1 // SPDX-License-Identifier: GPL-2.0
3 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
5 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
7 * Interactivity improvements by Mike Galbraith
8 * (C) 2007 Mike Galbraith <efault@gmx.de>
10 * Various enhancements by Dmitry Adamushko.
11 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
13 * Group scheduling enhancements by Srivatsa Vaddagiri
14 * Copyright IBM Corporation, 2007
15 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
17 * Scaled math optimizations by Thomas Gleixner
18 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
20 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
21 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
25 #include <trace/events/sched.h>
28 * Targeted preemption latency for CPU-bound tasks:
30 * NOTE: this latency value is not the same as the concept of
31 * 'timeslice length' - timeslices in CFS are of variable length
32 * and have no persistent notion like in traditional, time-slice
33 * based scheduling concepts.
35 * (to see the precise effective timeslice length of your workload,
36 * run vmstat and monitor the context-switches (cs) field)
38 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
40 unsigned int sysctl_sched_latency = 6000000ULL;
41 static unsigned int normalized_sysctl_sched_latency = 6000000ULL;
44 * The initial- and re-scaling of tunables is configurable
48 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
49 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
50 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
52 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
54 enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
57 * Minimal preemption granularity for CPU-bound tasks:
59 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
61 unsigned int sysctl_sched_min_granularity = 750000ULL;
62 static unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
65 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
67 static unsigned int sched_nr_latency = 8;
70 * After fork, child runs first. If set to 0 (default) then
71 * parent will (try to) run first.
73 unsigned int sysctl_sched_child_runs_first __read_mostly;
76 * SCHED_OTHER wake-up granularity.
78 * This option delays the preemption effects of decoupled workloads
79 * and reduces their over-scheduling. Synchronous workloads will still
80 * have immediate wakeup/sleep latencies.
82 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
84 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
85 static unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
87 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
91 * For asym packing, by default the lower numbered CPU has higher priority.
93 int __weak arch_asym_cpu_priority(int cpu)
99 * The margin used when comparing utilization with CPU capacity:
100 * util * margin < capacity * 1024
104 static unsigned int capacity_margin = 1280;
107 #ifdef CONFIG_CFS_BANDWIDTH
109 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
110 * each time a cfs_rq requests quota.
112 * Note: in the case that the slice exceeds the runtime remaining (either due
113 * to consumption or the quota being specified to be smaller than the slice)
114 * we will always only issue the remaining available time.
116 * (default: 5 msec, units: microseconds)
118 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
121 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
127 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
133 static inline void update_load_set(struct load_weight *lw, unsigned long w)
140 * Increase the granularity value when there are more CPUs,
141 * because with more CPUs the 'effective latency' as visible
142 * to users decreases. But the relationship is not linear,
143 * so pick a second-best guess by going with the log2 of the
146 * This idea comes from the SD scheduler of Con Kolivas:
148 static unsigned int get_update_sysctl_factor(void)
150 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
153 switch (sysctl_sched_tunable_scaling) {
154 case SCHED_TUNABLESCALING_NONE:
157 case SCHED_TUNABLESCALING_LINEAR:
160 case SCHED_TUNABLESCALING_LOG:
162 factor = 1 + ilog2(cpus);
169 static void update_sysctl(void)
171 unsigned int factor = get_update_sysctl_factor();
173 #define SET_SYSCTL(name) \
174 (sysctl_##name = (factor) * normalized_sysctl_##name)
175 SET_SYSCTL(sched_min_granularity);
176 SET_SYSCTL(sched_latency);
177 SET_SYSCTL(sched_wakeup_granularity);
181 void sched_init_granularity(void)
186 #define WMULT_CONST (~0U)
187 #define WMULT_SHIFT 32
189 static void __update_inv_weight(struct load_weight *lw)
193 if (likely(lw->inv_weight))
196 w = scale_load_down(lw->weight);
198 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
200 else if (unlikely(!w))
201 lw->inv_weight = WMULT_CONST;
203 lw->inv_weight = WMULT_CONST / w;
207 * delta_exec * weight / lw.weight
209 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
211 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
212 * we're guaranteed shift stays positive because inv_weight is guaranteed to
213 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
215 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
216 * weight/lw.weight <= 1, and therefore our shift will also be positive.
218 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
220 u64 fact = scale_load_down(weight);
221 int shift = WMULT_SHIFT;
223 __update_inv_weight(lw);
225 if (unlikely(fact >> 32)) {
232 /* hint to use a 32x32->64 mul */
233 fact = (u64)(u32)fact * lw->inv_weight;
240 return mul_u64_u32_shr(delta_exec, fact, shift);
244 const struct sched_class fair_sched_class;
246 /**************************************************************
247 * CFS operations on generic schedulable entities:
250 #ifdef CONFIG_FAIR_GROUP_SCHED
251 static inline struct task_struct *task_of(struct sched_entity *se)
253 SCHED_WARN_ON(!entity_is_task(se));
254 return container_of(se, struct task_struct, se);
257 /* Walk up scheduling entities hierarchy */
258 #define for_each_sched_entity(se) \
259 for (; se; se = se->parent)
261 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
266 /* runqueue on which this entity is (to be) queued */
267 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
272 /* runqueue "owned" by this group */
273 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
278 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
280 struct rq *rq = rq_of(cfs_rq);
281 int cpu = cpu_of(rq);
284 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
289 * Ensure we either appear before our parent (if already
290 * enqueued) or force our parent to appear after us when it is
291 * enqueued. The fact that we always enqueue bottom-up
292 * reduces this to two cases and a special case for the root
293 * cfs_rq. Furthermore, it also means that we will always reset
294 * tmp_alone_branch either when the branch is connected
295 * to a tree or when we reach the top of the tree
297 if (cfs_rq->tg->parent &&
298 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
300 * If parent is already on the list, we add the child
301 * just before. Thanks to circular linked property of
302 * the list, this means to put the child at the tail
303 * of the list that starts by parent.
305 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
306 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
308 * The branch is now connected to its tree so we can
309 * reset tmp_alone_branch to the beginning of the
312 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
316 if (!cfs_rq->tg->parent) {
318 * cfs rq without parent should be put
319 * at the tail of the list.
321 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
322 &rq->leaf_cfs_rq_list);
324 * We have reach the top of a tree so we can reset
325 * tmp_alone_branch to the beginning of the list.
327 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
332 * The parent has not already been added so we want to
333 * make sure that it will be put after us.
334 * tmp_alone_branch points to the begin of the branch
335 * where we will add parent.
