrib(struct sched_entity
*se
)
2718 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2719 contrib
= se
->avg
.runnable_avg_sum
* scale_load_down(se
->load
.weight
);
2720 contrib
/= (se
->avg
.avg_period
+ 1);
2721 se
->avg
.load_avg_contrib
= scale_load(contrib
);
2724 /* Compute the current contribution to load_avg by se, return any delta */
2725 static long __update_entity_load_avg_contrib(struct sched_entity
*se
)
2727 long old_contrib
= se
->avg
.load_avg_contrib
;
2729 if (entity_is_task(se
)) {
2730 __update_task_entity_contrib(se
);
2732 __update_tg_runnable_avg(&se
->avg
, group_cfs_rq(se
));
2733 __update_group_entity_contrib(se
);
2736 return se
->avg
.load_avg_contrib
- old_contrib
;
2740 static inline void __update_task_entity_utilization(struct sched_entity
*se
)
2744 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2745 contrib
= se
->avg
.running_avg_sum
* scale_load_down(SCHED_LOAD_SCALE
);
2746 contrib
/= (se
->avg
.avg_period
+ 1);
2747 se
->avg
.utilization_avg_contrib
= scale_load(contrib
);
2750 static long __update_entity_utilization_avg_contrib(struct sched_entity
*se
)
2752 long old_contrib
= se
->avg
.utilization_avg_contrib
;
2754 if (entity_is_task(se
))
2755 __update_task_entity_utilization(se
);
2757 se
->avg
.utilization_avg_contrib
=
2758 group_cfs_rq(se
)->utilization_load_avg
;
2760 return se
->avg
.utilization_avg_contrib
- old_contrib
;
2763 static inline void subtract_blocked_load_contrib(struct cfs_rq
*cfs_rq
,
2766 if (likely(load_contrib
< cfs_rq
->blocked_load_avg
))
2767 cfs_rq
->blocked_load_avg
-= load_contrib
;
2769 cfs_rq
->blocked_load_avg
= 0;
2772 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
2774 /* Update a sched_entity's runnable average */
2775 static inline void update_entity_load_avg(struct sched_entity
*se
,
2778 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2779 long contrib_delta
, utilization_delta
;
2780 int cpu
= cpu_of(rq_of(cfs_rq
));
2784 * For a group entity we need to use their owned cfs_rq_clock_task() in
2785 * case they are the parent of a throttled hierarchy.
2787 if (entity_is_task(se
))
2788 now
= cfs_rq_clock_task(cfs_rq
);
2790 now
= cfs_rq_clock_task(group_cfs_rq(se
));
2792 if (!__update_entity_runnable_avg(now
, cpu
, &se
->avg
, se
->on_rq
,
2793 cfs_rq
->curr
== se
))
2796 contrib_delta
= __update_entity_load_avg_contrib(se
);
2797 utilization_delta
= __update_entity_utilization_avg_contrib(se
);
2803 cfs_rq
->runnable_load_avg
+= contrib_delta
;
2804 cfs_rq
->utilization_load_avg
+= utilization_delta
;
2806 subtract_blocked_load_contrib(cfs_rq
, -contrib_delta
);
2811 * Decay the load contributed by all blocked children and account this so that
2812 * their contribution may appropriately discounted when they wake up.
2814 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
, int force_update
)
2816 u64 now
= cfs_rq_clock_task(cfs_rq
) >> 20;
2819 decays
= now
- cfs_rq
->last_decay
;
2820 if (!decays
&& !force_update
)
2823 if (atomic_long_read(&cfs_rq
->removed_load
)) {
2824 unsigned long removed_load
;
2825 removed_load
= atomic_long_xchg(&cfs_rq
->removed_load
, 0);
2826 subtract_blocked_load_contrib(cfs_rq
, removed_load
);
2830 cfs_rq
->blocked_load_avg
= decay_load(cfs_rq
->blocked_load_avg
,
2832 atomic64_add(decays
, &cfs_rq
->decay_counter
);
2833 cfs_rq
->last_decay
= now
;
2836 __update_cfs_rq_tg_load_contrib(cfs_rq
, force_update
);
2839 /* Add the load generated by se into cfs_rq's child load-average */
2840 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2841 struct sched_entity
*se
,
2845 * We track migrations using entity decay_count <= 0, on a wake-up
2846 * migration we use a negative decay count to track the remote decays
2847 * accumulated while sleeping.
2849 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2850 * are seen by enqueue_entity_load_avg() as a migration with an already
2851 * constructed load_avg_contrib.
2853 if (unlikely(se
->avg
.decay_count
<= 0)) {
2854 se
->avg
.last_runnable_update
= rq_clock_task(rq_of(cfs_rq
));
2855 if (se
->avg
.decay_count
) {
2857 * In a wake-up migration we have to approximate the
2858 * time sleeping. This is because we can't synchronize
2859 * clock_task between the two cpus, and it is not
2860 * guaranteed to be read-safe. Instead, we can
2861 * approximate this using our carried decays, which are
2862 * explicitly atomically readable.
2864 se
->avg
.last_runnable_update
-= (-se
->avg
.decay_count
)
2866 update_entity_load_avg(se
, 0);
2867 /* Indicate that we're now synchronized and on-rq */
2868 se
->avg
.decay_count
= 0;
2872 __synchronize_entity_decay(se
);
2875 /* migrated tasks did not contribute to our blocked load */
2877 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
2878 update_entity_load_avg(se
, 0);
2881 cfs_rq
->runnable_load_avg
+= se
->avg
.load_avg_contrib
;
2882 cfs_rq
->utilization_load_avg
+= se
->avg
.utilization_avg_contrib
;
2883 /* we force update consideration on load-balancer moves */
2884 update_cfs_rq_blocked_load(cfs_rq
, !wakeup
);
2888 * Remove se's load from this cfs_rq child load-average, if the entity is
2889 * transitioning to a blocked state we track its projected decay using
2892 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2893 struct sched_entity
*se
,
2896 update_entity_load_avg(se
, 1);
2897 /* we force update consideration on load-balancer moves */
2898 update_cfs_rq_blocked_load(cfs_rq
, !sleep
);
2900 cfs_rq
->runnable_load_avg
-= se
->avg
.load_avg_contrib
;
2901 cfs_rq
->utilization_load_avg
-= se
->avg
.utilization_avg_contrib
;
2903 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
2904 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
2905 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2909 * Update the rq's load with the elapsed running time before entering
2910 * idle. if the last scheduled task is not a CFS task, idle_enter will
2911 * be the only way to update the runnable statistic.
2913 void idle_enter_fair(struct rq
*this_rq
)
2915 update_rq_runnable_avg(this_rq
, 1);
2919 * Update the rq's load with the elapsed idle time before a task is
2920 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2921 * be the only way to update the runnable statistic.
2923 void idle_exit_fair(struct rq
*this_rq
)
2925 update_rq_runnable_avg(this_rq
, 0);
2928 static int idle_balance(struct rq
*this_rq
);
2930 #else /* CONFIG_SMP */
2932 static inline void update_entity_load_avg(struct sched_entity
*se
,
2933 int update_cfs_rq
) {}
2934 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
2935 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2936 struct sched_entity
*se
,
2938 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
2939 struct sched_entity
*se
,
2941 static inline void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
2942 int force_update
) {}
2944 static inline int idle_balance(struct rq
*rq
)
2949 #endif /* CONFIG_SMP */
2951 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2953 #ifdef CONFIG_SCHEDSTATS
2954 struct task_struct
*tsk
= NULL
;
2956 if (entity_is_task(se
))
2959 if (se
->statistics
.sleep_start
) {
2960 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.sleep_start
;
2965 if (unlikely(delta
> se
->statistics
.sleep_max
))
2966 se
->statistics
.sleep_max
= delta
;
2968 se
->statistics
.sleep_start
= 0;
2969 se
->statistics
.sum_sleep_runtime
+= delta
;
2972 account_scheduler_latency(tsk
, delta
>> 10, 1);
2973 trace_sched_stat_sleep(tsk
, delta
);
2976 if (se
->statistics
.block_start
) {
2977 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.block_start
;
2982 if (unlikely(delta
> se
->statistics
.block_max
))
2983 se
->statistics
.block_max
= delta
;
2985 se
->statistics
.block_start
= 0;
2986 se
->statistics
.sum_sleep_runtime
+= delta
;
2989 if (tsk
->in_iowait
) {
2990 se
->statistics
.iowait_sum
+= delta
;
2991 se
->statistics
.iowait_count
++;
2992 trace_sched_stat_iowait(tsk
, delta
);
2995 trace_sched_stat_blocked(tsk
, delta
);
2998 * Blocking time is in units of nanosecs, so shift by
2999 * 20 to get a milliseconds-range estimation of the
3000 * amount of time that the task spent sleeping:
3002 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
3003 profile_hits(SLEEP_PROFILING
,
3004 (void *)get_wchan(tsk
),
3007 account_scheduler_latency(tsk
, delta
>> 10, 0);
3013 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3015 #ifdef CONFIG_SCHED_DEBUG
3016 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
3021 if (d
> 3*sysctl_sched_latency
)
3022 schedstat_inc(cfs_rq
, nr_spread_over
);
3027 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
3029 u64 vruntime
= cfs_rq
->min_vruntime
;
3032 * The 'current' period is already promised to the current tasks,
3033 * however the extra weight of the new task will slow them down a
3034 * little, place the new task so that it fits in the slot that
3035 * stays open at the end.
3037 if (initial
&& sched_feat(START_DEBIT
))
3038 vruntime
+= sched_vslice(cfs_rq
, se
);
3040 /* sleeps up to a single latency don't count. */
3042 unsigned long thresh
= sysctl_sched_latency
;
3045 * Halve their sleep time's effect, to allow
3046 * for a gentler effect of sleepers:
3048 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
3054 /* ensure we never gain time by being placed backwards. */
3055 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
3058 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
3061 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3064 * Update the normalized vruntime before updating min_vruntime
3065 * through calling update_curr().
3067 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
3068 se
->vruntime
+= cfs_rq
->min_vruntime
;
3071 * Update run-time statistics of the 'current'.
3073 update_curr(cfs_rq
);
3074 enqueue_entity_load_avg(cfs_rq
, se
, flags
& ENQUEUE_WAKEUP
);
3075 account_entity_enqueue(cfs_rq
, se
);
3076 update_cfs_shares(cfs_rq
);
3078 if (flags
& ENQUEUE_WAKEUP
) {
3079 place_entity(cfs_rq
, se
, 0);
3080 enqueue_sleeper(cfs_rq
, se
);
3083 update_stats_enqueue(cfs_rq
, se
);
3084 check_spread(cfs_rq
, se
);
3085 if (se
!= cfs_rq
->curr
)
3086 __enqueue_entity(cfs_rq
, se
);
3089 if (cfs_rq
->nr_running
== 1) {
3090 list_add_leaf_cfs_rq(cfs_rq
);
3091 check_enqueue_throttle(cfs_rq
);
3095 static void __clear_buddies_last(struct sched_entity
*se
)
3097 for_each_sched_entity(se
) {
3098 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3099 if (cfs_rq
->last
!= se
)
3102 cfs_rq
->last
= NULL
;
3106 static void __clear_buddies_next(struct sched_entity
*se
)
3108 for_each_sched_entity(se
) {
3109 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3110 if (cfs_rq
->next
!= se
)
3113 cfs_rq
->next
= NULL
;
3117 static void __clear_buddies_skip(struct sched_entity
*se
)
3119 for_each_sched_entity(se
) {
3120 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3121 if (cfs_rq
->skip
!= se
)
3124 cfs_rq
->skip
= NULL
;
3128 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3130 if (cfs_rq
->last
== se
)
3131 __clear_buddies_last(se
);
3133 if (cfs_rq
->next
== se
)
3134 __clear_buddies_next(se
);
3136 if (cfs_rq
->skip
== se
)
3137 __clear_buddies_skip(se
);
3140 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3143 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
3146 * Update run-time statistics of the 'current'.
3148 update_curr(cfs_rq
);
3149 dequeue_entity_load_avg(cfs_rq
, se
, flags
& DEQUEUE_SLEEP
);
3151 update_stats_dequeue(cfs_rq
, se
);
3152 if (flags
& DEQUEUE_SLEEP
) {
3153 #ifdef CONFIG_SCHEDSTATS
3154 if (entity_is_task(se
)) {
3155 struct task_struct
*tsk
= task_of(se
);
3157 if (tsk
->state
& TASK_INTERRUPTIBLE
)
3158 se
->statistics
.sleep_start
= rq_clock(rq_of(cfs_rq
));
3159 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
3160 se
->statistics
.block_start
= rq_clock(rq_of(cfs_rq
));
3165 clear_buddies(cfs_rq
, se
);
3167 if (se
!= cfs_rq
->curr
)
3168 __dequeue_entity(cfs_rq
, se
);
3170 account_entity_dequeue(cfs_rq
, se
);
3173 * Normalize the entity after updating the min_vruntime because the
3174 * update can refer to the ->curr item and we need to reflect this
3175 * movement in our normalized position.
3177 if (!(flags
& DEQUEUE_SLEEP
))
3178 se
->vruntime
-= cfs_rq
->min_vruntime
;
3180 /* return excess runtime on last dequeue */
3181 return_cfs_rq_runtime(cfs_rq
);
3183 update_min_vruntime(cfs_rq
);
3184 update_cfs_shares(cfs_rq
);
3188 * Preempt the current task with a newly woken task if needed:
3191 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
3193 unsigned long ideal_runtime
, delta_exec
;
3194 struct sched_entity
*se
;
3197 ideal_runtime
= sched_slice(cfs_rq
, curr
);
3198 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
3199 if (delta_exec
> ideal_runtime
) {
3200 resched_curr(rq_of(cfs_rq
));
3202 * The current task ran long enough, ensure it doesn't get
3203 * re-elected due to buddy favours.
3205 clear_buddies(cfs_rq
, curr
);
3210 * Ensure that a task that missed wakeup preemption by a
3211 * narrow margin doesn't have to wait for a full slice.
3212 * This also mitigates buddy induced latencies under load.
3214 if (delta_exec
< sysctl_sched_min_granularity
)
3217 se
= __pick_first_entity(cfs_rq
);
3218 delta
= curr
->vruntime
- se
->vruntime
;
3223 if (delta
> ideal_runtime
)
3224 resched_curr(rq_of(cfs_rq
));
3228 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
3230 /* 'current' is not kept within the tree. */
3233 * Any task has to be enqueued before it get to execute on
3234 * a CPU. So account for the time it spent waiting on the
3237 update_stats_wait_end(cfs_rq
, se
);
3238 __dequeue_entity(cfs_rq
, se
);
3239 update_entity_load_avg(se
, 1);
3242 update_stats_curr_start(cfs_rq
, se
);
3244 #ifdef CONFIG_SCHEDSTATS
3246 * Track our maximum slice length, if the CPU's load is at
3247 * least twice that of our own weight (i.e. dont track it
3248 * when there are only lesser-weight tasks around):
3250 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
3251 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
3252 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
3255 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
3259 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
3262 * Pick the next process, keeping these things in mind, in this order:
3263 * 1) keep things fair between processes/task groups
3264 * 2) pick the "next" process, since someone really wants that to run
3265 * 3) pick the "last" process, for cache locality
3266 * 4) do not run the "skip" process, if something else is available
3268 static struct sched_entity
*
3269 pick_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
3271 struct sched_entity
*left
= __pick_first_entity(cfs_rq
);
3272 struct sched_entity
*se
;
3275 * If curr is set we have to see if its left of the leftmost entity
3276 * still in the tree, provided there was anything in the tree at all.
3278 if (!left
|| (curr
&& entity_before(curr
, left
)))
3281 se
= left
; /* ideally we run the leftmost entity */
3284 * Avoid running the skip buddy, if running something else can
3285 * be done without getting too unfair.
3287 if (cfs_rq
->skip
== se
) {
3288 struct sched_entity
*second
;
3291 second
= __pick_first_entity(cfs_rq
);
3293 second
= __pick_next_entity(se
);
3294 if (!second
|| (curr
&& entity_before(curr
, second
)))
3298 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
3303 * Prefer last buddy, try to return the CPU to a preempted task.
3305 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
3309 * Someone really wants this to run. If it's not unfair, run it.
3311 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
3314 clear_buddies(cfs_rq
, se
);
3319 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
3321 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
3324 * If still on the runqueue then deactivate_task()
3325 * was not called and update_curr() has to be done:
3328 update_curr(cfs_rq
);
3330 /* throttle cfs_rqs exceeding runtime */
3331 check_cfs_rq_runtime(cfs_rq
);
3333 check_spread(cfs_rq
, prev
);
3335 update_stats_wait_start(cfs_rq
, prev
);
3336 /* Put 'current' back into the tree. */
3337 __enqueue_entity(cfs_rq
, prev
);
3338 /* in !on_rq case, update occurred at dequeue */
3339 update_entity_load_avg(prev
, 1);
3341 cfs_rq
->curr
= NULL
;
3345 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
3348 * Update run-time statistics of the 'current'.
3350 update_curr(cfs_rq
);
3353 * Ensure that runnable average is periodically updated.
3355 update_entity_load_avg(curr
, 1);
3356 update_cfs_rq_blocked_load(cfs_rq
, 1);
3357 update_cfs_shares(cfs_rq
);
3359 #ifdef CONFIG_SCHED_HRTICK
3361 * queued ticks are scheduled to match the slice, so don't bother
3362 * validating it and just reschedule.
