2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
38 * Targeted preemption latency for CPU-bound tasks:
39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41 * NOTE: this latency value is not the same as the concept of
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
49 unsigned int sysctl_sched_latency
= 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency
= 6000000ULL;
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG
;
65 * Minimal preemption granularity for CPU-bound tasks:
66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 unsigned int sysctl_sched_min_granularity
= 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity
= 750000ULL;
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74 static unsigned int sched_nr_latency
= 8;
77 * After fork, child runs first. If set to 0 (default) then
78 * parent will (try to) run first.
80 unsigned int sysctl_sched_child_runs_first __read_mostly
;
83 * SCHED_OTHER wake-up granularity.
84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
90 unsigned int sysctl_sched_wakeup_granularity
= 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity
= 1000000UL;
93 const_debug
unsigned int sysctl_sched_migration_cost
= 500000UL;
96 * The exponential sliding window over which load is averaged for shares
100 unsigned int __read_mostly sysctl_sched_shares_window
= 10000000UL;
102 #ifdef CONFIG_CFS_BANDWIDTH
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
111 * default: 5 msec, units: microseconds
113 unsigned int sysctl_sched_cfs_bandwidth_slice
= 5000UL;
116 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
122 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
128 static inline void update_load_set(struct load_weight
*lw
, unsigned long w
)
135 * Increase the granularity value when there are more CPUs,
136 * because with more CPUs the 'effective latency' as visible
137 * to users decreases. But the relationship is not linear,
138 * so pick a second-best guess by going with the log2 of the
141 * This idea comes from the SD scheduler of Con Kolivas:
143 static int get_update_sysctl_factor(void)
145 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
148 switch (sysctl_sched_tunable_scaling
) {
149 case SCHED_TUNABLESCALING_NONE
:
152 case SCHED_TUNABLESCALING_LINEAR
:
155 case SCHED_TUNABLESCALING_LOG
:
157 factor
= 1 + ilog2(cpus
);
164 static void update_sysctl(void)
166 unsigned int factor
= get_update_sysctl_factor();
168 #define SET_SYSCTL(name) \
169 (sysctl_##name = (factor) * normalized_sysctl_##name)
170 SET_SYSCTL(sched_min_granularity
);
171 SET_SYSCTL(sched_latency
);
172 SET_SYSCTL(sched_wakeup_granularity
);
176 void sched_init_granularity(void)
181 #if BITS_PER_LONG == 32
182 # define WMULT_CONST (~0UL)
184 # define WMULT_CONST (1UL << 32)
187 #define WMULT_SHIFT 32
190 * Shift right and round:
192 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
195 * delta *= weight / lw
198 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
199 struct load_weight
*lw
)
204 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
205 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
206 * 2^SCHED_LOAD_RESOLUTION.
208 if (likely(weight
> (1UL << SCHED_LOAD_RESOLUTION
)))
209 tmp
= (u64
)delta_exec
* scale_load_down(weight
);
211 tmp
= (u64
)delta_exec
;
213 if (!lw
->inv_weight
) {
214 unsigned long w
= scale_load_down(lw
->weight
);
216 if (BITS_PER_LONG
> 32 && unlikely(w
>= WMULT_CONST
))
218 else if (unlikely(!w
))
219 lw
->inv_weight
= WMULT_CONST
;
221 lw
->inv_weight
= WMULT_CONST
/ w
;
225 * Check whether we'd overflow the 64-bit multiplication:
227 if (unlikely(tmp
> WMULT_CONST
))
228 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
231 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
233 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
237 const struct sched_class fair_sched_class
;
239 /**************************************************************
240 * CFS operations on generic schedulable entities:
243 #ifdef CONFIG_FAIR_GROUP_SCHED
245 /* cpu runqueue to which this cfs_rq is attached */
246 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
251 /* An entity is a task if it doesn't "own" a runqueue */
252 #define entity_is_task(se) (!se->my_q)
254 static inline struct task_struct
*task_of(struct sched_entity
*se
)
256 #ifdef CONFIG_SCHED_DEBUG
257 WARN_ON_ONCE(!entity_is_task(se
));
259 return container_of(se
, struct task_struct
, se
);
262 /* Walk up scheduling entities hierarchy */
263 #define for_each_sched_entity(se) \
264 for (; se; se = se->parent)
266 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
271 /* runqueue on which this entity is (to be) queued */
272 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
277 /* runqueue "owned" by this group */
278 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
283 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
288 if (!cfs_rq
->on_list
) {
290 * Ensure we either appear before our parent (if already
291 * enqueued) or force our parent to appear after us when it is
292 * enqueued. The fact that we always enqueue bottom-up
293 * reduces this to two cases.
295 if (cfs_rq
->tg
->parent
&&
296 cfs_rq
->tg
->parent
->cfs_rq
[cpu_of(rq_of(cfs_rq
))]->on_list
) {
297 list_add_rcu(&cfs_rq
->leaf_cfs_rq_list
,
298 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
300 list_add_tail_rcu(&cfs_rq
->leaf_cfs_rq_list
,
301 &rq_of(cfs_rq
)->leaf_cfs_rq_list
);
305 /* We should have no load, but we need to update last_decay. */
306 update_cfs_rq_blocked_load(cfs_rq
, 0);
310 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
312 if (cfs_rq
->on_list
) {
313 list_del_rcu(&cfs_rq
->leaf_cfs_rq_list
);
318 /* Iterate thr' all leaf cfs_rq's on a runqueue */
319 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
320 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
322 /* Do the two (enqueued) entities belong to the same group ? */
324 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
326 if (se
->cfs_rq
== pse
->cfs_rq
)
332 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
337 /* return depth at which a sched entity is present in the hierarchy */
338 static inline int depth_se(struct sched_entity
*se
)
342 for_each_sched_entity(se
)
349 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
351 int se_depth
, pse_depth
;
354 * preemption test can be made between sibling entities who are in the
355 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
356 * both tasks until we find their ancestors who are siblings of common
360 /* First walk up until both entities are at same depth */
361 se_depth
= depth_se(*se
);
362 pse_depth
= depth_se(*pse
);
364 while (se_depth
> pse_depth
) {
366 *se
= parent_entity(*se
);
369 while (pse_depth
> se_depth
) {
371 *pse
= parent_entity(*pse
);
374 while (!is_same_group(*se
, *pse
)) {
375 *se
= parent_entity(*se
);
376 *pse
= parent_entity(*pse
);
380 #else /* !CONFIG_FAIR_GROUP_SCHED */
382 static inline struct task_struct
*task_of(struct sched_entity
*se
)
384 return container_of(se
, struct task_struct
, se
);
387 static inline struct rq
*rq_of(struct cfs_rq
*cfs_rq
)
389 return container_of(cfs_rq
, struct rq
, cfs
);
392 #define entity_is_task(se) 1
394 #define for_each_sched_entity(se) \
395 for (; se; se = NULL)
397 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
399 return &task_rq(p
)->cfs
;
402 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
404 struct task_struct
*p
= task_of(se
);
405 struct rq
*rq
= task_rq(p
);
410 /* runqueue "owned" by this group */
411 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
416 static inline void list_add_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
420 static inline void list_del_leaf_cfs_rq(struct cfs_rq
*cfs_rq
)
424 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
425 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
428 is_same_group(struct sched_entity
*se
, struct sched_entity
*pse
)
433 static inline struct sched_entity
*parent_entity(struct sched_entity
*se
)
439 find_matching_se(struct sched_entity
**se
, struct sched_entity
**pse
)
443 #endif /* CONFIG_FAIR_GROUP_SCHED */
445 static __always_inline
446 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, unsigned long delta_exec
);
448 /**************************************************************
449 * Scheduling class tree data structure manipulation methods:
452 static inline u64
max_vruntime(u64 max_vruntime
, u64 vruntime
)
454 s64 delta
= (s64
)(vruntime
- max_vruntime
);
456 max_vruntime
= vruntime
;
461 static inline u64
min_vruntime(u64 min_vruntime
, u64 vruntime
)
463 s64 delta
= (s64
)(vruntime
- min_vruntime
);
465 min_vruntime
= vruntime
;
470 static inline int entity_before(struct sched_entity
*a
,
471 struct sched_entity
*b
)
473 return (s64
)(a
->vruntime
- b
->vruntime
) < 0;
476 static void update_min_vruntime(struct cfs_rq
*cfs_rq
)
478 u64 vruntime
= cfs_rq
->min_vruntime
;
481 vruntime
= cfs_rq
->curr
->vruntime
;
483 if (cfs_rq
->rb_leftmost
) {
484 struct sched_entity
*se
= rb_entry(cfs_rq
->rb_leftmost
,
489 vruntime
= se
->vruntime
;
491 vruntime
= min_vruntime(vruntime
, se
->vruntime
);
494 /* ensure we never gain time by being placed backwards. */
495 cfs_rq
->min_vruntime
= max_vruntime(cfs_rq
->min_vruntime
, vruntime
);
498 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
503 * Enqueue an entity into the rb-tree:
505 static void __enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
507 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
508 struct rb_node
*parent
= NULL
;
509 struct sched_entity
*entry
;
513 * Find the right place in the rbtree:
517 entry
= rb_entry(parent
, struct sched_entity
, run_node
);
519 * We dont care about collisions. Nodes with
520 * the same key stay together.
522 if (entity_before(se
, entry
)) {
523 link
= &parent
->rb_left
;
525 link
= &parent
->rb_right
;
531 * Maintain a cache of leftmost tree entries (it is frequently
535 cfs_rq
->rb_leftmost
= &se
->run_node
;
537 rb_link_node(&se
->run_node
, parent
, link
);
538 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
541 static void __dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
543 if (cfs_rq
->rb_leftmost
== &se
->run_node
) {
544 struct rb_node
*next_node
;
546 next_node
= rb_next(&se
->run_node
);
547 cfs_rq
->rb_leftmost
= next_node
;
550 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
553 struct sched_entity
*__pick_first_entity(struct cfs_rq
*cfs_rq
)
555 struct rb_node
*left
= cfs_rq
->rb_leftmost
;
560 return rb_entry(left
, struct sched_entity
, run_node
);
563 static struct sched_entity
*__pick_next_entity(struct sched_entity
*se
)
565 struct rb_node
*next
= rb_next(&se
->run_node
);
570 return rb_entry(next
, struct sched_entity
, run_node
);
573 #ifdef CONFIG_SCHED_DEBUG
574 struct sched_entity
*__pick_last_entity(struct cfs_rq
*cfs_rq
)
576 struct rb_node
*last
= rb_last(&cfs_rq
->tasks_timeline
);
581 return rb_entry(last
, struct sched_entity
, run_node
);
584 /**************************************************************
585 * Scheduling class statistics methods:
588 int sched_proc_update_handler(struct ctl_table
*table
, int write
,
589 void __user
*buffer
, size_t *lenp
,
592 int ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
593 int factor
= get_update_sysctl_factor();
598 sched_nr_latency
= DIV_ROUND_UP(sysctl_sched_latency
,
599 sysctl_sched_min_granularity
);
601 #define WRT_SYSCTL(name) \
602 (normalized_sysctl_##name = sysctl_##name / (factor))
603 WRT_SYSCTL(sched_min_granularity
);
604 WRT_SYSCTL(sched_latency
);
605 WRT_SYSCTL(sched_wakeup_granularity
);
615 static inline unsigned long
616 calc_delta_fair(unsigned long delta
, struct sched_entity
*se
)
618 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
619 delta
= calc_delta_mine(delta
, NICE_0_LOAD
, &se
->load
);
625 * The idea is to set a period in which each task runs once.
627 * When there are too many tasks (sched_nr_latency) we have to stretch
628 * this period because otherwise the slices get too small.
630 * p = (nr <= nl) ? l : l*nr/nl
632 static u64
__sched_period(unsigned long nr_running
)
634 u64 period
= sysctl_sched_latency
;
635 unsigned long nr_latency
= sched_nr_latency
;
637 if (unlikely(nr_running
> nr_latency
)) {
638 period
= sysctl_sched_min_granularity
;
639 period
*= nr_running
;
646 * We calculate the wall-time slice from the period by taking a part
647 * proportional to the weight.
651 static u64
sched_slice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
653 u64 slice
= __sched_period(cfs_rq
->nr_running
+ !se
->on_rq
);
655 for_each_sched_entity(se
) {
656 struct load_weight
*load
;
657 struct load_weight lw
;
659 cfs_rq
= cfs_rq_of(se
);
660 load
= &cfs_rq
->load
;
662 if (unlikely(!se
->on_rq
)) {
665 update_load_add(&lw
, se
->load
.weight
);
668 slice
= calc_delta_mine(slice
, se
->load
.weight
, load
);
674 * We calculate the vruntime slice of a to-be-inserted task.
678 static u64
sched_vslice(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
680 return calc_delta_fair(sched_slice(cfs_rq
, se
), se
);
684 static inline void __update_task_entity_contrib(struct sched_entity
*se
);
686 /* Give new task start runnable values to heavy its load in infant time */
687 void init_task_runnable_average(struct task_struct
*p
)
691 p
->se
.avg
.decay_count
= 0;
692 slice
= sched_slice(task_cfs_rq(p
), &p
->se
) >> 10;
693 p
->se
.avg
.runnable_avg_sum
= slice
;
694 p
->se
.avg
.runnable_avg_period
= slice
;
695 __update_task_entity_contrib(&p
->se
);
698 void init_task_runnable_average(struct task_struct
*p
)
704 * Update the current task's runtime statistics. Skip current tasks that
705 * are not in our scheduling class.
708 __update_curr(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
,
709 unsigned long delta_exec
)
711 unsigned long delta_exec_weighted
;
713 schedstat_set(curr
->statistics
.exec_max
,
714 max((u64
)delta_exec
, curr
->statistics
.exec_max
));
716 curr
->sum_exec_runtime
+= delta_exec
;
717 schedstat_add(cfs_rq
, exec_clock
, delta_exec
);
718 delta_exec_weighted
= calc_delta_fair(delta_exec
, curr
);
720 curr
->vruntime
+= delta_exec_weighted
;
721 update_min_vruntime(cfs_rq
);
724 static void update_curr(struct cfs_rq
*cfs_rq
)
726 struct sched_entity
*curr
= cfs_rq
->curr
;
727 u64 now
= rq_clock_task(rq_of(cfs_rq
));
728 unsigned long delta_exec
;
734 * Get the amount of time the current task was running
735 * since the last time we changed load (this cannot
736 * overflow on 32 bits):
738 delta_exec
= (unsigned long)(now
- curr
->exec_start
);
742 __update_curr(cfs_rq
, curr
, delta_exec
);
743 curr
->exec_start
= now
;
745 if (entity_is_task(curr
)) {
746 struct task_struct
*curtask
= task_of(curr
);
748 trace_sched_stat_runtime(curtask
, delta_exec
, curr
->vruntime
);
749 cpuacct_charge(curtask
, delta_exec
);
750 account_group_exec_runtime(curtask
, delta_exec
);
753 account_cfs_rq_runtime(cfs_rq
, delta_exec
);
757 update_stats_wait_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
759 schedstat_set(se
->statistics
.wait_start
, rq_clock(rq_of(cfs_rq
)));
763 * Task is being enqueued - update stats:
765 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
768 * Are we enqueueing a waiting task? (for current tasks
769 * a dequeue/enqueue event is a NOP)
771 if (se
!= cfs_rq
->curr
)
772 update_stats_wait_start(cfs_rq
, se
);
776 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
778 schedstat_set(se
->statistics
.wait_max
, max(se
->statistics
.wait_max
,
779 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
));
780 schedstat_set(se
->statistics
.wait_count
, se
->statistics
.wait_count
+ 1);
781 schedstat_set(se
->statistics
.wait_sum
, se
->statistics
.wait_sum
+
782 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
783 #ifdef CONFIG_SCHEDSTATS
784 if (entity_is_task(se
)) {
785 trace_sched_stat_wait(task_of(se
),
786 rq_clock(rq_of(cfs_rq
)) - se
->statistics
.wait_start
);
789 schedstat_set(se
->statistics
.wait_start
, 0);
793 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
796 * Mark the end of the wait period if dequeueing a
799 if (se
!= cfs_rq
->curr
)
800 update_stats_wait_end(cfs_rq
, se
);
804 * We are picking a new current task - update its stats:
807 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
810 * We are starting a new run period:
812 se
->exec_start
= rq_clock_task(rq_of(cfs_rq
));
815 /**************************************************
816 * Scheduling class queueing methods:
819 #ifdef CONFIG_NUMA_BALANCING
821 * Approximate time to scan a full NUMA task in ms. The task scan period is
822 * calculated based on the tasks virtual memory size and
823 * numa_balancing_scan_size.
825 unsigned int sysctl_numa_balancing_scan_period_min
= 1000;
826 unsigned int sysctl_numa_balancing_scan_period_max
= 60000;
827 unsigned int sysctl_numa_balancing_scan_period_reset
= 60000;
829 /* Portion of address space to scan in MB */
830 unsigned int sysctl_numa_balancing_scan_size
= 256;
832 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
833 unsigned int sysctl_numa_balancing_scan_delay
= 1000;
835 static unsigned int task_nr_scan_windows(struct task_struct
*p
)
837 unsigned long rss
= 0;
838 unsigned long nr_scan_pages
;
841 * Calculations based on RSS as non-present and empty pages are skipped
842 * by the PTE scanner and NUMA hinting faults should be trapped based
845 nr_scan_pages
= sysctl_numa_balancing_scan_size
<< (20 - PAGE_SHIFT
);
846 rss
= get_mm_rss(p
->mm
);
850 rss
= round_up(rss
, nr_scan_pages
);
851 return rss
/ nr_scan_pages
;
854 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
855 #define MAX_SCAN_WINDOW 2560
857 static unsigned int task_scan_min(struct task_struct
*p
)
859 unsigned int scan
, floor
;
860 unsigned int windows
= 1;
862 if (sysctl_numa_balancing_scan_size
< MAX_SCAN_WINDOW
)
863 windows
= MAX_SCAN_WINDOW
/ sysctl_numa_balancing_scan_size
;
864 floor
= 1000 / windows
;
866 scan
= sysctl_numa_balancing_scan_period_min
/ task_nr_scan_windows(p
);
867 return max_t(unsigned int, floor
, scan
);
870 static unsigned int task_scan_max(struct task_struct
*p
)
872 unsigned int smin
= task_scan_min(p
);
875 /* Watch for min being lower than max due to floor calculations */
876 smax
= sysctl_numa_balancing_scan_period_max
/ task_nr_scan_windows(p
);
877 return max(smin
, smax
);
880 static void task_numa_placement(struct task_struct
*p
)
884 if (!p
->mm
) /* for example, ksmd faulting in a user's mm */
886 seq
= ACCESS_ONCE(p
->mm
->numa_scan_seq
);
887 if (p
->numa_scan_seq
== seq
)
889 p
->numa_scan_seq
= seq
;
890 p
->numa_scan_period_max
= task_scan_max(p
);
892 /* FIXME: Scheduling placement policy hints go here */
896 * Got a PROT_NONE fault for a page on @node.
898 void task_numa_fault(int node
, int pages
, bool migrated
)
900 struct task_struct
*p
= current
;
902 if (!numabalancing_enabled
)
905 /* Allocate buffer to track faults on a per-node basis */
906 if (unlikely(!p
->numa_faults
)) {
907 int size
= sizeof(*p
->numa_faults
) * nr_node_ids
;
909 p
->numa_faults
= kzalloc(size
, GFP_KERNEL
|__GFP_NOWARN
);
915 * If pages are properly placed (did not migrate) then scan slower.
916 * This is reset periodically in case of phase changes
919 /* Initialise if necessary */
920 if (!p
->numa_scan_period_max
)
921 p
->numa_scan_period_max
= task_scan_max(p
);
923 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
924 p
->numa_scan_period
+ 10);
927 task_numa_placement(p
);
929 p
->numa_faults
[node
] += pages
;
932 static void reset_ptenuma_scan(struct task_struct
*p
)
934 ACCESS_ONCE(p
->mm
->numa_scan_seq
)++;
935 p
->mm
->numa_scan_offset
= 0;
939 * The expensive part of numa migration is done from task_work context.
940 * Triggered from task_tick_numa().
942 void task_numa_work(struct callback_head
*work
)
944 unsigned long migrate
, next_scan
, now
= jiffies
;
945 struct task_struct
*p
= current
;
946 struct mm_struct
*mm
= p
->mm
;
947 struct vm_area_struct
*vma
;
948 unsigned long start
, end
;
949 unsigned long nr_pte_updates
= 0;
952 WARN_ON_ONCE(p
!= container_of(work
, struct task_struct
, numa_work
));
954 work
->next
= work
; /* protect against double add */
956 * Who cares about NUMA placement when they're dying.
958 * NOTE: make sure not to dereference p->mm before this check,
959 * exit_task_work() happens _after_ exit_mm() so we could be called
960 * without p->mm even though we still had it when we enqueued this
963 if (p
->flags
& PF_EXITING
)
966 if (!mm
->numa_next_reset
|| !mm
->numa_next_scan
) {
967 mm
->numa_next_scan
= now
+
968 msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
969 mm
->numa_next_reset
= now
+
970 msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset
);
974 * Reset the scan period if enough time has gone by. Objective is that
975 * scanning will be reduced if pages are properly placed. As tasks
976 * can enter different phases this needs to be re-examined. Lacking
977 * proper tracking of reference behaviour, this blunt hammer is used.
979 migrate
= mm
->numa_next_reset
;
980 if (time_after(now
, migrate
)) {
981 p
->numa_scan_period
= task_scan_min(p
);
982 next_scan
= now
+ msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset
);
983 xchg(&mm
->numa_next_reset
, next_scan
);
987 * Enforce maximal scan/migration frequency..