337 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
339 * update tmp_alone_branch to points to the new begin
342 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
346 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
348 if (cfs_rq->on_list) {
349 struct rq *rq = rq_of(cfs_rq);
352 * With cfs_rq being unthrottled/throttled during an enqueue,
353 * it can happen the tmp_alone_branch points the a leaf that
354 * we finally want to del. In this case, tmp_alone_branch moves
355 * to the prev element but it will point to rq->leaf_cfs_rq_list
356 * at the end of the enqueue.
358 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
359 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
361 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
366 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
368 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
371 /* Iterate through all cfs_rq's on a runqueue in bottom-up order */
372 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
373 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
375 /* Do the two (enqueued) entities belong to the same group ? */
376 static inline struct cfs_rq *
377 is_same_group(struct sched_entity *se, struct sched_entity *pse)
379 if (se->cfs_rq == pse->cfs_rq)
385 static inline struct sched_entity *parent_entity(struct sched_entity *se)
391 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
393 int se_depth, pse_depth;
396 * preemption test can be made between sibling entities who are in the
397 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
398 * both tasks until we find their ancestors who are siblings of common
402 /* First walk up until both entities are at same depth */
403 se_depth = (*se)->depth;
404 pse_depth = (*pse)->depth;
406 while (se_depth > pse_depth) {
408 *se = parent_entity(*se);
411 while (pse_depth > se_depth) {
413 *pse = parent_entity(*pse);
416 while (!is_same_group(*se, *pse)) {
417 *se = parent_entity(*se);
418 *pse = parent_entity(*pse);
422 #else /* !CONFIG_FAIR_GROUP_SCHED */
424 static inline struct task_struct *task_of(struct sched_entity *se)
426 return container_of(se, struct task_struct, se);
429 #define for_each_sched_entity(se) \
430 for (; se; se = NULL)
432 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
434 return &task_rq(p)->cfs;
437 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
439 struct task_struct *p = task_of(se);
440 struct rq *rq = task_rq(p);
445 /* runqueue "owned" by this group */
446 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
451 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
456 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
460 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
464 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
465 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
467 static inline struct sched_entity *parent_entity(struct sched_entity *se)
473 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
477 #endif /* CONFIG_FAIR_GROUP_SCHED */
479 static __always_inline
480 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
482 /**************************************************************
483 * Scheduling class tree data structure manipulation methods:
486 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
488 s64 delta = (s64)(vruntime - max_vruntime);
490 max_vruntime = vruntime;
495 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
497 s64 delta = (s64)(vruntime - min_vruntime);
499 min_vruntime = vruntime;
504 static inline int entity_before(struct sched_entity *a,
505 struct sched_entity *b)
507 return (s64)(a->vruntime - b->vruntime) < 0;
510 static void update_min_vruntime(struct cfs_rq *cfs_rq)
512 struct sched_entity *curr = cfs_rq->curr;
513 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
515 u64 vruntime = cfs_rq->min_vruntime;
519 vruntime = curr->vruntime;
524 if (leftmost) { /* non-empty tree */
525 struct sched_entity *se;
526 se = rb_entry(leftmost, struct sched_entity, run_node);
529 vruntime = se->vruntime;
531 vruntime = min_vruntime(vruntime, se->vruntime);
534 /* ensure we never gain time by being placed backwards. */
535 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
538 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
543 * Enqueue an entity into the rb-tree:
545 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
547 struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node;
548 struct rb_node *parent = NULL;
549 struct sched_entity *entry;
550 bool leftmost = true;
553 * Find the right place in the rbtree:
557 entry = rb_entry(parent, struct sched_entity, run_node);
559 * We dont care about collisions. Nodes with
560 * the same key stay together.
562 if (entity_before(se, entry)) {
563 link = &parent->rb_left;
565 link = &parent->rb_right;
570 rb_link_node(&se->run_node, parent, link);
571 rb_insert_color_cached(&se->run_node,
572 &cfs_rq->tasks_timeline, leftmost);
575 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
577 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
580 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
582 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
587 return rb_entry(left, struct sched_entity, run_node);
590 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
592 struct rb_node *next = rb_next(&se->run_node);
597 return rb_entry(next, struct sched_entity, run_node);
600 #ifdef CONFIG_SCHED_DEBUG
601 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
603 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
608 return rb_entry(last, struct sched_entity, run_node);
611 /**************************************************************
612 * Scheduling class statistics methods:
615 int sched_proc_update_handler(struct ctl_table *table, int write,
616 void __user *buffer, size_t *lenp,
619 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
620 unsigned int factor = get_update_sysctl_factor();
625 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
626 sysctl_sched_min_granularity);
628 #define WRT_SYSCTL(name) \
629 (normalized_sysctl_##name = sysctl_##name / (factor))
630 WRT_SYSCTL(sched_min_granularity);
631 WRT_SYSCTL(sched_latency);
632 WRT_SYSCTL(sched_wakeup_granularity);
642 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
644 if (unlikely(se->load.weight != NICE_0_LOAD))
645 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
651 * The idea is to set a period in which each task runs once.
653 * When there are too many tasks (sched_nr_latency) we have to stretch
654 * this period because otherwise the slices get too small.
656 * p = (nr <= nl) ? l : l*nr/nl
658 static u64 __sched_period(unsigned long nr_running)
660 if (unlikely(nr_running > sched_nr_latency))
661 return nr_running * sysctl_sched_min_granularity;
663 return sysctl_sched_latency;
667 * We calculate the wall-time slice from the period by taking a part
668 * proportional to the weight.
672 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
674 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
676 for_each_sched_entity(se) {
677 struct load_weight *load;
678 struct load_weight lw;
680 cfs_rq = cfs_rq_of(se);
681 load = &cfs_rq->load;
683 if (unlikely(!se->on_rq)) {
686 update_load_add(&lw, se->load.weight);
689 slice = __calc_delta(slice, se->load.weight, load);
695 * We calculate the vruntime slice of a to-be-inserted task.
699 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
701 return calc_delta_fair(sched_slice(cfs_rq, se), se);
707 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
708 static unsigned long task_h_load(struct task_struct *p);
709 static unsigned long capacity_of(int cpu);
711 /* Give new sched_entity start runnable values to heavy its load in infant time */
712 void init_entity_runnable_average(struct sched_entity *se)
714 struct sched_avg *sa = &se->avg;
716 memset(sa, 0, sizeof(*sa));
719 * Tasks are initialized with full load to be seen as heavy tasks until
720 * they get a chance to stabilize to their real load level.
721 * Group entities are initialized with zero load to reflect the fact that
722 * nothing has been attached to the task group yet.