3365 resched_curr(rq_of(cfs_rq
));
3369 * don't let the period tick interfere with the hrtick preemption
3371 if (!sched_feat(DOUBLE_TICK
) &&
3372 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
3376 if (cfs_rq
->nr_running
> 1)
3377 check_preempt_tick(cfs_rq
, curr
);
3381 /**************************************************
3382 * CFS bandwidth control machinery
3385 #ifdef CONFIG_CFS_BANDWIDTH
3387 #ifdef HAVE_JUMP_LABEL
3388 static struct static_key __cfs_bandwidth_used
;
3390 static inline bool cfs_bandwidth_used(void)
3392 return static_key_false(&__cfs_bandwidth_used
);
3395 void cfs_bandwidth_usage_inc(void)
3397 static_key_slow_inc(&__cfs_bandwidth_used
);
3400 void cfs_bandwidth_usage_dec(void)
3402 static_key_slow_dec(&__cfs_bandwidth_used
);
3404 #else /* HAVE_JUMP_LABEL */
3405 static bool cfs_bandwidth_used(void)
3410 void cfs_bandwidth_usage_inc(void) {}
3411 void cfs_bandwidth_usage_dec(void) {}
3412 #endif /* HAVE_JUMP_LABEL */
3415 * default period for cfs group bandwidth.
3416 * default: 0.1s, units: nanoseconds
3418 static inline u64
default_cfs_period(void)
3420 return 100000000ULL;
3423 static inline u64
sched_cfs_bandwidth_slice(void)
3425 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
3429 * Replenish runtime according to assigned quota and update expiration time.
3430 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3431 * additional synchronization around rq->lock.
3433 * requires cfs_b->lock
3435 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
3439 if (cfs_b
->quota
== RUNTIME_INF
)
3442 now
= sched_clock_cpu(smp_processor_id());
3443 cfs_b
->runtime
= cfs_b
->quota
;
3444 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
3447 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
3449 return &tg
->cfs_bandwidth
;
3452 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3453 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
3455 if (unlikely(cfs_rq
->throttle_count
))
3456 return cfs_rq
->throttled_clock_task
;
3458 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
3461 /* returns 0 on failure to allocate runtime */
3462 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3464 struct task_group
*tg
= cfs_rq
->tg
;
3465 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
3466 u64 amount
= 0, min_amount
, expires
;
3468 /* note: this is a positive sum as runtime_remaining <= 0 */
3469 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
3471 raw_spin_lock(&cfs_b
->lock
);
3472 if (cfs_b
->quota
== RUNTIME_INF
)
3473 amount
= min_amount
;
3476 * If the bandwidth pool has become inactive, then at least one
3477 * period must have elapsed since the last consumption.
3478 * Refresh the global state and ensure bandwidth timer becomes
3481 if (!cfs_b
->timer_active
) {
3482 __refill_cfs_bandwidth_runtime(cfs_b
);
3483 __start_cfs_bandwidth(cfs_b
, false);
3486 if (cfs_b
->runtime
> 0) {
3487 amount
= min(cfs_b
->runtime
, min_amount
);
3488 cfs_b
->runtime
-= amount
;
3492 expires
= cfs_b
->runtime_expires
;
3493 raw_spin_unlock(&cfs_b
->lock
);
3495 cfs_rq
->runtime_remaining
+= amount
;
3497 * we may have advanced our local expiration to account for allowed
3498 * spread between our sched_clock and the one on which runtime was
3501 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
3502 cfs_rq
->runtime_expires
= expires
;
3504 return cfs_rq
->runtime_remaining
> 0;
3508 * Note: This depends on the synchronization provided by sched_clock and the
3509 * fact that rq->clock snapshots this value.
3511 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3513 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3515 /* if the deadline is ahead of our clock, nothing to do */
3516 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
3519 if (cfs_rq
->runtime_remaining
< 0)
3523 * If the local deadline has passed we have to consider the
3524 * possibility that our sched_clock is 'fast' and the global deadline
3525 * has not truly expired.
3527 * Fortunately we can check determine whether this the case by checking
3528 * whether the global deadline has advanced. It is valid to compare
3529 * cfs_b->runtime_expires without any locks since we only care about
3530 * exact equality, so a partial write will still work.
3533 if (cfs_rq
->runtime_expires
!= cfs_b
->runtime_expires
) {
3534 /* extend local deadline, drift is bounded above by 2 ticks */
3535 cfs_rq
->runtime_expires
+= TICK_NSEC
;
3537 /* global deadline is ahead, expiration has passed */
3538 cfs_rq
->runtime_remaining
= 0;
3542 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3544 /* dock delta_exec before expiring quota (as it could span periods) */
3545 cfs_rq
->runtime_remaining
-= delta_exec
;
3546 expire_cfs_rq_runtime(cfs_rq
);
3548 if (likely(cfs_rq
->runtime_remaining
> 0))
3552 * if we're unable to extend our runtime we resched so that the active
3553 * hierarchy can be throttled
3555 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
3556 resched_curr(rq_of(cfs_rq
));
3559 static __always_inline
3560 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
)
3562 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
3565 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
3568 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
3570 return cfs_bandwidth_used() && cfs_rq
->throttled
;
3573 /* check whether cfs_rq, or any parent, is throttled */
3574 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
3576 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
3580 * Ensure that neither of the group entities corresponding to src_cpu or
3581 * dest_cpu are members of a throttled hierarchy when performing group
3582 * load-balance operations.
3584 static inline int throttled_lb_pair(struct task_group
*tg
,
3585 int src_cpu
, int dest_cpu
)
3587 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
3589 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
3590 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
3592 return throttled_hierarchy(src_cfs_rq
) ||
3593 throttled_hierarchy(dest_cfs_rq
);
3596 /* updated child weight may affect parent so we have to do this bottom up */
3597 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
3599 struct rq
*rq
= data
;
3600 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3602 cfs_rq
->throttle_count
--;
3604 if (!cfs_rq
->throttle_count
) {
3605 /* adjust cfs_rq_clock_task() */
3606 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
3607 cfs_rq
->throttled_clock_task
;
3614 static int tg_throttle_down(struct task_group
*tg
, void *data
)
3616 struct rq
*rq
= data
;
3617 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
3619 /* group is entering throttled state, stop time */
3620 if (!cfs_rq
->throttle_count
)
3621 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
3622 cfs_rq
->throttle_count
++;
3627 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
3629 struct rq
*rq
= rq_of(cfs_rq
);
3630 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3631 struct sched_entity
*se
;
3632 long task_delta
, dequeue
= 1;
3634 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
3636 /* freeze hierarchy runnable averages while throttled */
3638 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
3641 task_delta
= cfs_rq
->h_nr_running
;
3642 for_each_sched_entity(se
) {
3643 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
3644 /* throttled entity or throttle-on-deactivate */
3649 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
3650 qcfs_rq
->h_nr_running
-= task_delta
;
3652 if (qcfs_rq
->load
.weight
)
3657 sub_nr_running(rq
, task_delta
);
3659 cfs_rq
->throttled
= 1;
3660 cfs_rq
->throttled_clock
= rq_clock(rq
);
3661 raw_spin_lock(&cfs_b
->lock
);
3663 * Add to the _head_ of the list, so that an already-started
3664 * distribute_cfs_runtime will not see us
3666 list_add_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
3667 if (!cfs_b
->timer_active
)
3668 __start_cfs_bandwidth(cfs_b
, false);
3669 raw_spin_unlock(&cfs_b
->lock
);
3672 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
3674 struct rq
*rq
= rq_of(cfs_rq
);
3675 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3676 struct sched_entity
*se
;
3680 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
3682 cfs_rq
->throttled
= 0;
3684 update_rq_clock(rq
);
3686 raw_spin_lock(&cfs_b
->lock
);
3687 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
3688 list_del_rcu(&cfs_rq
->throttled_list
);
3689 raw_spin_unlock(&cfs_b
->lock
);
3691 /* update hierarchical throttle state */
3692 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
3694 if (!cfs_rq
->load
.weight
)
3697 task_delta
= cfs_rq
->h_nr_running
;
3698 for_each_sched_entity(se
) {
3702 cfs_rq
= cfs_rq_of(se
);
3704 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
3705 cfs_rq
->h_nr_running
+= task_delta
;
3707 if (cfs_rq_throttled(cfs_rq
))
3712 add_nr_running(rq
, task_delta
);
3714 /* determine whether we need to wake up potentially idle cpu */
3715 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
3719 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
3720 u64 remaining
, u64 expires
)
3722 struct cfs_rq
*cfs_rq
;
3724 u64 starting_runtime
= remaining
;
3727 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
3729 struct rq
*rq
= rq_of(cfs_rq
);
3731 raw_spin_lock(&rq
->lock
);
3732 if (!cfs_rq_throttled(cfs_rq
))
3735 runtime
= -cfs_rq
->runtime_remaining
+ 1;
3736 if (runtime
> remaining
)
3737 runtime
= remaining
;
3738 remaining
-= runtime
;
3740 cfs_rq
->runtime_remaining
+= runtime
;
3741 cfs_rq
->runtime_expires
= expires
;
3743 /* we check whether we're throttled above */
3744 if (cfs_rq
->runtime_remaining
> 0)
3745 unthrottle_cfs_rq(cfs_rq
);
3748 raw_spin_unlock(&rq
->lock
);
3755 return starting_runtime
- remaining
;
3759 * Responsible for refilling a task_group's bandwidth and unthrottling its
3760 * cfs_rqs as appropriate. If there has been no activity within the last
3761 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3762 * used to track this state.
3764 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
3766 u64 runtime
, runtime_expires
;
3769 /* no need to continue the timer with no bandwidth constraint */
3770 if (cfs_b
->quota
== RUNTIME_INF
)
3771 goto out_deactivate
;
3773 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3774 cfs_b
->nr_periods
+= overrun
;
3777 * idle depends on !throttled (for the case of a large deficit), and if
3778 * we're going inactive then everything else can be deferred
3780 if (cfs_b
->idle
&& !throttled
)
3781 goto out_deactivate
;
3784 * if we have relooped after returning idle once, we need to update our
3785 * status as actually running, so that other cpus doing
3786 * __start_cfs_bandwidth will stop trying to cancel us.
3788 cfs_b
->timer_active
= 1;
3790 __refill_cfs_bandwidth_runtime(cfs_b
);
3793 /* mark as potentially idle for the upcoming period */
3798 /* account preceding periods in which throttling occurred */
3799 cfs_b
->nr_throttled
+= overrun
;
3801 runtime_expires
= cfs_b
->runtime_expires
;
3804 * This check is repeated as we are holding onto the new bandwidth while
3805 * we unthrottle. This can potentially race with an unthrottled group
3806 * trying to acquire new bandwidth from the global pool. This can result
3807 * in us over-using our runtime if it is all used during this loop, but
3808 * only by limited amounts in that extreme case.
3810 while (throttled
&& cfs_b
->runtime
> 0) {
3811 runtime
= cfs_b
->runtime
;
3812 raw_spin_unlock(&cfs_b
->lock
);
3813 /* we can't nest cfs_b->lock while distributing bandwidth */
3814 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
3816 raw_spin_lock(&cfs_b
->lock
);
3818 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
3820 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
3824 * While we are ensured activity in the period following an
3825 * unthrottle, this also covers the case in which the new bandwidth is
3826 * insufficient to cover the existing bandwidth deficit. (Forcing the
3827 * timer to remain active while there are any throttled entities.)
3834 cfs_b
->timer_active
= 0;
3838 /* a cfs_rq won't donate quota below this amount */
3839 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
3840 /* minimum remaining period time to redistribute slack quota */
3841 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
3842 /* how long we wait to gather additional slack before distributing */
3843 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
3846 * Are we near the end of the current quota period?
3848 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3849 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3850 * migrate_hrtimers, base is never cleared, so we are fine.
3852 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
3854 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
3857 /* if the call-back is running a quota refresh is already occurring */
3858 if (hrtimer_callback_running(refresh_timer
))
3861 /* is a quota refresh about to occur? */
3862 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
3863 if (remaining
< min_expire
)
3869 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
3871 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
3873 /* if there's a quota refresh soon don't bother with slack */
3874 if (runtime_refresh_within(cfs_b
, min_left
))
3877 start_bandwidth_timer(&cfs_b
->slack_timer
,
3878 ns_to_ktime(cfs_bandwidth_slack_period
));
3881 /* we know any runtime found here is valid as update_curr() precedes return */
3882 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3884 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
3885 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
3887 if (slack_runtime
<= 0)
3890 raw_spin_lock(&cfs_b
->lock
);
3891 if (cfs_b
->quota
!= RUNTIME_INF
&&
3892 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
3893 cfs_b
->runtime
+= slack_runtime
;
3895 /* we are under rq->lock, defer unthrottling using a timer */
3896 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
3897 !list_empty(&cfs_b
->throttled_cfs_rq
))
3898 start_cfs_slack_bandwidth(cfs_b
);
3900 raw_spin_unlock(&cfs_b
->lock
);
3902 /* even if it's not valid for return we don't want to try again */
3903 cfs_rq
->runtime_remaining
-= slack_runtime
;
3906 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3908 if (!cfs_bandwidth_used())
3911 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
3914 __return_cfs_rq_runtime(cfs_rq
);
3918 * This is done with a timer (instead of inline with bandwidth return) since
3919 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3921 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
3923 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
3926 /* confirm we're still not at a refresh boundary */
3927 raw_spin_lock(&cfs_b
->lock
);
3928 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
)) {
3929 raw_spin_unlock(&cfs_b
->lock
);
3933 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
)
3934 runtime
= cfs_b
->runtime
;
3936 expires
= cfs_b
->runtime_expires
;
3937 raw_spin_unlock(&cfs_b
->lock
);
3942 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
3944 raw_spin_lock(&cfs_b
->lock
);
3945 if (expires
== cfs_b
->runtime_expires
)
3946 cfs_b
->runtime
-= min(runtime
, cfs_b
->runtime
);
3947 raw_spin_unlock(&cfs_b
->lock
);
3951 * When a group wakes up we want to make sure that its quota is not already
3952 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3953 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3955 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
3957 if (!cfs_bandwidth_used())
3960 /* an active group must be handled by the update_curr()->put() path */
3961 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
3964 /* ensure the group is not already throttled */
3965 if (cfs_rq_throttled(cfs_rq
))
3968 /* update runtime allocation */
3969 account_cfs_rq_runtime(cfs_rq
, 0);
3970 if (cfs_rq
->runtime_remaining
<= 0)
3971 throttle_cfs_rq(cfs_rq
);
3974 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3975 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
3977 if (!cfs_bandwidth_used())
3980 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
3984 * it's possible for a throttled entity to be forced into a running
3985 * state (e.g. set_curr_task), in this case we're finished.
3987 if (cfs_rq_throttled(cfs_rq
))
3990 throttle_cfs_rq(cfs_rq
);
3994 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
3996 struct cfs_bandwidth
*cfs_b
=
3997 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
3998 do_sched_cfs_slack_timer(cfs_b
);
4000 return HRTIMER_NORESTART
;
4003 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
4005 struct cfs_bandwidth
*cfs_b
=
4006 container_of(timer
, struct cfs_bandwidth
, period_timer
);
4011 raw_spin_lock(&cfs_b
->lock
);
4013 now
= hrtimer_cb_get_time(timer
);
4014 overrun
= hrtimer_forward(timer
, now
, cfs_b
->period
);
4019 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
4021 raw_spin_unlock(&cfs_b
->lock
);
4023 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
4026 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
4028 raw_spin_lock_init(&cfs_b
->lock
);
4030 cfs_b
->quota
= RUNTIME_INF
;
4031 cfs_b
->period
= ns_to_ktime(default_cfs_period());
4033 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
4034 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
4035 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
4036 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
4037 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
4040 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
4042 cfs_rq
->runtime_enabled
= 0;
4043 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
4046 /* requires cfs_b->lock, may release to reprogram timer */
4047 void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
, bool force
)
4050 * The timer may be active because we're trying to set a new bandwidth
4051 * period or because we're racing with the tear-down path
4052 * (timer_active==0 becomes visible before the hrtimer call-back
4053 * terminates). In either case we ensure that it's re-programmed
4055 while (unlikely(hrtimer_active(&cfs_b
->period_timer
)) &&
4056 hrtimer_try_to_cancel(&cfs_b
->period_timer
) < 0) {
4057 /* bounce the lock to allow do_sched_cfs_period_timer to run */
4058 raw_spin_unlock(&cfs_b
->lock
);
4060 raw_spin_lock(&cfs_b
->lock
);
4061 /* if someone else restarted the timer then we're done */
4062 if (!force
&& cfs_b
->timer_active
)
4066 cfs_b
->timer_active
= 1;
4067 start_bandwidth_timer(&cfs_b
->period_timer
, cfs_b
->period
);
4070 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
4072 /* init_cfs_bandwidth() was not called */
4073 if (!cfs_b
->throttled_cfs_rq
.next
)
4076 hrtimer_cancel(&cfs_b
->period_timer
);
4077 hrtimer_cancel(&cfs_b
->slack_timer
);
4080 static void __maybe_unused
update_runtime_enabled(struct rq
*rq
)
4082 struct cfs_rq
*cfs_rq
;
4084 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
4085 struct cfs_bandwidth
*cfs_b
= &cfs_rq
->tg
->cfs_bandwidth
;
4087 raw_spin_lock(&cfs_b
->lock
);
4088 cfs_rq
->runtime_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
4089 raw_spin_unlock(&cfs_b
->lock
);
4093 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
4095 struct cfs_rq
*cfs_rq
;
4097 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
4098 if (!cfs_rq
->runtime_enabled
)
4102 * clock_task is not advancing so we just need to make sure
4103 * there's some valid quota amount
4105 cfs_rq
->runtime_remaining
= 1;
4107 * Offline rq is schedulable till cpu is completely disabled
4108 * in take_cpu_down(), so we prevent new cfs throttling here.