989 migrate
= mm
->numa_next_scan
;
990 if (time_before(now
, migrate
))
993 if (p
->numa_scan_period
== 0) {
994 p
->numa_scan_period_max
= task_scan_max(p
);
995 p
->numa_scan_period
= task_scan_min(p
);
998 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
999 if (cmpxchg(&mm
->numa_next_scan
, migrate
, next_scan
) != migrate
)
1003 * Delay this task enough that another task of this mm will likely win
1004 * the next time around.
1006 p
->node_stamp
+= 2 * TICK_NSEC
;
1008 start
= mm
->numa_scan_offset
;
1009 pages
= sysctl_numa_balancing_scan_size
;
1010 pages
<<= 20 - PAGE_SHIFT
; /* MB in pages */
1014 down_read(&mm
->mmap_sem
);
1015 vma
= find_vma(mm
, start
);
1017 reset_ptenuma_scan(p
);
1021 for (; vma
; vma
= vma
->vm_next
) {
1022 if (!vma_migratable(vma
))
1025 /* Skip small VMAs. They are not likely to be of relevance */
1026 if (vma
->vm_end
- vma
->vm_start
< HPAGE_SIZE
)
1030 start
= max(start
, vma
->vm_start
);
1031 end
= ALIGN(start
+ (pages
<< PAGE_SHIFT
), HPAGE_SIZE
);
1032 end
= min(end
, vma
->vm_end
);
1033 nr_pte_updates
+= change_prot_numa(vma
, start
, end
);
1036 * Scan sysctl_numa_balancing_scan_size but ensure that
1037 * at least one PTE is updated so that unused virtual
1038 * address space is quickly skipped.
1041 pages
-= (end
- start
) >> PAGE_SHIFT
;
1046 } while (end
!= vma
->vm_end
);
1051 * If the whole process was scanned without updates then no NUMA
1052 * hinting faults are being recorded and scan rate should be lower.
1054 if (mm
->numa_scan_offset
== 0 && !nr_pte_updates
) {
1055 p
->numa_scan_period
= min(p
->numa_scan_period_max
,
1056 p
->numa_scan_period
<< 1);
1058 next_scan
= now
+ msecs_to_jiffies(p
->numa_scan_period
);
1059 mm
->numa_next_scan
= next_scan
;
1063 * It is possible to reach the end of the VMA list but the last few
1064 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1065 * would find the !migratable VMA on the next scan but not reset the
1066 * scanner to the start so check it now.
1069 mm
->numa_scan_offset
= start
;
1071 reset_ptenuma_scan(p
);
1072 up_read(&mm
->mmap_sem
);
1076 * Drive the periodic memory faults..
1078 void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
1080 struct callback_head
*work
= &curr
->numa_work
;
1084 * We don't care about NUMA placement if we don't have memory.
1086 if (!curr
->mm
|| (curr
->flags
& PF_EXITING
) || work
->next
!= work
)
1090 * Using runtime rather than walltime has the dual advantage that
1091 * we (mostly) drive the selection from busy threads and that the
1092 * task needs to have done some actual work before we bother with
1095 now
= curr
->se
.sum_exec_runtime
;
1096 period
= (u64
)curr
->numa_scan_period
* NSEC_PER_MSEC
;
1098 if (now
- curr
->node_stamp
> period
) {
1099 if (!curr
->node_stamp
)
1100 curr
->numa_scan_period
= task_scan_min(curr
);
1101 curr
->node_stamp
+= period
;
1103 if (!time_before(jiffies
, curr
->mm
->numa_next_scan
)) {
1104 init_task_work(work
, task_numa_work
); /* TODO: move this into sched_fork() */
1105 task_work_add(curr
, work
, true);
1110 static void task_tick_numa(struct rq
*rq
, struct task_struct
*curr
)
1113 #endif /* CONFIG_NUMA_BALANCING */
1116 account_entity_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1118 update_load_add(&cfs_rq
->load
, se
->load
.weight
);
1119 if (!parent_entity(se
))
1120 update_load_add(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
1122 if (entity_is_task(se
))
1123 list_add(&se
->group_node
, &rq_of(cfs_rq
)->cfs_tasks
);
1125 cfs_rq
->nr_running
++;
1129 account_entity_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1131 update_load_sub(&cfs_rq
->load
, se
->load
.weight
);
1132 if (!parent_entity(se
))
1133 update_load_sub(&rq_of(cfs_rq
)->load
, se
->load
.weight
);
1134 if (entity_is_task(se
))
1135 list_del_init(&se
->group_node
);
1136 cfs_rq
->nr_running
--;
1139 #ifdef CONFIG_FAIR_GROUP_SCHED
1141 static inline long calc_tg_weight(struct task_group
*tg
, struct cfs_rq
*cfs_rq
)
1146 * Use this CPU's actual weight instead of the last load_contribution
1147 * to gain a more accurate current total weight. See
1148 * update_cfs_rq_load_contribution().
1150 tg_weight
= atomic_long_read(&tg
->load_avg
);
1151 tg_weight
-= cfs_rq
->tg_load_contrib
;
1152 tg_weight
+= cfs_rq
->load
.weight
;
1157 static long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
1159 long tg_weight
, load
, shares
;
1161 tg_weight
= calc_tg_weight(tg
, cfs_rq
);
1162 load
= cfs_rq
->load
.weight
;
1164 shares
= (tg
->shares
* load
);
1166 shares
/= tg_weight
;
1168 if (shares
< MIN_SHARES
)
1169 shares
= MIN_SHARES
;
1170 if (shares
> tg
->shares
)
1171 shares
= tg
->shares
;
1175 # else /* CONFIG_SMP */
1176 static inline long calc_cfs_shares(struct cfs_rq
*cfs_rq
, struct task_group
*tg
)
1180 # endif /* CONFIG_SMP */
1181 static void reweight_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
1182 unsigned long weight
)
1185 /* commit outstanding execution time */
1186 if (cfs_rq
->curr
== se
)
1187 update_curr(cfs_rq
);
1188 account_entity_dequeue(cfs_rq
, se
);
1191 update_load_set(&se
->load
, weight
);
1194 account_entity_enqueue(cfs_rq
, se
);
1197 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
);
1199 static void update_cfs_shares(struct cfs_rq
*cfs_rq
)
1201 struct task_group
*tg
;
1202 struct sched_entity
*se
;
1206 se
= tg
->se
[cpu_of(rq_of(cfs_rq
))];
1207 if (!se
|| throttled_hierarchy(cfs_rq
))
1210 if (likely(se
->load
.weight
== tg
->shares
))
1213 shares
= calc_cfs_shares(cfs_rq
, tg
);
1215 reweight_entity(cfs_rq_of(se
), se
, shares
);
1217 #else /* CONFIG_FAIR_GROUP_SCHED */
1218 static inline void update_cfs_shares(struct cfs_rq
*cfs_rq
)
1221 #endif /* CONFIG_FAIR_GROUP_SCHED */
1225 * We choose a half-life close to 1 scheduling period.
1226 * Note: The tables below are dependent on this value.
1228 #define LOAD_AVG_PERIOD 32
1229 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1230 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1232 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1233 static const u32 runnable_avg_yN_inv
[] = {
1234 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1235 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1236 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1237 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1238 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1239 0x85aac367, 0x82cd8698,
1243 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1244 * over-estimates when re-combining.
1246 static const u32 runnable_avg_yN_sum
[] = {
1247 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1248 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1249 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1254 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1256 static __always_inline u64
decay_load(u64 val
, u64 n
)
1258 unsigned int local_n
;
1262 else if (unlikely(n
> LOAD_AVG_PERIOD
* 63))
1265 /* after bounds checking we can collapse to 32-bit */
1269 * As y^PERIOD = 1/2, we can combine
1270 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1271 * With a look-up table which covers k^n (n<PERIOD)
1273 * To achieve constant time decay_load.
1275 if (unlikely(local_n
>= LOAD_AVG_PERIOD
)) {
1276 val
>>= local_n
/ LOAD_AVG_PERIOD
;
1277 local_n
%= LOAD_AVG_PERIOD
;
1280 val
*= runnable_avg_yN_inv
[local_n
];
1281 /* We don't use SRR here since we always want to round down. */
1286 * For updates fully spanning n periods, the contribution to runnable
1287 * average will be: \Sum 1024*y^n
1289 * We can compute this reasonably efficiently by combining:
1290 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1292 static u32
__compute_runnable_contrib(u64 n
)
1296 if (likely(n
<= LOAD_AVG_PERIOD
))
1297 return runnable_avg_yN_sum
[n
];
1298 else if (unlikely(n
>= LOAD_AVG_MAX_N
))
1299 return LOAD_AVG_MAX
;
1301 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1303 contrib
/= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1304 contrib
+= runnable_avg_yN_sum
[LOAD_AVG_PERIOD
];
1306 n
-= LOAD_AVG_PERIOD
;
1307 } while (n
> LOAD_AVG_PERIOD
);
1309 contrib
= decay_load(contrib
, n
);
1310 return contrib
+ runnable_avg_yN_sum
[n
];
1314 * We can represent the historical contribution to runnable average as the
1315 * coefficients of a geometric series. To do this we sub-divide our runnable
1316 * history into segments of approximately 1ms (1024us); label the segment that
1317 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1319 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1321 * (now) (~1ms ago) (~2ms ago)
1323 * Let u_i denote the fraction of p_i that the entity was runnable.
1325 * We then designate the fractions u_i as our co-efficients, yielding the
1326 * following representation of historical load:
1327 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1329 * We choose y based on the with of a reasonably scheduling period, fixing:
1332 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1333 * approximately half as much as the contribution to load within the last ms
1336 * When a period "rolls over" and we have new u_0`, multiplying the previous
1337 * sum again by y is sufficient to update:
1338 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1339 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1341 static __always_inline
int __update_entity_runnable_avg(u64 now
,
1342 struct sched_avg
*sa
,
1346 u32 runnable_contrib
;
1347 int delta_w
, decayed
= 0;
1349 delta
= now
- sa
->last_runnable_update
;
1351 * This should only happen when time goes backwards, which it
1352 * unfortunately does during sched clock init when we swap over to TSC.
1354 if ((s64
)delta
< 0) {
1355 sa
->last_runnable_update
= now
;
1360 * Use 1024ns as the unit of measurement since it's a reasonable
1361 * approximation of 1us and fast to compute.
1366 sa
->last_runnable_update
= now
;
1368 /* delta_w is the amount already accumulated against our next period */
1369 delta_w
= sa
->runnable_avg_period
% 1024;
1370 if (delta
+ delta_w
>= 1024) {
1371 /* period roll-over */
1375 * Now that we know we're crossing a period boundary, figure
1376 * out how much from delta we need to complete the current
1377 * period and accrue it.
1379 delta_w
= 1024 - delta_w
;
1381 sa
->runnable_avg_sum
+= delta_w
;
1382 sa
->runnable_avg_period
+= delta_w
;
1386 /* Figure out how many additional periods this update spans */
1387 periods
= delta
/ 1024;
1390 sa
->runnable_avg_sum
= decay_load(sa
->runnable_avg_sum
,
1392 sa
->runnable_avg_period
= decay_load(sa
->runnable_avg_period
,
1395 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1396 runnable_contrib
= __compute_runnable_contrib(periods
);
1398 sa
->runnable_avg_sum
+= runnable_contrib
;
1399 sa
->runnable_avg_period
+= runnable_contrib
;
1402 /* Remainder of delta accrued against u_0` */
1404 sa
->runnable_avg_sum
+= delta
;
1405 sa
->runnable_avg_period
+= delta
;
1410 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1411 static inline u64
__synchronize_entity_decay(struct sched_entity
*se
)
1413 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1414 u64 decays
= atomic64_read(&cfs_rq
->decay_counter
);
1416 decays
-= se
->avg
.decay_count
;
1420 se
->avg
.load_avg_contrib
= decay_load(se
->avg
.load_avg_contrib
, decays
);
1421 se
->avg
.decay_count
= 0;
1426 #ifdef CONFIG_FAIR_GROUP_SCHED
1427 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
1430 struct task_group
*tg
= cfs_rq
->tg
;
1433 tg_contrib
= cfs_rq
->runnable_load_avg
+ cfs_rq
->blocked_load_avg
;
1434 tg_contrib
-= cfs_rq
->tg_load_contrib
;
1436 if (force_update
|| abs(tg_contrib
) > cfs_rq
->tg_load_contrib
/ 8) {
1437 atomic_long_add(tg_contrib
, &tg
->load_avg
);
1438 cfs_rq
->tg_load_contrib
+= tg_contrib
;
1443 * Aggregate cfs_rq runnable averages into an equivalent task_group
1444 * representation for computing load contributions.
1446 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
1447 struct cfs_rq
*cfs_rq
)
1449 struct task_group
*tg
= cfs_rq
->tg
;
1452 /* The fraction of a cpu used by this cfs_rq */
1453 contrib
= div_u64(sa
->runnable_avg_sum
<< NICE_0_SHIFT
,
1454 sa
->runnable_avg_period
+ 1);
1455 contrib
-= cfs_rq
->tg_runnable_contrib
;
1457 if (abs(contrib
) > cfs_rq
->tg_runnable_contrib
/ 64) {
1458 atomic_add(contrib
, &tg
->runnable_avg
);
1459 cfs_rq
->tg_runnable_contrib
+= contrib
;
1463 static inline void __update_group_entity_contrib(struct sched_entity
*se
)
1465 struct cfs_rq
*cfs_rq
= group_cfs_rq(se
);
1466 struct task_group
*tg
= cfs_rq
->tg
;
1471 contrib
= cfs_rq
->tg_load_contrib
* tg
->shares
;
1472 se
->avg
.load_avg_contrib
= div_u64(contrib
,
1473 atomic_long_read(&tg
->load_avg
) + 1);
1476 * For group entities we need to compute a correction term in the case
1477 * that they are consuming <1 cpu so that we would contribute the same
1478 * load as a task of equal weight.
1480 * Explicitly co-ordinating this measurement would be expensive, but
1481 * fortunately the sum of each cpus contribution forms a usable
1482 * lower-bound on the true value.
1484 * Consider the aggregate of 2 contributions. Either they are disjoint
1485 * (and the sum represents true value) or they are disjoint and we are
1486 * understating by the aggregate of their overlap.
1488 * Extending this to N cpus, for a given overlap, the maximum amount we
1489 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1490 * cpus that overlap for this interval and w_i is the interval width.
1492 * On a small machine; the first term is well-bounded which bounds the
1493 * total error since w_i is a subset of the period. Whereas on a
1494 * larger machine, while this first term can be larger, if w_i is the
1495 * of consequential size guaranteed to see n_i*w_i quickly converge to
1496 * our upper bound of 1-cpu.
1498 runnable_avg
= atomic_read(&tg
->runnable_avg
);
1499 if (runnable_avg
< NICE_0_LOAD
) {
1500 se
->avg
.load_avg_contrib
*= runnable_avg
;
1501 se
->avg
.load_avg_contrib
>>= NICE_0_SHIFT
;
1505 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq
*cfs_rq
,
1506 int force_update
) {}
1507 static inline void __update_tg_runnable_avg(struct sched_avg
*sa
,
1508 struct cfs_rq
*cfs_rq
) {}
1509 static inline void __update_group_entity_contrib(struct sched_entity
*se
) {}
1512 static inline void __update_task_entity_contrib(struct sched_entity
*se
)
1516 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1517 contrib
= se
->avg
.runnable_avg_sum
* scale_load_down(se
->load
.weight
);
1518 contrib
/= (se
->avg
.runnable_avg_period
+ 1);
1519 se
->avg
.load_avg_contrib
= scale_load(contrib
);
1522 /* Compute the current contribution to load_avg by se, return any delta */
1523 static long __update_entity_load_avg_contrib(struct sched_entity
*se
)
1525 long old_contrib
= se
->avg
.load_avg_contrib
;
1527 if (entity_is_task(se
)) {
1528 __update_task_entity_contrib(se
);
1530 __update_tg_runnable_avg(&se
->avg
, group_cfs_rq(se
));
1531 __update_group_entity_contrib(se
);
1534 return se
->avg
.load_avg_contrib
- old_contrib
;
1537 static inline void subtract_blocked_load_contrib(struct cfs_rq
*cfs_rq
,
1540 if (likely(load_contrib
< cfs_rq
->blocked_load_avg
))
1541 cfs_rq
->blocked_load_avg
-= load_contrib
;
1543 cfs_rq
->blocked_load_avg
= 0;
1546 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
);
1548 /* Update a sched_entity's runnable average */
1549 static inline void update_entity_load_avg(struct sched_entity
*se
,
1552 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1557 * For a group entity we need to use their owned cfs_rq_clock_task() in
1558 * case they are the parent of a throttled hierarchy.
1560 if (entity_is_task(se
))
1561 now
= cfs_rq_clock_task(cfs_rq
);
1563 now
= cfs_rq_clock_task(group_cfs_rq(se
));
1565 if (!__update_entity_runnable_avg(now
, &se
->avg
, se
->on_rq
))
1568 contrib_delta
= __update_entity_load_avg_contrib(se
);
1574 cfs_rq
->runnable_load_avg
+= contrib_delta
;
1576 subtract_blocked_load_contrib(cfs_rq
, -contrib_delta
);
1580 * Decay the load contributed by all blocked children and account this so that
1581 * their contribution may appropriately discounted when they wake up.
1583 static void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
, int force_update
)
1585 u64 now
= cfs_rq_clock_task(cfs_rq
) >> 20;
1588 decays
= now
- cfs_rq
->last_decay
;
1589 if (!decays
&& !force_update
)
1592 if (atomic_long_read(&cfs_rq
->removed_load
)) {
1593 unsigned long removed_load
;
1594 removed_load
= atomic_long_xchg(&cfs_rq
->removed_load
, 0);
1595 subtract_blocked_load_contrib(cfs_rq
, removed_load
);
1599 cfs_rq
->blocked_load_avg
= decay_load(cfs_rq
->blocked_load_avg
,
1601 atomic64_add(decays
, &cfs_rq
->decay_counter
);
1602 cfs_rq
->last_decay
= now
;
1605 __update_cfs_rq_tg_load_contrib(cfs_rq
, force_update
);
1608 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
)
1610 __update_entity_runnable_avg(rq_clock_task(rq
), &rq
->avg
, runnable
);
1611 __update_tg_runnable_avg(&rq
->avg
, &rq
->cfs
);
1614 /* Add the load generated by se into cfs_rq's child load-average */
1615 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
1616 struct sched_entity
*se
,
1620 * We track migrations using entity decay_count <= 0, on a wake-up
1621 * migration we use a negative decay count to track the remote decays
1622 * accumulated while sleeping.
1624 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
1625 * are seen by enqueue_entity_load_avg() as a migration with an already
1626 * constructed load_avg_contrib.
1628 if (unlikely(se
->avg
.decay_count
<= 0)) {
1629 se
->avg
.last_runnable_update
= rq_clock_task(rq_of(cfs_rq
));
1630 if (se
->avg
.decay_count
) {
1632 * In a wake-up migration we have to approximate the
1633 * time sleeping. This is because we can't synchronize
1634 * clock_task between the two cpus, and it is not
1635 * guaranteed to be read-safe. Instead, we can
1636 * approximate this using our carried decays, which are
1637 * explicitly atomically readable.
1639 se
->avg
.last_runnable_update
-= (-se
->avg
.decay_count
)
1641 update_entity_load_avg(se
, 0);
1642 /* Indicate that we're now synchronized and on-rq */
1643 se
->avg
.decay_count
= 0;
1648 * Task re-woke on same cpu (or else migrate_task_rq_fair()
1649 * would have made count negative); we must be careful to avoid
1650 * double-accounting blocked time after synchronizing decays.
1652 se
->avg
.last_runnable_update
+= __synchronize_entity_decay(se
)
1656 /* migrated tasks did not contribute to our blocked load */
1658 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
1659 update_entity_load_avg(se
, 0);
1662 cfs_rq
->runnable_load_avg
+= se
->avg
.load_avg_contrib
;
1663 /* we force update consideration on load-balancer moves */
1664 update_cfs_rq_blocked_load(cfs_rq
, !wakeup
);
1668 * Remove se's load from this cfs_rq child load-average, if the entity is
1669 * transitioning to a blocked state we track its projected decay using
1672 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
1673 struct sched_entity
*se
,
1676 update_entity_load_avg(se
, 1);
1677 /* we force update consideration on load-balancer moves */
1678 update_cfs_rq_blocked_load(cfs_rq
, !sleep
);
1680 cfs_rq
->runnable_load_avg
-= se
->avg
.load_avg_contrib
;
1682 cfs_rq
->blocked_load_avg
+= se
->avg
.load_avg_contrib
;
1683 se
->avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
1684 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1688 * Update the rq's load with the elapsed running time before entering
1689 * idle. if the last scheduled task is not a CFS task, idle_enter will
1690 * be the only way to update the runnable statistic.
1692 void idle_enter_fair(struct rq
*this_rq
)
1694 update_rq_runnable_avg(this_rq
, 1);
1698 * Update the rq's load with the elapsed idle time before a task is
1699 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1700 * be the only way to update the runnable statistic.