724 if (entity_is_task(se))
725 sa->runnable_load_avg = sa->load_avg = scale_load_down(se->load.weight);
727 se->runnable_weight = se->load.weight;
729 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
732 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
733 static void attach_entity_cfs_rq(struct sched_entity *se);
736 * With new tasks being created, their initial util_avgs are extrapolated
737 * based on the cfs_rq's current util_avg:
739 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
741 * However, in many cases, the above util_avg does not give a desired
742 * value. Moreover, the sum of the util_avgs may be divergent, such
743 * as when the series is a harmonic series.
745 * To solve this problem, we also cap the util_avg of successive tasks to
746 * only 1/2 of the left utilization budget:
748 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
750 * where n denotes the nth task and cpu_scale the CPU capacity.
752 * For example, for a CPU with 1024 of capacity, a simplest series from
753 * the beginning would be like:
755 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
756 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
758 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
759 * if util_avg > util_avg_cap.
761 void post_init_entity_util_avg(struct sched_entity *se)
763 struct cfs_rq *cfs_rq = cfs_rq_of(se);
764 struct sched_avg *sa = &se->avg;
765 long cpu_scale = arch_scale_cpu_capacity(NULL, cpu_of(rq_of(cfs_rq)));
766 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
769 if (cfs_rq->avg.util_avg != 0) {
770 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
771 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
773 if (sa->util_avg > cap)
780 if (entity_is_task(se)) {
781 struct task_struct *p = task_of(se);
782 if (p->sched_class != &fair_sched_class) {
784 * For !fair tasks do:
786 update_cfs_rq_load_avg(now, cfs_rq);
787 attach_entity_load_avg(cfs_rq, se, 0);
788 switched_from_fair(rq, p);
790 * such that the next switched_to_fair() has the
793 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
798 attach_entity_cfs_rq(se);
801 #else /* !CONFIG_SMP */
802 void init_entity_runnable_average(struct sched_entity *se)
805 void post_init_entity_util_avg(struct sched_entity *se)
808 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
811 #endif /* CONFIG_SMP */
814 * Update the current task's runtime statistics.
816 static void update_curr(struct cfs_rq *cfs_rq)
818 struct sched_entity *curr = cfs_rq->curr;
819 u64 now = rq_clock_task(rq_of(cfs_rq));
825 delta_exec = now - curr->exec_start;
826 if (unlikely((s64)delta_exec <= 0))
829 curr->exec_start = now;
831 schedstat_set(curr->statistics.exec_max,
832 max(delta_exec, curr->statistics.exec_max));
834 curr->sum_exec_runtime += delta_exec;
835 schedstat_add(cfs_rq->exec_clock, delta_exec);
837 curr->vruntime += calc_delta_fair(delta_exec, curr);
838 update_min_vruntime(cfs_rq);
840 if (entity_is_task(curr)) {
841 struct task_struct *curtask = task_of(curr);
843 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
844 cgroup_account_cputime(curtask, delta_exec);
845 account_group_exec_runtime(curtask, delta_exec);
848 account_cfs_rq_runtime(cfs_rq, delta_exec);
851 static void update_curr_fair(struct rq *rq)
853 update_curr(cfs_rq_of(&rq->curr->se));
857 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
859 u64 wait_start, prev_wait_start;
861 if (!schedstat_enabled())
864 wait_start = rq_clock(rq_of(cfs_rq));
865 prev_wait_start = schedstat_val(se->statistics.wait_start);
867 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
868 likely(wait_start > prev_wait_start))
869 wait_start -= prev_wait_start;
871 __schedstat_set(se->statistics.wait_start, wait_start);
875 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
877 struct task_struct *p;
880 if (!schedstat_enabled())
883 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
885 if (entity_is_task(se)) {
887 if (task_on_rq_migrating(p)) {
889 * Preserve migrating task's wait time so wait_start
890 * time stamp can be adjusted to accumulate wait time
891 * prior to migration.
893 __schedstat_set(se->statistics.wait_start, delta);
896 trace_sched_stat_wait(p, delta);
899 __schedstat_set(se->statistics.wait_max,
900 max(schedstat_val(se->statistics.wait_max), delta));
901 __schedstat_inc(se->statistics.wait_count);
902 __schedstat_add(se->statistics.wait_sum, delta);
903 __schedstat_set(se->statistics.wait_start, 0);
907 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
909 struct task_struct *tsk = NULL;
910 u64 sleep_start, block_start;
912 if (!schedstat_enabled())
915 sleep_start = schedstat_val(se->statistics.sleep_start);
916 block_start = schedstat_val(se->statistics.block_start);
918 if (entity_is_task(se))
922 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
927 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
928 __schedstat_set(se->statistics.sleep_max, delta);
930 __schedstat_set(se->statistics.sleep_start, 0);
931 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
934 account_scheduler_latency(tsk, delta >> 10, 1);
935 trace_sched_stat_sleep(tsk, delta);
939 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
944 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
945 __schedstat_set(se->statistics.block_max, delta);
947 __schedstat_set(se->statistics.block_start, 0);
948 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
951 if (tsk->in_iowait) {
952 __schedstat_add(se->statistics.iowait_sum, delta);
953 __schedstat_inc(se->statistics.iowait_count);
954 trace_sched_stat_iowait(tsk, delta);
957 trace_sched_stat_blocked(tsk, delta);
960 * Blocking time is in units of nanosecs, so shift by
961 * 20 to get a milliseconds-range estimation of the
962 * amount of time that the task spent sleeping:
964 if (unlikely(prof_on == SLEEP_PROFILING)) {
965 profile_hits(SLEEP_PROFILING,
966 (void *)get_wchan(tsk),
969 account_scheduler_latency(tsk, delta >> 10, 0);
975 * Task is being enqueued - update stats:
978 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
980 if (!schedstat_enabled())
984 * Are we enqueueing a waiting task? (for current tasks
985 * a dequeue/enqueue event is a NOP)
987 if (se != cfs_rq->curr)
988 update_stats_wait_start(cfs_rq, se);
990 if (flags & ENQUEUE_WAKEUP)
991 update_stats_enqueue_sleeper(cfs_rq, se);
995 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
998 if (!schedstat_enabled())
1002 * Mark the end of the wait period if dequeueing a
1005 if (se != cfs_rq->curr)
1006 update_stats_wait_end(cfs_rq, se);
1008 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1009 struct task_struct *tsk = task_of(se);
1011 if (tsk->state & TASK_INTERRUPTIBLE)
1012 __schedstat_set(se->statistics.sleep_start,
1013 rq_clock(rq_of(cfs_rq)));
1014 if (tsk->state & TASK_UNINTERRUPTIBLE)
1015 __schedstat_set(se->statistics.block_start,
1016 rq_clock(rq_of(cfs_rq)));
1021 * We are picking a new current task - update its stats:
1024 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1027 * We are starting a new run period:
1029 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1032 /**************************************************
1033 * Scheduling class queueing methods:
1036 #ifdef CONFIG_NUMA_BALANCING
1038 * Approximate time to scan a full NUMA task in ms. The task scan period is
1039 * calculated based on the tasks virtual memory size and
1040 * numa_balancing_scan_size.