4110 cfs_rq
->runtime_enabled
= 0;
4112 if (cfs_rq_throttled(cfs_rq
))
4113 unthrottle_cfs_rq(cfs_rq
);
4117 #else /* CONFIG_CFS_BANDWIDTH */
4118 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
4120 return rq_clock_task(rq_of(cfs_rq
));
4123 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, u64 delta_exec
) {}
4124 static bool check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) { return false; }
4125 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
4126 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
4128 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
4133 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
4138 static inline int throttled_lb_pair(struct task_group
*tg
,
4139 int src_cpu
, int dest_cpu
)
4144 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
4146 #ifdef CONFIG_FAIR_GROUP_SCHED
4147 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
4150 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
4154 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
4155 static inline void update_runtime_enabled(struct rq
*rq
) {}
4156 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
4158 #endif /* CONFIG_CFS_BANDWIDTH */
4160 /**************************************************
4161 * CFS operations on tasks:
4164 #ifdef CONFIG_SCHED_HRTICK
4165 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
4167 struct sched_entity
*se
= &p
->se
;
4168 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4170 WARN_ON(task_rq(p
) != rq
);
4172 if (cfs_rq
->nr_running
> 1) {
4173 u64 slice
= sched_slice(cfs_rq
, se
);
4174 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
4175 s64 delta
= slice
- ran
;
4182 hrtick_start(rq
, delta
);
4187 * called from enqueue/dequeue and updates the hrtick when the
4188 * current task is from our class and nr_running is low enough
4191 static void hrtick_update(struct rq
*rq
)
4193 struct task_struct
*curr
= rq
->curr
;
4195 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
4198 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
4199 hrtick_start_fair(rq
, curr
);
4201 #else /* !CONFIG_SCHED_HRTICK */
4203 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
4207 static inline void hrtick_update(struct rq
*rq
)
4213 * The enqueue_task method is called before nr_running is
4214 * increased. Here we update the fair scheduling stats and
4215 * then put the task into the rbtree:
4218 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
4220 struct cfs_rq
*cfs_rq
;
4221 struct sched_entity
*se
= &p
->se
;
4223 for_each_sched_entity(se
) {
4226 cfs_rq
= cfs_rq_of(se
);
4227 enqueue_entity(cfs_rq
, se
, flags
);
4230 * end evaluation on encountering a throttled cfs_rq
4232 * note: in the case of encountering a throttled cfs_rq we will
4233 * post the final h_nr_running increment below.
4235 if (cfs_rq_throttled(cfs_rq
))
4237 cfs_rq
->h_nr_running
++;
4239 flags
= ENQUEUE_WAKEUP
;
4242 for_each_sched_entity(se
) {
4243 cfs_rq
= cfs_rq_of(se
);
4244 cfs_rq
->h_nr_running
++;
4246 if (cfs_rq_throttled(cfs_rq
))
4249 update_cfs_shares(cfs_rq
);
4250 update_entity_load_avg(se
, 1);
4254 update_rq_runnable_avg(rq
, rq
->nr_running
);
4255 add_nr_running(rq
, 1);
4260 static void set_next_buddy(struct sched_entity
*se
);
4263 * The dequeue_task method is called before nr_running is
4264 * decreased. We remove the task from the rbtree and
4265 * update the fair scheduling stats:
4267 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
4269 struct cfs_rq
*cfs_rq
;
4270 struct sched_entity
*se
= &p
->se
;
4271 int task_sleep
= flags
& DEQUEUE_SLEEP
;
4273 for_each_sched_entity(se
) {
4274 cfs_rq
= cfs_rq_of(se
);
4275 dequeue_entity(cfs_rq
, se
, flags
);
4278 * end evaluation on encountering a throttled cfs_rq
4280 * note: in the case of encountering a throttled cfs_rq we will
4281 * post the final h_nr_running decrement below.
4283 if (cfs_rq_throttled(cfs_rq
))
4285 cfs_rq
->h_nr_running
--;
4287 /* Don't dequeue parent if it has other entities besides us */
4288 if (cfs_rq
->load
.weight
) {
4290 * Bias pick_next to pick a task from this cfs_rq, as
4291 * p is sleeping when it is within its sched_slice.
4293 if (task_sleep
&& parent_entity(se
))
4294 set_next_buddy(parent_entity(se
));
4296 /* avoid re-evaluating load for this entity */
4297 se
= parent_entity(se
);
4300 flags
|= DEQUEUE_SLEEP
;
4303 for_each_sched_entity(se
) {
4304 cfs_rq
= cfs_rq_of(se
);
4305 cfs_rq
->h_nr_running
--;
4307 if (cfs_rq_throttled(cfs_rq
))
4310 update_cfs_shares(cfs_rq
);
4311 update_entity_load_avg(se
, 1);
4315 sub_nr_running(rq
, 1);
4316 update_rq_runnable_avg(rq
, 1);
4322 /* Used instead of source_load when we know the type == 0 */
4323 static unsigned long weighted_cpuload(const int cpu
)
4325 return cpu_rq(cpu
)->cfs
.runnable_load_avg
;
4329 * Return a low guess at the load of a migration-source cpu weighted
4330 * according to the scheduling class and "nice" value.
4332 * We want to under-estimate the load of migration sources, to
4333 * balance conservatively.
4335 static unsigned long source_load(int cpu
, int type
)
4337 struct rq
*rq
= cpu_rq(cpu
);
4338 unsigned long total
= weighted_cpuload(cpu
);
4340 if (type
== 0 || !sched_feat(LB_BIAS
))
4343 return min(rq
->cpu_load
[type
-1], total
);
4347 * Return a high guess at the load of a migration-target cpu weighted
4348 * according to the scheduling class and "nice" value.
4350 static unsigned long target_load(int cpu
, int type
)
4352 struct rq
*rq
= cpu_rq(cpu
);
4353 unsigned long total
= weighted_cpuload(cpu
);
4355 if (type
== 0 || !sched_feat(LB_BIAS
))
4358 return max(rq
->cpu_load
[type
-1], total
);
4361 static unsigned long capacity_of(int cpu
)
4363 return cpu_rq(cpu
)->cpu_capacity
;
4366 static unsigned long capacity_orig_of(int cpu
)
4368 return cpu_rq(cpu
)->cpu_capacity_orig
;
4371 static unsigned long cpu_avg_load_per_task(int cpu
)
4373 struct rq
*rq
= cpu_rq(cpu
);
4374 unsigned long nr_running
= ACCESS_ONCE(rq
->cfs
.h_nr_running
);
4375 unsigned long load_avg
= rq
->cfs
.runnable_load_avg
;
4378 return load_avg
/ nr_running
;
4383 static void record_wakee(struct task_struct
*p
)
4386 * Rough decay (wiping) for cost saving, don't worry
4387 * about the boundary, really active task won't care
4390 if (time_after(jiffies
, current
->wakee_flip_decay_ts
+ HZ
)) {
4391 current
->wakee_flips
>>= 1;
4392 current
->wakee_flip_decay_ts
= jiffies
;
4395 if (current
->last_wakee
!= p
) {
4396 current
->last_wakee
= p
;
4397 current
->wakee_flips
++;
4401 static void task_waking_fair(struct task_struct
*p
)
4403 struct sched_entity
*se
= &p
->se
;
4404 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4407 #ifndef CONFIG_64BIT
4408 u64 min_vruntime_copy
;
4411 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
4413 min_vruntime
= cfs_rq
->min_vruntime
;
4414 } while (min_vruntime
!= min_vruntime_copy
);
4416 min_vruntime
= cfs_rq
->min_vruntime
;
4419 se
->vruntime
-= min_vruntime
;
4423 #ifdef CONFIG_FAIR_GROUP_SCHED
4425 * effective_load() calculates the load change as seen from the root_task_group
4427 * Adding load to a group doesn't make a group heavier, but can cause movement
4428 * of group shares between cpus. Assuming the shares were perfectly aligned one
4429 * can calculate the shift in shares.
4431 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4432 * on this @cpu and results in a total addition (subtraction) of @wg to the
4433 * total group weight.
4435 * Given a runqueue weight distribution (rw_i) we can compute a shares
4436 * distribution (s_i) using:
4438 * s_i = rw_i / \Sum rw_j (1)
4440 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4441 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4442 * shares distribution (s_i):
4444 * rw_i = { 2, 4, 1, 0 }
4445 * s_i = { 2/7, 4/7, 1/7, 0 }
4447 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4448 * task used to run on and the CPU the waker is running on), we need to
4449 * compute the effect of waking a task on either CPU and, in case of a sync
4450 * wakeup, compute the effect of the current task going to sleep.
4452 * So for a change of @wl to the local @cpu with an overall group weight change
4453 * of @wl we can compute the new shares distribution (s'_i) using:
4455 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4457 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4458 * differences in waking a task to CPU 0. The additional task changes the
4459 * weight and shares distributions like:
4461 * rw'_i = { 3, 4, 1, 0 }
4462 * s'_i = { 3/8, 4/8, 1/8, 0 }
4464 * We can then compute the difference in effective weight by using:
4466 * dw_i = S * (s'_i - s_i) (3)
4468 * Where 'S' is the group weight as seen by its parent.
4470 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4471 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4472 * 4/7) times the weight of the group.
4474 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4476 struct sched_entity
*se
= tg
->se
[cpu
];
4478 if (!tg
->parent
) /* the trivial, non-cgroup case */
4481 for_each_sched_entity(se
) {
4487 * W = @wg + \Sum rw_j
4489 W
= wg
+ calc_tg_weight(tg
, se
->my_q
);
4494 w
= se
->my_q
->load
.weight
+ wl
;
4497 * wl = S * s'_i; see (2)
4500 wl
= (w
* (long)tg
->shares
) / W
;
4505 * Per the above, wl is the new se->load.weight value; since
4506 * those are clipped to [MIN_SHARES, ...) do so now. See
4507 * calc_cfs_shares().
4509 if (wl
< MIN_SHARES
)
4513 * wl = dw_i = S * (s'_i - s_i); see (3)
4515 wl
-= se
->load
.weight
;
4518 * Recursively apply this logic to all parent groups to compute
4519 * the final effective load change on the root group. Since
4520 * only the @tg group gets extra weight, all parent groups can
4521 * only redistribute existing shares. @wl is the shift in shares
4522 * resulting from this level per the above.
4531 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
4538 static int wake_wide(struct task_struct
*p
)
4540 int factor
= this_cpu_read(sd_llc_size
);
4543 * Yeah, it's the switching-frequency, could means many wakee or
4544 * rapidly switch, use factor here will just help to automatically
4545 * adjust the loose-degree, so bigger node will lead to more pull.
4547 if (p
->wakee_flips
> factor
) {
4549 * wakee is somewhat hot, it needs certain amount of cpu
4550 * resource, so if waker is far more hot, prefer to leave
4553 if (current
->wakee_flips
> (factor
* p
->wakee_flips
))
4560 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
4562 s64 this_load
, load
;
4563 s64 this_eff_load
, prev_eff_load
;
4564 int idx
, this_cpu
, prev_cpu
;
4565 struct task_group
*tg
;
4566 unsigned long weight
;
4570 * If we wake multiple tasks be careful to not bounce
4571 * ourselves around too much.
4577 this_cpu
= smp_processor_id();
4578 prev_cpu
= task_cpu(p
);
4579 load
= source_load(prev_cpu
, idx
);
4580 this_load
= target_load(this_cpu
, idx
);
4583 * If sync wakeup then subtract the (maximum possible)
4584 * effect of the currently running task from the load
4585 * of the current CPU:
4588 tg
= task_group(current
);
4589 weight
= current
->se
.load
.weight
;
4591 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
4592 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
4596 weight
= p
->se
.load
.weight
;
4599 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4600 * due to the sync cause above having dropped this_load to 0, we'll
4601 * always have an imbalance, but there's really nothing you can do
4602 * about that, so that's good too.
4604 * Otherwise check if either cpus are near enough in load to allow this
4605 * task to be woken on this_cpu.
4607 this_eff_load
= 100;
4608 this_eff_load
*= capacity_of(prev_cpu
);
4610 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
4611 prev_eff_load
*= capacity_of(this_cpu
);
4613 if (this_load
> 0) {
4614 this_eff_load
*= this_load
+
4615 effective_load(tg
, this_cpu
, weight
, weight
);
4617 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
4620 balanced
= this_eff_load
<= prev_eff_load
;
4622 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
4627 schedstat_inc(sd
, ttwu_move_affine
);
4628 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
4634 * find_idlest_group finds and returns the least busy CPU group within the
4637 static struct sched_group
*
4638 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
4639 int this_cpu
, int sd_flag
)
4641 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
4642 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
4643 int load_idx
= sd
->forkexec_idx
;
4644 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
4646 if (sd_flag
& SD_BALANCE_WAKE
)
4647 load_idx
= sd
->wake_idx
;
4650 unsigned long load
, avg_load
;
4654 /* Skip over this group if it has no CPUs allowed */
4655 if (!cpumask_intersects(sched_group_cpus(group
),
4656 tsk_cpus_allowed(p
)))
4659 local_group
= cpumask_test_cpu(this_cpu
,
4660 sched_group_cpus(group
));
4662 /* Tally up the load of all CPUs in the group */
4665 for_each_cpu(i
, sched_group_cpus(group
)) {
4666 /* Bias balancing toward cpus of our domain */
4668 load
= source_load(i
, load_idx
);
4670 load
= target_load(i
, load_idx
);
4675 /* Adjust by relative CPU capacity of the group */
4676 avg_load
= (avg_load
* SCHED_CAPACITY_SCALE
) / group
->sgc
->capacity
;
4679 this_load
= avg_load
;
4680 } else if (avg_load
< min_load
) {
4681 min_load
= avg_load
;
4684 } while (group
= group
->next
, group
!= sd
->groups
);
4686 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
4692 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4695 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
4697 unsigned long load
, min_load
= ULONG_MAX
;
4698 unsigned int min_exit_latency
= UINT_MAX
;
4699 u64 latest_idle_timestamp
= 0;
4700 int least_loaded_cpu
= this_cpu
;
4701 int shallowest_idle_cpu
= -1;
4704 /* Traverse only the allowed CPUs */
4705 for_each_cpu_and(i
, sched_group_cpus(group
), tsk_cpus_allowed(p
)) {
4707 struct rq
*rq
= cpu_rq(i
);
4708 struct cpuidle_state
*idle
= idle_get_state(rq
);
4709 if (idle
&& idle
->exit_latency
< min_exit_latency
) {
4711 * We give priority to a CPU whose idle state
4712 * has the smallest exit latency irrespective
4713 * of any idle timestamp.
4715 min_exit_latency
= idle
->exit_latency
;
4716 latest_idle_timestamp
= rq
->idle_stamp
;
4717 shallowest_idle_cpu
= i
;
4718 } else if ((!idle
|| idle
->exit_latency
== min_exit_latency
) &&
4719 rq
->idle_stamp
> latest_idle_timestamp
) {
4721 * If equal or no active idle state, then
4722 * the most recently idled CPU might have
4725 latest_idle_timestamp
= rq
->idle_stamp
;
4726 shallowest_idle_cpu
= i
;
4728 } else if (shallowest_idle_cpu
== -1) {
4729 load
= weighted_cpuload(i
);
4730 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
4732 least_loaded_cpu
= i
;
4737 return shallowest_idle_cpu
!= -1 ? shallowest_idle_cpu
: least_loaded_cpu
;
4741 * Try and locate an idle CPU in the sched_domain.
4743 static int select_idle_sibling(struct task_struct
*p
, int target
)
4745 struct sched_domain
*sd
;
4746 struct sched_group
*sg
;
4747 int i
= task_cpu(p
);
4749 if (idle_cpu(target
))
4753 * If the prevous cpu is cache affine and idle, don't be stupid.
4755 if (i
!= target
&& cpus_share_cache(i
, target
) && idle_cpu(i
))
4759 * Otherwise, iterate the domains and find an elegible idle cpu.
4761 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
4762 for_each_lower_domain(sd
) {
4765 if (!cpumask_intersects(sched_group_cpus(sg
),
4766 tsk_cpus_allowed(p
)))
4769 for_each_cpu(i
, sched_group_cpus(sg
)) {
4770 if (i
== target
|| !idle_cpu(i
))
4774 target
= cpumask_first_and(sched_group_cpus(sg
),
4775 tsk_cpus_allowed(p
));
4779 } while (sg
!= sd
->groups
);
4785 * get_cpu_usage returns the amount of capacity of a CPU that is used by CFS
4786 * tasks. The unit of the return value must be the one of capacity so we can
4787 * compare the usage with the capacity of the CPU that is available for CFS
4788 * task (ie cpu_capacity).
4789 * cfs.utilization_load_avg is the sum of running time of runnable tasks on a
4790 * CPU. It represents the amount of utilization of a CPU in the range
4791 * [0..SCHED_LOAD_SCALE]. The usage of a CPU can't be higher than the full
4792 * capacity of the CPU because it's about the running time on this CPU.
4793 * Nevertheless, cfs.utilization_load_avg can be higher than SCHED_LOAD_SCALE
4794 * because of unfortunate rounding in avg_period and running_load_avg or just
4795 * after migrating tasks until the average stabilizes with the new running
4796 * time. So we need to check that the usage stays into the range
4797 * [0..cpu_capacity_orig] and cap if necessary.