1702 void idle_exit_fair(struct rq
*this_rq
)
1704 update_rq_runnable_avg(this_rq
, 0);
1708 static inline void update_entity_load_avg(struct sched_entity
*se
,
1709 int update_cfs_rq
) {}
1710 static inline void update_rq_runnable_avg(struct rq
*rq
, int runnable
) {}
1711 static inline void enqueue_entity_load_avg(struct cfs_rq
*cfs_rq
,
1712 struct sched_entity
*se
,
1714 static inline void dequeue_entity_load_avg(struct cfs_rq
*cfs_rq
,
1715 struct sched_entity
*se
,
1717 static inline void update_cfs_rq_blocked_load(struct cfs_rq
*cfs_rq
,
1718 int force_update
) {}
1721 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1723 #ifdef CONFIG_SCHEDSTATS
1724 struct task_struct
*tsk
= NULL
;
1726 if (entity_is_task(se
))
1729 if (se
->statistics
.sleep_start
) {
1730 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.sleep_start
;
1735 if (unlikely(delta
> se
->statistics
.sleep_max
))
1736 se
->statistics
.sleep_max
= delta
;
1738 se
->statistics
.sleep_start
= 0;
1739 se
->statistics
.sum_sleep_runtime
+= delta
;
1742 account_scheduler_latency(tsk
, delta
>> 10, 1);
1743 trace_sched_stat_sleep(tsk
, delta
);
1746 if (se
->statistics
.block_start
) {
1747 u64 delta
= rq_clock(rq_of(cfs_rq
)) - se
->statistics
.block_start
;
1752 if (unlikely(delta
> se
->statistics
.block_max
))
1753 se
->statistics
.block_max
= delta
;
1755 se
->statistics
.block_start
= 0;
1756 se
->statistics
.sum_sleep_runtime
+= delta
;
1759 if (tsk
->in_iowait
) {
1760 se
->statistics
.iowait_sum
+= delta
;
1761 se
->statistics
.iowait_count
++;
1762 trace_sched_stat_iowait(tsk
, delta
);
1765 trace_sched_stat_blocked(tsk
, delta
);
1768 * Blocking time is in units of nanosecs, so shift by
1769 * 20 to get a milliseconds-range estimation of the
1770 * amount of time that the task spent sleeping:
1772 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
1773 profile_hits(SLEEP_PROFILING
,
1774 (void *)get_wchan(tsk
),
1777 account_scheduler_latency(tsk
, delta
>> 10, 0);
1783 static void check_spread(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1785 #ifdef CONFIG_SCHED_DEBUG
1786 s64 d
= se
->vruntime
- cfs_rq
->min_vruntime
;
1791 if (d
> 3*sysctl_sched_latency
)
1792 schedstat_inc(cfs_rq
, nr_spread_over
);
1797 place_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int initial
)
1799 u64 vruntime
= cfs_rq
->min_vruntime
;
1802 * The 'current' period is already promised to the current tasks,
1803 * however the extra weight of the new task will slow them down a
1804 * little, place the new task so that it fits in the slot that
1805 * stays open at the end.
1807 if (initial
&& sched_feat(START_DEBIT
))
1808 vruntime
+= sched_vslice(cfs_rq
, se
);
1810 /* sleeps up to a single latency don't count. */
1812 unsigned long thresh
= sysctl_sched_latency
;
1815 * Halve their sleep time's effect, to allow
1816 * for a gentler effect of sleepers:
1818 if (sched_feat(GENTLE_FAIR_SLEEPERS
))
1824 /* ensure we never gain time by being placed backwards. */
1825 se
->vruntime
= max_vruntime(se
->vruntime
, vruntime
);
1828 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
);
1831 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
1834 * Update the normalized vruntime before updating min_vruntime
1835 * through calling update_curr().
1837 if (!(flags
& ENQUEUE_WAKEUP
) || (flags
& ENQUEUE_WAKING
))
1838 se
->vruntime
+= cfs_rq
->min_vruntime
;
1841 * Update run-time statistics of the 'current'.
1843 update_curr(cfs_rq
);
1844 enqueue_entity_load_avg(cfs_rq
, se
, flags
& ENQUEUE_WAKEUP
);
1845 account_entity_enqueue(cfs_rq
, se
);
1846 update_cfs_shares(cfs_rq
);
1848 if (flags
& ENQUEUE_WAKEUP
) {
1849 place_entity(cfs_rq
, se
, 0);
1850 enqueue_sleeper(cfs_rq
, se
);
1853 update_stats_enqueue(cfs_rq
, se
);
1854 check_spread(cfs_rq
, se
);
1855 if (se
!= cfs_rq
->curr
)
1856 __enqueue_entity(cfs_rq
, se
);
1859 if (cfs_rq
->nr_running
== 1) {
1860 list_add_leaf_cfs_rq(cfs_rq
);
1861 check_enqueue_throttle(cfs_rq
);
1865 static void __clear_buddies_last(struct sched_entity
*se
)
1867 for_each_sched_entity(se
) {
1868 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1869 if (cfs_rq
->last
== se
)
1870 cfs_rq
->last
= NULL
;
1876 static void __clear_buddies_next(struct sched_entity
*se
)
1878 for_each_sched_entity(se
) {
1879 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1880 if (cfs_rq
->next
== se
)
1881 cfs_rq
->next
= NULL
;
1887 static void __clear_buddies_skip(struct sched_entity
*se
)
1889 for_each_sched_entity(se
) {
1890 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
1891 if (cfs_rq
->skip
== se
)
1892 cfs_rq
->skip
= NULL
;
1898 static void clear_buddies(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
1900 if (cfs_rq
->last
== se
)
1901 __clear_buddies_last(se
);
1903 if (cfs_rq
->next
== se
)
1904 __clear_buddies_next(se
);
1906 if (cfs_rq
->skip
== se
)
1907 __clear_buddies_skip(se
);
1910 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
1913 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int flags
)
1916 * Update run-time statistics of the 'current'.
1918 update_curr(cfs_rq
);
1919 dequeue_entity_load_avg(cfs_rq
, se
, flags
& DEQUEUE_SLEEP
);
1921 update_stats_dequeue(cfs_rq
, se
);
1922 if (flags
& DEQUEUE_SLEEP
) {
1923 #ifdef CONFIG_SCHEDSTATS
1924 if (entity_is_task(se
)) {
1925 struct task_struct
*tsk
= task_of(se
);
1927 if (tsk
->state
& TASK_INTERRUPTIBLE
)
1928 se
->statistics
.sleep_start
= rq_clock(rq_of(cfs_rq
));
1929 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
1930 se
->statistics
.block_start
= rq_clock(rq_of(cfs_rq
));
1935 clear_buddies(cfs_rq
, se
);
1937 if (se
!= cfs_rq
->curr
)
1938 __dequeue_entity(cfs_rq
, se
);
1940 account_entity_dequeue(cfs_rq
, se
);
1943 * Normalize the entity after updating the min_vruntime because the
1944 * update can refer to the ->curr item and we need to reflect this
1945 * movement in our normalized position.
1947 if (!(flags
& DEQUEUE_SLEEP
))
1948 se
->vruntime
-= cfs_rq
->min_vruntime
;
1950 /* return excess runtime on last dequeue */
1951 return_cfs_rq_runtime(cfs_rq
);
1953 update_min_vruntime(cfs_rq
);
1954 update_cfs_shares(cfs_rq
);
1958 * Preempt the current task with a newly woken task if needed:
1961 check_preempt_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
1963 unsigned long ideal_runtime
, delta_exec
;
1964 struct sched_entity
*se
;
1967 ideal_runtime
= sched_slice(cfs_rq
, curr
);
1968 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
1969 if (delta_exec
> ideal_runtime
) {
1970 resched_task(rq_of(cfs_rq
)->curr
);
1972 * The current task ran long enough, ensure it doesn't get
1973 * re-elected due to buddy favours.
1975 clear_buddies(cfs_rq
, curr
);
1980 * Ensure that a task that missed wakeup preemption by a
1981 * narrow margin doesn't have to wait for a full slice.
1982 * This also mitigates buddy induced latencies under load.
1984 if (delta_exec
< sysctl_sched_min_granularity
)
1987 se
= __pick_first_entity(cfs_rq
);
1988 delta
= curr
->vruntime
- se
->vruntime
;
1993 if (delta
> ideal_runtime
)
1994 resched_task(rq_of(cfs_rq
)->curr
);
1998 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
2000 /* 'current' is not kept within the tree. */
2003 * Any task has to be enqueued before it get to execute on
2004 * a CPU. So account for the time it spent waiting on the
2007 update_stats_wait_end(cfs_rq
, se
);
2008 __dequeue_entity(cfs_rq
, se
);
2011 update_stats_curr_start(cfs_rq
, se
);
2013 #ifdef CONFIG_SCHEDSTATS
2015 * Track our maximum slice length, if the CPU's load is at
2016 * least twice that of our own weight (i.e. dont track it
2017 * when there are only lesser-weight tasks around):
2019 if (rq_of(cfs_rq
)->load
.weight
>= 2*se
->load
.weight
) {
2020 se
->statistics
.slice_max
= max(se
->statistics
.slice_max
,
2021 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
2024 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
2028 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
);
2031 * Pick the next process, keeping these things in mind, in this order:
2032 * 1) keep things fair between processes/task groups
2033 * 2) pick the "next" process, since someone really wants that to run
2034 * 3) pick the "last" process, for cache locality
2035 * 4) do not run the "skip" process, if something else is available
2037 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
2039 struct sched_entity
*se
= __pick_first_entity(cfs_rq
);
2040 struct sched_entity
*left
= se
;
2043 * Avoid running the skip buddy, if running something else can
2044 * be done without getting too unfair.
2046 if (cfs_rq
->skip
== se
) {
2047 struct sched_entity
*second
= __pick_next_entity(se
);
2048 if (second
&& wakeup_preempt_entity(second
, left
) < 1)
2053 * Prefer last buddy, try to return the CPU to a preempted task.
2055 if (cfs_rq
->last
&& wakeup_preempt_entity(cfs_rq
->last
, left
) < 1)
2059 * Someone really wants this to run. If it's not unfair, run it.
2061 if (cfs_rq
->next
&& wakeup_preempt_entity(cfs_rq
->next
, left
) < 1)
2064 clear_buddies(cfs_rq
, se
);
2069 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
);
2071 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
2074 * If still on the runqueue then deactivate_task()
2075 * was not called and update_curr() has to be done:
2078 update_curr(cfs_rq
);
2080 /* throttle cfs_rqs exceeding runtime */
2081 check_cfs_rq_runtime(cfs_rq
);
2083 check_spread(cfs_rq
, prev
);
2085 update_stats_wait_start(cfs_rq
, prev
);
2086 /* Put 'current' back into the tree. */
2087 __enqueue_entity(cfs_rq
, prev
);
2088 /* in !on_rq case, update occurred at dequeue */
2089 update_entity_load_avg(prev
, 1);
2091 cfs_rq
->curr
= NULL
;
2095 entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
, int queued
)
2098 * Update run-time statistics of the 'current'.
2100 update_curr(cfs_rq
);
2103 * Ensure that runnable average is periodically updated.
2105 update_entity_load_avg(curr
, 1);
2106 update_cfs_rq_blocked_load(cfs_rq
, 1);
2107 update_cfs_shares(cfs_rq
);
2109 #ifdef CONFIG_SCHED_HRTICK
2111 * queued ticks are scheduled to match the slice, so don't bother
2112 * validating it and just reschedule.
2115 resched_task(rq_of(cfs_rq
)->curr
);
2119 * don't let the period tick interfere with the hrtick preemption
2121 if (!sched_feat(DOUBLE_TICK
) &&
2122 hrtimer_active(&rq_of(cfs_rq
)->hrtick_timer
))
2126 if (cfs_rq
->nr_running
> 1)
2127 check_preempt_tick(cfs_rq
, curr
);
2131 /**************************************************
2132 * CFS bandwidth control machinery
2135 #ifdef CONFIG_CFS_BANDWIDTH
2137 #ifdef HAVE_JUMP_LABEL
2138 static struct static_key __cfs_bandwidth_used
;
2140 static inline bool cfs_bandwidth_used(void)
2142 return static_key_false(&__cfs_bandwidth_used
);
2145 void account_cfs_bandwidth_used(int enabled
, int was_enabled
)
2147 /* only need to count groups transitioning between enabled/!enabled */
2148 if (enabled
&& !was_enabled
)
2149 static_key_slow_inc(&__cfs_bandwidth_used
);
2150 else if (!enabled
&& was_enabled
)
2151 static_key_slow_dec(&__cfs_bandwidth_used
);
2153 #else /* HAVE_JUMP_LABEL */
2154 static bool cfs_bandwidth_used(void)
2159 void account_cfs_bandwidth_used(int enabled
, int was_enabled
) {}
2160 #endif /* HAVE_JUMP_LABEL */
2163 * default period for cfs group bandwidth.
2164 * default: 0.1s, units: nanoseconds
2166 static inline u64
default_cfs_period(void)
2168 return 100000000ULL;
2171 static inline u64
sched_cfs_bandwidth_slice(void)
2173 return (u64
)sysctl_sched_cfs_bandwidth_slice
* NSEC_PER_USEC
;
2177 * Replenish runtime according to assigned quota and update expiration time.
2178 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2179 * additional synchronization around rq->lock.
2181 * requires cfs_b->lock
2183 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth
*cfs_b
)
2187 if (cfs_b
->quota
== RUNTIME_INF
)
2190 now
= sched_clock_cpu(smp_processor_id());
2191 cfs_b
->runtime
= cfs_b
->quota
;
2192 cfs_b
->runtime_expires
= now
+ ktime_to_ns(cfs_b
->period
);
2195 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
2197 return &tg
->cfs_bandwidth
;
2200 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2201 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
2203 if (unlikely(cfs_rq
->throttle_count
))
2204 return cfs_rq
->throttled_clock_task
;
2206 return rq_clock_task(rq_of(cfs_rq
)) - cfs_rq
->throttled_clock_task_time
;
2209 /* returns 0 on failure to allocate runtime */
2210 static int assign_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2212 struct task_group
*tg
= cfs_rq
->tg
;
2213 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
2214 u64 amount
= 0, min_amount
, expires
;
2216 /* note: this is a positive sum as runtime_remaining <= 0 */
2217 min_amount
= sched_cfs_bandwidth_slice() - cfs_rq
->runtime_remaining
;
2219 raw_spin_lock(&cfs_b
->lock
);
2220 if (cfs_b
->quota
== RUNTIME_INF
)
2221 amount
= min_amount
;
2224 * If the bandwidth pool has become inactive, then at least one
2225 * period must have elapsed since the last consumption.
2226 * Refresh the global state and ensure bandwidth timer becomes
2229 if (!cfs_b
->timer_active
) {
2230 __refill_cfs_bandwidth_runtime(cfs_b
);
2231 __start_cfs_bandwidth(cfs_b
);
2234 if (cfs_b
->runtime
> 0) {
2235 amount
= min(cfs_b
->runtime
, min_amount
);
2236 cfs_b
->runtime
-= amount
;
2240 expires
= cfs_b
->runtime_expires
;
2241 raw_spin_unlock(&cfs_b
->lock
);
2243 cfs_rq
->runtime_remaining
+= amount
;
2245 * we may have advanced our local expiration to account for allowed
2246 * spread between our sched_clock and the one on which runtime was
2249 if ((s64
)(expires
- cfs_rq
->runtime_expires
) > 0)
2250 cfs_rq
->runtime_expires
= expires
;
2252 return cfs_rq
->runtime_remaining
> 0;
2256 * Note: This depends on the synchronization provided by sched_clock and the
2257 * fact that rq->clock snapshots this value.
2259 static void expire_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2261 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
2263 /* if the deadline is ahead of our clock, nothing to do */
2264 if (likely((s64
)(rq_clock(rq_of(cfs_rq
)) - cfs_rq
->runtime_expires
) < 0))
2267 if (cfs_rq
->runtime_remaining
< 0)
2271 * If the local deadline has passed we have to consider the
2272 * possibility that our sched_clock is 'fast' and the global deadline
2273 * has not truly expired.
2275 * Fortunately we can check determine whether this the case by checking
2276 * whether the global deadline has advanced.
2279 if ((s64
)(cfs_rq
->runtime_expires
- cfs_b
->runtime_expires
) >= 0) {
2280 /* extend local deadline, drift is bounded above by 2 ticks */
2281 cfs_rq
->runtime_expires
+= TICK_NSEC
;
2283 /* global deadline is ahead, expiration has passed */
2284 cfs_rq
->runtime_remaining
= 0;
2288 static void __account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
,
2289 unsigned long delta_exec
)
2291 /* dock delta_exec before expiring quota (as it could span periods) */
2292 cfs_rq
->runtime_remaining
-= delta_exec
;
2293 expire_cfs_rq_runtime(cfs_rq
);
2295 if (likely(cfs_rq
->runtime_remaining
> 0))
2299 * if we're unable to extend our runtime we resched so that the active
2300 * hierarchy can be throttled
2302 if (!assign_cfs_rq_runtime(cfs_rq
) && likely(cfs_rq
->curr
))
2303 resched_task(rq_of(cfs_rq
)->curr
);
2306 static __always_inline
2307 void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
, unsigned long delta_exec
)
2309 if (!cfs_bandwidth_used() || !cfs_rq
->runtime_enabled
)
2312 __account_cfs_rq_runtime(cfs_rq
, delta_exec
);
2315 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
2317 return cfs_bandwidth_used() && cfs_rq
->throttled
;
2320 /* check whether cfs_rq, or any parent, is throttled */
2321 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
2323 return cfs_bandwidth_used() && cfs_rq
->throttle_count
;
2327 * Ensure that neither of the group entities corresponding to src_cpu or
2328 * dest_cpu are members of a throttled hierarchy when performing group
2329 * load-balance operations.
2331 static inline int throttled_lb_pair(struct task_group
*tg
,
2332 int src_cpu
, int dest_cpu
)
2334 struct cfs_rq
*src_cfs_rq
, *dest_cfs_rq
;
2336 src_cfs_rq
= tg
->cfs_rq
[src_cpu
];
2337 dest_cfs_rq
= tg
->cfs_rq
[dest_cpu
];
2339 return throttled_hierarchy(src_cfs_rq
) ||
2340 throttled_hierarchy(dest_cfs_rq
);
2343 /* updated child weight may affect parent so we have to do this bottom up */
2344 static int tg_unthrottle_up(struct task_group
*tg
, void *data
)
2346 struct rq
*rq
= data
;
2347 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
2349 cfs_rq
->throttle_count
--;
2351 if (!cfs_rq
->throttle_count
) {
2352 /* adjust cfs_rq_clock_task() */
2353 cfs_rq
->throttled_clock_task_time
+= rq_clock_task(rq
) -
2354 cfs_rq
->throttled_clock_task
;
2361 static int tg_throttle_down(struct task_group
*tg
, void *data
)
2363 struct rq
*rq
= data
;
2364 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu_of(rq
)];
2366 /* group is entering throttled state, stop time */
2367 if (!cfs_rq
->throttle_count
)
2368 cfs_rq
->throttled_clock_task
= rq_clock_task(rq
);
2369 cfs_rq
->throttle_count
++;
2374 static void throttle_cfs_rq(struct cfs_rq
*cfs_rq
)
2376 struct rq
*rq
= rq_of(cfs_rq
);
2377 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
2378 struct sched_entity
*se
;
2379 long task_delta
, dequeue
= 1;
2381 se
= cfs_rq
->tg
->se
[cpu_of(rq_of(cfs_rq
))];
2383 /* freeze hierarchy runnable averages while throttled */
2385 walk_tg_tree_from(cfs_rq
->tg
, tg_throttle_down
, tg_nop
, (void *)rq
);
2388 task_delta
= cfs_rq
->h_nr_running
;
2389 for_each_sched_entity(se
) {
2390 struct cfs_rq
*qcfs_rq
= cfs_rq_of(se
);
2391 /* throttled entity or throttle-on-deactivate */
2396 dequeue_entity(qcfs_rq
, se
, DEQUEUE_SLEEP
);
2397 qcfs_rq
->h_nr_running
-= task_delta
;
2399 if (qcfs_rq
->load
.weight
)
2404 rq
->nr_running
-= task_delta
;
2406 cfs_rq
->throttled
= 1;
2407 cfs_rq
->throttled_clock
= rq_clock(rq
);
2408 raw_spin_lock(&cfs_b
->lock
);
2409 list_add_tail_rcu(&cfs_rq
->throttled_list
, &cfs_b
->throttled_cfs_rq
);
2410 raw_spin_unlock(&cfs_b
->lock
);
2413 void unthrottle_cfs_rq(struct cfs_rq
*cfs_rq
)
2415 struct rq
*rq
= rq_of(cfs_rq
);
2416 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
2417 struct sched_entity
*se
;
2421 se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
2423 cfs_rq
->throttled
= 0;
2425 update_rq_clock(rq
);
2427 raw_spin_lock(&cfs_b
->lock
);
2428 cfs_b
->throttled_time
+= rq_clock(rq
) - cfs_rq
->throttled_clock
;
2429 list_del_rcu(&cfs_rq
->throttled_list
);
2430 raw_spin_unlock(&cfs_b
->lock
);
2432 /* update hierarchical throttle state */
2433 walk_tg_tree_from(cfs_rq
->tg
, tg_nop
, tg_unthrottle_up
, (void *)rq
);
2435 if (!cfs_rq
->load
.weight
)
2438 task_delta
= cfs_rq
->h_nr_running
;
2439 for_each_sched_entity(se
) {
2443 cfs_rq
= cfs_rq_of(se
);
2445 enqueue_entity(cfs_rq
, se
, ENQUEUE_WAKEUP
);
2446 cfs_rq
->h_nr_running
+= task_delta
;
2448 if (cfs_rq_throttled(cfs_rq
))
2453 rq
->nr_running
+= task_delta
;
2455 /* determine whether we need to wake up potentially idle cpu */
2456 if (rq
->curr
== rq
->idle
&& rq
->cfs
.nr_running
)
2457 resched_task(rq
->curr
);
2460 static u64
distribute_cfs_runtime(struct cfs_bandwidth
*cfs_b
,
2461 u64 remaining
, u64 expires
)
2463 struct cfs_rq
*cfs_rq
;
2464 u64 runtime
= remaining
;
2467 list_for_each_entry_rcu(cfs_rq
, &cfs_b
->throttled_cfs_rq
,
2469 struct rq
*rq
= rq_of(cfs_rq
);
2471 raw_spin_lock(&rq
->lock
);
2472 if (!cfs_rq_throttled(cfs_rq
))
2475 runtime
= -cfs_rq
->runtime_remaining
+ 1;
2476 if (runtime
> remaining
)
2477 runtime
= remaining
;
2478 remaining
-= runtime
;
2480 cfs_rq
->runtime_remaining
+= runtime
;
2481 cfs_rq
->runtime_expires
= expires
;
2483 /* we check whether we're throttled above */
2484 if (cfs_rq
->runtime_remaining
> 0)
2485 unthrottle_cfs_rq(cfs_rq
);
2488 raw_spin_unlock(&rq
->lock
);
2499 * Responsible for refilling a task_group's bandwidth and unthrottling its
2500 * cfs_rqs as appropriate. If there has been no activity within the last
2501 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2502 * used to track this state.