1042 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1043 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1045 /* Portion of address space to scan in MB */
1046 unsigned int sysctl_numa_balancing_scan_size = 256;
1048 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1049 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1052 refcount_t refcount;
1054 spinlock_t lock; /* nr_tasks, tasks */
1059 struct rcu_head rcu;
1060 unsigned long total_faults;
1061 unsigned long max_faults_cpu;
1063 * Faults_cpu is used to decide whether memory should move
1064 * towards the CPU. As a consequence, these stats are weighted
1065 * more by CPU use than by memory faults.
1067 unsigned long *faults_cpu;
1068 unsigned long faults[0];
1071 static inline unsigned long group_faults_priv(struct numa_group *ng);
1072 static inline unsigned long group_faults_shared(struct numa_group *ng);
1074 static unsigned int task_nr_scan_windows(struct task_struct *p)
1076 unsigned long rss = 0;
1077 unsigned long nr_scan_pages;
1080 * Calculations based on RSS as non-present and empty pages are skipped
1081 * by the PTE scanner and NUMA hinting faults should be trapped based
1084 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1085 rss = get_mm_rss(p->mm);
1087 rss = nr_scan_pages;
1089 rss = round_up(rss, nr_scan_pages);
1090 return rss / nr_scan_pages;
1093 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1094 #define MAX_SCAN_WINDOW 2560
1096 static unsigned int task_scan_min(struct task_struct *p)
1098 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1099 unsigned int scan, floor;
1100 unsigned int windows = 1;
1102 if (scan_size < MAX_SCAN_WINDOW)
1103 windows = MAX_SCAN_WINDOW / scan_size;
1104 floor = 1000 / windows;
1106 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1107 return max_t(unsigned int, floor, scan);
1110 static unsigned int task_scan_start(struct task_struct *p)
1112 unsigned long smin = task_scan_min(p);
1113 unsigned long period = smin;
1115 /* Scale the maximum scan period with the amount of shared memory. */
1116 if (p->numa_group) {
1117 struct numa_group *ng = p->numa_group;
1118 unsigned long shared = group_faults_shared(ng);
1119 unsigned long private = group_faults_priv(ng);
1121 period *= refcount_read(&ng->refcount);
1122 period *= shared + 1;
1123 period /= private + shared + 1;
1126 return max(smin, period);
1129 static unsigned int task_scan_max(struct task_struct *p)
1131 unsigned long smin = task_scan_min(p);
1134 /* Watch for min being lower than max due to floor calculations */
1135 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1137 /* Scale the maximum scan period with the amount of shared memory. */
1138 if (p->numa_group) {
1139 struct numa_group *ng = p->numa_group;
1140 unsigned long shared = group_faults_shared(ng);
1141 unsigned long private = group_faults_priv(ng);
1142 unsigned long period = smax;
1144 period *= refcount_read(&ng->refcount);
1145 period *= shared + 1;
1146 period /= private + shared + 1;
1148 smax = max(smax, period);
1151 return max(smin, smax);
1154 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
1157 struct mm_struct *mm = p->mm;
1160 mm_users = atomic_read(&mm->mm_users);
1161 if (mm_users == 1) {
1162 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1163 mm->numa_scan_seq = 0;
1167 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
1168 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1169 p->numa_work.next = &p->numa_work;
1170 p->numa_faults = NULL;
1171 p->numa_group = NULL;
1172 p->last_task_numa_placement = 0;
1173 p->last_sum_exec_runtime = 0;
1175 /* New address space, reset the preferred nid */
1176 if (!(clone_flags & CLONE_VM)) {
1177 p->numa_preferred_nid = -1;
1182 * New thread, keep existing numa_preferred_nid which should be copied
1183 * already by arch_dup_task_struct but stagger when scans start.
1188 delay = min_t(unsigned int, task_scan_max(current),
1189 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
1190 delay += 2 * TICK_NSEC;
1191 p->node_stamp = delay;
1195 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1197 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1198 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1201 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1203 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1204 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1207 /* Shared or private faults. */
1208 #define NR_NUMA_HINT_FAULT_TYPES 2
1210 /* Memory and CPU locality */
1211 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1213 /* Averaged statistics, and temporary buffers. */
1214 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1216 pid_t task_numa_group_id(struct task_struct *p)
1218 return p->numa_group ? p->numa_group->gid : 0;
1222 * The averaged statistics, shared & private, memory & CPU,
1223 * occupy the first half of the array. The second half of the
1224 * array is for current counters, which are averaged into the
1225 * first set by task_numa_placement.
1227 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1229 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1232 static inline unsigned long task_faults(struct task_struct *p, int nid)
1234 if (!p->numa_faults)
1237 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1238 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1241 static inline unsigned long group_faults(struct task_struct *p, int nid)
1246 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1247 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1250 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1252 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1253 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1256 static inline unsigned long group_faults_priv(struct numa_group *ng)
1258 unsigned long faults = 0;
1261 for_each_online_node(node) {
1262 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1268 static inline unsigned long group_faults_shared(struct numa_group *ng)
1270 unsigned long faults = 0;
1273 for_each_online_node(node) {
1274 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1281 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1282 * considered part of a numa group's pseudo-interleaving set. Migrations
1283 * between these nodes are slowed down, to allow things to settle down.
1285 #define ACTIVE_NODE_FRACTION 3
1287 static bool numa_is_active_node(int nid, struct numa_group *ng)
1289 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1292 /* Handle placement on systems where not all nodes are directly connected. */
1293 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1294 int maxdist, bool task)
1296 unsigned long score = 0;
1300 * All nodes are directly connected, and the same distance
1301 * from each other. No need for fancy placement algorithms.
1303 if (sched_numa_topology_type == NUMA_DIRECT)
1307 * This code is called for each node, introducing N^2 complexity,
1308 * which should be ok given the number of nodes rarely exceeds 8.
1310 for_each_online_node(node) {
1311 unsigned long faults;
1312 int dist = node_distance(nid, node);
1315 * The furthest away nodes in the system are not interesting
1316 * for placement; nid was already counted.