4798 * Without capping the usage, a group could be seen as overloaded (CPU0 usage
4799 * at 121% + CPU1 usage at 80%) whereas CPU1 has 20% of available capacity
4801 static int get_cpu_usage(int cpu
)
4803 unsigned long usage
= cpu_rq(cpu
)->cfs
.utilization_load_avg
;
4804 unsigned long capacity
= capacity_orig_of(cpu
);
4806 if (usage
>= SCHED_LOAD_SCALE
)
4809 return (usage
* capacity
) >> SCHED_LOAD_SHIFT
;
4813 * select_task_rq_fair: Select target runqueue for the waking task in domains
4814 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4815 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4817 * Balances load by selecting the idlest cpu in the idlest group, or under
4818 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4820 * Returns the target cpu number.
4822 * preempt must be disabled.
4825 select_task_rq_fair(struct task_struct
*p
, int prev_cpu
, int sd_flag
, int wake_flags
)
4827 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
4828 int cpu
= smp_processor_id();
4830 int want_affine
= 0;
4831 int sync
= wake_flags
& WF_SYNC
;
4833 if (sd_flag
& SD_BALANCE_WAKE
)
4834 want_affine
= cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
));
4837 for_each_domain(cpu
, tmp
) {
4838 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
4842 * If both cpu and prev_cpu are part of this domain,
4843 * cpu is a valid SD_WAKE_AFFINE target.
4845 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
4846 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
4851 if (tmp
->flags
& sd_flag
)
4855 if (affine_sd
&& cpu
!= prev_cpu
&& wake_affine(affine_sd
, p
, sync
))
4858 if (sd_flag
& SD_BALANCE_WAKE
) {
4859 new_cpu
= select_idle_sibling(p
, prev_cpu
);
4864 struct sched_group
*group
;
4867 if (!(sd
->flags
& sd_flag
)) {
4872 group
= find_idlest_group(sd
, p
, cpu
, sd_flag
);
4878 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
4879 if (new_cpu
== -1 || new_cpu
== cpu
) {
4880 /* Now try balancing at a lower domain level of cpu */
4885 /* Now try balancing at a lower domain level of new_cpu */
4887 weight
= sd
->span_weight
;
4889 for_each_domain(cpu
, tmp
) {
4890 if (weight
<= tmp
->span_weight
)
4892 if (tmp
->flags
& sd_flag
)
4895 /* while loop will break here if sd == NULL */
4904 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4905 * cfs_rq_of(p) references at time of call are still valid and identify the
4906 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4907 * other assumptions, including the state of rq->lock, should be made.
4910 migrate_task_rq_fair(struct task_struct
*p
, int next_cpu
)
4912 struct sched_entity
*se
= &p
->se
;
4913 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
4916 * Load tracking: accumulate removed load so that it can be processed
4917 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4918 * to blocked load iff they have a positive decay-count. It can never
4919 * be negative here since on-rq tasks have decay-count == 0.
4921 if (se
->avg
.decay_count
) {
4922 se
->avg
.decay_count
= -__synchronize_entity_decay(se
);
4923 atomic_long_add(se
->avg
.load_avg_contrib
,
4924 &cfs_rq
->removed_load
);
4927 /* We have migrated, no longer consider this task hot */
4930 #endif /* CONFIG_SMP */
4932 static unsigned long
4933 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
4935 unsigned long gran
= sysctl_sched_wakeup_granularity
;
4938 * Since its curr running now, convert the gran from real-time
4939 * to virtual-time in his units.
4941 * By using 'se' instead of 'curr' we penalize light tasks, so
4942 * they get preempted easier. That is, if 'se' < 'curr' then
4943 * the resulting gran will be larger, therefore penalizing the
4944 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4945 * be smaller, again penalizing the lighter task.
4947 * This is especially important for buddies when the leftmost
4948 * task is higher priority than the buddy.
4950 return calc_delta_fair(gran
, se
);
4954 * Should 'se' preempt 'curr'.
4968 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
4970 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
4975 gran
= wakeup_gran(curr
, se
);
4982 static void set_last_buddy(struct sched_entity
*se
)
4984 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4987 for_each_sched_entity(se
)
4988 cfs_rq_of(se
)->last
= se
;
4991 static void set_next_buddy(struct sched_entity
*se
)
4993 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
4996 for_each_sched_entity(se
)
4997 cfs_rq_of(se
)->next
= se
;
5000 static void set_skip_buddy(struct sched_entity
*se
)
5002 for_each_sched_entity(se
)
5003 cfs_rq_of(se
)->skip
= se
;
5007 * Preempt the current task with a newly woken task if needed:
5009 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
5011 struct task_struct
*curr
= rq
->curr
;
5012 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
5013 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
5014 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
5015 int next_buddy_marked
= 0;
5017 if (unlikely(se
== pse
))
5021 * This is possible from callers such as attach_tasks(), in which we
5022 * unconditionally check_prempt_curr() after an enqueue (which may have
5023 * lead to a throttle). This both saves work and prevents false
5024 * next-buddy nomination below.
5026 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
5029 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
5030 set_next_buddy(pse
);
5031 next_buddy_marked
= 1;
5035 * We can come here with TIF_NEED_RESCHED already set from new task
5038 * Note: this also catches the edge-case of curr being in a throttled
5039 * group (e.g. via set_curr_task), since update_curr() (in the
5040 * enqueue of curr) will have resulted in resched being set. This
5041 * prevents us from potentially nominating it as a false LAST_BUDDY
5044 if (test_tsk_need_resched(curr
))
5047 /* Idle tasks are by definition preempted by non-idle tasks. */
5048 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
5049 likely(p
->policy
!= SCHED_IDLE
))
5053 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5054 * is driven by the tick):
5056 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
5059 find_matching_se(&se
, &pse
);
5060 update_curr(cfs_rq_of(se
));
5062 if (wakeup_preempt_entity(se
, pse
) == 1) {
5064 * Bias pick_next to pick the sched entity that is
5065 * triggering this preemption.
5067 if (!next_buddy_marked
)
5068 set_next_buddy(pse
);
5077 * Only set the backward buddy when the current task is still
5078 * on the rq. This can happen when a wakeup gets interleaved
5079 * with schedule on the ->pre_schedule() or idle_balance()
5080 * point, either of which can * drop the rq lock.
5082 * Also, during early boot the idle thread is in the fair class,
5083 * for obvious reasons its a bad idea to schedule back to it.
5085 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
5088 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
5092 static struct task_struct
*
5093 pick_next_task_fair(struct rq
*rq
, struct task_struct
*prev
)
5095 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
5096 struct sched_entity
*se
;
5097 struct task_struct
*p
;
5101 #ifdef CONFIG_FAIR_GROUP_SCHED
5102 if (!cfs_rq
->nr_running
)
5105 if (prev
->sched_class
!= &fair_sched_class
)
5109 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5110 * likely that a next task is from the same cgroup as the current.
5112 * Therefore attempt to avoid putting and setting the entire cgroup
5113 * hierarchy, only change the part that actually changes.
5117 struct sched_entity
*curr
= cfs_rq
->curr
;
5120 * Since we got here without doing put_prev_entity() we also
5121 * have to consider cfs_rq->curr. If it is still a runnable
5122 * entity, update_curr() will update its vruntime, otherwise
5123 * forget we've ever seen it.
5125 if (curr
&& curr
->on_rq
)
5126 update_curr(cfs_rq
);
5131 * This call to check_cfs_rq_runtime() will do the throttle and
5132 * dequeue its entity in the parent(s). Therefore the 'simple'
5133 * nr_running test will indeed be correct.
5135 if (unlikely(check_cfs_rq_runtime(cfs_rq
)))
5138 se
= pick_next_entity(cfs_rq
, curr
);
5139 cfs_rq
= group_cfs_rq(se
);
5145 * Since we haven't yet done put_prev_entity and if the selected task
5146 * is a different task than we started out with, try and touch the
5147 * least amount of cfs_rqs.
5150 struct sched_entity
*pse
= &prev
->se
;
5152 while (!(cfs_rq
= is_same_group(se
, pse
))) {
5153 int se_depth
= se
->depth
;
5154 int pse_depth
= pse
->depth
;
5156 if (se_depth
<= pse_depth
) {
5157 put_prev_entity(cfs_rq_of(pse
), pse
);
5158 pse
= parent_entity(pse
);
5160 if (se_depth
>= pse_depth
) {
5161 set_next_entity(cfs_rq_of(se
), se
);
5162 se
= parent_entity(se
);
5166 put_prev_entity(cfs_rq
, pse
);
5167 set_next_entity(cfs_rq
, se
);
5170 if (hrtick_enabled(rq
))
5171 hrtick_start_fair(rq
, p
);
5178 if (!cfs_rq
->nr_running
)
5181 put_prev_task(rq
, prev
);
5184 se
= pick_next_entity(cfs_rq
, NULL
);
5185 set_next_entity(cfs_rq
, se
);
5186 cfs_rq
= group_cfs_rq(se
);
5191 if (hrtick_enabled(rq
))
5192 hrtick_start_fair(rq
, p
);
5197 new_tasks
= idle_balance(rq
);
5199 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5200 * possible for any higher priority task to appear. In that case we
5201 * must re-start the pick_next_entity() loop.
5213 * Account for a descheduled task:
5215 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
5217 struct sched_entity
*se
= &prev
->se
;
5218 struct cfs_rq
*cfs_rq
;
5220 for_each_sched_entity(se
) {
5221 cfs_rq
= cfs_rq_of(se
);
5222 put_prev_entity(cfs_rq
, se
);
5227 * sched_yield() is very simple
5229 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5231 static void yield_task_fair(struct rq
*rq
)
5233 struct task_struct
*curr
= rq
->curr
;
5234 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
5235 struct sched_entity
*se
= &curr
->se
;
5238 * Are we the only task in the tree?
5240 if (unlikely(rq
->nr_running
== 1))
5243 clear_buddies(cfs_rq
, se
);
5245 if (curr
->policy
!= SCHED_BATCH
) {
5246 update_rq_clock(rq
);
5248 * Update run-time statistics of the 'current'.
5250 update_curr(cfs_rq
);
5252 * Tell update_rq_clock() that we've just updated,
5253 * so we don't do microscopic update in schedule()
5254 * and double the fastpath cost.
5256 rq_clock_skip_update(rq
, true);
5262 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
5264 struct sched_entity
*se
= &p
->se
;
5266 /* throttled hierarchies are not runnable */
5267 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
5270 /* Tell the scheduler that we'd really like pse to run next. */
5273 yield_task_fair(rq
);
5279 /**************************************************
5280 * Fair scheduling class load-balancing methods.
5284 * The purpose of load-balancing is to achieve the same basic fairness the
5285 * per-cpu scheduler provides, namely provide a proportional amount of compute
5286 * time to each task. This is expressed in the following equation:
5288 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5290 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5291 * W_i,0 is defined as:
5293 * W_i,0 = \Sum_j w_i,j (2)
5295 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5296 * is derived from the nice value as per prio_to_weight[].
5298 * The weight average is an exponential decay average of the instantaneous
5301 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5303 * C_i is the compute capacity of cpu i, typically it is the
5304 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5305 * can also include other factors [XXX].
5307 * To achieve this balance we define a measure of imbalance which follows
5308 * directly from (1):
5310 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5312 * We them move tasks around to minimize the imbalance. In the continuous
5313 * function space it is obvious this converges, in the discrete case we get
5314 * a few fun cases generally called infeasible weight scenarios.
5317 * - infeasible weights;
5318 * - local vs global optima in the discrete case. ]
5323 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5324 * for all i,j solution, we create a tree of cpus that follows the hardware
5325 * topology where each level pairs two lower groups (or better). This results
5326 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5327 * tree to only the first of the previous level and we decrease the frequency
5328 * of load-balance at each level inv. proportional to the number of cpus in
5334 * \Sum { --- * --- * 2^i } = O(n) (5)
5336 * `- size of each group
5337 * | | `- number of cpus doing load-balance
5339 * `- sum over all levels
5341 * Coupled with a limit on how many tasks we can migrate every balance pass,
5342 * this makes (5) the runtime complexity of the balancer.
5344 * An important property here is that each CPU is still (indirectly) connected
5345 * to every other cpu in at most O(log n) steps:
5347 * The adjacency matrix of the resulting graph is given by:
5350 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5353 * And you'll find that:
5355 * A^(log_2 n)_i,j != 0 for all i,j (7)
5357 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5358 * The task movement gives a factor of O(m), giving a convergence complexity
5361 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5366 * In order to avoid CPUs going idle while there's still work to do, new idle
5367 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5368 * tree itself instead of relying on other CPUs to bring it work.
5370 * This adds some complexity to both (5) and (8) but it reduces the total idle
5378 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5381 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5386 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5388 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5390 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5393 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5394 * rewrite all of this once again.]
5397 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
5399 enum fbq_type
{ regular
, remote
, all
};
5401 #define LBF_ALL_PINNED 0x01
5402 #define LBF_NEED_BREAK 0x02
5403 #define LBF_DST_PINNED 0x04
5404 #define LBF_SOME_PINNED 0x08
5407 struct sched_domain
*sd
;
5415 struct cpumask
*dst_grpmask
;
5417 enum cpu_idle_type idle
;
5419 /* The set of CPUs under consideration for load-balancing */
5420 struct cpumask
*cpus
;
5425 unsigned int loop_break
;
5426 unsigned int loop_max
;
5428 enum fbq_type fbq_type
;
5429 struct list_head tasks
;
5433 * Is this task likely cache-hot:
5435 static int task_hot(struct task_struct
*p
, struct lb_env
*env
)
5439 lockdep_assert_held(&env
->src_rq
->lock
);
5441 if (p
->sched_class
!= &fair_sched_class
)
5444 if (unlikely(p
->policy
== SCHED_IDLE
))
5448 * Buddy candidates are cache hot:
5450 if (sched_feat(CACHE_HOT_BUDDY
) && env
->dst_rq
->nr_running
&&
5451 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
5452 &p
->se
== cfs_rq_of(&p
->se
)->last
))
5455 if (sysctl_sched_migration_cost
== -1)
5457 if (sysctl_sched_migration_cost
== 0)
5460 delta
= rq_clock_task(env
->src_rq
) - p
->se
.exec_start
;
5462 return delta
< (s64
)sysctl_sched_migration_cost
;
5465 #ifdef CONFIG_NUMA_BALANCING
5466 /* Returns true if the destination node has incurred more faults */
5467 static bool migrate_improves_locality(struct task_struct
*p
, struct lb_env
*env
)
5469 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
5470 int src_nid
, dst_nid
;
5472 if (!sched_feat(NUMA_FAVOUR_HIGHER
) || !p
->numa_faults
||
5473 !(env
->sd
->flags
& SD_NUMA
)) {
5477 src_nid
= cpu_to_node(env
->src_cpu
);
5478 dst_nid
= cpu_to_node(env
->dst_cpu
);
5480 if (src_nid
== dst_nid
)
5484 /* Task is already in the group's interleave set. */
5485 if (node_isset(src_nid
, numa_group
->active_nodes
))
5488 /* Task is moving into the group's interleave set. */
5489 if (node_isset(dst_nid
, numa_group
->active_nodes
))
5492 return group_faults(p
, dst_nid
) > group_faults(p
, src_nid
);
5495 /* Encourage migration to the preferred node. */
5496 if (dst_nid
== p
->numa_preferred_nid
)
5499 return task_faults(p
, dst_nid
) > task_faults(p
, src_nid
);
5503 static bool migrate_degrades_locality(struct task_struct
*p
, struct lb_env
*env
)
5505 struct numa_group
*numa_group
= rcu_dereference(p
->numa_group
);
5506 int src_nid
, dst_nid
;
5508 if (!sched_feat(NUMA
) || !sched_feat(NUMA_RESIST_LOWER
))
5511 if (!p
->numa_faults
|| !(env
->sd
->flags
& SD_NUMA
))
5514 src_nid
= cpu_to_node(env
->src_cpu
);
5515 dst_nid
= cpu_to_node(env
->dst_cpu
);
5517 if (src_nid
== dst_nid
)
5521 /* Task is moving within/into the group's interleave set. */
5522 if (node_isset(dst_nid
, numa_group
->active_nodes
))
5525 /* Task is moving out of the group's interleave set. */
5526 if (node_isset(src_nid
, numa_group
->active_nodes
))
5529 return group_faults(p
, dst_nid
) < group_faults(p
, src_nid
);
5532 /* Migrating away from the preferred node is always bad. */
5533 if (src_nid
== p
->numa_preferred_nid
)
5536 return task_faults(p
, dst_nid
) < task_faults(p
, src_nid
);
5540 static inline bool migrate_improves_locality(struct task_struct
*p
,
5546 static inline bool migrate_degrades_locality(struct task_struct
*p
,
5554 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5557 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
5559 int tsk_cache_hot
= 0;
5561 lockdep_assert_held(&env
->src_rq
->lock
);
5564 * We do not migrate tasks that are:
5565 * 1) throttled_lb_pair, or
5566 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5567 * 3) running (obviously), or
5568 * 4) are cache-hot on their current CPU.
5570 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
5573 if (!cpumask_test_cpu(env
->dst_cpu
, tsk_cpus_allowed(p
))) {
5576 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
5578 env
->flags
|= LBF_SOME_PINNED
;
5581 * Remember if this task can be migrated to any other cpu in
5582 * our sched_group. We may want to revisit it if we couldn't
5583 * meet load balance goals by pulling other tasks on src_cpu.
5585 * Also avoid computing new_dst_cpu if we have already computed
5586 * one in current iteration.
5588 if (!env
->dst_grpmask
|| (env
->flags
& LBF_DST_PINNED
))
5591 /* Prevent to re-select dst_cpu via env's cpus */
5592 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
5593 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
))) {
5594 env
->flags
|= LBF_DST_PINNED
;
5595 env
->new_dst_cpu
= cpu
;
5603 /* Record that we found atleast one task that could run on dst_cpu */
5604 env
->flags
&= ~LBF_ALL_PINNED
;
5606 if (task_running(env
->src_rq
, p
)) {
5607 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
5612 * Aggressive migration if:
5613 * 1) destination numa is preferred
5614 * 2) task is cache cold, or
5615 * 3) too many balance attempts have failed.