2504 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
)
2506 u64 runtime
, runtime_expires
;
2507 int idle
= 1, throttled
;
2509 raw_spin_lock(&cfs_b
->lock
);
2510 /* no need to continue the timer with no bandwidth constraint */
2511 if (cfs_b
->quota
== RUNTIME_INF
)
2514 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
2515 /* idle depends on !throttled (for the case of a large deficit) */
2516 idle
= cfs_b
->idle
&& !throttled
;
2517 cfs_b
->nr_periods
+= overrun
;
2519 /* if we're going inactive then everything else can be deferred */
2523 __refill_cfs_bandwidth_runtime(cfs_b
);
2526 /* mark as potentially idle for the upcoming period */
2531 /* account preceding periods in which throttling occurred */
2532 cfs_b
->nr_throttled
+= overrun
;
2535 * There are throttled entities so we must first use the new bandwidth
2536 * to unthrottle them before making it generally available. This
2537 * ensures that all existing debts will be paid before a new cfs_rq is
2540 runtime
= cfs_b
->runtime
;
2541 runtime_expires
= cfs_b
->runtime_expires
;
2545 * This check is repeated as we are holding onto the new bandwidth
2546 * while we unthrottle. This can potentially race with an unthrottled
2547 * group trying to acquire new bandwidth from the global pool.
2549 while (throttled
&& runtime
> 0) {
2550 raw_spin_unlock(&cfs_b
->lock
);
2551 /* we can't nest cfs_b->lock while distributing bandwidth */
2552 runtime
= distribute_cfs_runtime(cfs_b
, runtime
,
2554 raw_spin_lock(&cfs_b
->lock
);
2556 throttled
= !list_empty(&cfs_b
->throttled_cfs_rq
);
2559 /* return (any) remaining runtime */
2560 cfs_b
->runtime
= runtime
;
2562 * While we are ensured activity in the period following an
2563 * unthrottle, this also covers the case in which the new bandwidth is
2564 * insufficient to cover the existing bandwidth deficit. (Forcing the
2565 * timer to remain active while there are any throttled entities.)
2570 cfs_b
->timer_active
= 0;
2571 raw_spin_unlock(&cfs_b
->lock
);
2576 /* a cfs_rq won't donate quota below this amount */
2577 static const u64 min_cfs_rq_runtime
= 1 * NSEC_PER_MSEC
;
2578 /* minimum remaining period time to redistribute slack quota */
2579 static const u64 min_bandwidth_expiration
= 2 * NSEC_PER_MSEC
;
2580 /* how long we wait to gather additional slack before distributing */
2581 static const u64 cfs_bandwidth_slack_period
= 5 * NSEC_PER_MSEC
;
2583 /* are we near the end of the current quota period? */
2584 static int runtime_refresh_within(struct cfs_bandwidth
*cfs_b
, u64 min_expire
)
2586 struct hrtimer
*refresh_timer
= &cfs_b
->period_timer
;
2589 /* if the call-back is running a quota refresh is already occurring */
2590 if (hrtimer_callback_running(refresh_timer
))
2593 /* is a quota refresh about to occur? */
2594 remaining
= ktime_to_ns(hrtimer_expires_remaining(refresh_timer
));
2595 if (remaining
< min_expire
)
2601 static void start_cfs_slack_bandwidth(struct cfs_bandwidth
*cfs_b
)
2603 u64 min_left
= cfs_bandwidth_slack_period
+ min_bandwidth_expiration
;
2605 /* if there's a quota refresh soon don't bother with slack */
2606 if (runtime_refresh_within(cfs_b
, min_left
))
2609 start_bandwidth_timer(&cfs_b
->slack_timer
,
2610 ns_to_ktime(cfs_bandwidth_slack_period
));
2613 /* we know any runtime found here is valid as update_curr() precedes return */
2614 static void __return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2616 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
2617 s64 slack_runtime
= cfs_rq
->runtime_remaining
- min_cfs_rq_runtime
;
2619 if (slack_runtime
<= 0)
2622 raw_spin_lock(&cfs_b
->lock
);
2623 if (cfs_b
->quota
!= RUNTIME_INF
&&
2624 cfs_rq
->runtime_expires
== cfs_b
->runtime_expires
) {
2625 cfs_b
->runtime
+= slack_runtime
;
2627 /* we are under rq->lock, defer unthrottling using a timer */
2628 if (cfs_b
->runtime
> sched_cfs_bandwidth_slice() &&
2629 !list_empty(&cfs_b
->throttled_cfs_rq
))
2630 start_cfs_slack_bandwidth(cfs_b
);
2632 raw_spin_unlock(&cfs_b
->lock
);
2634 /* even if it's not valid for return we don't want to try again */
2635 cfs_rq
->runtime_remaining
-= slack_runtime
;
2638 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2640 if (!cfs_bandwidth_used())
2643 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->nr_running
)
2646 __return_cfs_rq_runtime(cfs_rq
);
2650 * This is done with a timer (instead of inline with bandwidth return) since
2651 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2653 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
)
2655 u64 runtime
= 0, slice
= sched_cfs_bandwidth_slice();
2658 /* confirm we're still not at a refresh boundary */
2659 if (runtime_refresh_within(cfs_b
, min_bandwidth_expiration
))
2662 raw_spin_lock(&cfs_b
->lock
);
2663 if (cfs_b
->quota
!= RUNTIME_INF
&& cfs_b
->runtime
> slice
) {
2664 runtime
= cfs_b
->runtime
;
2667 expires
= cfs_b
->runtime_expires
;
2668 raw_spin_unlock(&cfs_b
->lock
);
2673 runtime
= distribute_cfs_runtime(cfs_b
, runtime
, expires
);
2675 raw_spin_lock(&cfs_b
->lock
);
2676 if (expires
== cfs_b
->runtime_expires
)
2677 cfs_b
->runtime
= runtime
;
2678 raw_spin_unlock(&cfs_b
->lock
);
2682 * When a group wakes up we want to make sure that its quota is not already
2683 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2684 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2686 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
)
2688 if (!cfs_bandwidth_used())
2691 /* an active group must be handled by the update_curr()->put() path */
2692 if (!cfs_rq
->runtime_enabled
|| cfs_rq
->curr
)
2695 /* ensure the group is not already throttled */
2696 if (cfs_rq_throttled(cfs_rq
))
2699 /* update runtime allocation */
2700 account_cfs_rq_runtime(cfs_rq
, 0);
2701 if (cfs_rq
->runtime_remaining
<= 0)
2702 throttle_cfs_rq(cfs_rq
);
2705 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2706 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2708 if (!cfs_bandwidth_used())
2711 if (likely(!cfs_rq
->runtime_enabled
|| cfs_rq
->runtime_remaining
> 0))
2715 * it's possible for a throttled entity to be forced into a running
2716 * state (e.g. set_curr_task), in this case we're finished.
2718 if (cfs_rq_throttled(cfs_rq
))
2721 throttle_cfs_rq(cfs_rq
);
2724 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
2726 struct cfs_bandwidth
*cfs_b
=
2727 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
2728 do_sched_cfs_slack_timer(cfs_b
);
2730 return HRTIMER_NORESTART
;
2733 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
2735 struct cfs_bandwidth
*cfs_b
=
2736 container_of(timer
, struct cfs_bandwidth
, period_timer
);
2742 now
= hrtimer_cb_get_time(timer
);
2743 overrun
= hrtimer_forward(timer
, now
, cfs_b
->period
);
2748 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
2751 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
2754 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
2756 raw_spin_lock_init(&cfs_b
->lock
);
2758 cfs_b
->quota
= RUNTIME_INF
;
2759 cfs_b
->period
= ns_to_ktime(default_cfs_period());
2761 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
2762 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
2763 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
2764 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
2765 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
2768 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
2770 cfs_rq
->runtime_enabled
= 0;
2771 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
2774 /* requires cfs_b->lock, may release to reprogram timer */
2775 void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
2778 * The timer may be active because we're trying to set a new bandwidth
2779 * period or because we're racing with the tear-down path
2780 * (timer_active==0 becomes visible before the hrtimer call-back
2781 * terminates). In either case we ensure that it's re-programmed
2783 while (unlikely(hrtimer_active(&cfs_b
->period_timer
))) {
2784 raw_spin_unlock(&cfs_b
->lock
);
2785 /* ensure cfs_b->lock is available while we wait */
2786 hrtimer_cancel(&cfs_b
->period_timer
);
2788 raw_spin_lock(&cfs_b
->lock
);
2789 /* if someone else restarted the timer then we're done */
2790 if (cfs_b
->timer_active
)
2794 cfs_b
->timer_active
= 1;
2795 start_bandwidth_timer(&cfs_b
->period_timer
, cfs_b
->period
);
2798 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
2800 hrtimer_cancel(&cfs_b
->period_timer
);
2801 hrtimer_cancel(&cfs_b
->slack_timer
);
2804 static void __maybe_unused
unthrottle_offline_cfs_rqs(struct rq
*rq
)
2806 struct cfs_rq
*cfs_rq
;
2808 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
2809 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
2811 if (!cfs_rq
->runtime_enabled
)
2815 * clock_task is not advancing so we just need to make sure
2816 * there's some valid quota amount
2818 cfs_rq
->runtime_remaining
= cfs_b
->quota
;
2819 if (cfs_rq_throttled(cfs_rq
))
2820 unthrottle_cfs_rq(cfs_rq
);
2824 #else /* CONFIG_CFS_BANDWIDTH */
2825 static inline u64
cfs_rq_clock_task(struct cfs_rq
*cfs_rq
)
2827 return rq_clock_task(rq_of(cfs_rq
));
2830 static void account_cfs_rq_runtime(struct cfs_rq
*cfs_rq
,
2831 unsigned long delta_exec
) {}
2832 static void check_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
2833 static void check_enqueue_throttle(struct cfs_rq
*cfs_rq
) {}
2834 static __always_inline
void return_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
2836 static inline int cfs_rq_throttled(struct cfs_rq
*cfs_rq
)
2841 static inline int throttled_hierarchy(struct cfs_rq
*cfs_rq
)
2846 static inline int throttled_lb_pair(struct task_group
*tg
,
2847 int src_cpu
, int dest_cpu
)
2852 void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
2854 #ifdef CONFIG_FAIR_GROUP_SCHED
2855 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
2858 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
2862 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
2863 static inline void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
2865 #endif /* CONFIG_CFS_BANDWIDTH */
2867 /**************************************************
2868 * CFS operations on tasks:
2871 #ifdef CONFIG_SCHED_HRTICK
2872 static void hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
2874 struct sched_entity
*se
= &p
->se
;
2875 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
2877 WARN_ON(task_rq(p
) != rq
);
2879 if (cfs_rq
->nr_running
> 1) {
2880 u64 slice
= sched_slice(cfs_rq
, se
);
2881 u64 ran
= se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
;
2882 s64 delta
= slice
- ran
;
2891 * Don't schedule slices shorter than 10000ns, that just
2892 * doesn't make sense. Rely on vruntime for fairness.
2895 delta
= max_t(s64
, 10000LL, delta
);
2897 hrtick_start(rq
, delta
);
2902 * called from enqueue/dequeue and updates the hrtick when the
2903 * current task is from our class and nr_running is low enough
2906 static void hrtick_update(struct rq
*rq
)
2908 struct task_struct
*curr
= rq
->curr
;
2910 if (!hrtick_enabled(rq
) || curr
->sched_class
!= &fair_sched_class
)
2913 if (cfs_rq_of(&curr
->se
)->nr_running
< sched_nr_latency
)
2914 hrtick_start_fair(rq
, curr
);
2916 #else /* !CONFIG_SCHED_HRTICK */
2918 hrtick_start_fair(struct rq
*rq
, struct task_struct
*p
)
2922 static inline void hrtick_update(struct rq
*rq
)
2928 * The enqueue_task method is called before nr_running is
2929 * increased. Here we update the fair scheduling stats and
2930 * then put the task into the rbtree:
2933 enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
2935 struct cfs_rq
*cfs_rq
;
2936 struct sched_entity
*se
= &p
->se
;
2938 for_each_sched_entity(se
) {
2941 cfs_rq
= cfs_rq_of(se
);
2942 enqueue_entity(cfs_rq
, se
, flags
);
2945 * end evaluation on encountering a throttled cfs_rq
2947 * note: in the case of encountering a throttled cfs_rq we will
2948 * post the final h_nr_running increment below.
2950 if (cfs_rq_throttled(cfs_rq
))
2952 cfs_rq
->h_nr_running
++;
2954 flags
= ENQUEUE_WAKEUP
;
2957 for_each_sched_entity(se
) {
2958 cfs_rq
= cfs_rq_of(se
);
2959 cfs_rq
->h_nr_running
++;
2961 if (cfs_rq_throttled(cfs_rq
))
2964 update_cfs_shares(cfs_rq
);
2965 update_entity_load_avg(se
, 1);
2969 update_rq_runnable_avg(rq
, rq
->nr_running
);
2975 static void set_next_buddy(struct sched_entity
*se
);
2978 * The dequeue_task method is called before nr_running is
2979 * decreased. We remove the task from the rbtree and
2980 * update the fair scheduling stats:
2982 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int flags
)
2984 struct cfs_rq
*cfs_rq
;
2985 struct sched_entity
*se
= &p
->se
;
2986 int task_sleep
= flags
& DEQUEUE_SLEEP
;
2988 for_each_sched_entity(se
) {
2989 cfs_rq
= cfs_rq_of(se
);
2990 dequeue_entity(cfs_rq
, se
, flags
);
2993 * end evaluation on encountering a throttled cfs_rq
2995 * note: in the case of encountering a throttled cfs_rq we will
2996 * post the final h_nr_running decrement below.
2998 if (cfs_rq_throttled(cfs_rq
))
3000 cfs_rq
->h_nr_running
--;
3002 /* Don't dequeue parent if it has other entities besides us */
3003 if (cfs_rq
->load
.weight
) {
3005 * Bias pick_next to pick a task from this cfs_rq, as
3006 * p is sleeping when it is within its sched_slice.
3008 if (task_sleep
&& parent_entity(se
))
3009 set_next_buddy(parent_entity(se
));
3011 /* avoid re-evaluating load for this entity */
3012 se
= parent_entity(se
);
3015 flags
|= DEQUEUE_SLEEP
;
3018 for_each_sched_entity(se
) {
3019 cfs_rq
= cfs_rq_of(se
);
3020 cfs_rq
->h_nr_running
--;
3022 if (cfs_rq_throttled(cfs_rq
))
3025 update_cfs_shares(cfs_rq
);
3026 update_entity_load_avg(se
, 1);
3031 update_rq_runnable_avg(rq
, 1);
3037 /* Used instead of source_load when we know the type == 0 */
3038 static unsigned long weighted_cpuload(const int cpu
)
3040 return cpu_rq(cpu
)->cfs
.runnable_load_avg
;
3044 * Return a low guess at the load of a migration-source cpu weighted
3045 * according to the scheduling class and "nice" value.
3047 * We want to under-estimate the load of migration sources, to
3048 * balance conservatively.
3050 static unsigned long source_load(int cpu
, int type
)
3052 struct rq
*rq
= cpu_rq(cpu
);
3053 unsigned long total
= weighted_cpuload(cpu
);
3055 if (type
== 0 || !sched_feat(LB_BIAS
))
3058 return min(rq
->cpu_load
[type
-1], total
);
3062 * Return a high guess at the load of a migration-target cpu weighted
3063 * according to the scheduling class and "nice" value.
3065 static unsigned long target_load(int cpu
, int type
)
3067 struct rq
*rq
= cpu_rq(cpu
);
3068 unsigned long total
= weighted_cpuload(cpu
);
3070 if (type
== 0 || !sched_feat(LB_BIAS
))
3073 return max(rq
->cpu_load
[type
-1], total
);
3076 static unsigned long power_of(int cpu
)
3078 return cpu_rq(cpu
)->cpu_power
;
3081 static unsigned long cpu_avg_load_per_task(int cpu
)
3083 struct rq
*rq
= cpu_rq(cpu
);
3084 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
3085 unsigned long load_avg
= rq
->cfs
.runnable_load_avg
;
3088 return load_avg
/ nr_running
;
3093 static void record_wakee(struct task_struct
*p
)
3096 * Rough decay (wiping) for cost saving, don't worry
3097 * about the boundary, really active task won't care
3100 if (jiffies
> current
->wakee_flip_decay_ts
+ HZ
) {
3101 current
->wakee_flips
= 0;
3102 current
->wakee_flip_decay_ts
= jiffies
;
3105 if (current
->last_wakee
!= p
) {
3106 current
->last_wakee
= p
;
3107 current
->wakee_flips
++;
3111 static void task_waking_fair(struct task_struct
*p
)
3113 struct sched_entity
*se
= &p
->se
;
3114 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3117 #ifndef CONFIG_64BIT
3118 u64 min_vruntime_copy
;
3121 min_vruntime_copy
= cfs_rq
->min_vruntime_copy
;
3123 min_vruntime
= cfs_rq
->min_vruntime
;
3124 } while (min_vruntime
!= min_vruntime_copy
);
3126 min_vruntime
= cfs_rq
->min_vruntime
;
3129 se
->vruntime
-= min_vruntime
;
3133 #ifdef CONFIG_FAIR_GROUP_SCHED
3135 * effective_load() calculates the load change as seen from the root_task_group
3137 * Adding load to a group doesn't make a group heavier, but can cause movement
3138 * of group shares between cpus. Assuming the shares were perfectly aligned one
3139 * can calculate the shift in shares.
3141 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3142 * on this @cpu and results in a total addition (subtraction) of @wg to the
3143 * total group weight.
3145 * Given a runqueue weight distribution (rw_i) we can compute a shares
3146 * distribution (s_i) using:
3148 * s_i = rw_i / \Sum rw_j (1)
3150 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3151 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3152 * shares distribution (s_i):
3154 * rw_i = { 2, 4, 1, 0 }
3155 * s_i = { 2/7, 4/7, 1/7, 0 }
3157 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3158 * task used to run on and the CPU the waker is running on), we need to
3159 * compute the effect of waking a task on either CPU and, in case of a sync
3160 * wakeup, compute the effect of the current task going to sleep.
3162 * So for a change of @wl to the local @cpu with an overall group weight change
3163 * of @wl we can compute the new shares distribution (s'_i) using:
3165 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3167 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3168 * differences in waking a task to CPU 0. The additional task changes the
3169 * weight and shares distributions like:
3171 * rw'_i = { 3, 4, 1, 0 }
3172 * s'_i = { 3/8, 4/8, 1/8, 0 }
3174 * We can then compute the difference in effective weight by using:
3176 * dw_i = S * (s'_i - s_i) (3)
3178 * Where 'S' is the group weight as seen by its parent.
3180 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3181 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3182 * 4/7) times the weight of the group.
3184 static long effective_load(struct task_group
*tg
, int cpu
, long wl
, long wg
)
3186 struct sched_entity
*se
= tg
->se
[cpu
];
3188 if (!tg
->parent
) /* the trivial, non-cgroup case */
3191 for_each_sched_entity(se
) {
3197 * W = @wg + \Sum rw_j
3199 W
= wg
+ calc_tg_weight(tg
, se
->my_q
);
3204 w
= se
->my_q
->load
.weight
+ wl
;
3207 * wl = S * s'_i; see (2)
3210 wl
= (w
* tg
->shares
) / W
;
3215 * Per the above, wl is the new se->load.weight value; since
3216 * those are clipped to [MIN_SHARES, ...) do so now. See
3217 * calc_cfs_shares().
3219 if (wl
< MIN_SHARES
)
3223 * wl = dw_i = S * (s'_i - s_i); see (3)
3225 wl
-= se
->load
.weight
;
3228 * Recursively apply this logic to all parent groups to compute
3229 * the final effective load change on the root group. Since
3230 * only the @tg group gets extra weight, all parent groups can
3231 * only redistribute existing shares. @wl is the shift in shares
3232 * resulting from this level per the above.
3241 static inline unsigned long effective_load(struct task_group
*tg
, int cpu
,
3242 unsigned long wl
, unsigned long wg
)
3249 static int wake_wide(struct task_struct
*p
)
3251 int factor
= this_cpu_read(sd_llc_size
);
3254 * Yeah, it's the switching-frequency, could means many wakee or
3255 * rapidly switch, use factor here will just help to automatically
3256 * adjust the loose-degree, so bigger node will lead to more pull.