1318 if (dist == sched_max_numa_distance || node == nid)
1322 * On systems with a backplane NUMA topology, compare groups
1323 * of nodes, and move tasks towards the group with the most
1324 * memory accesses. When comparing two nodes at distance
1325 * "hoplimit", only nodes closer by than "hoplimit" are part
1326 * of each group. Skip other nodes.
1328 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1332 /* Add up the faults from nearby nodes. */
1334 faults = task_faults(p, node);
1336 faults = group_faults(p, node);
1339 * On systems with a glueless mesh NUMA topology, there are
1340 * no fixed "groups of nodes". Instead, nodes that are not
1341 * directly connected bounce traffic through intermediate
1342 * nodes; a numa_group can occupy any set of nodes.
1343 * The further away a node is, the less the faults count.
1344 * This seems to result in good task placement.
1346 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1347 faults *= (sched_max_numa_distance - dist);
1348 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1358 * These return the fraction of accesses done by a particular task, or
1359 * task group, on a particular numa node. The group weight is given a
1360 * larger multiplier, in order to group tasks together that are almost
1361 * evenly spread out between numa nodes.
1363 static inline unsigned long task_weight(struct task_struct *p, int nid,
1366 unsigned long faults, total_faults;
1368 if (!p->numa_faults)
1371 total_faults = p->total_numa_faults;
1376 faults = task_faults(p, nid);
1377 faults += score_nearby_nodes(p, nid, dist, true);
1379 return 1000 * faults / total_faults;
1382 static inline unsigned long group_weight(struct task_struct *p, int nid,
1385 unsigned long faults, total_faults;
1390 total_faults = p->numa_group->total_faults;
1395 faults = group_faults(p, nid);
1396 faults += score_nearby_nodes(p, nid, dist, false);
1398 return 1000 * faults / total_faults;
1401 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1402 int src_nid, int dst_cpu)
1404 struct numa_group *ng = p->numa_group;
1405 int dst_nid = cpu_to_node(dst_cpu);
1406 int last_cpupid, this_cpupid;
1408 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1409 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1412 * Allow first faults or private faults to migrate immediately early in
1413 * the lifetime of a task. The magic number 4 is based on waiting for
1414 * two full passes of the "multi-stage node selection" test that is
1417 if ((p->numa_preferred_nid == -1 || p->numa_scan_seq <= 4) &&
1418 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1422 * Multi-stage node selection is used in conjunction with a periodic
1423 * migration fault to build a temporal task<->page relation. By using
1424 * a two-stage filter we remove short/unlikely relations.
1426 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1427 * a task's usage of a particular page (n_p) per total usage of this
1428 * page (n_t) (in a given time-span) to a probability.
1430 * Our periodic faults will sample this probability and getting the
1431 * same result twice in a row, given these samples are fully
1432 * independent, is then given by P(n)^2, provided our sample period
1433 * is sufficiently short compared to the usage pattern.
1435 * This quadric squishes small probabilities, making it less likely we
1436 * act on an unlikely task<->page relation.
1438 if (!cpupid_pid_unset(last_cpupid) &&
1439 cpupid_to_nid(last_cpupid) != dst_nid)
1442 /* Always allow migrate on private faults */
1443 if (cpupid_match_pid(p, last_cpupid))
1446 /* A shared fault, but p->numa_group has not been set up yet. */
1451 * Destination node is much more heavily used than the source
1452 * node? Allow migration.
1454 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1455 ACTIVE_NODE_FRACTION)
1459 * Distribute memory according to CPU & memory use on each node,
1460 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1462 * faults_cpu(dst) 3 faults_cpu(src)
1463 * --------------- * - > ---------------
1464 * faults_mem(dst) 4 faults_mem(src)
1466 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1467 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1470 static unsigned long weighted_cpuload(struct rq *rq);
1471 static unsigned long source_load(int cpu, int type);
1472 static unsigned long target_load(int cpu, int type);
1474 /* Cached statistics for all CPUs within a node */
1478 /* Total compute capacity of CPUs on a node */
1479 unsigned long compute_capacity;
1483 * XXX borrowed from update_sg_lb_stats
1485 static void update_numa_stats(struct numa_stats *ns, int nid)
1489 memset(ns, 0, sizeof(*ns));
1490 for_each_cpu(cpu, cpumask_of_node(nid)) {
1491 struct rq *rq = cpu_rq(cpu);
1493 ns->load += weighted_cpuload(rq);
1494 ns->compute_capacity += capacity_of(cpu);
1499 struct task_numa_env {
1500 struct task_struct *p;
1502 int src_cpu, src_nid;
1503 int dst_cpu, dst_nid;
1505 struct numa_stats src_stats, dst_stats;
1510 struct task_struct *best_task;
1515 static void task_numa_assign(struct task_numa_env *env,
1516 struct task_struct *p, long imp)
1518 struct rq *rq = cpu_rq(env->dst_cpu);
1520 /* Bail out if run-queue part of active NUMA balance. */
1521 if (xchg(&rq->numa_migrate_on, 1))
1525 * Clear previous best_cpu/rq numa-migrate flag, since task now
1526 * found a better CPU to move/swap.
1528 if (env->best_cpu != -1) {
1529 rq = cpu_rq(env->best_cpu);
1530 WRITE_ONCE(rq->numa_migrate_on, 0);
1534 put_task_struct(env->best_task);
1539 env->best_imp = imp;
1540 env->best_cpu = env->dst_cpu;
1543 static bool load_too_imbalanced(long src_load, long dst_load,
1544 struct task_numa_env *env)
1547 long orig_src_load, orig_dst_load;
1548 long src_capacity, dst_capacity;
1551 * The load is corrected for the CPU capacity available on each node.
1554 * ------------ vs ---------
1555 * src_capacity dst_capacity
1557 src_capacity = env->src_stats.compute_capacity;
1558 dst_capacity = env->dst_stats.compute_capacity;
1560 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1562 orig_src_load = env->src_stats.load;
1563 orig_dst_load = env->dst_stats.load;
1565 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1567 /* Would this change make things worse? */
1568 return (imb > old_imb);
1572 * Maximum NUMA importance can be 1998 (2*999);
1573 * SMALLIMP @ 30 would be close to 1998/64.
1574 * Used to deter task migration.
1579 * This checks if the overall compute and NUMA accesses of the system would
1580 * be improved if the source tasks was migrated to the target dst_cpu taking
1581 * into account that it might be best if task running on the dst_cpu should
1582 * be exchanged with the source task
1584 static void task_numa_compare(struct task_numa_env *env,
1585 long taskimp, long groupimp, bool maymove)
1587 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1588 struct task_struct *cur;
1589 long src_load, dst_load;
1591 long imp = env->p->numa_group ? groupimp : taskimp;
1593 int dist = env->dist;
1595 if (READ_ONCE(dst_rq->numa_migrate_on))
1599 cur = task_rcu_dereference(&dst_rq->curr);
1600 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1604 * Because we have preemption enabled we can get migrated around and
1605 * end try selecting ourselves (current == env->p) as a swap candidate.