5617 tsk_cache_hot
= task_hot(p
, env
);
5619 tsk_cache_hot
= migrate_degrades_locality(p
, env
);
5621 if (migrate_improves_locality(p
, env
) || !tsk_cache_hot
||
5622 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
5623 if (tsk_cache_hot
) {
5624 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
5625 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
5630 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
5635 * detach_task() -- detach the task for the migration specified in env
5637 static void detach_task(struct task_struct
*p
, struct lb_env
*env
)
5639 lockdep_assert_held(&env
->src_rq
->lock
);
5641 deactivate_task(env
->src_rq
, p
, 0);
5642 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
5643 set_task_cpu(p
, env
->dst_cpu
);
5647 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5648 * part of active balancing operations within "domain".
5650 * Returns a task if successful and NULL otherwise.
5652 static struct task_struct
*detach_one_task(struct lb_env
*env
)
5654 struct task_struct
*p
, *n
;
5656 lockdep_assert_held(&env
->src_rq
->lock
);
5658 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
5659 if (!can_migrate_task(p
, env
))
5662 detach_task(p
, env
);
5665 * Right now, this is only the second place where
5666 * lb_gained[env->idle] is updated (other is detach_tasks)
5667 * so we can safely collect stats here rather than
5668 * inside detach_tasks().
5670 schedstat_inc(env
->sd
, lb_gained
[env
->idle
]);
5676 static const unsigned int sched_nr_migrate_break
= 32;
5679 * detach_tasks() -- tries to detach up to imbalance weighted load from
5680 * busiest_rq, as part of a balancing operation within domain "sd".
5682 * Returns number of detached tasks if successful and 0 otherwise.
5684 static int detach_tasks(struct lb_env
*env
)
5686 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
5687 struct task_struct
*p
;
5691 lockdep_assert_held(&env
->src_rq
->lock
);
5693 if (env
->imbalance
<= 0)
5696 while (!list_empty(tasks
)) {
5697 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
5700 /* We've more or less seen every task there is, call it quits */
5701 if (env
->loop
> env
->loop_max
)
5704 /* take a breather every nr_migrate tasks */
5705 if (env
->loop
> env
->loop_break
) {
5706 env
->loop_break
+= sched_nr_migrate_break
;
5707 env
->flags
|= LBF_NEED_BREAK
;
5711 if (!can_migrate_task(p
, env
))
5714 load
= task_h_load(p
);
5716 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
5719 if ((load
/ 2) > env
->imbalance
)
5722 detach_task(p
, env
);
5723 list_add(&p
->se
.group_node
, &env
->tasks
);
5726 env
->imbalance
-= load
;
5728 #ifdef CONFIG_PREEMPT
5730 * NEWIDLE balancing is a source of latency, so preemptible
5731 * kernels will stop after the first task is detached to minimize
5732 * the critical section.
5734 if (env
->idle
== CPU_NEWLY_IDLE
)
5739 * We only want to steal up to the prescribed amount of
5742 if (env
->imbalance
<= 0)
5747 list_move_tail(&p
->se
.group_node
, tasks
);
5751 * Right now, this is one of only two places we collect this stat
5752 * so we can safely collect detach_one_task() stats here rather
5753 * than inside detach_one_task().
5755 schedstat_add(env
->sd
, lb_gained
[env
->idle
], detached
);
5761 * attach_task() -- attach the task detached by detach_task() to its new rq.
5763 static void attach_task(struct rq
*rq
, struct task_struct
*p
)
5765 lockdep_assert_held(&rq
->lock
);
5767 BUG_ON(task_rq(p
) != rq
);
5768 p
->on_rq
= TASK_ON_RQ_QUEUED
;
5769 activate_task(rq
, p
, 0);
5770 check_preempt_curr(rq
, p
, 0);
5774 * attach_one_task() -- attaches the task returned from detach_one_task() to
5777 static void attach_one_task(struct rq
*rq
, struct task_struct
*p
)
5779 raw_spin_lock(&rq
->lock
);
5781 raw_spin_unlock(&rq
->lock
);
5785 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5788 static void attach_tasks(struct lb_env
*env
)
5790 struct list_head
*tasks
= &env
->tasks
;
5791 struct task_struct
*p
;
5793 raw_spin_lock(&env
->dst_rq
->lock
);
5795 while (!list_empty(tasks
)) {
5796 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
5797 list_del_init(&p
->se
.group_node
);
5799 attach_task(env
->dst_rq
, p
);
5802 raw_spin_unlock(&env
->dst_rq
->lock
);
5805 #ifdef CONFIG_FAIR_GROUP_SCHED
5807 * update tg->load_weight by folding this cpu's load_avg
5809 static void __update_blocked_averages_cpu(struct task_group
*tg
, int cpu
)
5811 struct sched_entity
*se
= tg
->se
[cpu
];
5812 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu
];
5814 /* throttled entities do not contribute to load */
5815 if (throttled_hierarchy(cfs_rq
))
5818 update_cfs_rq_blocked_load(cfs_rq
, 1);
5821 update_entity_load_avg(se
, 1);
5823 * We pivot on our runnable average having decayed to zero for
5824 * list removal. This generally implies that all our children
5825 * have also been removed (modulo rounding error or bandwidth
5826 * control); however, such cases are rare and we can fix these
5829 * TODO: fix up out-of-order children on enqueue.
5831 if (!se
->avg
.runnable_avg_sum
&& !cfs_rq
->nr_running
)
5832 list_del_leaf_cfs_rq(cfs_rq
);
5834 struct rq
*rq
= rq_of(cfs_rq
);
5835 update_rq_runnable_avg(rq
, rq
->nr_running
);
5839 static void update_blocked_averages(int cpu
)
5841 struct rq
*rq
= cpu_rq(cpu
);
5842 struct cfs_rq
*cfs_rq
;
5843 unsigned long flags
;
5845 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5846 update_rq_clock(rq
);
5848 * Iterates the task_group tree in a bottom up fashion, see
5849 * list_add_leaf_cfs_rq() for details.
5851 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
5853 * Note: We may want to consider periodically releasing
5854 * rq->lock about these updates so that creating many task
5855 * groups does not result in continually extending hold time.
5857 __update_blocked_averages_cpu(cfs_rq
->tg
, rq
->cpu
);
5860 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5864 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5865 * This needs to be done in a top-down fashion because the load of a child
5866 * group is a fraction of its parents load.
5868 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
5870 struct rq
*rq
= rq_of(cfs_rq
);
5871 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
5872 unsigned long now
= jiffies
;
5875 if (cfs_rq
->last_h_load_update
== now
)
5878 cfs_rq
->h_load_next
= NULL
;
5879 for_each_sched_entity(se
) {
5880 cfs_rq
= cfs_rq_of(se
);
5881 cfs_rq
->h_load_next
= se
;
5882 if (cfs_rq
->last_h_load_update
== now
)
5887 cfs_rq
->h_load
= cfs_rq
->runnable_load_avg
;
5888 cfs_rq
->last_h_load_update
= now
;
5891 while ((se
= cfs_rq
->h_load_next
) != NULL
) {
5892 load
= cfs_rq
->h_load
;
5893 load
= div64_ul(load
* se
->avg
.load_avg_contrib
,
5894 cfs_rq
->runnable_load_avg
+ 1);
5895 cfs_rq
= group_cfs_rq(se
);
5896 cfs_rq
->h_load
= load
;
5897 cfs_rq
->last_h_load_update
= now
;
5901 static unsigned long task_h_load(struct task_struct
*p
)
5903 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
5905 update_cfs_rq_h_load(cfs_rq
);
5906 return div64_ul(p
->se
.avg
.load_avg_contrib
* cfs_rq
->h_load
,
5907 cfs_rq
->runnable_load_avg
+ 1);
5910 static inline void update_blocked_averages(int cpu
)
5914 static unsigned long task_h_load(struct task_struct
*p
)
5916 return p
->se
.avg
.load_avg_contrib
;
5920 /********** Helpers for find_busiest_group ************************/
5929 * sg_lb_stats - stats of a sched_group required for load_balancing
5931 struct sg_lb_stats
{
5932 unsigned long avg_load
; /*Avg load across the CPUs of the group */
5933 unsigned long group_load
; /* Total load over the CPUs of the group */
5934 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
5935 unsigned long load_per_task
;
5936 unsigned long group_capacity
;
5937 unsigned long group_usage
; /* Total usage of the group */
5938 unsigned int sum_nr_running
; /* Nr tasks running in the group */
5939 unsigned int group_capacity_factor
;
5940 unsigned int idle_cpus
;
5941 unsigned int group_weight
;
5942 enum group_type group_type
;
5943 int group_has_free_capacity
;
5944 #ifdef CONFIG_NUMA_BALANCING
5945 unsigned int nr_numa_running
;
5946 unsigned int nr_preferred_running
;
5951 * sd_lb_stats - Structure to store the statistics of a sched_domain
5952 * during load balancing.
5954 struct sd_lb_stats
{
5955 struct sched_group
*busiest
; /* Busiest group in this sd */
5956 struct sched_group
*local
; /* Local group in this sd */
5957 unsigned long total_load
; /* Total load of all groups in sd */
5958 unsigned long total_capacity
; /* Total capacity of all groups in sd */
5959 unsigned long avg_load
; /* Average load across all groups in sd */
5961 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
5962 struct sg_lb_stats local_stat
; /* Statistics of the local group */
5965 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
5968 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5969 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5970 * We must however clear busiest_stat::avg_load because
5971 * update_sd_pick_busiest() reads this before assignment.
5973 *sds
= (struct sd_lb_stats
){
5977 .total_capacity
= 0UL,
5980 .sum_nr_running
= 0,
5981 .group_type
= group_other
,
5987 * get_sd_load_idx - Obtain the load index for a given sched domain.
5988 * @sd: The sched_domain whose load_idx is to be obtained.
5989 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5991 * Return: The load index.
5993 static inline int get_sd_load_idx(struct sched_domain
*sd
,
5994 enum cpu_idle_type idle
)
6000 load_idx
= sd
->busy_idx
;
6003 case CPU_NEWLY_IDLE
:
6004 load_idx
= sd
->newidle_idx
;
6007 load_idx
= sd
->idle_idx
;
6014 static unsigned long default_scale_capacity(struct sched_domain
*sd
, int cpu
)
6016 return SCHED_CAPACITY_SCALE
;
6019 unsigned long __weak
arch_scale_freq_capacity(struct sched_domain
*sd
, int cpu
)
6021 return default_scale_capacity(sd
, cpu
);
6024 static unsigned long default_scale_cpu_capacity(struct sched_domain
*sd
, int cpu
)
6026 if ((sd
->flags
& SD_SHARE_CPUCAPACITY
) && (sd
->span_weight
> 1))
6027 return sd
->smt_gain
/ sd
->span_weight
;
6029 return SCHED_CAPACITY_SCALE
;
6032 unsigned long __weak
arch_scale_cpu_capacity(struct sched_domain
*sd
, int cpu
)
6034 return default_scale_cpu_capacity(sd
, cpu
);
6037 static unsigned long scale_rt_capacity(int cpu
)
6039 struct rq
*rq
= cpu_rq(cpu
);
6040 u64 total
, used
, age_stamp
, avg
;
6044 * Since we're reading these variables without serialization make sure
6045 * we read them once before doing sanity checks on them.
6047 age_stamp
= ACCESS_ONCE(rq
->age_stamp
);
6048 avg
= ACCESS_ONCE(rq
->rt_avg
);
6049 delta
= __rq_clock_broken(rq
) - age_stamp
;
6051 if (unlikely(delta
< 0))
6054 total
= sched_avg_period() + delta
;
6056 used
= div_u64(avg
, total
);
6058 if (likely(used
< SCHED_CAPACITY_SCALE
))
6059 return SCHED_CAPACITY_SCALE
- used
;
6064 static void update_cpu_capacity(struct sched_domain
*sd
, int cpu
)
6066 unsigned long capacity
= SCHED_CAPACITY_SCALE
;
6067 struct sched_group
*sdg
= sd
->groups
;
6069 if (sched_feat(ARCH_CAPACITY
))
6070 capacity
*= arch_scale_cpu_capacity(sd
, cpu
);
6072 capacity
*= default_scale_cpu_capacity(sd
, cpu
);
6074 capacity
>>= SCHED_CAPACITY_SHIFT
;
6076 cpu_rq(cpu
)->cpu_capacity_orig
= capacity
;
6077 sdg
->sgc
->capacity_orig
= capacity
;
6079 capacity
*= scale_rt_capacity(cpu
);
6080 capacity
>>= SCHED_CAPACITY_SHIFT
;
6085 cpu_rq(cpu
)->cpu_capacity
= capacity
;
6086 sdg
->sgc
->capacity
= capacity
;
6089 void update_group_capacity(struct sched_domain
*sd
, int cpu
)
6091 struct sched_domain
*child
= sd
->child
;
6092 struct sched_group
*group
, *sdg
= sd
->groups
;
6093 unsigned long capacity
, capacity_orig
;
6094 unsigned long interval
;
6096 interval
= msecs_to_jiffies(sd
->balance_interval
);
6097 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
6098 sdg
->sgc
->next_update
= jiffies
+ interval
;
6101 update_cpu_capacity(sd
, cpu
);
6105 capacity_orig
= capacity
= 0;
6107 if (child
->flags
& SD_OVERLAP
) {
6109 * SD_OVERLAP domains cannot assume that child groups
6110 * span the current group.
6113 for_each_cpu(cpu
, sched_group_cpus(sdg
)) {
6114 struct sched_group_capacity
*sgc
;
6115 struct rq
*rq
= cpu_rq(cpu
);
6118 * build_sched_domains() -> init_sched_groups_capacity()
6119 * gets here before we've attached the domains to the
6122 * Use capacity_of(), which is set irrespective of domains
6123 * in update_cpu_capacity().
6125 * This avoids capacity/capacity_orig from being 0 and
6126 * causing divide-by-zero issues on boot.
6128 * Runtime updates will correct capacity_orig.
6130 if (unlikely(!rq
->sd
)) {
6131 capacity_orig
+= capacity_orig_of(cpu
);
6132 capacity
+= capacity_of(cpu
);
6136 sgc
= rq
->sd
->groups
->sgc
;
6137 capacity_orig
+= sgc
->capacity_orig
;
6138 capacity
+= sgc
->capacity
;
6142 * !SD_OVERLAP domains can assume that child groups
6143 * span the current group.
6146 group
= child
->groups
;
6148 capacity_orig
+= group
->sgc
->capacity_orig
;
6149 capacity
+= group
->sgc
->capacity
;
6150 group
= group
->next
;
6151 } while (group
!= child
->groups
);
6154 sdg
->sgc
->capacity_orig
= capacity_orig
;
6155 sdg
->sgc
->capacity
= capacity
;
6159 * Try and fix up capacity for tiny siblings, this is needed when
6160 * things like SD_ASYM_PACKING need f_b_g to select another sibling
6161 * which on its own isn't powerful enough.
6163 * See update_sd_pick_busiest() and check_asym_packing().
6166 fix_small_capacity(struct sched_domain
*sd
, struct sched_group
*group
)
6169 * Only siblings can have significantly less than SCHED_CAPACITY_SCALE
6171 if (!(sd
->flags
& SD_SHARE_CPUCAPACITY
))
6175 * If ~90% of the cpu_capacity is still there, we're good.
6177 if (group
->sgc
->capacity
* 32 > group
->sgc
->capacity_orig
* 29)
6184 * Group imbalance indicates (and tries to solve) the problem where balancing
6185 * groups is inadequate due to tsk_cpus_allowed() constraints.
6187 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6188 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6191 * { 0 1 2 3 } { 4 5 6 7 }
6194 * If we were to balance group-wise we'd place two tasks in the first group and
6195 * two tasks in the second group. Clearly this is undesired as it will overload
6196 * cpu 3 and leave one of the cpus in the second group unused.
6198 * The current solution to this issue is detecting the skew in the first group
6199 * by noticing the lower domain failed to reach balance and had difficulty
6200 * moving tasks due to affinity constraints.
6202 * When this is so detected; this group becomes a candidate for busiest; see
6203 * update_sd_pick_busiest(). And calculate_imbalance() and
6204 * find_busiest_group() avoid some of the usual balance conditions to allow it
6205 * to create an effective group imbalance.
6207 * This is a somewhat tricky proposition since the next run might not find the
6208 * group imbalance and decide the groups need to be balanced again. A most
6209 * subtle and fragile situation.
6212 static inline int sg_imbalanced(struct sched_group
*group
)
6214 return group
->sgc
->imbalance
;
6218 * Compute the group capacity factor.
6220 * Avoid the issue where N*frac(smt_capacity) >= 1 creates 'phantom' cores by
6221 * first dividing out the smt factor and computing the actual number of cores
6222 * and limit unit capacity with that.