3258 if (p
->wakee_flips
> factor
) {
3260 * wakee is somewhat hot, it needs certain amount of cpu
3261 * resource, so if waker is far more hot, prefer to leave
3264 if (current
->wakee_flips
> (factor
* p
->wakee_flips
))
3271 static int wake_affine(struct sched_domain
*sd
, struct task_struct
*p
, int sync
)
3273 s64 this_load
, load
;
3274 int idx
, this_cpu
, prev_cpu
;
3275 unsigned long tl_per_task
;
3276 struct task_group
*tg
;
3277 unsigned long weight
;
3281 * If we wake multiple tasks be careful to not bounce
3282 * ourselves around too much.
3288 this_cpu
= smp_processor_id();
3289 prev_cpu
= task_cpu(p
);
3290 load
= source_load(prev_cpu
, idx
);
3291 this_load
= target_load(this_cpu
, idx
);
3294 * If sync wakeup then subtract the (maximum possible)
3295 * effect of the currently running task from the load
3296 * of the current CPU:
3299 tg
= task_group(current
);
3300 weight
= current
->se
.load
.weight
;
3302 this_load
+= effective_load(tg
, this_cpu
, -weight
, -weight
);
3303 load
+= effective_load(tg
, prev_cpu
, 0, -weight
);
3307 weight
= p
->se
.load
.weight
;
3310 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3311 * due to the sync cause above having dropped this_load to 0, we'll
3312 * always have an imbalance, but there's really nothing you can do
3313 * about that, so that's good too.
3315 * Otherwise check if either cpus are near enough in load to allow this
3316 * task to be woken on this_cpu.
3318 if (this_load
> 0) {
3319 s64 this_eff_load
, prev_eff_load
;
3321 this_eff_load
= 100;
3322 this_eff_load
*= power_of(prev_cpu
);
3323 this_eff_load
*= this_load
+
3324 effective_load(tg
, this_cpu
, weight
, weight
);
3326 prev_eff_load
= 100 + (sd
->imbalance_pct
- 100) / 2;
3327 prev_eff_load
*= power_of(this_cpu
);
3328 prev_eff_load
*= load
+ effective_load(tg
, prev_cpu
, 0, weight
);
3330 balanced
= this_eff_load
<= prev_eff_load
;
3335 * If the currently running task will sleep within
3336 * a reasonable amount of time then attract this newly
3339 if (sync
&& balanced
)
3342 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine_attempts
);
3343 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
3346 (this_load
<= load
&&
3347 this_load
+ target_load(prev_cpu
, idx
) <= tl_per_task
)) {
3349 * This domain has SD_WAKE_AFFINE and
3350 * p is cache cold in this domain, and
3351 * there is no bad imbalance.
3353 schedstat_inc(sd
, ttwu_move_affine
);
3354 schedstat_inc(p
, se
.statistics
.nr_wakeups_affine
);
3362 * find_idlest_group finds and returns the least busy CPU group within the
3365 static struct sched_group
*
3366 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
,
3367 int this_cpu
, int load_idx
)
3369 struct sched_group
*idlest
= NULL
, *group
= sd
->groups
;
3370 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
3371 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
3374 unsigned long load
, avg_load
;
3378 /* Skip over this group if it has no CPUs allowed */
3379 if (!cpumask_intersects(sched_group_cpus(group
),
3380 tsk_cpus_allowed(p
)))
3383 local_group
= cpumask_test_cpu(this_cpu
,
3384 sched_group_cpus(group
));
3386 /* Tally up the load of all CPUs in the group */
3389 for_each_cpu(i
, sched_group_cpus(group
)) {
3390 /* Bias balancing toward cpus of our domain */
3392 load
= source_load(i
, load_idx
);
3394 load
= target_load(i
, load_idx
);
3399 /* Adjust by relative CPU power of the group */
3400 avg_load
= (avg_load
* SCHED_POWER_SCALE
) / group
->sgp
->power
;
3403 this_load
= avg_load
;
3404 } else if (avg_load
< min_load
) {
3405 min_load
= avg_load
;
3408 } while (group
= group
->next
, group
!= sd
->groups
);
3410 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
3416 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3419 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
3421 unsigned long load
, min_load
= ULONG_MAX
;
3425 /* Traverse only the allowed CPUs */
3426 for_each_cpu_and(i
, sched_group_cpus(group
), tsk_cpus_allowed(p
)) {
3427 load
= weighted_cpuload(i
);
3429 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
3439 * Try and locate an idle CPU in the sched_domain.
3441 static int select_idle_sibling(struct task_struct
*p
, int target
)
3443 struct sched_domain
*sd
;
3444 struct sched_group
*sg
;
3445 int i
= task_cpu(p
);
3447 if (idle_cpu(target
))
3451 * If the prevous cpu is cache affine and idle, don't be stupid.
3453 if (i
!= target
&& cpus_share_cache(i
, target
) && idle_cpu(i
))
3457 * Otherwise, iterate the domains and find an elegible idle cpu.
3459 sd
= rcu_dereference(per_cpu(sd_llc
, target
));
3460 for_each_lower_domain(sd
) {
3463 if (!cpumask_intersects(sched_group_cpus(sg
),
3464 tsk_cpus_allowed(p
)))
3467 for_each_cpu(i
, sched_group_cpus(sg
)) {
3468 if (i
== target
|| !idle_cpu(i
))
3472 target
= cpumask_first_and(sched_group_cpus(sg
),
3473 tsk_cpus_allowed(p
));
3477 } while (sg
!= sd
->groups
);
3484 * sched_balance_self: balance the current task (running on cpu) in domains
3485 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3488 * Balance, ie. select the least loaded group.
3490 * Returns the target CPU number, or the same CPU if no balancing is needed.
3492 * preempt must be disabled.
3495 select_task_rq_fair(struct task_struct
*p
, int sd_flag
, int wake_flags
)
3497 struct sched_domain
*tmp
, *affine_sd
= NULL
, *sd
= NULL
;
3498 int cpu
= smp_processor_id();
3499 int prev_cpu
= task_cpu(p
);
3501 int want_affine
= 0;
3502 int sync
= wake_flags
& WF_SYNC
;
3504 if (p
->nr_cpus_allowed
== 1)
3507 if (sd_flag
& SD_BALANCE_WAKE
) {
3508 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)))
3514 for_each_domain(cpu
, tmp
) {
3515 if (!(tmp
->flags
& SD_LOAD_BALANCE
))
3519 * If both cpu and prev_cpu are part of this domain,
3520 * cpu is a valid SD_WAKE_AFFINE target.
3522 if (want_affine
&& (tmp
->flags
& SD_WAKE_AFFINE
) &&
3523 cpumask_test_cpu(prev_cpu
, sched_domain_span(tmp
))) {
3528 if (tmp
->flags
& sd_flag
)
3533 if (cpu
!= prev_cpu
&& wake_affine(affine_sd
, p
, sync
))
3536 new_cpu
= select_idle_sibling(p
, prev_cpu
);
3541 int load_idx
= sd
->forkexec_idx
;
3542 struct sched_group
*group
;
3545 if (!(sd
->flags
& sd_flag
)) {
3550 if (sd_flag
& SD_BALANCE_WAKE
)
3551 load_idx
= sd
->wake_idx
;
3553 group
= find_idlest_group(sd
, p
, cpu
, load_idx
);
3559 new_cpu
= find_idlest_cpu(group
, p
, cpu
);
3560 if (new_cpu
== -1 || new_cpu
== cpu
) {
3561 /* Now try balancing at a lower domain level of cpu */
3566 /* Now try balancing at a lower domain level of new_cpu */
3568 weight
= sd
->span_weight
;
3570 for_each_domain(cpu
, tmp
) {
3571 if (weight
<= tmp
->span_weight
)
3573 if (tmp
->flags
& sd_flag
)
3576 /* while loop will break here if sd == NULL */
3585 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3586 * cfs_rq_of(p) references at time of call are still valid and identify the
3587 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3588 * other assumptions, including the state of rq->lock, should be made.
3591 migrate_task_rq_fair(struct task_struct
*p
, int next_cpu
)
3593 struct sched_entity
*se
= &p
->se
;
3594 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
3597 * Load tracking: accumulate removed load so that it can be processed
3598 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3599 * to blocked load iff they have a positive decay-count. It can never
3600 * be negative here since on-rq tasks have decay-count == 0.
3602 if (se
->avg
.decay_count
) {
3603 se
->avg
.decay_count
= -__synchronize_entity_decay(se
);
3604 atomic_long_add(se
->avg
.load_avg_contrib
,
3605 &cfs_rq
->removed_load
);
3608 #endif /* CONFIG_SMP */
3610 static unsigned long
3611 wakeup_gran(struct sched_entity
*curr
, struct sched_entity
*se
)
3613 unsigned long gran
= sysctl_sched_wakeup_granularity
;
3616 * Since its curr running now, convert the gran from real-time
3617 * to virtual-time in his units.
3619 * By using 'se' instead of 'curr' we penalize light tasks, so
3620 * they get preempted easier. That is, if 'se' < 'curr' then
3621 * the resulting gran will be larger, therefore penalizing the
3622 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3623 * be smaller, again penalizing the lighter task.
3625 * This is especially important for buddies when the leftmost
3626 * task is higher priority than the buddy.
3628 return calc_delta_fair(gran
, se
);
3632 * Should 'se' preempt 'curr'.
3646 wakeup_preempt_entity(struct sched_entity
*curr
, struct sched_entity
*se
)
3648 s64 gran
, vdiff
= curr
->vruntime
- se
->vruntime
;
3653 gran
= wakeup_gran(curr
, se
);
3660 static void set_last_buddy(struct sched_entity
*se
)
3662 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
3665 for_each_sched_entity(se
)
3666 cfs_rq_of(se
)->last
= se
;
3669 static void set_next_buddy(struct sched_entity
*se
)
3671 if (entity_is_task(se
) && unlikely(task_of(se
)->policy
== SCHED_IDLE
))
3674 for_each_sched_entity(se
)
3675 cfs_rq_of(se
)->next
= se
;
3678 static void set_skip_buddy(struct sched_entity
*se
)
3680 for_each_sched_entity(se
)
3681 cfs_rq_of(se
)->skip
= se
;
3685 * Preempt the current task with a newly woken task if needed:
3687 static void check_preempt_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
3689 struct task_struct
*curr
= rq
->curr
;
3690 struct sched_entity
*se
= &curr
->se
, *pse
= &p
->se
;
3691 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
3692 int scale
= cfs_rq
->nr_running
>= sched_nr_latency
;
3693 int next_buddy_marked
= 0;
3695 if (unlikely(se
== pse
))
3699 * This is possible from callers such as move_task(), in which we
3700 * unconditionally check_prempt_curr() after an enqueue (which may have
3701 * lead to a throttle). This both saves work and prevents false
3702 * next-buddy nomination below.
3704 if (unlikely(throttled_hierarchy(cfs_rq_of(pse
))))
3707 if (sched_feat(NEXT_BUDDY
) && scale
&& !(wake_flags
& WF_FORK
)) {
3708 set_next_buddy(pse
);
3709 next_buddy_marked
= 1;
3713 * We can come here with TIF_NEED_RESCHED already set from new task
3716 * Note: this also catches the edge-case of curr being in a throttled
3717 * group (e.g. via set_curr_task), since update_curr() (in the
3718 * enqueue of curr) will have resulted in resched being set. This
3719 * prevents us from potentially nominating it as a false LAST_BUDDY
3722 if (test_tsk_need_resched(curr
))
3725 /* Idle tasks are by definition preempted by non-idle tasks. */
3726 if (unlikely(curr
->policy
== SCHED_IDLE
) &&
3727 likely(p
->policy
!= SCHED_IDLE
))
3731 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3732 * is driven by the tick):
3734 if (unlikely(p
->policy
!= SCHED_NORMAL
) || !sched_feat(WAKEUP_PREEMPTION
))
3737 find_matching_se(&se
, &pse
);
3738 update_curr(cfs_rq_of(se
));
3740 if (wakeup_preempt_entity(se
, pse
) == 1) {
3742 * Bias pick_next to pick the sched entity that is
3743 * triggering this preemption.
3745 if (!next_buddy_marked
)
3746 set_next_buddy(pse
);
3755 * Only set the backward buddy when the current task is still
3756 * on the rq. This can happen when a wakeup gets interleaved
3757 * with schedule on the ->pre_schedule() or idle_balance()
3758 * point, either of which can * drop the rq lock.
3760 * Also, during early boot the idle thread is in the fair class,
3761 * for obvious reasons its a bad idea to schedule back to it.
3763 if (unlikely(!se
->on_rq
|| curr
== rq
->idle
))
3766 if (sched_feat(LAST_BUDDY
) && scale
&& entity_is_task(se
))
3770 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
3772 struct task_struct
*p
;
3773 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
3774 struct sched_entity
*se
;
3776 if (!cfs_rq
->nr_running
)
3780 se
= pick_next_entity(cfs_rq
);
3781 set_next_entity(cfs_rq
, se
);
3782 cfs_rq
= group_cfs_rq(se
);
3786 if (hrtick_enabled(rq
))
3787 hrtick_start_fair(rq
, p
);
3793 * Account for a descheduled task:
3795 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
3797 struct sched_entity
*se
= &prev
->se
;
3798 struct cfs_rq
*cfs_rq
;
3800 for_each_sched_entity(se
) {
3801 cfs_rq
= cfs_rq_of(se
);
3802 put_prev_entity(cfs_rq
, se
);
3807 * sched_yield() is very simple
3809 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3811 static void yield_task_fair(struct rq
*rq
)
3813 struct task_struct
*curr
= rq
->curr
;
3814 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
3815 struct sched_entity
*se
= &curr
->se
;
3818 * Are we the only task in the tree?
3820 if (unlikely(rq
->nr_running
== 1))
3823 clear_buddies(cfs_rq
, se
);
3825 if (curr
->policy
!= SCHED_BATCH
) {
3826 update_rq_clock(rq
);
3828 * Update run-time statistics of the 'current'.
3830 update_curr(cfs_rq
);
3832 * Tell update_rq_clock() that we've just updated,
3833 * so we don't do microscopic update in schedule()
3834 * and double the fastpath cost.
3836 rq
->skip_clock_update
= 1;
3842 static bool yield_to_task_fair(struct rq
*rq
, struct task_struct
*p
, bool preempt
)
3844 struct sched_entity
*se
= &p
->se
;
3846 /* throttled hierarchies are not runnable */
3847 if (!se
->on_rq
|| throttled_hierarchy(cfs_rq_of(se
)))
3850 /* Tell the scheduler that we'd really like pse to run next. */
3853 yield_task_fair(rq
);
3859 /**************************************************
3860 * Fair scheduling class load-balancing methods.
3864 * The purpose of load-balancing is to achieve the same basic fairness the
3865 * per-cpu scheduler provides, namely provide a proportional amount of compute
3866 * time to each task. This is expressed in the following equation:
3868 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
3870 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
3871 * W_i,0 is defined as:
3873 * W_i,0 = \Sum_j w_i,j (2)
3875 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
3876 * is derived from the nice value as per prio_to_weight[].
3878 * The weight average is an exponential decay average of the instantaneous
3881 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
3883 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
3884 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
3885 * can also include other factors [XXX].
3887 * To achieve this balance we define a measure of imbalance which follows
3888 * directly from (1):
3890 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
3892 * We them move tasks around to minimize the imbalance. In the continuous
3893 * function space it is obvious this converges, in the discrete case we get
3894 * a few fun cases generally called infeasible weight scenarios.
3897 * - infeasible weights;
3898 * - local vs global optima in the discrete case. ]
3903 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
3904 * for all i,j solution, we create a tree of cpus that follows the hardware
3905 * topology where each level pairs two lower groups (or better). This results
3906 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
3907 * tree to only the first of the previous level and we decrease the frequency
3908 * of load-balance at each level inv. proportional to the number of cpus in
3914 * \Sum { --- * --- * 2^i } = O(n) (5)
3916 * `- size of each group
3917 * | | `- number of cpus doing load-balance
3919 * `- sum over all levels
3921 * Coupled with a limit on how many tasks we can migrate every balance pass,
3922 * this makes (5) the runtime complexity of the balancer.
3924 * An important property here is that each CPU is still (indirectly) connected
3925 * to every other cpu in at most O(log n) steps:
3927 * The adjacency matrix of the resulting graph is given by:
3930 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
3933 * And you'll find that:
3935 * A^(log_2 n)_i,j != 0 for all i,j (7)
3937 * Showing there's indeed a path between every cpu in at most O(log n) steps.
3938 * The task movement gives a factor of O(m), giving a convergence complexity
3941 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
3946 * In order to avoid CPUs going idle while there's still work to do, new idle
3947 * balancing is more aggressive and has the newly idle cpu iterate up the domain
3948 * tree itself instead of relying on other CPUs to bring it work.
3950 * This adds some complexity to both (5) and (8) but it reduces the total idle
3958 * Cgroups make a horror show out of (2), instead of a simple sum we get:
3961 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
3966 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
3968 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
3970 * The big problem is S_k, its a global sum needed to compute a local (W_i)
3973 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
3974 * rewrite all of this once again.]
3977 static unsigned long __read_mostly max_load_balance_interval
= HZ
/10;
3979 #define LBF_ALL_PINNED 0x01
3980 #define LBF_NEED_BREAK 0x02
3981 #define LBF_DST_PINNED 0x04
3982 #define LBF_SOME_PINNED 0x08
3985 struct sched_domain
*sd
;
3993 struct cpumask
*dst_grpmask
;
3995 enum cpu_idle_type idle
;
3997 /* The set of CPUs under consideration for load-balancing */
3998 struct cpumask
*cpus
;
4003 unsigned int loop_break
;
4004 unsigned int loop_max
;
4008 * move_task - move a task from one runqueue to another runqueue.
4009 * Both runqueues must be locked.
4011 static void move_task(struct task_struct
*p
, struct lb_env
*env
)
4013 deactivate_task(env
->src_rq
, p
, 0);
4014 set_task_cpu(p
, env
->dst_cpu
);
4015 activate_task(env
->dst_rq
, p
, 0);
4016 check_preempt_curr(env
->dst_rq
, p
, 0);
4020 * Is this task likely cache-hot:
4023 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
4027 if (p
->sched_class
!= &fair_sched_class
)
4030 if (unlikely(p
->policy
== SCHED_IDLE
))
4034 * Buddy candidates are cache hot:
4036 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
4037 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
4038 &p
->se
== cfs_rq_of(&p
->se
)->last
))
4041 if (sysctl_sched_migration_cost
== -1)
4043 if (sysctl_sched_migration_cost
== 0)
4046 delta
= now
- p
->se
.exec_start
;
4048 return delta
< (s64
)sysctl_sched_migration_cost
;
4052 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4055 int can_migrate_task(struct task_struct
*p
, struct lb_env
*env
)
4057 int tsk_cache_hot
= 0;
4059 * We do not migrate tasks that are:
4060 * 1) throttled_lb_pair, or
4061 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4062 * 3) running (obviously), or
4063 * 4) are cache-hot on their current CPU.
4065 if (throttled_lb_pair(task_group(p
), env
->src_cpu
, env
->dst_cpu
))
4068 if (!cpumask_test_cpu(env
->dst_cpu
, tsk_cpus_allowed(p
))) {
4071 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_affine
);
4073 env
->flags
|= LBF_SOME_PINNED
;
4076 * Remember if this task can be migrated to any other cpu in
4077 * our sched_group. We may want to revisit it if we couldn't
4078 * meet load balance goals by pulling other tasks on src_cpu.
4080 * Also avoid computing new_dst_cpu if we have already computed
4081 * one in current iteration.
4083 if (!env
->dst_grpmask
|| (env
->flags
& LBF_DST_PINNED
))
4086 /* Prevent to re-select dst_cpu via env's cpus */
4087 for_each_cpu_and(cpu
, env
->dst_grpmask
, env
->cpus
) {
4088 if (cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
))) {
4089 env
->flags
|= LBF_DST_PINNED
;
4090 env
->new_dst_cpu
= cpu
;
4098 /* Record that we found atleast one task that could run on dst_cpu */
4099 env
->flags
&= ~LBF_ALL_PINNED
;
4101 if (task_running(env
->src_rq
, p
)) {
4102 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_running
);
4107 * Aggressive migration if:
4108 * 1) task is cache cold, or
4109 * 2) too many balance attempts have failed.
4112 tsk_cache_hot
= task_hot(p
, rq_clock_task(env
->src_rq
), env
->sd
);
4113 if (!tsk_cache_hot
||
4114 env
->sd
->nr_balance_failed
> env
->sd
->cache_nice_tries
) {
4116 if (tsk_cache_hot
) {
4117 schedstat_inc(env
->sd
, lb_hot_gained
[env
->idle
]);
4118 schedstat_inc(p
, se
.statistics
.nr_forced_migrations
);
4124 schedstat_inc(p
, se
.statistics
.nr_failed_migrations_hot
);
4129 * move_one_task tries to move exactly one task from busiest to this_rq, as
4130 * part of active balancing operations within "domain".
4131 * Returns 1 if successful and 0 otherwise.
4133 * Called with both runqueues locked.
4135 static int move_one_task(struct lb_env
*env
)
4137 struct task_struct
*p
, *n
;
4139 list_for_each_entry_safe(p
, n
, &env
->src_rq
->cfs_tasks
, se
.group_node
) {
4140 if (!can_migrate_task(p
, env
))
4145 * Right now, this is only the second place move_task()
4146 * is called, so we can safely collect move_task()
4147 * stats here rather than inside move_task().
4149 schedstat_inc(env
->sd
, lb_gained
[env
->idle
]);
4155 static unsigned long task_h_load(struct task_struct
*p
);
4157 static const unsigned int sched_nr_migrate_break
= 32;
4160 * move_tasks tries to move up to imbalance weighted load from busiest to
4161 * this_rq, as part of a balancing operation within domain "sd".
4162 * Returns 1 if successful and 0 otherwise.
4164 * Called with both runqueues locked.