1611 if (maymove && moveimp >= env->best_imp)
1618 * "imp" is the fault differential for the source task between the
1619 * source and destination node. Calculate the total differential for
1620 * the source task and potential destination task. The more negative
1621 * the value is, the more remote accesses that would be expected to
1622 * be incurred if the tasks were swapped.
1624 /* Skip this swap candidate if cannot move to the source cpu */
1625 if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed))
1629 * If dst and source tasks are in the same NUMA group, or not
1630 * in any group then look only at task weights.
1632 if (cur->numa_group == env->p->numa_group) {
1633 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1634 task_weight(cur, env->dst_nid, dist);
1636 * Add some hysteresis to prevent swapping the
1637 * tasks within a group over tiny differences.
1639 if (cur->numa_group)
1643 * Compare the group weights. If a task is all by itself
1644 * (not part of a group), use the task weight instead.
1646 if (cur->numa_group && env->p->numa_group)
1647 imp += group_weight(cur, env->src_nid, dist) -
1648 group_weight(cur, env->dst_nid, dist);
1650 imp += task_weight(cur, env->src_nid, dist) -
1651 task_weight(cur, env->dst_nid, dist);
1654 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1661 * If the NUMA importance is less than SMALLIMP,
1662 * task migration might only result in ping pong
1663 * of tasks and also hurt performance due to cache
1666 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1670 * In the overloaded case, try and keep the load balanced.
1672 load = task_h_load(env->p) - task_h_load(cur);
1676 dst_load = env->dst_stats.load + load;
1677 src_load = env->src_stats.load - load;
1679 if (load_too_imbalanced(src_load, dst_load, env))
1684 * One idle CPU per node is evaluated for a task numa move.
1685 * Call select_idle_sibling to maybe find a better one.
1689 * select_idle_siblings() uses an per-CPU cpumask that
1690 * can be used from IRQ context.
1692 local_irq_disable();
1693 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1698 task_numa_assign(env, cur, imp);
1703 static void task_numa_find_cpu(struct task_numa_env *env,
1704 long taskimp, long groupimp)
1706 long src_load, dst_load, load;
1707 bool maymove = false;
1710 load = task_h_load(env->p);
1711 dst_load = env->dst_stats.load + load;
1712 src_load = env->src_stats.load - load;
1715 * If the improvement from just moving env->p direction is better
1716 * than swapping tasks around, check if a move is possible.
1718 maymove = !load_too_imbalanced(src_load, dst_load, env);
1720 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1721 /* Skip this CPU if the source task cannot migrate */
1722 if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
1726 task_numa_compare(env, taskimp, groupimp, maymove);
1730 static int task_numa_migrate(struct task_struct *p)
1732 struct task_numa_env env = {
1735 .src_cpu = task_cpu(p),
1736 .src_nid = task_node(p),
1738 .imbalance_pct = 112,
1744 struct sched_domain *sd;
1746 unsigned long taskweight, groupweight;
1748 long taskimp, groupimp;
1751 * Pick the lowest SD_NUMA domain, as that would have the smallest
1752 * imbalance and would be the first to start moving tasks about.
1754 * And we want to avoid any moving of tasks about, as that would create
1755 * random movement of tasks -- counter the numa conditions we're trying
1759 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1761 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1765 * Cpusets can break the scheduler domain tree into smaller
1766 * balance domains, some of which do not cross NUMA boundaries.
1767 * Tasks that are "trapped" in such domains cannot be migrated
1768 * elsewhere, so there is no point in (re)trying.
1770 if (unlikely(!sd)) {
1771 sched_setnuma(p, task_node(p));
1775 env.dst_nid = p->numa_preferred_nid;
1776 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1777 taskweight = task_weight(p, env.src_nid, dist);
1778 groupweight = group_weight(p, env.src_nid, dist);
1779 update_numa_stats(&env.src_stats, env.src_nid);
1780 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1781 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1782 update_numa_stats(&env.dst_stats, env.dst_nid);
1784 /* Try to find a spot on the preferred nid. */
1785 task_numa_find_cpu(&env, taskimp, groupimp);
1788 * Look at other nodes in these cases:
1789 * - there is no space available on the preferred_nid
1790 * - the task is part of a numa_group that is interleaved across
1791 * multiple NUMA nodes; in order to better consolidate the group,
1792 * we need to check other locations.
1794 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1795 for_each_online_node(nid) {
1796 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1799 dist = node_distance(env.src_nid, env.dst_nid);
1800 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1802 taskweight = task_weight(p, env.src_nid, dist);
1803 groupweight = group_weight(p, env.src_nid, dist);
1806 /* Only consider nodes where both task and groups benefit */
1807 taskimp = task_weight(p, nid, dist) - taskweight;
1808 groupimp = group_weight(p, nid, dist) - groupweight;
1809 if (taskimp < 0 && groupimp < 0)
1814 update_numa_stats(&env.dst_stats, env.dst_nid);
1815 task_numa_find_cpu(&env, taskimp, groupimp);
1820 * If the task is part of a workload that spans multiple NUMA nodes,
1821 * and is migrating into one of the workload's active nodes, remember
1822 * this node as the task's preferred numa node, so the workload can
1824 * A task that migrated to a second choice node will be better off
1825 * trying for a better one later. Do not set the preferred node here.
1827 if (p->numa_group) {
1828 if (env.best_cpu == -1)
1831 nid = cpu_to_node(env.best_cpu);
1833 if (nid != p->numa_preferred_nid)
1834 sched_setnuma(p, nid);
1837 /* No better CPU than the current one was found. */
1838 if (env.best_cpu == -1)
1841 best_rq = cpu_rq(env.best_cpu);
1842 if (env.best_task == NULL) {
1843 ret = migrate_task_to(p, env.best_cpu);
1844 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1846 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1850 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
1851 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1854 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1855 put_task_struct(env.best_task);
1859 /* Attempt to migrate a task to a CPU on the preferred node. */
1860 static void numa_migrate_preferred(struct task_struct *p)
1862 unsigned long interval = HZ;
1864 /* This task has no NUMA fault statistics yet */
1865 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1868 /* Periodically retry migrating the task to the preferred node */
1869 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1870 p->numa_migrate_retry = jiffies + interval;
1872 /* Success if task is already running on preferred CPU */
1873 if (task_node(p) == p->numa_preferred_nid)
1876 /* Otherwise, try migrate to a CPU on the preferred node */
1877 task_numa_migrate(p);
1881 * Find out how many nodes on the workload is actively running on. Do this by
1882 * tracking the nodes from which NUMA hinting faults are triggered. This can
1883 * be different from the set of nodes where the workload's memory is currently
1886 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1888 unsigned long faults, max_faults = 0;
1889 int nid, active_nodes = 0;
1891 for_each_online_node(nid) {
1892 faults = group_faults_cpu(numa_group, nid);
1893 if (faults > max_faults)
1894 max_faults = faults;
1897 for_each_online_node(nid) {
1898 faults = group_faults_cpu(numa_group, nid);
1899 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1903 numa_group->max_faults_cpu = max_faults;
1904 numa_group->active_nodes = active_nodes;
1908 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1909 * increments. The more local the fault statistics are, the higher the scan
1910 * period will be for the next scan window. If local/(local+remote) ratio is
1911 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1912 * the scan period will decrease. Aim for 70% local accesses.