6224 static inline int sg_capacity_factor(struct lb_env
*env
, struct sched_group
*group
)
6226 unsigned int capacity_factor
, smt
, cpus
;
6227 unsigned int capacity
, capacity_orig
;
6229 capacity
= group
->sgc
->capacity
;
6230 capacity_orig
= group
->sgc
->capacity_orig
;
6231 cpus
= group
->group_weight
;
6233 /* smt := ceil(cpus / capacity), assumes: 1 < smt_capacity < 2 */
6234 smt
= DIV_ROUND_UP(SCHED_CAPACITY_SCALE
* cpus
, capacity_orig
);
6235 capacity_factor
= cpus
/ smt
; /* cores */
6237 capacity_factor
= min_t(unsigned,
6238 capacity_factor
, DIV_ROUND_CLOSEST(capacity
, SCHED_CAPACITY_SCALE
));
6239 if (!capacity_factor
)
6240 capacity_factor
= fix_small_capacity(env
->sd
, group
);
6242 return capacity_factor
;
6245 static enum group_type
6246 group_classify(struct sched_group
*group
, struct sg_lb_stats
*sgs
)
6248 if (sgs
->sum_nr_running
> sgs
->group_capacity_factor
)
6249 return group_overloaded
;
6251 if (sg_imbalanced(group
))
6252 return group_imbalanced
;
6258 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6259 * @env: The load balancing environment.
6260 * @group: sched_group whose statistics are to be updated.
6261 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6262 * @local_group: Does group contain this_cpu.
6263 * @sgs: variable to hold the statistics for this group.
6264 * @overload: Indicate more than one runnable task for any CPU.
6266 static inline void update_sg_lb_stats(struct lb_env
*env
,
6267 struct sched_group
*group
, int load_idx
,
6268 int local_group
, struct sg_lb_stats
*sgs
,
6274 memset(sgs
, 0, sizeof(*sgs
));
6276 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
6277 struct rq
*rq
= cpu_rq(i
);
6279 /* Bias balancing toward cpus of our domain */
6281 load
= target_load(i
, load_idx
);
6283 load
= source_load(i
, load_idx
);
6285 sgs
->group_load
+= load
;
6286 sgs
->group_usage
+= get_cpu_usage(i
);
6287 sgs
->sum_nr_running
+= rq
->cfs
.h_nr_running
;
6289 if (rq
->nr_running
> 1)
6292 #ifdef CONFIG_NUMA_BALANCING
6293 sgs
->nr_numa_running
+= rq
->nr_numa_running
;
6294 sgs
->nr_preferred_running
+= rq
->nr_preferred_running
;
6296 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
6301 /* Adjust by relative CPU capacity of the group */
6302 sgs
->group_capacity
= group
->sgc
->capacity
;
6303 sgs
->avg_load
= (sgs
->group_load
*SCHED_CAPACITY_SCALE
) / sgs
->group_capacity
;
6305 if (sgs
->sum_nr_running
)
6306 sgs
->load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
6308 sgs
->group_weight
= group
->group_weight
;
6309 sgs
->group_capacity_factor
= sg_capacity_factor(env
, group
);
6310 sgs
->group_type
= group_classify(group
, sgs
);
6312 if (sgs
->group_capacity_factor
> sgs
->sum_nr_running
)
6313 sgs
->group_has_free_capacity
= 1;
6317 * update_sd_pick_busiest - return 1 on busiest group
6318 * @env: The load balancing environment.
6319 * @sds: sched_domain statistics
6320 * @sg: sched_group candidate to be checked for being the busiest
6321 * @sgs: sched_group statistics
6323 * Determine if @sg is a busier group than the previously selected
6326 * Return: %true if @sg is a busier group than the previously selected
6327 * busiest group. %false otherwise.
6329 static bool update_sd_pick_busiest(struct lb_env
*env
,
6330 struct sd_lb_stats
*sds
,
6331 struct sched_group
*sg
,
6332 struct sg_lb_stats
*sgs
)
6334 struct sg_lb_stats
*busiest
= &sds
->busiest_stat
;
6336 if (sgs
->group_type
> busiest
->group_type
)
6339 if (sgs
->group_type
< busiest
->group_type
)
6342 if (sgs
->avg_load
<= busiest
->avg_load
)
6345 /* This is the busiest node in its class. */
6346 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
6350 * ASYM_PACKING needs to move all the work to the lowest
6351 * numbered CPUs in the group, therefore mark all groups
6352 * higher than ourself as busy.
6354 if (sgs
->sum_nr_running
&& env
->dst_cpu
< group_first_cpu(sg
)) {
6358 if (group_first_cpu(sds
->busiest
) > group_first_cpu(sg
))
6365 #ifdef CONFIG_NUMA_BALANCING
6366 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
6368 if (sgs
->sum_nr_running
> sgs
->nr_numa_running
)
6370 if (sgs
->sum_nr_running
> sgs
->nr_preferred_running
)
6375 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
6377 if (rq
->nr_running
> rq
->nr_numa_running
)
6379 if (rq
->nr_running
> rq
->nr_preferred_running
)
6384 static inline enum fbq_type
fbq_classify_group(struct sg_lb_stats
*sgs
)
6389 static inline enum fbq_type
fbq_classify_rq(struct rq
*rq
)
6393 #endif /* CONFIG_NUMA_BALANCING */
6396 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6397 * @env: The load balancing environment.
6398 * @sds: variable to hold the statistics for this sched_domain.
6400 static inline void update_sd_lb_stats(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6402 struct sched_domain
*child
= env
->sd
->child
;
6403 struct sched_group
*sg
= env
->sd
->groups
;
6404 struct sg_lb_stats tmp_sgs
;
6405 int load_idx
, prefer_sibling
= 0;
6406 bool overload
= false;
6408 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
6411 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
6414 struct sg_lb_stats
*sgs
= &tmp_sgs
;
6417 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_cpus(sg
));
6420 sgs
= &sds
->local_stat
;
6422 if (env
->idle
!= CPU_NEWLY_IDLE
||
6423 time_after_eq(jiffies
, sg
->sgc
->next_update
))
6424 update_group_capacity(env
->sd
, env
->dst_cpu
);
6427 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, sgs
,
6434 * In case the child domain prefers tasks go to siblings
6435 * first, lower the sg capacity factor to one so that we'll try
6436 * and move all the excess tasks away. We lower the capacity
6437 * of a group only if the local group has the capacity to fit
6438 * these excess tasks, i.e. nr_running < group_capacity_factor. The
6439 * extra check prevents the case where you always pull from the
6440 * heaviest group when it is already under-utilized (possible
6441 * with a large weight task outweighs the tasks on the system).
6443 if (prefer_sibling
&& sds
->local
&&
6444 sds
->local_stat
.group_has_free_capacity
) {
6445 sgs
->group_capacity_factor
= min(sgs
->group_capacity_factor
, 1U);
6446 sgs
->group_type
= group_classify(sg
, sgs
);
6449 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
6451 sds
->busiest_stat
= *sgs
;
6455 /* Now, start updating sd_lb_stats */
6456 sds
->total_load
+= sgs
->group_load
;
6457 sds
->total_capacity
+= sgs
->group_capacity
;
6460 } while (sg
!= env
->sd
->groups
);
6462 if (env
->sd
->flags
& SD_NUMA
)
6463 env
->fbq_type
= fbq_classify_group(&sds
->busiest_stat
);
6465 if (!env
->sd
->parent
) {
6466 /* update overload indicator if we are at root domain */
6467 if (env
->dst_rq
->rd
->overload
!= overload
)
6468 env
->dst_rq
->rd
->overload
= overload
;
6474 * check_asym_packing - Check to see if the group is packed into the
6477 * This is primarily intended to used at the sibling level. Some
6478 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6479 * case of POWER7, it can move to lower SMT modes only when higher
6480 * threads are idle. When in lower SMT modes, the threads will
6481 * perform better since they share less core resources. Hence when we
6482 * have idle threads, we want them to be the higher ones.
6484 * This packing function is run on idle threads. It checks to see if
6485 * the busiest CPU in this domain (core in the P7 case) has a higher
6486 * CPU number than the packing function is being run on. Here we are
6487 * assuming lower CPU number will be equivalent to lower a SMT thread
6490 * Return: 1 when packing is required and a task should be moved to
6491 * this CPU. The amount of the imbalance is returned in *imbalance.
6493 * @env: The load balancing environment.
6494 * @sds: Statistics of the sched_domain which is to be packed
6496 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6500 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
6506 busiest_cpu
= group_first_cpu(sds
->busiest
);
6507 if (env
->dst_cpu
> busiest_cpu
)
6510 env
->imbalance
= DIV_ROUND_CLOSEST(
6511 sds
->busiest_stat
.avg_load
* sds
->busiest_stat
.group_capacity
,
6512 SCHED_CAPACITY_SCALE
);
6518 * fix_small_imbalance - Calculate the minor imbalance that exists
6519 * amongst the groups of a sched_domain, during
6521 * @env: The load balancing environment.
6522 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6525 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6527 unsigned long tmp
, capa_now
= 0, capa_move
= 0;
6528 unsigned int imbn
= 2;
6529 unsigned long scaled_busy_load_per_task
;
6530 struct sg_lb_stats
*local
, *busiest
;
6532 local
= &sds
->local_stat
;
6533 busiest
= &sds
->busiest_stat
;
6535 if (!local
->sum_nr_running
)
6536 local
->load_per_task
= cpu_avg_load_per_task(env
->dst_cpu
);
6537 else if (busiest
->load_per_task
> local
->load_per_task
)
6540 scaled_busy_load_per_task
=
6541 (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
6542 busiest
->group_capacity
;
6544 if (busiest
->avg_load
+ scaled_busy_load_per_task
>=
6545 local
->avg_load
+ (scaled_busy_load_per_task
* imbn
)) {
6546 env
->imbalance
= busiest
->load_per_task
;
6551 * OK, we don't have enough imbalance to justify moving tasks,
6552 * however we may be able to increase total CPU capacity used by
6556 capa_now
+= busiest
->group_capacity
*
6557 min(busiest
->load_per_task
, busiest
->avg_load
);
6558 capa_now
+= local
->group_capacity
*
6559 min(local
->load_per_task
, local
->avg_load
);
6560 capa_now
/= SCHED_CAPACITY_SCALE
;
6562 /* Amount of load we'd subtract */
6563 if (busiest
->avg_load
> scaled_busy_load_per_task
) {
6564 capa_move
+= busiest
->group_capacity
*
6565 min(busiest
->load_per_task
,
6566 busiest
->avg_load
- scaled_busy_load_per_task
);
6569 /* Amount of load we'd add */
6570 if (busiest
->avg_load
* busiest
->group_capacity
<
6571 busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) {
6572 tmp
= (busiest
->avg_load
* busiest
->group_capacity
) /
6573 local
->group_capacity
;
6575 tmp
= (busiest
->load_per_task
* SCHED_CAPACITY_SCALE
) /
6576 local
->group_capacity
;
6578 capa_move
+= local
->group_capacity
*
6579 min(local
->load_per_task
, local
->avg_load
+ tmp
);
6580 capa_move
/= SCHED_CAPACITY_SCALE
;
6582 /* Move if we gain throughput */
6583 if (capa_move
> capa_now
)
6584 env
->imbalance
= busiest
->load_per_task
;
6588 * calculate_imbalance - Calculate the amount of imbalance present within the
6589 * groups of a given sched_domain during load balance.
6590 * @env: load balance environment
6591 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6593 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
6595 unsigned long max_pull
, load_above_capacity
= ~0UL;
6596 struct sg_lb_stats
*local
, *busiest
;
6598 local
= &sds
->local_stat
;
6599 busiest
= &sds
->busiest_stat
;
6601 if (busiest
->group_type
== group_imbalanced
) {
6603 * In the group_imb case we cannot rely on group-wide averages
6604 * to ensure cpu-load equilibrium, look at wider averages. XXX
6606 busiest
->load_per_task
=
6607 min(busiest
->load_per_task
, sds
->avg_load
);
6611 * In the presence of smp nice balancing, certain scenarios can have
6612 * max load less than avg load(as we skip the groups at or below
6613 * its cpu_capacity, while calculating max_load..)
6615 if (busiest
->avg_load
<= sds
->avg_load
||
6616 local
->avg_load
>= sds
->avg_load
) {
6618 return fix_small_imbalance(env
, sds
);
6622 * If there aren't any idle cpus, avoid creating some.
6624 if (busiest
->group_type
== group_overloaded
&&
6625 local
->group_type
== group_overloaded
) {
6626 load_above_capacity
=
6627 (busiest
->sum_nr_running
- busiest
->group_capacity_factor
);
6629 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_CAPACITY_SCALE
);
6630 load_above_capacity
/= busiest
->group_capacity
;
6634 * We're trying to get all the cpus to the average_load, so we don't
6635 * want to push ourselves above the average load, nor do we wish to
6636 * reduce the max loaded cpu below the average load. At the same time,
6637 * we also don't want to reduce the group load below the group capacity
6638 * (so that we can implement power-savings policies etc). Thus we look
6639 * for the minimum possible imbalance.
6641 max_pull
= min(busiest
->avg_load
- sds
->avg_load
, load_above_capacity
);
6643 /* How much load to actually move to equalise the imbalance */
6644 env
->imbalance
= min(
6645 max_pull
* busiest
->group_capacity
,
6646 (sds
->avg_load
- local
->avg_load
) * local
->group_capacity
6647 ) / SCHED_CAPACITY_SCALE
;
6650 * if *imbalance is less than the average load per runnable task
6651 * there is no guarantee that any tasks will be moved so we'll have
6652 * a think about bumping its value to force at least one task to be
6655 if (env
->imbalance
< busiest
->load_per_task
)
6656 return fix_small_imbalance(env
, sds
);
6659 /******* find_busiest_group() helpers end here *********************/
6662 * find_busiest_group - Returns the busiest group within the sched_domain
6663 * if there is an imbalance. If there isn't an imbalance, and
6664 * the user has opted for power-savings, it returns a group whose
6665 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6666 * such a group exists.
6668 * Also calculates the amount of weighted load which should be moved
6669 * to restore balance.
6671 * @env: The load balancing environment.
6673 * Return: - The busiest group if imbalance exists.
6674 * - If no imbalance and user has opted for power-savings balance,
6675 * return the least loaded group whose CPUs can be
6676 * put to idle by rebalancing its tasks onto our group.
6678 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
6680 struct sg_lb_stats
*local
, *busiest
;
6681 struct sd_lb_stats sds
;
6683 init_sd_lb_stats(&sds
);
6686 * Compute the various statistics relavent for load balancing at
6689 update_sd_lb_stats(env
, &sds
);
6690 local
= &sds
.local_stat
;
6691 busiest
= &sds
.busiest_stat
;
6693 if ((env
->idle
== CPU_IDLE
|| env
->idle
== CPU_NEWLY_IDLE
) &&
6694 check_asym_packing(env
, &sds
))
6697 /* There is no busy sibling group to pull tasks from */
6698 if (!sds
.busiest
|| busiest
->sum_nr_running
== 0)
6701 sds
.avg_load
= (SCHED_CAPACITY_SCALE
* sds
.total_load
)
6702 / sds
.total_capacity
;
6705 * If the busiest group is imbalanced the below checks don't
6706 * work because they assume all things are equal, which typically
6707 * isn't true due to cpus_allowed constraints and the like.
6709 if (busiest
->group_type
== group_imbalanced
)
6712 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6713 if (env
->idle
== CPU_NEWLY_IDLE
&& local
->group_has_free_capacity
&&
6714 !busiest
->group_has_free_capacity
)
6718 * If the local group is busier than the selected busiest group
6719 * don't try and pull any tasks.
6721 if (local
->avg_load
>= busiest
->avg_load
)
6725 * Don't pull any tasks if this group is already above the domain
6728 if (local
->avg_load
>= sds
.avg_load
)
6731 if (env
->idle
== CPU_IDLE
) {
6733 * This cpu is idle. If the busiest group is not overloaded
6734 * and there is no imbalance between this and busiest group
6735 * wrt idle cpus, it is balanced. The imbalance becomes
6736 * significant if the diff is greater than 1 otherwise we
6737 * might end up to just move the imbalance on another group
6739 if ((busiest
->group_type
!= group_overloaded
) &&
6740 (local
->idle_cpus
<= (busiest
->idle_cpus
+ 1)))
6744 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6745 * imbalance_pct to be conservative.
6747 if (100 * busiest
->avg_load
<=
6748 env
->sd
->imbalance_pct
* local
->avg_load
)
6753 /* Looks like there is an imbalance. Compute it */
6754 calculate_imbalance(env
, &sds
);
6763 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6765 static struct rq
*find_busiest_queue(struct lb_env
*env
,
6766 struct sched_group
*group
)
6768 struct rq
*busiest
= NULL
, *rq
;
6769 unsigned long busiest_load
= 0, busiest_capacity
= 1;
6772 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
6773 unsigned long capacity
, capacity_factor
, wl
;
6777 rt
= fbq_classify_rq(rq
);
6780 * We classify groups/runqueues into three groups:
6781 * - regular: there are !numa tasks
6782 * - remote: there are numa tasks that run on the 'wrong' node
6783 * - all: there is no distinction
6785 * In order to avoid migrating ideally placed numa tasks,
6786 * ignore those when there's better options.
6788 * If we ignore the actual busiest queue to migrate another
6789 * task, the next balance pass can still reduce the busiest
6790 * queue by moving tasks around inside the node.
6792 * If we cannot move enough load due to this classification
6793 * the next pass will adjust the group classification and
6794 * allow migration of more tasks.
6796 * Both cases only affect the total convergence complexity.
6798 if (rt
> env
->fbq_type
)
6801 capacity
= capacity_of(i
);
6802 capacity_factor
= DIV_ROUND_CLOSEST(capacity
, SCHED_CAPACITY_SCALE
);
6803 if (!capacity_factor
)
6804 capacity_factor
= fix_small_capacity(env
->sd
, group
);
6806 wl
= weighted_cpuload(i
);
6809 * When comparing with imbalance, use weighted_cpuload()
6810 * which is not scaled with the cpu capacity.
6812 if (capacity_factor
&& rq
->nr_running
== 1 && wl
> env
->imbalance
)
6816 * For the load comparisons with the other cpu's, consider
6817 * the weighted_cpuload() scaled with the cpu capacity, so
6818 * that the load can be moved away from the cpu that is
6819 * potentially running at a lower capacity.