4166 static int move_tasks(struct lb_env
*env
)
4168 struct list_head
*tasks
= &env
->src_rq
->cfs_tasks
;
4169 struct task_struct
*p
;
4173 if (env
->imbalance
<= 0)
4176 while (!list_empty(tasks
)) {
4177 p
= list_first_entry(tasks
, struct task_struct
, se
.group_node
);
4180 /* We've more or less seen every task there is, call it quits */
4181 if (env
->loop
> env
->loop_max
)
4184 /* take a breather every nr_migrate tasks */
4185 if (env
->loop
> env
->loop_break
) {
4186 env
->loop_break
+= sched_nr_migrate_break
;
4187 env
->flags
|= LBF_NEED_BREAK
;
4191 if (!can_migrate_task(p
, env
))
4194 load
= task_h_load(p
);
4196 if (sched_feat(LB_MIN
) && load
< 16 && !env
->sd
->nr_balance_failed
)
4199 if ((load
/ 2) > env
->imbalance
)
4204 env
->imbalance
-= load
;
4206 #ifdef CONFIG_PREEMPT
4208 * NEWIDLE balancing is a source of latency, so preemptible
4209 * kernels will stop after the first task is pulled to minimize
4210 * the critical section.
4212 if (env
->idle
== CPU_NEWLY_IDLE
)
4217 * We only want to steal up to the prescribed amount of
4220 if (env
->imbalance
<= 0)
4225 list_move_tail(&p
->se
.group_node
, tasks
);
4229 * Right now, this is one of only two places move_task() is called,
4230 * so we can safely collect move_task() stats here rather than
4231 * inside move_task().
4233 schedstat_add(env
->sd
, lb_gained
[env
->idle
], pulled
);
4238 #ifdef CONFIG_FAIR_GROUP_SCHED
4240 * update tg->load_weight by folding this cpu's load_avg
4242 static void __update_blocked_averages_cpu(struct task_group
*tg
, int cpu
)
4244 struct sched_entity
*se
= tg
->se
[cpu
];
4245 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[cpu
];
4247 /* throttled entities do not contribute to load */
4248 if (throttled_hierarchy(cfs_rq
))
4251 update_cfs_rq_blocked_load(cfs_rq
, 1);
4254 update_entity_load_avg(se
, 1);
4256 * We pivot on our runnable average having decayed to zero for
4257 * list removal. This generally implies that all our children
4258 * have also been removed (modulo rounding error or bandwidth
4259 * control); however, such cases are rare and we can fix these
4262 * TODO: fix up out-of-order children on enqueue.
4264 if (!se
->avg
.runnable_avg_sum
&& !cfs_rq
->nr_running
)
4265 list_del_leaf_cfs_rq(cfs_rq
);
4267 struct rq
*rq
= rq_of(cfs_rq
);
4268 update_rq_runnable_avg(rq
, rq
->nr_running
);
4272 static void update_blocked_averages(int cpu
)
4274 struct rq
*rq
= cpu_rq(cpu
);
4275 struct cfs_rq
*cfs_rq
;
4276 unsigned long flags
;
4278 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4279 update_rq_clock(rq
);
4281 * Iterates the task_group tree in a bottom up fashion, see
4282 * list_add_leaf_cfs_rq() for details.
4284 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
4286 * Note: We may want to consider periodically releasing
4287 * rq->lock about these updates so that creating many task
4288 * groups does not result in continually extending hold time.
4290 __update_blocked_averages_cpu(cfs_rq
->tg
, rq
->cpu
);
4293 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4297 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
4298 * This needs to be done in a top-down fashion because the load of a child
4299 * group is a fraction of its parents load.
4301 static void update_cfs_rq_h_load(struct cfs_rq
*cfs_rq
)
4303 struct rq
*rq
= rq_of(cfs_rq
);
4304 struct sched_entity
*se
= cfs_rq
->tg
->se
[cpu_of(rq
)];
4305 unsigned long now
= jiffies
;
4308 if (cfs_rq
->last_h_load_update
== now
)
4311 cfs_rq
->h_load_next
= NULL
;
4312 for_each_sched_entity(se
) {
4313 cfs_rq
= cfs_rq_of(se
);
4314 cfs_rq
->h_load_next
= se
;
4315 if (cfs_rq
->last_h_load_update
== now
)
4320 cfs_rq
->h_load
= cfs_rq
->runnable_load_avg
;
4321 cfs_rq
->last_h_load_update
= now
;
4324 while ((se
= cfs_rq
->h_load_next
) != NULL
) {
4325 load
= cfs_rq
->h_load
;
4326 load
= div64_ul(load
* se
->avg
.load_avg_contrib
,
4327 cfs_rq
->runnable_load_avg
+ 1);
4328 cfs_rq
= group_cfs_rq(se
);
4329 cfs_rq
->h_load
= load
;
4330 cfs_rq
->last_h_load_update
= now
;
4334 static unsigned long task_h_load(struct task_struct
*p
)
4336 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
4338 update_cfs_rq_h_load(cfs_rq
);
4339 return div64_ul(p
->se
.avg
.load_avg_contrib
* cfs_rq
->h_load
,
4340 cfs_rq
->runnable_load_avg
+ 1);
4343 static inline void update_blocked_averages(int cpu
)
4347 static unsigned long task_h_load(struct task_struct
*p
)
4349 return p
->se
.avg
.load_avg_contrib
;
4353 /********** Helpers for find_busiest_group ************************/
4355 * sg_lb_stats - stats of a sched_group required for load_balancing
4357 struct sg_lb_stats
{
4358 unsigned long avg_load
; /*Avg load across the CPUs of the group */
4359 unsigned long group_load
; /* Total load over the CPUs of the group */
4360 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
4361 unsigned long load_per_task
;
4362 unsigned long group_power
;
4363 unsigned int sum_nr_running
; /* Nr tasks running in the group */
4364 unsigned int group_capacity
;
4365 unsigned int idle_cpus
;
4366 unsigned int group_weight
;
4367 int group_imb
; /* Is there an imbalance in the group ? */
4368 int group_has_capacity
; /* Is there extra capacity in the group? */
4372 * sd_lb_stats - Structure to store the statistics of a sched_domain
4373 * during load balancing.
4375 struct sd_lb_stats
{
4376 struct sched_group
*busiest
; /* Busiest group in this sd */
4377 struct sched_group
*local
; /* Local group in this sd */
4378 unsigned long total_load
; /* Total load of all groups in sd */
4379 unsigned long total_pwr
; /* Total power of all groups in sd */
4380 unsigned long avg_load
; /* Average load across all groups in sd */
4382 struct sg_lb_stats busiest_stat
;/* Statistics of the busiest group */
4383 struct sg_lb_stats local_stat
; /* Statistics of the local group */
4386 static inline void init_sd_lb_stats(struct sd_lb_stats
*sds
)
4389 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
4390 * local_stat because update_sg_lb_stats() does a full clear/assignment.
4391 * We must however clear busiest_stat::avg_load because
4392 * update_sd_pick_busiest() reads this before assignment.
4394 *sds
= (struct sd_lb_stats
){
4406 * get_sd_load_idx - Obtain the load index for a given sched domain.
4407 * @sd: The sched_domain whose load_idx is to be obtained.
4408 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4410 * Return: The load index.
4412 static inline int get_sd_load_idx(struct sched_domain
*sd
,
4413 enum cpu_idle_type idle
)
4419 load_idx
= sd
->busy_idx
;
4422 case CPU_NEWLY_IDLE
:
4423 load_idx
= sd
->newidle_idx
;
4426 load_idx
= sd
->idle_idx
;
4433 static unsigned long default_scale_freq_power(struct sched_domain
*sd
, int cpu
)
4435 return SCHED_POWER_SCALE
;
4438 unsigned long __weak
arch_scale_freq_power(struct sched_domain
*sd
, int cpu
)
4440 return default_scale_freq_power(sd
, cpu
);
4443 static unsigned long default_scale_smt_power(struct sched_domain
*sd
, int cpu
)
4445 unsigned long weight
= sd
->span_weight
;
4446 unsigned long smt_gain
= sd
->smt_gain
;
4453 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
4455 return default_scale_smt_power(sd
, cpu
);
4458 static unsigned long scale_rt_power(int cpu
)
4460 struct rq
*rq
= cpu_rq(cpu
);
4461 u64 total
, available
, age_stamp
, avg
;
4464 * Since we're reading these variables without serialization make sure
4465 * we read them once before doing sanity checks on them.
4467 age_stamp
= ACCESS_ONCE(rq
->age_stamp
);
4468 avg
= ACCESS_ONCE(rq
->rt_avg
);
4470 total
= sched_avg_period() + (rq_clock(rq
) - age_stamp
);
4472 if (unlikely(total
< avg
)) {
4473 /* Ensures that power won't end up being negative */
4476 available
= total
- avg
;
4479 if (unlikely((s64
)total
< SCHED_POWER_SCALE
))
4480 total
= SCHED_POWER_SCALE
;
4482 total
>>= SCHED_POWER_SHIFT
;
4484 return div_u64(available
, total
);
4487 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
4489 unsigned long weight
= sd
->span_weight
;
4490 unsigned long power
= SCHED_POWER_SCALE
;
4491 struct sched_group
*sdg
= sd
->groups
;
4493 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
4494 if (sched_feat(ARCH_POWER
))
4495 power
*= arch_scale_smt_power(sd
, cpu
);
4497 power
*= default_scale_smt_power(sd
, cpu
);
4499 power
>>= SCHED_POWER_SHIFT
;
4502 sdg
->sgp
->power_orig
= power
;
4504 if (sched_feat(ARCH_POWER
))
4505 power
*= arch_scale_freq_power(sd
, cpu
);
4507 power
*= default_scale_freq_power(sd
, cpu
);
4509 power
>>= SCHED_POWER_SHIFT
;
4511 power
*= scale_rt_power(cpu
);
4512 power
>>= SCHED_POWER_SHIFT
;
4517 cpu_rq(cpu
)->cpu_power
= power
;
4518 sdg
->sgp
->power
= power
;
4521 void update_group_power(struct sched_domain
*sd
, int cpu
)
4523 struct sched_domain
*child
= sd
->child
;
4524 struct sched_group
*group
, *sdg
= sd
->groups
;
4525 unsigned long power
, power_orig
;
4526 unsigned long interval
;
4528 interval
= msecs_to_jiffies(sd
->balance_interval
);
4529 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
4530 sdg
->sgp
->next_update
= jiffies
+ interval
;
4533 update_cpu_power(sd
, cpu
);
4537 power_orig
= power
= 0;
4539 if (child
->flags
& SD_OVERLAP
) {
4541 * SD_OVERLAP domains cannot assume that child groups
4542 * span the current group.
4545 for_each_cpu(cpu
, sched_group_cpus(sdg
)) {
4546 struct sched_group
*sg
= cpu_rq(cpu
)->sd
->groups
;
4548 power_orig
+= sg
->sgp
->power_orig
;
4549 power
+= sg
->sgp
->power
;
4553 * !SD_OVERLAP domains can assume that child groups
4554 * span the current group.
4557 group
= child
->groups
;
4559 power_orig
+= group
->sgp
->power_orig
;
4560 power
+= group
->sgp
->power
;
4561 group
= group
->next
;
4562 } while (group
!= child
->groups
);
4565 sdg
->sgp
->power_orig
= power_orig
;
4566 sdg
->sgp
->power
= power
;
4570 * Try and fix up capacity for tiny siblings, this is needed when
4571 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4572 * which on its own isn't powerful enough.
4574 * See update_sd_pick_busiest() and check_asym_packing().
4577 fix_small_capacity(struct sched_domain
*sd
, struct sched_group
*group
)
4580 * Only siblings can have significantly less than SCHED_POWER_SCALE
4582 if (!(sd
->flags
& SD_SHARE_CPUPOWER
))
4586 * If ~90% of the cpu_power is still there, we're good.
4588 if (group
->sgp
->power
* 32 > group
->sgp
->power_orig
* 29)
4595 * Group imbalance indicates (and tries to solve) the problem where balancing
4596 * groups is inadequate due to tsk_cpus_allowed() constraints.
4598 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
4599 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
4602 * { 0 1 2 3 } { 4 5 6 7 }
4605 * If we were to balance group-wise we'd place two tasks in the first group and
4606 * two tasks in the second group. Clearly this is undesired as it will overload
4607 * cpu 3 and leave one of the cpus in the second group unused.
4609 * The current solution to this issue is detecting the skew in the first group
4610 * by noticing the lower domain failed to reach balance and had difficulty
4611 * moving tasks due to affinity constraints.
4613 * When this is so detected; this group becomes a candidate for busiest; see
4614 * update_sd_pick_busiest(). And calculcate_imbalance() and
4615 * find_busiest_group() avoid some of the usual balance conditions to allow it
4616 * to create an effective group imbalance.
4618 * This is a somewhat tricky proposition since the next run might not find the
4619 * group imbalance and decide the groups need to be balanced again. A most
4620 * subtle and fragile situation.
4623 static inline int sg_imbalanced(struct sched_group
*group
)
4625 return group
->sgp
->imbalance
;
4629 * Compute the group capacity.
4631 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
4632 * first dividing out the smt factor and computing the actual number of cores
4633 * and limit power unit capacity with that.
4635 static inline int sg_capacity(struct lb_env
*env
, struct sched_group
*group
)
4637 unsigned int capacity
, smt
, cpus
;
4638 unsigned int power
, power_orig
;
4640 power
= group
->sgp
->power
;
4641 power_orig
= group
->sgp
->power_orig
;
4642 cpus
= group
->group_weight
;
4644 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
4645 smt
= DIV_ROUND_UP(SCHED_POWER_SCALE
* cpus
, power_orig
);
4646 capacity
= cpus
/ smt
; /* cores */
4648 capacity
= min_t(unsigned, capacity
, DIV_ROUND_CLOSEST(power
, SCHED_POWER_SCALE
));
4650 capacity
= fix_small_capacity(env
->sd
, group
);
4656 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4657 * @env: The load balancing environment.
4658 * @group: sched_group whose statistics are to be updated.
4659 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4660 * @local_group: Does group contain this_cpu.
4661 * @sgs: variable to hold the statistics for this group.
4663 static inline void update_sg_lb_stats(struct lb_env
*env
,
4664 struct sched_group
*group
, int load_idx
,
4665 int local_group
, struct sg_lb_stats
*sgs
)
4667 unsigned long nr_running
;
4671 memset(sgs
, 0, sizeof(*sgs
));
4673 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
4674 struct rq
*rq
= cpu_rq(i
);
4676 nr_running
= rq
->nr_running
;
4678 /* Bias balancing toward cpus of our domain */
4680 load
= target_load(i
, load_idx
);
4682 load
= source_load(i
, load_idx
);
4684 sgs
->group_load
+= load
;
4685 sgs
->sum_nr_running
+= nr_running
;
4686 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
4691 /* Adjust by relative CPU power of the group */
4692 sgs
->group_power
= group
->sgp
->power
;
4693 sgs
->avg_load
= (sgs
->group_load
*SCHED_POWER_SCALE
) / sgs
->group_power
;
4695 if (sgs
->sum_nr_running
)
4696 sgs
->load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
4698 sgs
->group_weight
= group
->group_weight
;
4700 sgs
->group_imb
= sg_imbalanced(group
);
4701 sgs
->group_capacity
= sg_capacity(env
, group
);
4703 if (sgs
->group_capacity
> sgs
->sum_nr_running
)
4704 sgs
->group_has_capacity
= 1;
4708 * update_sd_pick_busiest - return 1 on busiest group
4709 * @env: The load balancing environment.
4710 * @sds: sched_domain statistics
4711 * @sg: sched_group candidate to be checked for being the busiest
4712 * @sgs: sched_group statistics
4714 * Determine if @sg is a busier group than the previously selected
4717 * Return: %true if @sg is a busier group than the previously selected
4718 * busiest group. %false otherwise.
4720 static bool update_sd_pick_busiest(struct lb_env
*env
,
4721 struct sd_lb_stats
*sds
,
4722 struct sched_group
*sg
,
4723 struct sg_lb_stats
*sgs
)
4725 if (sgs
->avg_load
<= sds
->busiest_stat
.avg_load
)
4728 if (sgs
->sum_nr_running
> sgs
->group_capacity
)
4735 * ASYM_PACKING needs to move all the work to the lowest
4736 * numbered CPUs in the group, therefore mark all groups
4737 * higher than ourself as busy.
4739 if ((env
->sd
->flags
& SD_ASYM_PACKING
) && sgs
->sum_nr_running
&&
4740 env
->dst_cpu
< group_first_cpu(sg
)) {
4744 if (group_first_cpu(sds
->busiest
) > group_first_cpu(sg
))
4752 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4753 * @env: The load balancing environment.
4754 * @balance: Should we balance.
4755 * @sds: variable to hold the statistics for this sched_domain.
4757 static inline void update_sd_lb_stats(struct lb_env
*env
,
4758 struct sd_lb_stats
*sds
)
4760 struct sched_domain
*child
= env
->sd
->child
;
4761 struct sched_group
*sg
= env
->sd
->groups
;
4762 struct sg_lb_stats tmp_sgs
;
4763 int load_idx
, prefer_sibling
= 0;
4765 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
4768 load_idx
= get_sd_load_idx(env
->sd
, env
->idle
);
4771 struct sg_lb_stats
*sgs
= &tmp_sgs
;
4774 local_group
= cpumask_test_cpu(env
->dst_cpu
, sched_group_cpus(sg
));
4777 sgs
= &sds
->local_stat
;
4779 if (env
->idle
!= CPU_NEWLY_IDLE
||
4780 time_after_eq(jiffies
, sg
->sgp
->next_update
))
4781 update_group_power(env
->sd
, env
->dst_cpu
);
4784 update_sg_lb_stats(env
, sg
, load_idx
, local_group
, sgs
);
4790 * In case the child domain prefers tasks go to siblings
4791 * first, lower the sg capacity to one so that we'll try
4792 * and move all the excess tasks away. We lower the capacity
4793 * of a group only if the local group has the capacity to fit
4794 * these excess tasks, i.e. nr_running < group_capacity. The
4795 * extra check prevents the case where you always pull from the
4796 * heaviest group when it is already under-utilized (possible
4797 * with a large weight task outweighs the tasks on the system).
4799 if (prefer_sibling
&& sds
->local
&&
4800 sds
->local_stat
.group_has_capacity
)
4801 sgs
->group_capacity
= min(sgs
->group_capacity
, 1U);
4803 if (update_sd_pick_busiest(env
, sds
, sg
, sgs
)) {
4805 sds
->busiest_stat
= *sgs
;
4809 /* Now, start updating sd_lb_stats */
4810 sds
->total_load
+= sgs
->group_load
;
4811 sds
->total_pwr
+= sgs
->group_power
;
4814 } while (sg
!= env
->sd
->groups
);
4818 * check_asym_packing - Check to see if the group is packed into the
4821 * This is primarily intended to used at the sibling level. Some
4822 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4823 * case of POWER7, it can move to lower SMT modes only when higher
4824 * threads are idle. When in lower SMT modes, the threads will
4825 * perform better since they share less core resources. Hence when we
4826 * have idle threads, we want them to be the higher ones.
4828 * This packing function is run on idle threads. It checks to see if
4829 * the busiest CPU in this domain (core in the P7 case) has a higher
4830 * CPU number than the packing function is being run on. Here we are
4831 * assuming lower CPU number will be equivalent to lower a SMT thread
4834 * Return: 1 when packing is required and a task should be moved to
4835 * this CPU. The amount of the imbalance is returned in *imbalance.
4837 * @env: The load balancing environment.
4838 * @sds: Statistics of the sched_domain which is to be packed
4840 static int check_asym_packing(struct lb_env
*env
, struct sd_lb_stats
*sds
)
4844 if (!(env
->sd
->flags
& SD_ASYM_PACKING
))
4850 busiest_cpu
= group_first_cpu(sds
->busiest
);
4851 if (env
->dst_cpu
> busiest_cpu
)
4854 env
->imbalance
= DIV_ROUND_CLOSEST(
4855 sds
->busiest_stat
.avg_load
* sds
->busiest_stat
.group_power
,
4862 * fix_small_imbalance - Calculate the minor imbalance that exists
4863 * amongst the groups of a sched_domain, during
4865 * @env: The load balancing environment.
4866 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4869 void fix_small_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
4871 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
4872 unsigned int imbn
= 2;
4873 unsigned long scaled_busy_load_per_task
;
4874 struct sg_lb_stats
*local
, *busiest
;
4876 local
= &sds
->local_stat
;
4877 busiest
= &sds
->busiest_stat
;
4879 if (!local
->sum_nr_running
)
4880 local
->load_per_task
= cpu_avg_load_per_task(env
->dst_cpu
);
4881 else if (busiest
->load_per_task
> local
->load_per_task
)
4884 scaled_busy_load_per_task
=
4885 (busiest
->load_per_task
* SCHED_POWER_SCALE
) /
4886 busiest
->group_power
;
4888 if (busiest
->avg_load
+ scaled_busy_load_per_task
>=
4889 local
->avg_load
+ (scaled_busy_load_per_task
* imbn
)) {
4890 env
->imbalance
= busiest
->load_per_task
;
4895 * OK, we don't have enough imbalance to justify moving tasks,
4896 * however we may be able to increase total CPU power used by
4900 pwr_now
+= busiest
->group_power
*
4901 min(busiest
->load_per_task
, busiest
->avg_load
);
4902 pwr_now
+= local
->group_power
*
4903 min(local
->load_per_task
, local
->avg_load
);
4904 pwr_now
/= SCHED_POWER_SCALE
;
4906 /* Amount of load we'd subtract */
4907 tmp
= (busiest
->load_per_task
* SCHED_POWER_SCALE
) /
4908 busiest
->group_power
;
4909 if (busiest
->avg_load
> tmp
) {
4910 pwr_move
+= busiest
->group_power
*
4911 min(busiest
->load_per_task
,
4912 busiest
->avg_load
- tmp
);
4915 /* Amount of load we'd add */
4916 if (busiest
->avg_load
* busiest
->group_power
<
4917 busiest
->load_per_task
* SCHED_POWER_SCALE
) {
4918 tmp
= (busiest
->avg_load
* busiest
->group_power
) /
4921 tmp
= (busiest
->load_per_task
* SCHED_POWER_SCALE
) /
4924 pwr_move
+= local
->group_power
*
4925 min(local
->load_per_task
, local
->avg_load
+ tmp
);
4926 pwr_move
/= SCHED_POWER_SCALE
;
4928 /* Move if we gain throughput */
4929 if (pwr_move
> pwr_now
)
4930 env
->imbalance
= busiest
->load_per_task
;
4934 * calculate_imbalance - Calculate the amount of imbalance present within the
4935 * groups of a given sched_domain during load balance.