1914 #define NUMA_PERIOD_SLOTS 10
1915 #define NUMA_PERIOD_THRESHOLD 7
1918 * Increase the scan period (slow down scanning) if the majority of
1919 * our memory is already on our local node, or if the majority of
1920 * the page accesses are shared with other processes.
1921 * Otherwise, decrease the scan period.
1923 static void update_task_scan_period(struct task_struct *p,
1924 unsigned long shared, unsigned long private)
1926 unsigned int period_slot;
1927 int lr_ratio, ps_ratio;
1930 unsigned long remote = p->numa_faults_locality[0];
1931 unsigned long local = p->numa_faults_locality[1];
1934 * If there were no record hinting faults then either the task is
1935 * completely idle or all activity is areas that are not of interest
1936 * to automatic numa balancing. Related to that, if there were failed
1937 * migration then it implies we are migrating too quickly or the local
1938 * node is overloaded. In either case, scan slower
1940 if (local + shared == 0 || p->numa_faults_locality[2]) {
1941 p->numa_scan_period = min(p->numa_scan_period_max,
1942 p->numa_scan_period << 1);
1944 p->mm->numa_next_scan = jiffies +
1945 msecs_to_jiffies(p->numa_scan_period);
1951 * Prepare to scale scan period relative to the current period.
1952 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1953 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1954 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1956 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1957 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1958 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1960 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1962 * Most memory accesses are local. There is no need to
1963 * do fast NUMA scanning, since memory is already local.
1965 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
1968 diff = slot * period_slot;
1969 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
1971 * Most memory accesses are shared with other tasks.
1972 * There is no point in continuing fast NUMA scanning,
1973 * since other tasks may just move the memory elsewhere.
1975 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
1978 diff = slot * period_slot;
1981 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
1982 * yet they are not on the local NUMA node. Speed up
1983 * NUMA scanning to get the memory moved over.
1985 int ratio = max(lr_ratio, ps_ratio);
1986 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1989 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1990 task_scan_min(p), task_scan_max(p));
1991 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1995 * Get the fraction of time the task has been running since the last
1996 * NUMA placement cycle. The scheduler keeps similar statistics, but
1997 * decays those on a 32ms period, which is orders of magnitude off
1998 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1999 * stats only if the task is so new there are no NUMA statistics yet.
2001 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2003 u64 runtime, delta, now;
2004 /* Use the start of this time slice to avoid calculations. */
2005 now = p->se.exec_start;
2006 runtime = p->se.sum_exec_runtime;
2008 if (p->last_task_numa_placement) {
2009 delta = runtime - p->last_sum_exec_runtime;
2010 *period = now - p->last_task_numa_placement;
2012 delta = p->se.avg.load_sum;
2013 *period = LOAD_AVG_MAX;
2016 p->last_sum_exec_runtime = runtime;
2017 p->last_task_numa_placement = now;
2023 * Determine the preferred nid for a task in a numa_group. This needs to
2024 * be done in a way that produces consistent results with group_weight,
2025 * otherwise workloads might not converge.
2027 static int preferred_group_nid(struct task_struct *p, int nid)
2032 /* Direct connections between all NUMA nodes. */
2033 if (sched_numa_topology_type == NUMA_DIRECT)
2037 * On a system with glueless mesh NUMA topology, group_weight
2038 * scores nodes according to the number of NUMA hinting faults on
2039 * both the node itself, and on nearby nodes.
2041 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2042 unsigned long score, max_score = 0;
2043 int node, max_node = nid;
2045 dist = sched_max_numa_distance;
2047 for_each_online_node(node) {
2048 score = group_weight(p, node, dist);
2049 if (score > max_score) {
2058 * Finding the preferred nid in a system with NUMA backplane
2059 * interconnect topology is more involved. The goal is to locate
2060 * tasks from numa_groups near each other in the system, and
2061 * untangle workloads from different sides of the system. This requires
2062 * searching down the hierarchy of node groups, recursively searching
2063 * inside the highest scoring group of nodes. The nodemask tricks
2064 * keep the complexity of the search down.
2066 nodes = node_online_map;
2067 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2068 unsigned long max_faults = 0;
2069 nodemask_t max_group = NODE_MASK_NONE;
2072 /* Are there nodes at this distance from each other? */
2073 if (!find_numa_distance(dist))
2076 for_each_node_mask(a, nodes) {
2077 unsigned long faults = 0;
2078 nodemask_t this_group;
2079 nodes_clear(this_group);
2081 /* Sum group's NUMA faults; includes a==b case. */
2082 for_each_node_mask(b, nodes) {
2083 if (node_distance(a, b) < dist) {
2084 faults += group_faults(p, b);
2085 node_set(b, this_group);
2086 node_clear(b, nodes);
2090 /* Remember the top group. */
2091 if (faults > max_faults) {
2092 max_faults = faults;
2093 max_group = this_group;
2095 * subtle: at the smallest distance there is
2096 * just one node left in each "group", the
2097 * winner is the preferred nid.