6821 * Thus we're looking for max(wl_i / capacity_i), crosswise
6822 * multiplication to rid ourselves of the division works out
6823 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6824 * our previous maximum.
6826 if (wl
* busiest_capacity
> busiest_load
* capacity
) {
6828 busiest_capacity
= capacity
;
6837 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6838 * so long as it is large enough.
6840 #define MAX_PINNED_INTERVAL 512
6842 /* Working cpumask for load_balance and load_balance_newidle. */
6843 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6845 static int need_active_balance(struct lb_env
*env
)
6847 struct sched_domain
*sd
= env
->sd
;
6849 if (env
->idle
== CPU_NEWLY_IDLE
) {
6852 * ASYM_PACKING needs to force migrate tasks from busy but
6853 * higher numbered CPUs in order to pack all tasks in the
6854 * lowest numbered CPUs.
6856 if ((sd
->flags
& SD_ASYM_PACKING
) && env
->src_cpu
> env
->dst_cpu
)
6860 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
6863 static int active_load_balance_cpu_stop(void *data
);
6865 static int should_we_balance(struct lb_env
*env
)
6867 struct sched_group
*sg
= env
->sd
->groups
;
6868 struct cpumask
*sg_cpus
, *sg_mask
;
6869 int cpu
, balance_cpu
= -1;
6872 * In the newly idle case, we will allow all the cpu's
6873 * to do the newly idle load balance.
6875 if (env
->idle
== CPU_NEWLY_IDLE
)
6878 sg_cpus
= sched_group_cpus(sg
);
6879 sg_mask
= sched_group_mask(sg
);
6880 /* Try to find first idle cpu */
6881 for_each_cpu_and(cpu
, sg_cpus
, env
->cpus
) {
6882 if (!cpumask_test_cpu(cpu
, sg_mask
) || !idle_cpu(cpu
))
6889 if (balance_cpu
== -1)
6890 balance_cpu
= group_balance_cpu(sg
);
6893 * First idle cpu or the first cpu(busiest) in this sched group
6894 * is eligible for doing load balancing at this and above domains.
6896 return balance_cpu
== env
->dst_cpu
;
6900 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6901 * tasks if there is an imbalance.
6903 static int load_balance(int this_cpu
, struct rq
*this_rq
,
6904 struct sched_domain
*sd
, enum cpu_idle_type idle
,
6905 int *continue_balancing
)
6907 int ld_moved
, cur_ld_moved
, active_balance
= 0;
6908 struct sched_domain
*sd_parent
= sd
->parent
;
6909 struct sched_group
*group
;
6911 unsigned long flags
;
6912 struct cpumask
*cpus
= this_cpu_cpumask_var_ptr(load_balance_mask
);
6914 struct lb_env env
= {
6916 .dst_cpu
= this_cpu
,
6918 .dst_grpmask
= sched_group_cpus(sd
->groups
),
6920 .loop_break
= sched_nr_migrate_break
,
6923 .tasks
= LIST_HEAD_INIT(env
.tasks
),
6927 * For NEWLY_IDLE load_balancing, we don't need to consider
6928 * other cpus in our group
6930 if (idle
== CPU_NEWLY_IDLE
)
6931 env
.dst_grpmask
= NULL
;
6933 cpumask_copy(cpus
, cpu_active_mask
);
6935 schedstat_inc(sd
, lb_count
[idle
]);
6938 if (!should_we_balance(&env
)) {
6939 *continue_balancing
= 0;
6943 group
= find_busiest_group(&env
);
6945 schedstat_inc(sd
, lb_nobusyg
[idle
]);
6949 busiest
= find_busiest_queue(&env
, group
);
6951 schedstat_inc(sd
, lb_nobusyq
[idle
]);
6955 BUG_ON(busiest
== env
.dst_rq
);
6957 schedstat_add(sd
, lb_imbalance
[idle
], env
.imbalance
);
6960 if (busiest
->nr_running
> 1) {
6962 * Attempt to move tasks. If find_busiest_group has found
6963 * an imbalance but busiest->nr_running <= 1, the group is
6964 * still unbalanced. ld_moved simply stays zero, so it is
6965 * correctly treated as an imbalance.
6967 env
.flags
|= LBF_ALL_PINNED
;
6968 env
.src_cpu
= busiest
->cpu
;
6969 env
.src_rq
= busiest
;
6970 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
6973 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
6976 * cur_ld_moved - load moved in current iteration
6977 * ld_moved - cumulative load moved across iterations
6979 cur_ld_moved
= detach_tasks(&env
);
6982 * We've detached some tasks from busiest_rq. Every
6983 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
6984 * unlock busiest->lock, and we are able to be sure
6985 * that nobody can manipulate the tasks in parallel.
6986 * See task_rq_lock() family for the details.
6989 raw_spin_unlock(&busiest
->lock
);
6993 ld_moved
+= cur_ld_moved
;
6996 local_irq_restore(flags
);
6998 if (env
.flags
& LBF_NEED_BREAK
) {
6999 env
.flags
&= ~LBF_NEED_BREAK
;
7004 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7005 * us and move them to an alternate dst_cpu in our sched_group
7006 * where they can run. The upper limit on how many times we
7007 * iterate on same src_cpu is dependent on number of cpus in our
7010 * This changes load balance semantics a bit on who can move
7011 * load to a given_cpu. In addition to the given_cpu itself
7012 * (or a ilb_cpu acting on its behalf where given_cpu is
7013 * nohz-idle), we now have balance_cpu in a position to move
7014 * load to given_cpu. In rare situations, this may cause
7015 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7016 * _independently_ and at _same_ time to move some load to
7017 * given_cpu) causing exceess load to be moved to given_cpu.
7018 * This however should not happen so much in practice and
7019 * moreover subsequent load balance cycles should correct the
7020 * excess load moved.
7022 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
7024 /* Prevent to re-select dst_cpu via env's cpus */
7025 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
7027 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
7028 env
.dst_cpu
= env
.new_dst_cpu
;
7029 env
.flags
&= ~LBF_DST_PINNED
;
7031 env
.loop_break
= sched_nr_migrate_break
;
7034 * Go back to "more_balance" rather than "redo" since we
7035 * need to continue with same src_cpu.
7041 * We failed to reach balance because of affinity.
7044 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
7046 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0)
7047 *group_imbalance
= 1;
7050 /* All tasks on this runqueue were pinned by CPU affinity */
7051 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
7052 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
7053 if (!cpumask_empty(cpus
)) {
7055 env
.loop_break
= sched_nr_migrate_break
;
7058 goto out_all_pinned
;
7063 schedstat_inc(sd
, lb_failed
[idle
]);
7065 * Increment the failure counter only on periodic balance.
7066 * We do not want newidle balance, which can be very
7067 * frequent, pollute the failure counter causing
7068 * excessive cache_hot migrations and active balances.
7070 if (idle
!= CPU_NEWLY_IDLE
)
7071 sd
->nr_balance_failed
++;
7073 if (need_active_balance(&env
)) {
7074 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
7076 /* don't kick the active_load_balance_cpu_stop,
7077 * if the curr task on busiest cpu can't be
7080 if (!cpumask_test_cpu(this_cpu
,
7081 tsk_cpus_allowed(busiest
->curr
))) {
7082 raw_spin_unlock_irqrestore(&busiest
->lock
,
7084 env
.flags
|= LBF_ALL_PINNED
;
7085 goto out_one_pinned
;
7089 * ->active_balance synchronizes accesses to
7090 * ->active_balance_work. Once set, it's cleared
7091 * only after active load balance is finished.
7093 if (!busiest
->active_balance
) {
7094 busiest
->active_balance
= 1;
7095 busiest
->push_cpu
= this_cpu
;
7098 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
7100 if (active_balance
) {
7101 stop_one_cpu_nowait(cpu_of(busiest
),
7102 active_load_balance_cpu_stop
, busiest
,
7103 &busiest
->active_balance_work
);
7107 * We've kicked active balancing, reset the failure
7110 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
7113 sd
->nr_balance_failed
= 0;
7115 if (likely(!active_balance
)) {
7116 /* We were unbalanced, so reset the balancing interval */
7117 sd
->balance_interval
= sd
->min_interval
;
7120 * If we've begun active balancing, start to back off. This
7121 * case may not be covered by the all_pinned logic if there
7122 * is only 1 task on the busy runqueue (because we don't call
7125 if (sd
->balance_interval
< sd
->max_interval
)
7126 sd
->balance_interval
*= 2;
7133 * We reach balance although we may have faced some affinity
7134 * constraints. Clear the imbalance flag if it was set.
7137 int *group_imbalance
= &sd_parent
->groups
->sgc
->imbalance
;
7139 if (*group_imbalance
)
7140 *group_imbalance
= 0;
7145 * We reach balance because all tasks are pinned at this level so
7146 * we can't migrate them. Let the imbalance flag set so parent level
7147 * can try to migrate them.
7149 schedstat_inc(sd
, lb_balanced
[idle
]);
7151 sd
->nr_balance_failed
= 0;
7154 /* tune up the balancing interval */
7155 if (((env
.flags
& LBF_ALL_PINNED
) &&
7156 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
7157 (sd
->balance_interval
< sd
->max_interval
))
7158 sd
->balance_interval
*= 2;
7165 static inline unsigned long
7166 get_sd_balance_interval(struct sched_domain
*sd
, int cpu_busy
)
7168 unsigned long interval
= sd
->balance_interval
;
7171 interval
*= sd
->busy_factor
;
7173 /* scale ms to jiffies */
7174 interval
= msecs_to_jiffies(interval
);
7175 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
7181 update_next_balance(struct sched_domain
*sd
, int cpu_busy
, unsigned long *next_balance
)
7183 unsigned long interval
, next
;
7185 interval
= get_sd_balance_interval(sd
, cpu_busy
);
7186 next
= sd
->last_balance
+ interval
;
7188 if (time_after(*next_balance
, next
))
7189 *next_balance
= next
;
7193 * idle_balance is called by schedule() if this_cpu is about to become
7194 * idle. Attempts to pull tasks from other CPUs.
7196 static int idle_balance(struct rq
*this_rq
)
7198 unsigned long next_balance
= jiffies
+ HZ
;
7199 int this_cpu
= this_rq
->cpu
;
7200 struct sched_domain
*sd
;
7201 int pulled_task
= 0;
7204 idle_enter_fair(this_rq
);
7207 * We must set idle_stamp _before_ calling idle_balance(), such that we
7208 * measure the duration of idle_balance() as idle time.
7210 this_rq
->idle_stamp
= rq_clock(this_rq
);
7212 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
||
7213 !this_rq
->rd
->overload
) {
7215 sd
= rcu_dereference_check_sched_domain(this_rq
->sd
);
7217 update_next_balance(sd
, 0, &next_balance
);
7224 * Drop the rq->lock, but keep IRQ/preempt disabled.
7226 raw_spin_unlock(&this_rq
->lock
);
7228 update_blocked_averages(this_cpu
);
7230 for_each_domain(this_cpu
, sd
) {
7231 int continue_balancing
= 1;
7232 u64 t0
, domain_cost
;
7234 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7237 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
) {
7238 update_next_balance(sd
, 0, &next_balance
);
7242 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
7243 t0
= sched_clock_cpu(this_cpu
);
7245 pulled_task
= load_balance(this_cpu
, this_rq
,
7247 &continue_balancing
);
7249 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
7250 if (domain_cost
> sd
->max_newidle_lb_cost
)
7251 sd
->max_newidle_lb_cost
= domain_cost
;
7253 curr_cost
+= domain_cost
;
7256 update_next_balance(sd
, 0, &next_balance
);
7259 * Stop searching for tasks to pull if there are
7260 * now runnable tasks on this rq.
7262 if (pulled_task
|| this_rq
->nr_running
> 0)
7267 raw_spin_lock(&this_rq
->lock
);
7269 if (curr_cost
> this_rq
->max_idle_balance_cost
)
7270 this_rq
->max_idle_balance_cost
= curr_cost
;
7273 * While browsing the domains, we released the rq lock, a task could
7274 * have been enqueued in the meantime. Since we're not going idle,
7275 * pretend we pulled a task.
7277 if (this_rq
->cfs
.h_nr_running
&& !pulled_task
)
7281 /* Move the next balance forward */
7282 if (time_after(this_rq
->next_balance
, next_balance
))
7283 this_rq
->next_balance
= next_balance
;
7285 /* Is there a task of a high priority class? */
7286 if (this_rq
->nr_running
!= this_rq
->cfs
.h_nr_running
)
7290 idle_exit_fair(this_rq
);
7291 this_rq
->idle_stamp
= 0;
7298 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7299 * running tasks off the busiest CPU onto idle CPUs. It requires at
7300 * least 1 task to be running on each physical CPU where possible, and
7301 * avoids physical / logical imbalances.
7303 static int active_load_balance_cpu_stop(void *data
)
7305 struct rq
*busiest_rq
= data
;
7306 int busiest_cpu
= cpu_of(busiest_rq
);
7307 int target_cpu
= busiest_rq
->push_cpu
;
7308 struct rq
*target_rq
= cpu_rq(target_cpu
);
7309 struct sched_domain
*sd
;
7310 struct task_struct
*p
= NULL
;
7312 raw_spin_lock_irq(&busiest_rq
->lock
);
7314 /* make sure the requested cpu hasn't gone down in the meantime */
7315 if (unlikely(busiest_cpu
!= smp_processor_id() ||
7316 !busiest_rq
->active_balance
))
7319 /* Is there any task to move? */
7320 if (busiest_rq
->nr_running
<= 1)
7324 * This condition is "impossible", if it occurs
7325 * we need to fix it. Originally reported by
7326 * Bjorn Helgaas on a 128-cpu setup.
7328 BUG_ON(busiest_rq
== target_rq
);
7330 /* Search for an sd spanning us and the target CPU. */
7332 for_each_domain(target_cpu
, sd
) {
7333 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
7334 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
7339 struct lb_env env
= {
7341 .dst_cpu
= target_cpu
,
7342 .dst_rq
= target_rq
,
7343 .src_cpu
= busiest_rq
->cpu
,
7344 .src_rq
= busiest_rq
,
7348 schedstat_inc(sd
, alb_count
);
7350 p
= detach_one_task(&env
);
7352 schedstat_inc(sd
, alb_pushed
);
7354 schedstat_inc(sd
, alb_failed
);
7358 busiest_rq
->active_balance
= 0;
7359 raw_spin_unlock(&busiest_rq
->lock
);
7362 attach_one_task(target_rq
, p
);
7369 static inline int on_null_domain(struct rq
*rq
)
7371 return unlikely(!rcu_dereference_sched(rq
->sd
));
7374 #ifdef CONFIG_NO_HZ_COMMON
7376 * idle load balancing details
7377 * - When one of the busy CPUs notice that there may be an idle rebalancing
7378 * needed, they will kick the idle load balancer, which then does idle
7379 * load balancing for all the idle CPUs.
7382 cpumask_var_t idle_cpus_mask
;
7384 unsigned long next_balance
; /* in jiffy units */
7385 } nohz ____cacheline_aligned
;
7387 static inline int find_new_ilb(void)
7389 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
7391 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
7398 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7399 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7400 * CPU (if there is one).
7402 static void nohz_balancer_kick(void)
7406 nohz
.next_balance
++;
7408 ilb_cpu
= find_new_ilb();
7410 if (ilb_cpu
>= nr_cpu_ids
)
7413 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
7416 * Use smp_send_reschedule() instead of resched_cpu().
7417 * This way we generate a sched IPI on the target cpu which
7418 * is idle. And the softirq performing nohz idle load balance
7419 * will be run before returning from the IPI.
7421 smp_send_reschedule(ilb_cpu
);
7425 static inline void nohz_balance_exit_idle(int cpu
)
7427 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
7429 * Completely isolated CPUs don't ever set, so we must test.
7431 if (likely(cpumask_test_cpu(cpu
, nohz
.idle_cpus_mask
))) {
7432 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
7433 atomic_dec(&nohz
.nr_cpus
);
7435 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
7439 static inline void set_cpu_sd_state_busy(void)
7441 struct sched_domain
*sd
;
7442 int cpu
= smp_processor_id();
7445 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7447 if (!sd
|| !sd
->nohz_idle
)
7451 atomic_inc(&sd
->groups
->sgc
->nr_busy_cpus
);
7456 void set_cpu_sd_state_idle(void)
7458 struct sched_domain
*sd
;
7459 int cpu
= smp_processor_id();
7462 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7464 if (!sd
|| sd
->nohz_idle
)
7468 atomic_dec(&sd
->groups
->sgc
->nr_busy_cpus
);
7474 * This routine will record that the cpu is going idle with tick stopped.
7475 * This info will be used in performing idle load balancing in the future.
7477 void nohz_balance_enter_idle(int cpu
)
7480 * If this cpu is going down, then nothing needs to be done.
7482 if (!cpu_active(cpu
))
7485 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
7489 * If we're a completely isolated CPU, we don't play.
7491 if (on_null_domain(cpu_rq(cpu
)))
7494 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
7495 atomic_inc(&nohz
.nr_cpus
);
7496 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
7499 static int sched_ilb_notifier(struct notifier_block
*nfb
,
7500 unsigned long action
, void *hcpu
)
7502 switch (action
& ~CPU_TASKS_FROZEN
) {
7504 nohz_balance_exit_idle(smp_processor_id());
7512 static DEFINE_SPINLOCK(balancing
);
7515 * Scale the max load_balance interval with the number of CPUs in the system.
7516 * This trades load-balance latency on larger machines for less cross talk.