4936 * @env: load balance environment
4937 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4939 static inline void calculate_imbalance(struct lb_env
*env
, struct sd_lb_stats
*sds
)
4941 unsigned long max_pull
, load_above_capacity
= ~0UL;
4942 struct sg_lb_stats
*local
, *busiest
;
4944 local
= &sds
->local_stat
;
4945 busiest
= &sds
->busiest_stat
;
4947 if (busiest
->group_imb
) {
4949 * In the group_imb case we cannot rely on group-wide averages
4950 * to ensure cpu-load equilibrium, look at wider averages. XXX
4952 busiest
->load_per_task
=
4953 min(busiest
->load_per_task
, sds
->avg_load
);
4957 * In the presence of smp nice balancing, certain scenarios can have
4958 * max load less than avg load(as we skip the groups at or below
4959 * its cpu_power, while calculating max_load..)
4961 if (busiest
->avg_load
<= sds
->avg_load
||
4962 local
->avg_load
>= sds
->avg_load
) {
4964 return fix_small_imbalance(env
, sds
);
4967 if (!busiest
->group_imb
) {
4969 * Don't want to pull so many tasks that a group would go idle.
4970 * Except of course for the group_imb case, since then we might
4971 * have to drop below capacity to reach cpu-load equilibrium.
4973 load_above_capacity
=
4974 (busiest
->sum_nr_running
- busiest
->group_capacity
);
4976 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_POWER_SCALE
);
4977 load_above_capacity
/= busiest
->group_power
;
4981 * We're trying to get all the cpus to the average_load, so we don't
4982 * want to push ourselves above the average load, nor do we wish to
4983 * reduce the max loaded cpu below the average load. At the same time,
4984 * we also don't want to reduce the group load below the group capacity
4985 * (so that we can implement power-savings policies etc). Thus we look
4986 * for the minimum possible imbalance.
4988 max_pull
= min(busiest
->avg_load
- sds
->avg_load
, load_above_capacity
);
4990 /* How much load to actually move to equalise the imbalance */
4991 env
->imbalance
= min(
4992 max_pull
* busiest
->group_power
,
4993 (sds
->avg_load
- local
->avg_load
) * local
->group_power
4994 ) / SCHED_POWER_SCALE
;
4997 * if *imbalance is less than the average load per runnable task
4998 * there is no guarantee that any tasks will be moved so we'll have
4999 * a think about bumping its value to force at least one task to be
5002 if (env
->imbalance
< busiest
->load_per_task
)
5003 return fix_small_imbalance(env
, sds
);
5006 /******* find_busiest_group() helpers end here *********************/
5009 * find_busiest_group - Returns the busiest group within the sched_domain
5010 * if there is an imbalance. If there isn't an imbalance, and
5011 * the user has opted for power-savings, it returns a group whose
5012 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5013 * such a group exists.
5015 * Also calculates the amount of weighted load which should be moved
5016 * to restore balance.
5018 * @env: The load balancing environment.
5020 * Return: - The busiest group if imbalance exists.
5021 * - If no imbalance and user has opted for power-savings balance,
5022 * return the least loaded group whose CPUs can be
5023 * put to idle by rebalancing its tasks onto our group.
5025 static struct sched_group
*find_busiest_group(struct lb_env
*env
)
5027 struct sg_lb_stats
*local
, *busiest
;
5028 struct sd_lb_stats sds
;
5030 init_sd_lb_stats(&sds
);
5033 * Compute the various statistics relavent for load balancing at
5036 update_sd_lb_stats(env
, &sds
);
5037 local
= &sds
.local_stat
;
5038 busiest
= &sds
.busiest_stat
;
5040 if ((env
->idle
== CPU_IDLE
|| env
->idle
== CPU_NEWLY_IDLE
) &&
5041 check_asym_packing(env
, &sds
))
5044 /* There is no busy sibling group to pull tasks from */
5045 if (!sds
.busiest
|| busiest
->sum_nr_running
== 0)
5048 sds
.avg_load
= (SCHED_POWER_SCALE
* sds
.total_load
) / sds
.total_pwr
;
5051 * If the busiest group is imbalanced the below checks don't
5052 * work because they assume all things are equal, which typically
5053 * isn't true due to cpus_allowed constraints and the like.
5055 if (busiest
->group_imb
)
5058 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5059 if (env
->idle
== CPU_NEWLY_IDLE
&& local
->group_has_capacity
&&
5060 !busiest
->group_has_capacity
)
5064 * If the local group is more busy than the selected busiest group
5065 * don't try and pull any tasks.
5067 if (local
->avg_load
>= busiest
->avg_load
)
5071 * Don't pull any tasks if this group is already above the domain
5074 if (local
->avg_load
>= sds
.avg_load
)
5077 if (env
->idle
== CPU_IDLE
) {
5079 * This cpu is idle. If the busiest group load doesn't
5080 * have more tasks than the number of available cpu's and
5081 * there is no imbalance between this and busiest group
5082 * wrt to idle cpu's, it is balanced.
5084 if ((local
->idle_cpus
< busiest
->idle_cpus
) &&
5085 busiest
->sum_nr_running
<= busiest
->group_weight
)
5089 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5090 * imbalance_pct to be conservative.
5092 if (100 * busiest
->avg_load
<=
5093 env
->sd
->imbalance_pct
* local
->avg_load
)
5098 /* Looks like there is an imbalance. Compute it */
5099 calculate_imbalance(env
, &sds
);
5108 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5110 static struct rq
*find_busiest_queue(struct lb_env
*env
,
5111 struct sched_group
*group
)
5113 struct rq
*busiest
= NULL
, *rq
;
5114 unsigned long busiest_load
= 0, busiest_power
= 1;
5117 for_each_cpu_and(i
, sched_group_cpus(group
), env
->cpus
) {
5118 unsigned long power
= power_of(i
);
5119 unsigned long capacity
= DIV_ROUND_CLOSEST(power
,
5124 capacity
= fix_small_capacity(env
->sd
, group
);
5127 wl
= weighted_cpuload(i
);
5130 * When comparing with imbalance, use weighted_cpuload()
5131 * which is not scaled with the cpu power.
5133 if (capacity
&& rq
->nr_running
== 1 && wl
> env
->imbalance
)
5137 * For the load comparisons with the other cpu's, consider
5138 * the weighted_cpuload() scaled with the cpu power, so that
5139 * the load can be moved away from the cpu that is potentially
5140 * running at a lower capacity.
5142 * Thus we're looking for max(wl_i / power_i), crosswise
5143 * multiplication to rid ourselves of the division works out
5144 * to: wl_i * power_j > wl_j * power_i; where j is our
5147 if (wl
* busiest_power
> busiest_load
* power
) {
5149 busiest_power
= power
;
5158 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5159 * so long as it is large enough.
5161 #define MAX_PINNED_INTERVAL 512
5163 /* Working cpumask for load_balance and load_balance_newidle. */
5164 DEFINE_PER_CPU(cpumask_var_t
, load_balance_mask
);
5166 static int need_active_balance(struct lb_env
*env
)
5168 struct sched_domain
*sd
= env
->sd
;
5170 if (env
->idle
== CPU_NEWLY_IDLE
) {
5173 * ASYM_PACKING needs to force migrate tasks from busy but
5174 * higher numbered CPUs in order to pack all tasks in the
5175 * lowest numbered CPUs.
5177 if ((sd
->flags
& SD_ASYM_PACKING
) && env
->src_cpu
> env
->dst_cpu
)
5181 return unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2);
5184 static int active_load_balance_cpu_stop(void *data
);
5186 static int should_we_balance(struct lb_env
*env
)
5188 struct sched_group
*sg
= env
->sd
->groups
;
5189 struct cpumask
*sg_cpus
, *sg_mask
;
5190 int cpu
, balance_cpu
= -1;
5193 * In the newly idle case, we will allow all the cpu's
5194 * to do the newly idle load balance.
5196 if (env
->idle
== CPU_NEWLY_IDLE
)
5199 sg_cpus
= sched_group_cpus(sg
);
5200 sg_mask
= sched_group_mask(sg
);
5201 /* Try to find first idle cpu */
5202 for_each_cpu_and(cpu
, sg_cpus
, env
->cpus
) {
5203 if (!cpumask_test_cpu(cpu
, sg_mask
) || !idle_cpu(cpu
))
5210 if (balance_cpu
== -1)
5211 balance_cpu
= group_balance_cpu(sg
);
5214 * First idle cpu or the first cpu(busiest) in this sched group
5215 * is eligible for doing load balancing at this and above domains.
5217 return balance_cpu
== env
->dst_cpu
;
5221 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5222 * tasks if there is an imbalance.
5224 static int load_balance(int this_cpu
, struct rq
*this_rq
,
5225 struct sched_domain
*sd
, enum cpu_idle_type idle
,
5226 int *continue_balancing
)
5228 int ld_moved
, cur_ld_moved
, active_balance
= 0;
5229 struct sched_domain
*sd_parent
= sd
->parent
;
5230 struct sched_group
*group
;
5232 unsigned long flags
;
5233 struct cpumask
*cpus
= __get_cpu_var(load_balance_mask
);
5235 struct lb_env env
= {
5237 .dst_cpu
= this_cpu
,
5239 .dst_grpmask
= sched_group_cpus(sd
->groups
),
5241 .loop_break
= sched_nr_migrate_break
,
5246 * For NEWLY_IDLE load_balancing, we don't need to consider
5247 * other cpus in our group
5249 if (idle
== CPU_NEWLY_IDLE
)
5250 env
.dst_grpmask
= NULL
;
5252 cpumask_copy(cpus
, cpu_active_mask
);
5254 schedstat_inc(sd
, lb_count
[idle
]);
5257 if (!should_we_balance(&env
)) {
5258 *continue_balancing
= 0;
5262 group
= find_busiest_group(&env
);
5264 schedstat_inc(sd
, lb_nobusyg
[idle
]);
5268 busiest
= find_busiest_queue(&env
, group
);
5270 schedstat_inc(sd
, lb_nobusyq
[idle
]);
5274 BUG_ON(busiest
== env
.dst_rq
);
5276 schedstat_add(sd
, lb_imbalance
[idle
], env
.imbalance
);
5279 if (busiest
->nr_running
> 1) {
5281 * Attempt to move tasks. If find_busiest_group has found
5282 * an imbalance but busiest->nr_running <= 1, the group is
5283 * still unbalanced. ld_moved simply stays zero, so it is
5284 * correctly treated as an imbalance.
5286 env
.flags
|= LBF_ALL_PINNED
;
5287 env
.src_cpu
= busiest
->cpu
;
5288 env
.src_rq
= busiest
;
5289 env
.loop_max
= min(sysctl_sched_nr_migrate
, busiest
->nr_running
);
5292 local_irq_save(flags
);
5293 double_rq_lock(env
.dst_rq
, busiest
);
5296 * cur_ld_moved - load moved in current iteration
5297 * ld_moved - cumulative load moved across iterations
5299 cur_ld_moved
= move_tasks(&env
);
5300 ld_moved
+= cur_ld_moved
;
5301 double_rq_unlock(env
.dst_rq
, busiest
);
5302 local_irq_restore(flags
);
5305 * some other cpu did the load balance for us.
5307 if (cur_ld_moved
&& env
.dst_cpu
!= smp_processor_id())
5308 resched_cpu(env
.dst_cpu
);
5310 if (env
.flags
& LBF_NEED_BREAK
) {
5311 env
.flags
&= ~LBF_NEED_BREAK
;
5316 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5317 * us and move them to an alternate dst_cpu in our sched_group
5318 * where they can run. The upper limit on how many times we
5319 * iterate on same src_cpu is dependent on number of cpus in our
5322 * This changes load balance semantics a bit on who can move
5323 * load to a given_cpu. In addition to the given_cpu itself
5324 * (or a ilb_cpu acting on its behalf where given_cpu is
5325 * nohz-idle), we now have balance_cpu in a position to move
5326 * load to given_cpu. In rare situations, this may cause
5327 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5328 * _independently_ and at _same_ time to move some load to
5329 * given_cpu) causing exceess load to be moved to given_cpu.
5330 * This however should not happen so much in practice and
5331 * moreover subsequent load balance cycles should correct the
5332 * excess load moved.
5334 if ((env
.flags
& LBF_DST_PINNED
) && env
.imbalance
> 0) {
5336 /* Prevent to re-select dst_cpu via env's cpus */
5337 cpumask_clear_cpu(env
.dst_cpu
, env
.cpus
);
5339 env
.dst_rq
= cpu_rq(env
.new_dst_cpu
);
5340 env
.dst_cpu
= env
.new_dst_cpu
;
5341 env
.flags
&= ~LBF_DST_PINNED
;
5343 env
.loop_break
= sched_nr_migrate_break
;
5346 * Go back to "more_balance" rather than "redo" since we
5347 * need to continue with same src_cpu.
5353 * We failed to reach balance because of affinity.
5356 int *group_imbalance
= &sd_parent
->groups
->sgp
->imbalance
;
5358 if ((env
.flags
& LBF_SOME_PINNED
) && env
.imbalance
> 0) {
5359 *group_imbalance
= 1;
5360 } else if (*group_imbalance
)
5361 *group_imbalance
= 0;
5364 /* All tasks on this runqueue were pinned by CPU affinity */
5365 if (unlikely(env
.flags
& LBF_ALL_PINNED
)) {
5366 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
5367 if (!cpumask_empty(cpus
)) {
5369 env
.loop_break
= sched_nr_migrate_break
;
5377 schedstat_inc(sd
, lb_failed
[idle
]);
5379 * Increment the failure counter only on periodic balance.
5380 * We do not want newidle balance, which can be very
5381 * frequent, pollute the failure counter causing
5382 * excessive cache_hot migrations and active balances.
5384 if (idle
!= CPU_NEWLY_IDLE
)
5385 sd
->nr_balance_failed
++;
5387 if (need_active_balance(&env
)) {
5388 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
5390 /* don't kick the active_load_balance_cpu_stop,
5391 * if the curr task on busiest cpu can't be
5394 if (!cpumask_test_cpu(this_cpu
,
5395 tsk_cpus_allowed(busiest
->curr
))) {
5396 raw_spin_unlock_irqrestore(&busiest
->lock
,
5398 env
.flags
|= LBF_ALL_PINNED
;
5399 goto out_one_pinned
;
5403 * ->active_balance synchronizes accesses to
5404 * ->active_balance_work. Once set, it's cleared
5405 * only after active load balance is finished.
5407 if (!busiest
->active_balance
) {
5408 busiest
->active_balance
= 1;
5409 busiest
->push_cpu
= this_cpu
;
5412 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
5414 if (active_balance
) {
5415 stop_one_cpu_nowait(cpu_of(busiest
),
5416 active_load_balance_cpu_stop
, busiest
,
5417 &busiest
->active_balance_work
);
5421 * We've kicked active balancing, reset the failure
5424 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
5427 sd
->nr_balance_failed
= 0;
5429 if (likely(!active_balance
)) {
5430 /* We were unbalanced, so reset the balancing interval */
5431 sd
->balance_interval
= sd
->min_interval
;
5434 * If we've begun active balancing, start to back off. This
5435 * case may not be covered by the all_pinned logic if there
5436 * is only 1 task on the busy runqueue (because we don't call
5439 if (sd
->balance_interval
< sd
->max_interval
)
5440 sd
->balance_interval
*= 2;
5446 schedstat_inc(sd
, lb_balanced
[idle
]);
5448 sd
->nr_balance_failed
= 0;
5451 /* tune up the balancing interval */
5452 if (((env
.flags
& LBF_ALL_PINNED
) &&
5453 sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
5454 (sd
->balance_interval
< sd
->max_interval
))
5455 sd
->balance_interval
*= 2;
5463 * idle_balance is called by schedule() if this_cpu is about to become
5464 * idle. Attempts to pull tasks from other CPUs.
5466 void idle_balance(int this_cpu
, struct rq
*this_rq
)
5468 struct sched_domain
*sd
;
5469 int pulled_task
= 0;
5470 unsigned long next_balance
= jiffies
+ HZ
;
5473 this_rq
->idle_stamp
= rq_clock(this_rq
);
5475 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
5479 * Drop the rq->lock, but keep IRQ/preempt disabled.
5481 raw_spin_unlock(&this_rq
->lock
);
5483 update_blocked_averages(this_cpu
);
5485 for_each_domain(this_cpu
, sd
) {
5486 unsigned long interval
;
5487 int continue_balancing
= 1;
5488 u64 t0
, domain_cost
;
5490 if (!(sd
->flags
& SD_LOAD_BALANCE
))
5493 if (this_rq
->avg_idle
< curr_cost
+ sd
->max_newidle_lb_cost
)
5496 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
5497 t0
= sched_clock_cpu(this_cpu
);
5499 /* If we've pulled tasks over stop searching: */
5500 pulled_task
= load_balance(this_cpu
, this_rq
,
5502 &continue_balancing
);
5504 domain_cost
= sched_clock_cpu(this_cpu
) - t0
;
5505 if (domain_cost
> sd
->max_newidle_lb_cost
)
5506 sd
->max_newidle_lb_cost
= domain_cost
;
5508 curr_cost
+= domain_cost
;
5511 interval
= msecs_to_jiffies(sd
->balance_interval
);
5512 if (time_after(next_balance
, sd
->last_balance
+ interval
))
5513 next_balance
= sd
->last_balance
+ interval
;
5515 this_rq
->idle_stamp
= 0;
5521 raw_spin_lock(&this_rq
->lock
);
5523 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
5525 * We are going idle. next_balance may be set based on
5526 * a busy processor. So reset next_balance.
5528 this_rq
->next_balance
= next_balance
;
5531 if (curr_cost
> this_rq
->max_idle_balance_cost
)
5532 this_rq
->max_idle_balance_cost
= curr_cost
;
5536 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5537 * running tasks off the busiest CPU onto idle CPUs. It requires at
5538 * least 1 task to be running on each physical CPU where possible, and
5539 * avoids physical / logical imbalances.
5541 static int active_load_balance_cpu_stop(void *data
)
5543 struct rq
*busiest_rq
= data
;
5544 int busiest_cpu
= cpu_of(busiest_rq
);
5545 int target_cpu
= busiest_rq
->push_cpu
;
5546 struct rq
*target_rq
= cpu_rq(target_cpu
);
5547 struct sched_domain
*sd
;
5549 raw_spin_lock_irq(&busiest_rq
->lock
);
5551 /* make sure the requested cpu hasn't gone down in the meantime */
5552 if (unlikely(busiest_cpu
!= smp_processor_id() ||
5553 !busiest_rq
->active_balance
))
5556 /* Is there any task to move? */
5557 if (busiest_rq
->nr_running
<= 1)
5561 * This condition is "impossible", if it occurs
5562 * we need to fix it. Originally reported by
5563 * Bjorn Helgaas on a 128-cpu setup.
5565 BUG_ON(busiest_rq
== target_rq
);
5567 /* move a task from busiest_rq to target_rq */
5568 double_lock_balance(busiest_rq
, target_rq
);
5570 /* Search for an sd spanning us and the target CPU. */
5572 for_each_domain(target_cpu
, sd
) {
5573 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
5574 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
5579 struct lb_env env
= {
5581 .dst_cpu
= target_cpu
,
5582 .dst_rq
= target_rq
,
5583 .src_cpu
= busiest_rq
->cpu
,
5584 .src_rq
= busiest_rq
,
5588 schedstat_inc(sd
, alb_count
);
5590 if (move_one_task(&env
))
5591 schedstat_inc(sd
, alb_pushed
);
5593 schedstat_inc(sd
, alb_failed
);
5596 double_unlock_balance(busiest_rq
, target_rq
);
5598 busiest_rq
->active_balance
= 0;
5599 raw_spin_unlock_irq(&busiest_rq
->lock
);
5603 #ifdef CONFIG_NO_HZ_COMMON
5605 * idle load balancing details
5606 * - When one of the busy CPUs notice that there may be an idle rebalancing
5607 * needed, they will kick the idle load balancer, which then does idle
5608 * load balancing for all the idle CPUs.
5611 cpumask_var_t idle_cpus_mask
;
5613 unsigned long next_balance
; /* in jiffy units */
5614 } nohz ____cacheline_aligned
;
5616 static inline int find_new_ilb(int call_cpu
)
5618 int ilb
= cpumask_first(nohz
.idle_cpus_mask
);
5620 if (ilb
< nr_cpu_ids
&& idle_cpu(ilb
))
5627 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5628 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5629 * CPU (if there is one).
5631 static void nohz_balancer_kick(int cpu
)
5635 nohz
.next_balance
++;
5637 ilb_cpu
= find_new_ilb(cpu
);
5639 if (ilb_cpu
>= nr_cpu_ids
)
5642 if (test_and_set_bit(NOHZ_BALANCE_KICK
, nohz_flags(ilb_cpu
)))
5645 * Use smp_send_reschedule() instead of resched_cpu().
5646 * This way we generate a sched IPI on the target cpu which
5647 * is idle. And the softirq performing nohz idle load balance
5648 * will be run before returning from the IPI.
5650 smp_send_reschedule(ilb_cpu
);
5654 static inline void nohz_balance_exit_idle(int cpu
)
5656 if (unlikely(test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))) {
5657 cpumask_clear_cpu(cpu
, nohz
.idle_cpus_mask
);
5658 atomic_dec(&nohz
.nr_cpus
);
5659 clear_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
5663 static inline void set_cpu_sd_state_busy(void)
5665 struct sched_domain
*sd
;
5668 sd
= rcu_dereference_check_sched_domain(this_rq()->sd
);
5670 if (!sd
|| !sd
->nohz_idle
)
5674 for (; sd
; sd
= sd
->parent
)
5675 atomic_inc(&sd
->groups
->sgp
->nr_busy_cpus
);
5680 void set_cpu_sd_state_idle(void)
5682 struct sched_domain
*sd
;
5685 sd
= rcu_dereference_check_sched_domain(this_rq()->sd
);
5687 if (!sd
|| sd
->nohz_idle
)
5691 for (; sd
; sd
= sd
->parent
)
5692 atomic_dec(&sd
->groups
->sgp
->nr_busy_cpus
);
5698 * This routine will record that the cpu is going idle with tick stopped.
5699 * This info will be used in performing idle load balancing in the future.
5701 void nohz_balance_enter_idle(int cpu
)
5704 * If this cpu is going down, then nothing needs to be done.
5706 if (!cpu_active(cpu
))
5709 if (test_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
)))
5712 cpumask_set_cpu(cpu
, nohz
.idle_cpus_mask
);
5713 atomic_inc(&nohz
.nr_cpus
);
5714 set_bit(NOHZ_TICK_STOPPED
, nohz_flags(cpu
));
5717 static int sched_ilb_notifier(struct notifier_block
*nfb
,
5718 unsigned long action
, void *hcpu
)
5720 switch (action
& ~CPU_TASKS_FROZEN
) {
5722 nohz_balance_exit_idle(smp_processor_id());
5730 static DEFINE_SPINLOCK(balancing
);
5733 * Scale the max load_balance interval with the number of CPUs in the system.
5734 * This trades load-balance latency on larger machines for less cross talk.
5736 void update_max_interval(void)
5738 max_load_balance_interval
= HZ
*num_online_cpus()/10;
5742 * It checks each scheduling domain to see if it is due to be balanced,
5743 * and initiates a balancing operation if so.
5745 * Balancing parameters are set up in init_sched_domains.
5747 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
5749 int continue_balancing
= 1;
5750 struct rq
*rq
= cpu_rq(cpu
);
5751 unsigned long interval
;
5752 struct sched_domain
*sd
;
5753 /* Earliest time when we have to do rebalance again */
5754 unsigned long next_balance
= jiffies
+ 60*HZ
;
5755 int update_next_balance
= 0;
5756 int need_serialize
, need_decay
= 0;
5759 update_blocked_averages(cpu
);
5762 for_each_domain(cpu
, sd
) {
5764 * Decay the newidle max times here because this is a regular
5765 * visit to all the domains. Decay ~1% per second.
5767 if (time_after(jiffies
, sd
->next_decay_max_lb_cost
)) {
5768 sd
->max_newidle_lb_cost
=
5769 (sd
->max_newidle_lb_cost
* 253) / 256;
5770 sd
->next_decay_max_lb_cost
= jiffies
+ HZ
;
5773 max_cost
+= sd
->max_newidle_lb_cost
;
5775 if (!(sd
->flags
& SD_LOAD_BALANCE
))
5779 * Stop the load balance at this level. There is another
5780 * CPU in our sched group which is doing load balancing more
5783 if (!continue_balancing
) {
5789 interval
= sd
->balance_interval
;
5790 if (idle
!= CPU_IDLE
)
5791 interval
*= sd
->busy_factor
;
5793 /* scale ms to jiffies */
5794 interval
= msecs_to_jiffies(interval
);
5795 interval
= clamp(interval
, 1UL, max_load_balance_interval
);
5797 need_serialize
= sd
->flags
& SD_SERIALIZE
;
5799 if (need_serialize
) {
5800 if (!spin_trylock(&balancing
))
5804 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
5805 if (load_balance(cpu
, rq
, sd
, idle
, &continue_balancing
)) {
5807 * The LBF_DST_PINNED logic could have changed
5808 * env->dst_cpu, so we can't know our idle
5809 * state even if we migrated tasks. Update it.
5811 idle
= idle_cpu(cpu
) ? CPU_IDLE
: CPU_NOT_IDLE
;
5813 sd
->last_balance
= jiffies
;
5816 spin_unlock(&balancing
);
5818 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
5819 next_balance
= sd
->last_balance
+ interval
;
5820 update_next_balance
= 1;
5825 * Ensure the rq-wide value also decays but keep it at a
5826 * reasonable floor to avoid funnies with rq->avg_idle.
5828 rq
->max_idle_balance_cost
=
5829 max((u64
)sysctl_sched_migration_cost
, max_cost
);
5834 * next_balance will be updated only when there is a need.
5835 * When the cpu is attached to null domain for ex, it will not be
5838 if (likely(update_next_balance
))
5839 rq
->next_balance
= next_balance
;
5842 #ifdef CONFIG_NO_HZ_COMMON
5844 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
5845 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5847 static void nohz_idle_balance(int this_cpu
, enum cpu_idle_type idle
)
5849 struct rq
*this_rq
= cpu_rq(this_cpu
);
5853 if (idle
!= CPU_IDLE
||
5854 !test_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
)))
5857 for_each_cpu(balance_cpu
, nohz
.idle_cpus_mask
) {
5858 if (balance_cpu
== this_cpu
|| !idle_cpu(balance_cpu
))
5862 * If this cpu gets work to do, stop the load balancing
5863 * work being done for other cpus. Next load
5864 * balancing owner will pick it up.
5869 rq
= cpu_rq(balance_cpu
);
5871 raw_spin_lock_irq(&rq
->lock
);
5872 update_rq_clock(rq
);
5873 update_idle_cpu_load(rq
);
5874 raw_spin_unlock_irq(&rq
->lock
);
5876 rebalance_domains(balance_cpu
, CPU_IDLE
);
5878 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
5879 this_rq
->next_balance
= rq
->next_balance
;
5881 nohz
.next_balance
= this_rq
->next_balance
;
5883 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(this_cpu
));
5887 * Current heuristic for kicking the idle load balancer in the presence
5888 * of an idle cpu is the system.
5889 * - This rq has more than one task.
5890 * - At any scheduler domain level, this cpu's scheduler group has multiple
5891 * busy cpu's exceeding the group's power.
5892 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5893 * domain span are idle.
5895 static inline int nohz_kick_needed(struct rq
*rq
, int cpu
)
5897 unsigned long now
= jiffies
;
5898 struct sched_domain
*sd
;
5900 if (unlikely(idle_cpu(cpu
)))
5904 * We may be recently in ticked or tickless idle mode. At the first
5905 * busy tick after returning from idle, we will update the busy stats.
5907 set_cpu_sd_state_busy();
5908 nohz_balance_exit_idle(cpu
);
5911 * None are in tickless mode and hence no need for NOHZ idle load
5914 if (likely(!atomic_read(&nohz
.nr_cpus
)))
5917 if (time_before(now
, nohz
.next_balance
))
5920 if (rq
->nr_running
>= 2)
5924 for_each_domain(cpu
, sd
) {
5925 struct sched_group
*sg
= sd
->groups
;
5926 struct sched_group_power
*sgp
= sg
->sgp
;
5927 int nr_busy
= atomic_read(&sgp
->nr_busy_cpus
);
5929 if (sd
->flags
& SD_SHARE_PKG_RESOURCES
&& nr_busy
> 1)
5930 goto need_kick_unlock
;
5932 if (sd
->flags
& SD_ASYM_PACKING
&& nr_busy
!= sg
->group_weight
5933 && (cpumask_first_and(nohz
.idle_cpus_mask
,
5934 sched_domain_span(sd
)) < cpu
))
5935 goto need_kick_unlock
;
5937 if (!(sd
->flags
& (SD_SHARE_PKG_RESOURCES
| SD_ASYM_PACKING
)))
5949 static void nohz_idle_balance(int this_cpu
, enum cpu_idle_type idle
) { }
5953 * run_rebalance_domains is triggered when needed from the scheduler tick.
5954 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5956 static void run_rebalance_domains(struct softirq_action
*h
)
5958 int this_cpu
= smp_processor_id();
5959 struct rq
*this_rq
= cpu_rq(this_cpu
);
5960 enum cpu_idle_type idle
= this_rq
->idle_balance
?
5961 CPU_IDLE
: CPU_NOT_IDLE
;
5963 rebalance_domains(this_cpu
, idle
);
5966 * If this cpu has a pending nohz_balance_kick, then do the
5967 * balancing on behalf of the other idle cpus whose ticks are
5970 nohz_idle_balance(this_cpu
, idle
);
5973 static inline int on_null_domain(int cpu
)
5975 return !rcu_dereference_sched(cpu_rq(cpu
)->sd
);
5979 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5981 void trigger_load_balance(struct rq
*rq
, int cpu
)
5983 /* Don't need to rebalance while attached to NULL domain */
5984 if (time_after_eq(jiffies
, rq
->next_balance
) &&
5985 likely(!on_null_domain(cpu
)))
5986 raise_softirq(SCHED_SOFTIRQ
);
5987 #ifdef CONFIG_NO_HZ_COMMON
5988 if (nohz_kick_needed(rq
, cpu
) && likely(!on_null_domain(cpu
)))
5989 nohz_balancer_kick(cpu
);
5993 static void rq_online_fair(struct rq
*rq
)
5998 static void rq_offline_fair(struct rq
*rq
)
6002 /* Ensure any throttled groups are reachable by pick_next_task */
6003 unthrottle_offline_cfs_rqs(rq
);
6006 #endif /* CONFIG_SMP */
6009 * scheduler tick hitting a task of our scheduling class:
6011 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
, int queued
)
6013 struct cfs_rq
*cfs_rq
;
6014 struct sched_entity
*se
= &curr
->se
;
6016 for_each_sched_entity(se
) {
6017 cfs_rq
= cfs_rq_of(se
);
6018 entity_tick(cfs_rq
, se
, queued
);
6021 if (numabalancing_enabled
)
6022 task_tick_numa(rq
, curr
);
6024 update_rq_runnable_avg(rq
, 1);
6028 * called on fork with the child task as argument from the parent's context
6029 * - child not yet on the tasklist
6030 * - preemption disabled
6032 static void task_fork_fair(struct task_struct
*p
)
6034 struct cfs_rq
*cfs_rq
;
6035 struct sched_entity
*se
= &p
->se
, *curr
;
6036 int this_cpu
= smp_processor_id();
6037 struct rq
*rq
= this_rq();
6038 unsigned long flags
;
6040 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6042 update_rq_clock(rq
);
6044 cfs_rq
= task_cfs_rq(current
);
6045 curr
= cfs_rq
->curr
;
6048 * Not only the cpu but also the task_group of the parent might have
6049 * been changed after parent->se.parent,cfs_rq were copied to
6050 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6051 * of child point to valid ones.
6054 __set_task_cpu(p
, this_cpu
);
6057 update_curr(cfs_rq
);
6060 se
->vruntime
= curr
->vruntime
;
6061 place_entity(cfs_rq
, se
, 1);
6063 if (sysctl_sched_child_runs_first
&& curr
&& entity_before(curr
, se
)) {
6065 * Upon rescheduling, sched_class::put_prev_task() will place
6066 * 'current' within the tree based on its new key value.
6068 swap(curr
->vruntime
, se
->vruntime
);
6069 resched_task(rq
->curr
);
6072 se
->vruntime
-= cfs_rq
->min_vruntime
;
6074 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6078 * Priority of the task has changed. Check to see if we preempt
6082 prio_changed_fair(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
6088 * Reschedule if we are currently running on this runqueue and
6089 * our priority decreased, or if we are not currently running on
6090 * this runqueue and our priority is higher than the current's
6092 if (rq
->curr
== p
) {
6093 if (p
->prio
> oldprio
)
6094 resched_task(rq
->curr
);
6096 check_preempt_curr(rq
, p
, 0);
6099 static void switched_from_fair(struct rq
*rq
, struct task_struct
*p
)
6101 struct sched_entity
*se
= &p
->se
;
6102 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
6105 * Ensure the task's vruntime is normalized, so that when its
6106 * switched back to the fair class the enqueue_entity(.flags=0) will
6107 * do the right thing.
6109 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6110 * have normalized the vruntime, if it was !on_rq, then only when
6111 * the task is sleeping will it still have non-normalized vruntime.
6113 if (!se
->on_rq
&& p
->state
!= TASK_RUNNING
) {
6115 * Fix up our vruntime so that the current sleep doesn't
6116 * cause 'unlimited' sleep bonus.
6118 place_entity(cfs_rq
, se
, 0);
6119 se
->vruntime
-= cfs_rq
->min_vruntime
;
6124 * Remove our load from contribution when we leave sched_fair
6125 * and ensure we don't carry in an old decay_count if we
6128 if (se
->avg
.decay_count
) {
6129 __synchronize_entity_decay(se
);
6130 subtract_blocked_load_contrib(cfs_rq
, se
->avg
.load_avg_contrib
);
6136 * We switched to the sched_fair class.
6138 static void switched_to_fair(struct rq
*rq
, struct task_struct
*p
)
6144 * We were most likely switched from sched_rt, so
6145 * kick off the schedule if running, otherwise just see
6146 * if we can still preempt the current task.
6149 resched_task(rq
->curr
);
6151 check_preempt_curr(rq
, p
, 0);
6154 /* Account for a task changing its policy or group.
6156 * This routine is mostly called to set cfs_rq->curr field when a task
6157 * migrates between groups/classes.
6159 static void set_curr_task_fair(struct rq
*rq
)
6161 struct sched_entity
*se
= &rq
->curr
->se
;
6163 for_each_sched_entity(se
) {
6164 struct cfs_rq
*cfs_rq
= cfs_rq_of(se
);
6166 set_next_entity(cfs_rq
, se
);
6167 /* ensure bandwidth has been allocated on our new cfs_rq */
6168 account_cfs_rq_runtime(cfs_rq
, 0);
6172 void init_cfs_rq(struct cfs_rq
*cfs_rq
)
6174 cfs_rq
->tasks_timeline
= RB_ROOT
;
6175 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
6176 #ifndef CONFIG_64BIT
6177 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
6180 atomic64_set(&cfs_rq
->decay_counter
, 1);
6181 atomic_long_set(&cfs_rq
->removed_load
, 0);
6185 #ifdef CONFIG_FAIR_GROUP_SCHED
6186 static void task_move_group_fair(struct task_struct
*p
, int on_rq
)
6188 struct cfs_rq
*cfs_rq
;
6190 * If the task was not on the rq at the time of this cgroup movement
6191 * it must have been asleep, sleeping tasks keep their ->vruntime
6192 * absolute on their old rq until wakeup (needed for the fair sleeper
6193 * bonus in place_entity()).
6195 * If it was on the rq, we've just 'preempted' it, which does convert
6196 * ->vruntime to a relative base.
6198 * Make sure both cases convert their relative position when migrating
6199 * to another cgroup's rq. This does somewhat interfere with the
6200 * fair sleeper stuff for the first placement, but who cares.
6203 * When !on_rq, vruntime of the task has usually NOT been normalized.
6204 * But there are some cases where it has already been normalized:
6206 * - Moving a forked child which is waiting for being woken up by
6207 * wake_up_new_task().
6208 * - Moving a task which has been woken up by try_to_wake_up() and
6209 * waiting for actually being woken up by sched_ttwu_pending().
6211 * To prevent boost or penalty in the new cfs_rq caused by delta
6212 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6214 if (!on_rq
&& (!p
->se
.sum_exec_runtime
|| p
->state
== TASK_WAKING
))
6218 p
->se
.vruntime
-= cfs_rq_of(&p
->se
)->min_vruntime
;
6219 set_task_rq(p
, task_cpu(p
));
6221 cfs_rq
= cfs_rq_of(&p
->se
);
6222 p
->se
.vruntime
+= cfs_rq
->min_vruntime
;
6225 * migrate_task_rq_fair() will have removed our previous
6226 * contribution, but we must synchronize for ongoing future
6229 p
->se
.avg
.decay_count
= atomic64_read(&cfs_rq
->decay_counter
);
6230 cfs_rq
->blocked_load_avg
+= p
->se
.avg
.load_avg_contrib
;
6235 void free_fair_sched_group(struct task_group
*tg
)
6239 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
6241 for_each_possible_cpu(i
) {
6243 kfree(tg
->cfs_rq
[i
]);
6252 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
6254 struct cfs_rq
*cfs_rq
;
6255 struct sched_entity
*se
;
6258 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
6261 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
6265 tg
->shares
= NICE_0_LOAD
;
6267 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
6269 for_each_possible_cpu(i
) {
6270 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
6271 GFP_KERNEL
, cpu_to_node(i
));
6275 se
= kzalloc_node(sizeof(struct sched_entity
),
6276 GFP_KERNEL
, cpu_to_node(i
));
6280 init_cfs_rq(cfs_rq
);
6281 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
6292 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
6294 struct rq
*rq
= cpu_rq(cpu
);
6295 unsigned long flags
;
6298 * Only empty task groups can be destroyed; so we can speculatively
6299 * check on_list without danger of it being re-added.
6301 if (!tg
->cfs_rq
[cpu
]->on_list
)
6304 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6305 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
6306 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6309 void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
6310 struct sched_entity
*se
, int cpu
,
6311 struct sched_entity
*parent
)
6313 struct rq
*rq
= cpu_rq(cpu
);
6317 init_cfs_rq_runtime(cfs_rq
);
6319 tg
->cfs_rq
[cpu
] = cfs_rq
;
6322 /* se could be NULL for root_task_group */
6327 se
->cfs_rq
= &rq
->cfs
;
6329 se
->cfs_rq
= parent
->my_q
;
6332 update_load_set(&se
->load
, 0);
6333 se
->parent
= parent
;
6336 static DEFINE_MUTEX(shares_mutex
);
6338 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
6341 unsigned long flags
;
6344 * We can't change the weight of the root cgroup.
6349 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
6351 mutex_lock(&shares_mutex
);
6352 if (tg
->shares
== shares
)
6355 tg
->shares
= shares
;
6356 for_each_possible_cpu(i
) {
6357 struct rq
*rq
= cpu_rq(i
);
6358 struct sched_entity
*se
;
6361 /* Propagate contribution to hierarchy */
6362 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6364 /* Possible calls to update_curr() need rq clock */
6365 update_rq_clock(rq
);
6366 for_each_sched_entity(se
)
6367 update_cfs_shares(group_cfs_rq(se
));
6368 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6372 mutex_unlock(&shares_mutex
);
6375 #else /* CONFIG_FAIR_GROUP_SCHED */
6377 void free_fair_sched_group(struct task_group
*tg
) { }
6379 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
6384 void unregister_fair_sched_group(struct task_group
*tg
, int cpu
) { }
6386 #endif /* CONFIG_FAIR_GROUP_SCHED */
6389 static unsigned int get_rr_interval_fair(struct rq
*rq
, struct task_struct
*task
)
6391 struct sched_entity
*se
= &task
->se
;
6392 unsigned int rr_interval
= 0;
6395 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6398 if (rq
->cfs
.load
.weight
)
6399 rr_interval
= NS_TO_JIFFIES(sched_slice(cfs_rq_of(se
), se
));
6405 * All the scheduling class methods:
6407 const struct sched_class fair_sched_class
= {
6408 .next
= &idle_sched_class
,
6409 .enqueue_task
= enqueue_task_fair
,
6410 .dequeue_task
= dequeue_task_fair
,
6411 .yield_task
= yield_task_fair
,
6412 .yield_to_task
= yield_to_task_fair
,
6414 .check_preempt_curr
= check_preempt_wakeup
,
6416 .pick_next_task
= pick_next_task_fair
,
6417 .put_prev_task
= put_prev_task_fair
,
6420 .select_task_rq
= select_task_rq_fair
,
6421 .migrate_task_rq
= migrate_task_rq_fair
,
6423 .rq_online
= rq_online_fair
,
6424 .rq_offline
= rq_offline_fair
,
6426 .task_waking
= task_waking_fair
,
6429 .set_curr_task
= set_curr_task_fair
,
6430 .task_tick
= task_tick_fair
,
6431 .task_fork
= task_fork_fair
,
6433 .prio_changed
= prio_changed_fair
,
6434 .switched_from
= switched_from_fair
,
6435 .switched_to
= switched_to_fair
,
6437 .get_rr_interval
= get_rr_interval_fair
,
6439 #ifdef CONFIG_FAIR_GROUP_SCHED
6440 .task_move_group
= task_move_group_fair
,
6444 #ifdef CONFIG_SCHED_DEBUG
6445 void print_cfs_stats(struct seq_file
*m
, int cpu
)
6447 struct cfs_rq
*cfs_rq
;
6450 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
6451 print_cfs_rq(m
, cpu
, cfs_rq
);
6456 __init
void init_sched_fair_class(void)
6459 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
6461 #ifdef CONFIG_NO_HZ_COMMON
6462 nohz
.next_balance
= jiffies
;
6463 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
6464 cpu_notifier(sched_ilb_notifier
, 0);