7518 void update_max_interval(void)
7520 max_load_balance_interval
= HZ
*num_online_cpus()/10;
7524 * It checks each scheduling domain to see if it is due to be balanced,
7525 * and initiates a balancing operation if so.
7527 * Balancing parameters are set up in init_sched_domains.
7529 static void rebalance_domains(struct rq
*rq
, enum cpu_idle_type idle
)
7531 int continue_balancing
= 1;
7533 unsigned long interval
;
7534 struct sched_domain
*sd
;
7535 /* Earliest time when we have to do rebalance again */
7536 unsigned long next_balance
= jiffies
+ 60*HZ
;
7537 int update_next_balance
= 0;
7538 int need_serialize
, need_decay
= 0;
7541 update_blocked_averages(cpu
);
7544 for_each_domain(cpu
, sd
) {
7546 * Decay the newidle max times here because this is a regular
7547 * visit to all the domains. Decay ~1% per second.
7549 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
7550 sd
->max_newidle_lb_cost
=
7551 (sd
->max_newidle_lb_cost
* 253) / 256;
7552 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
7555 max_cost
+= sd
->max_newidle_lb_cost
;
7557 if (!(sd
->flags
& SD_LOAD_BALANCE
))
7561 * Stop the load balance at this level. There is another
7562 * CPU in our sched group which is doing load balancing more
7565 if (!continue_balancing
) {
7571 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7573 need_serialize
= sd
->flags
& SD_SERIALIZE
;
7574 if (need_serialize
) {
7575 if (!spin_trylock(&balancing
))
7579 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
7580 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
7582 * The LBF_DST_PINNED logic could have changed
7583 * env->dst_cpu, so we can't know our idle
7584 * state even if we migrated tasks. Update it.
7586 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
7588 sd
->last_balance
= jiffies
;
7589 interval
= get_sd_balance_interval(sd
, idle
!= CPU_IDLE
);
7592 spin_unlock(&balancing
);
7594 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
7595 next_balance
= sd
->last_balance
+ interval
;
7596 update_next_balance
= 1;
7601 * Ensure the rq-wide value also decays but keep it at a
7602 * reasonable floor to avoid funnies with rq->avg_idle.
7604 rq
->max_idle_balance_cost
=
7605 max((u64
)sysctl_sched_migration_cost
, max_cost
);
7610 * next_balance will be updated only when there is a need.
7611 * When the cpu is attached to null domain for ex, it will not be
7614 if (likely(update_next_balance
))
7615 rq
->next_balance
= next_balance
;
7618 #ifdef CONFIG_NO_HZ_COMMON
7620 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7621 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7623 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
)
7625 int this_cpu
= this_rq
->cpu
;
7629 if (idle
!= CPU_IDLE
||
7630 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
7633 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
7634 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
7638 * If this cpu gets work to do, stop the load balancing
7639 * work being done for other cpus. Next load
7640 * balancing owner will pick it up.
7645 rq
= cpu_rq(balance_cpu
);
7648 * If time for next balance is due,
7651 if (time_after_eq(jiffies
, rq
->next_balance
)) {
7652 raw_spin_lock_irq(&rq
->lock
);
7653 update_rq_clock(rq
);
7654 update_idle_cpu_load(rq
);
7655 raw_spin_unlock_irq(&rq
->lock
);
7656 rebalance_domains(rq
, CPU_IDLE
);
7659 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
7660 this_rq
->next_balance
= rq
->next_balance
;
7662 nohz
.next_balance
= this_rq
->next_balance
;
7664 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
7668 * Current heuristic for kicking the idle load balancer in the presence
7669 * of an idle cpu is the system.
7670 * - This rq has more than one task.
7671 * - At any scheduler domain level, this cpu's scheduler group has multiple
7672 * busy cpu's exceeding the group's capacity.
7673 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7674 * domain span are idle.
7676 static inline int nohz_kick_needed(struct rq
*rq
)
7678 unsigned long now
= jiffies
;
7679 struct sched_domain
*sd
;
7680 struct sched_group_capacity
*sgc
;
7681 int nr_busy
, cpu
= rq
->cpu
;
7683 if (unlikely(rq
->idle_balance
))
7687 * We may be recently in ticked or tickless idle mode. At the first
7688 * busy tick after returning from idle, we will update the busy stats.
7690 set_cpu_sd_state_busy();
7691 nohz_balance_exit_idle(cpu
);
7694 * None are in tickless mode and hence no need for NOHZ idle load
7697 if (likely(!atomic_read(&nohz
.nr_cpus
)))
7700 if (time_before(now
, nohz
.next_balance
))
7703 if (rq
->nr_running
>= 2)
7707 sd
= rcu_dereference(per_cpu(sd_busy
, cpu
));
7710 sgc
= sd
->groups
->sgc
;
7711 nr_busy
= atomic_read(&sgc
->nr_busy_cpus
);
7714 goto need_kick_unlock
;
7717 sd
= rcu_dereference(per_cpu(sd_asym
, cpu
));
7719 if (sd
&& (cpumask_first_and(nohz
.idle_cpus_mask
,
7720 sched_domain_span(sd
)) < cpu
))
7721 goto need_kick_unlock
;
7732 static void nohz_idle_balance(struct rq
*this_rq
, enum cpu_idle_type idle
) { }
7736 * run_rebalance_domains is triggered when needed from the scheduler tick.
7737 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7739 static void run_rebalance_domains(struct softirq_action
*h
)
7741 struct rq
*this_rq
= this_rq();
7742 enum cpu_idle_type idle
= this_rq
->idle_balance
?
7743 CPU_IDLE
: CPU_NOT_IDLE
;
7745 rebalance_domains(this_rq
, idle
);
7748 * If this cpu has a pending nohz_balance_kick, then do the
7749 * balancing on behalf of the other idle cpus whose ticks are
7752 nohz_idle_balance(this_rq
, idle
);
7756 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7758 void trigger_load_balance(struct rq
*rq
)
7760 /* Don't need to rebalance while attached to NULL domain */
7761 if (unlikely(on_null_domain(rq
)))
7764 if (time_after_eq(jiffies
, rq
->next_balance
))
7765 raise_softirq(SCHED_SOFTIRQ
);
7766 #ifdef CONFIG_NO_HZ_COMMON
7767 if (nohz_kick_needed(rq
))
7768 nohz_balancer_kick();
7772 static void rq_online_fair(struct rq
*rq
)
7776 update_runtime_enabled(rq
);
7779 static void rq_offline_fair(struct rq
*rq
)
7783 /* Ensure any throttled groups are reachable by pick_next_task */
7784 unthrottle_offline_cfs_rqs(rq
);
7787 #endif /* CONFIG_SMP */
7790 * scheduler tick hitting a task of our scheduling class:
7792 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
7794 struct cfs_rq
*cfs_rq
;
7795 struct sched_entity
*se
= &curr
->se
;
7797 for_each_sched_entity(se
) {
7798 cfs_rq
= cfs_rq_of(se
);
7799 entity_tick(cfs_rq
, se
, queued
);
7802 if (numabalancing_enabled
)
7803 task_tick_numa(rq
, curr
);
7805 update_rq_runnable_avg(rq
, 1);
7809 * called on fork with the child task as argument from the parent's context
7810 * - child not yet on the tasklist
7811 * - preemption disabled
7813 static void task_fork_fair(struct task_struct
*p
)
7815 struct cfs_rq
*cfs_rq
;
7816 struct sched_entity
*se
= &p
->se
, *curr
;
7817 int this_cpu
= smp_processor_id();
7818 struct rq
*rq
= this_rq();
7819 unsigned long flags
;
7821 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7823 update_rq_clock(rq
);
7825 cfs_rq
= task_cfs_rq(current
);
7826 curr
= cfs_rq
->curr
;
7829 * Not only the cpu but also the task_group of the parent might have
7830 * been changed after parent->se.parent,cfs_rq were copied to
7831 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7832 * of child point to valid ones.
7835 __set_task_cpu(p
, this_cpu
);
7838 update_curr(cfs_rq
);
7841 se
->vruntime
= curr
->vruntime
;
7842 place_entity(cfs_rq
, se
, 1);
7844 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
7846 * Upon rescheduling, sched_class::put_prev_task() will place
7847 * 'current' within the tree based on its new key value.
7849 swap(curr
->vruntime
, se
->vruntime
);
7853 se
->vruntime
-= cfs_rq
->min_vruntime
;
7855 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7859 * Priority of the task has changed. Check to see if we preempt
7863 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
7865 if (!task_on_rq_queued(p
))
7869 * Reschedule if we are currently running on this runqueue and
7870 * our priority decreased, or if we are not currently running on
7871 * this runqueue and our priority is higher than the current's
7873 if (rq
->curr
== p
) {
7874 if (p
->prio
> oldprio
)
7877 check_preempt_curr(rq
, p
, 0);
7880 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
7882 struct sched_entity
*se
= &p
->se
;
7883 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7886 * Ensure the task's vruntime is normalized, so that when it's
7887 * switched back to the fair class the enqueue_entity(.flags=0) will
7888 * do the right thing.
7890 * If it's queued, then the dequeue_entity(.flags=0) will already
7891 * have normalized the vruntime, if it's !queued, then only when
7892 * the task is sleeping will it still have non-normalized vruntime.
7894 if (!task_on_rq_queued(p
) && p
->state
!= TASK_RUNNING
) {
7896 * Fix up our vruntime so that the current sleep doesn't
7897 * cause 'unlimited' sleep bonus.
7899 place_entity(cfs_rq
, se
, 0);
7900 se
->vruntime
-= cfs_rq
->min_vruntime
;
7905 * Remove our load from contribution when we leave sched_fair
7906 * and ensure we don't carry in an old decay_count if we
7909 if (se
->avg
.decay_count
) {
7910 __synchronize_entity_decay(se
);
7911 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
7917 * We switched to the sched_fair class.
7919 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
7921 #ifdef CONFIG_FAIR_GROUP_SCHED
7922 struct sched_entity
*se
= &p
->se
;
7924 * Since the real-depth could have been changed (only FAIR
7925 * class maintain depth value), reset depth properly.
7927 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
7929 if (!task_on_rq_queued(p
))
7933 * We were most likely switched from sched_rt, so
7934 * kick off the schedule if running, otherwise just see
7935 * if we can still preempt the current task.
7940 check_preempt_curr(rq
, p
, 0);
7943 /* Account for a task changing its policy or group.
7945 * This routine is mostly called to set cfs_rq->curr field when a task
7946 * migrates between groups/classes.
7948 static void set_curr_task_fair(struct rq
*rq
)
7950 struct sched_entity
*se
= &rq
->curr
->se
;
7952 for_each_sched_entity(se
) {
7953 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
7955 set_next_entity(cfs_rq
, se
);
7956 /* ensure bandwidth has been allocated on our new cfs_rq */
7957 account_cfs_rq_runtime(cfs_rq
, 0);
7961 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
7963 cfs_rq
->tasks_timeline
= RB_ROOT
;
7964 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7965 #ifndef CONFIG_64BIT
7966 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
7969 atomic64_set(&cfs_rq
->decay_counter
, 1);
7970 atomic_long_set(&cfs_rq
->removed_load
, 0);
7974 #ifdef CONFIG_FAIR_GROUP_SCHED
7975 static void task_move_group_fair(struct task_struct
*p
, int queued
)
7977 struct sched_entity
*se
= &p
->se
;
7978 struct cfs_rq
*cfs_rq
;
7981 * If the task was not on the rq at the time of this cgroup movement
7982 * it must have been asleep, sleeping tasks keep their ->vruntime
7983 * absolute on their old rq until wakeup (needed for the fair sleeper
7984 * bonus in place_entity()).
7986 * If it was on the rq, we've just 'preempted' it, which does convert
7987 * ->vruntime to a relative base.
7989 * Make sure both cases convert their relative position when migrating
7990 * to another cgroup's rq. This does somewhat interfere with the
7991 * fair sleeper stuff for the first placement, but who cares.
7994 * When !queued, vruntime of the task has usually NOT been normalized.
7995 * But there are some cases where it has already been normalized:
7997 * - Moving a forked child which is waiting for being woken up by
7998 * wake_up_new_task().
7999 * - Moving a task which has been woken up by try_to_wake_up() and
8000 * waiting for actually being woken up by sched_ttwu_pending().
8002 * To prevent boost or penalty in the new cfs_rq caused by delta
8003 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
8005 if (!queued
&& (!se
->sum_exec_runtime
|| p
->state
== TASK_WAKING
))
8009 se
->vruntime
-= cfs_rq_of(se
)->min_vruntime
;
8010 set_task_rq(p
, task_cpu(p
));
8011 se
->depth
= se
->parent
? se
->parent
->depth
+ 1 : 0;
8013 cfs_rq
= cfs_rq_of(se
);
8014 se
->vruntime
+= cfs_rq
->min_vruntime
;
8017 * migrate_task_rq_fair() will have removed our previous
8018 * contribution, but we must synchronize for ongoing future
8021 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
8022 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
8027 void free_fair_sched_group(struct task_group
*tg
)
8031 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
8033 for_each_possible_cpu(i
) {
8035 kfree(tg
->cfs_rq
[i
]);
8044 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8046 struct cfs_rq
*cfs_rq
;
8047 struct sched_entity
*se
;
8050 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8053 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8057 tg
->shares
= NICE_0_LOAD
;
8059 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
8061 for_each_possible_cpu(i
) {
8062 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
8063 GFP_KERNEL
, cpu_to_node(i
));
8067 se
= kzalloc_node(sizeof(struct sched_entity
),
8068 GFP_KERNEL
, cpu_to_node(i
));
8072 init_cfs_rq(cfs_rq
);
8073 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
8084 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8086 struct rq
*rq
= cpu_rq(cpu
);
8087 unsigned long flags
;
8090 * Only empty task groups can be destroyed; so we can speculatively
8091 * check on_list without danger of it being re-added.
8093 if (!tg
->cfs_rq
[cpu
]->on_list
)
8096 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8097 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
8098 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8101 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8102 struct sched_entity
*se
, int cpu
,
8103 struct sched_entity
*parent
)
8105 struct rq
*rq
= cpu_rq(cpu
);
8109 init_cfs_rq_runtime(cfs_rq
);
8111 tg
->cfs_rq
[cpu
] = cfs_rq
;
8114 /* se could be NULL for root_task_group */
8119 se
->cfs_rq
= &rq
->cfs
;
8122 se
->cfs_rq
= parent
->my_q
;
8123 se
->depth
= parent
->depth
+ 1;
8127 /* guarantee group entities always have weight */
8128 update_load_set(&se
->load
, NICE_0_LOAD
);
8129 se
->parent
= parent
;
8132 static DEFINE_MUTEX(shares_mutex
);
8134 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8137 unsigned long flags
;
8140 * We can't change the weight of the root cgroup.
8145 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
8147 mutex_lock(&shares_mutex
);
8148 if (tg
->shares
== shares
)
8151 tg
->shares
= shares
;
8152 for_each_possible_cpu(i
) {
8153 struct rq
*rq
= cpu_rq(i
);
8154 struct sched_entity
*se
;
8157 /* Propagate contribution to hierarchy */
8158 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8160 /* Possible calls to update_curr() need rq clock */
8161 update_rq_clock(rq
);
8162 for_each_sched_entity(se
)
8163 update_cfs_shares(group_cfs_rq(se
));
8164 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8168 mutex_unlock(&shares_mutex
);
8171 #else /* CONFIG_FAIR_GROUP_SCHED */
8173 void free_fair_sched_group(struct task_group
*tg
) { }
8175 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8180 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
) { }
8182 #endif /* CONFIG_FAIR_GROUP_SCHED */
8185 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
8187 struct sched_entity
*se
= &task
->se
;
8188 unsigned int rr_interval
= 0;
8191 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8194 if (rq
->cfs
.load
.weight
)
8195 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
8201 * All the scheduling class methods:
8203 const struct sched_class fair_sched_class
= {
8204 .next
= &idle_sched_class
,
8205 .enqueue_task
= enqueue_task_fair
,
8206 .dequeue_task
= dequeue_task_fair
,
8207 .yield_task
= yield_task_fair
,
8208 .yield_to_task
= yield_to_task_fair
,
8210 .check_preempt_curr
= check_preempt_wakeup
,
8212 .pick_next_task
= pick_next_task_fair
,
8213 .put_prev_task
= put_prev_task_fair
,
8216 .select_task_rq
= select_task_rq_fair
,
8217 .migrate_task_rq
= migrate_task_rq_fair
,
8219 .rq_online
= rq_online_fair
,
8220 .rq_offline
= rq_offline_fair
,
8222 .task_waking
= task_waking_fair
,
8225 .set_curr_task
= set_curr_task_fair
,
8226 .task_tick
= task_tick_fair
,
8227 .task_fork
= task_fork_fair
,
8229 .prio_changed
= prio_changed_fair
,
8230 .switched_from
= switched_from_fair
,
8231 .switched_to
= switched_to_fair
,
8233 .get_rr_interval
= get_rr_interval_fair
,
8235 .update_curr
= update_curr_fair
,
8237 #ifdef CONFIG_FAIR_GROUP_SCHED
8238 .task_move_group
= task_move_group_fair
,
8242 #ifdef CONFIG_SCHED_DEBUG
8243 void print_cfs_stats(struct seq_file
*m
, int cpu
)
8245 struct cfs_rq
*cfs_rq
;
8248 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
8249 print_cfs_rq(m
, cpu
, cfs_rq
);
8254 __init
void init_sched_fair_class(void)
8257 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8259 #ifdef CONFIG_NO_HZ_COMMON
8260 nohz
.next_balance
= jiffies
;
8261 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
8262 cpu_notifier(sched_ilb_notifier
, 0);