2102 /* Next round, evaluate the nodes within max_group. */
2110 static void task_numa_placement(struct task_struct *p)
2112 int seq, nid, max_nid = -1;
2113 unsigned long max_faults = 0;
2114 unsigned long fault_types[2] = { 0, 0 };
2115 unsigned long total_faults;
2116 u64 runtime, period;
2117 spinlock_t *group_lock = NULL;
2120 * The p->mm->numa_scan_seq field gets updated without
2121 * exclusive access. Use READ_ONCE() here to ensure
2122 * that the field is read in a single access:
2124 seq = READ_ONCE(p->mm->numa_scan_seq);
2125 if (p->numa_scan_seq == seq)
2127 p->numa_scan_seq = seq;
2128 p->numa_scan_period_max = task_scan_max(p);
2130 total_faults = p->numa_faults_locality[0] +
2131 p->numa_faults_locality[1];
2132 runtime = numa_get_avg_runtime(p, &period);
2134 /* If the task is part of a group prevent parallel updates to group stats */
2135 if (p->numa_group) {
2136 group_lock = &p->numa_group->lock;
2137 spin_lock_irq(group_lock);
2140 /* Find the node with the highest number of faults */
2141 for_each_online_node(nid) {
2142 /* Keep track of the offsets in numa_faults array */
2143 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2144 unsigned long faults = 0, group_faults = 0;
2147 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2148 long diff, f_diff, f_weight;
2150 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2151 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2152 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2153 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2155 /* Decay existing window, copy faults since last scan */
2156 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2157 fault_types[priv] += p->numa_faults[membuf_idx];
2158 p->numa_faults[membuf_idx] = 0;
2161 * Normalize the faults_from, so all tasks in a group
2162 * count according to CPU use, instead of by the raw
2163 * number of faults. Tasks with little runtime have
2164 * little over-all impact on throughput, and thus their
2165 * faults are less important.
2167 f_weight = div64_u64(runtime << 16, period + 1);
2168 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2170 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2171 p->numa_faults[cpubuf_idx] = 0;
2173 p->numa_faults[mem_idx] += diff;
2174 p->numa_faults[cpu_idx] += f_diff;
2175 faults += p->numa_faults[mem_idx];
2176 p->total_numa_faults += diff;
2177 if (p->numa_group) {
2179 * safe because we can only change our own group
2181 * mem_idx represents the offset for a given
2182 * nid and priv in a specific region because it
2183 * is at the beginning of the numa_faults array.
2185 p->numa_group->faults[mem_idx] += diff;
2186 p->numa_group->faults_cpu[mem_idx] += f_diff;
2187 p->numa_group->total_faults += diff;
2188 group_faults += p->numa_group->faults[mem_idx];
2192 if (!p->numa_group) {
2193 if (faults > max_faults) {
2194 max_faults = faults;
2197 } else if (group_faults > max_faults) {
2198 max_faults = group_faults;
2203 if (p->numa_group) {
2204 numa_group_count_active_nodes(p->numa_group);
2205 spin_unlock_irq(group_lock);
2206 max_nid = preferred_group_nid(p, max_nid);
2210 /* Set the new preferred node */
2211 if (max_nid != p->numa_preferred_nid)
2212 sched_setnuma(p, max_nid);
2215 update_task_scan_period(p, fault_types[0], fault_types[1]);
2218 static inline int get_numa_group(struct numa_group *grp)
2220 return refcount_inc_not_zero(&grp->refcount);
2223 static inline void put_numa_group(struct numa_group *grp)
2225 if (refcount_dec_and_test(&grp->refcount))
2226 kfree_rcu(grp, rcu);
2229 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2232 struct numa_group *grp, *my_grp;
2233 struct task_struct *tsk;
2235 int cpu = cpupid_to_cpu(cpupid);
2238 if (unlikely(!p->numa_group)) {
2239 unsigned int size = sizeof(struct numa_group) +
2240 4*nr_node_ids*sizeof(unsigned long);
2242 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2246 refcount_set(&grp->refcount, 1);
2247 grp->active_nodes = 1;
2248 grp->max_faults_cpu = 0;
2249 spin_lock_init(&grp->lock);
2251 /* Second half of the array tracks nids where faults happen */
2252 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2255 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2256 grp->faults[i] = p->numa_faults[i];
2258 grp->total_faults = p->total_numa_faults;
2261 rcu_assign_pointer(p->numa_group, grp);
2265 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2267 if (!cpupid_match_pid(tsk, cpupid))
2270 grp = rcu_dereference(tsk->numa_group);
2274 my_grp = p->numa_group;
2279 * Only join the other group if its bigger; if we're the bigger group,
2280 * the other task will join us.
2282 if (my_grp->nr_tasks > grp->nr_tasks)
2286 * Tie-break on the grp address.
2288 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2291 /* Always join threads in the same process. */
2292 if (tsk->mm == current->mm)
2295 /* Simple filter to avoid false positives due to PID collisions */
2296 if (flags & TNF_SHARED)
2299 /* Update priv based on whether false sharing was detected */
2302 if (join && !get_numa_group(grp))
2310 BUG_ON(irqs_disabled());
2311 double_lock_irq(&my_grp->lock, &grp->lock);
2313 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2314 my_grp->faults[i] -= p->numa_faults[i];
2315 grp->faults[i] += p->numa_faults[i];
2317 my_grp->total_faults -= p->total_numa_faults;
2318 grp->total_faults += p->total_numa_faults;
2323 spin_unlock(&my_grp->lock);
2324 spin_unlock_irq(&grp->lock);
2326 rcu_assign_pointer(p->numa_group, grp);
2328 put_numa_group(my_grp);
2336 void task_numa_free(struct task_struct *p)
2338 struct numa_group *grp = p->numa_group;
2339 void *numa_faults = p->numa_faults;
2340 unsigned long flags;
2344 spin_lock_irqsave(&grp->lock, flags);
2345 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2346 grp->faults[i] -= p->numa_faults[i];
2347 grp->total_faults -= p->total_numa_faults;
2350 spin_unlock_irqrestore(&grp->lock, flags);
2351 RCU_INIT_POINTER(p->numa_group, NULL);
2352 put_numa_group(grp);
2355 p->numa_faults = NULL;
2360 * Got a PROT_NONE fault for a page on @node.
2362 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2364 struct task_struct *p = current;
2365 bool migrated = flags & TNF_MIGRATED;
2366 int cpu_node = task_node(current);
2367 int local = !!(flags & TNF_FAULT_LOCAL);
2368 struct numa_group *ng;
2371 if (!static_branch_likely(&sched_numa_balancing))
2374 /* for example, ksmd faulting in a user's mm */
2378 /* Allocate buffer to track faults on a per-node basis */
2379 if (unlikely(!p->numa_faults)) {
2380 int size = sizeof(*p->numa_faults) *
2381 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2383 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2384 if (!p->numa_faults)
2387 p->total_numa_faults = 0;
2388 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2392 * First accesses are treated as private, otherwise consider accesses
2393 * to be private if the accessing pid has not changed
2395 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2398 priv = cpupid_match_pid(p, last_cpupid);
2399 if (!priv && !(flags & TNF_NO_GROUP))
2400 task_numa_group(p, last_cpupid, flags, &priv);
2404 * If a workload spans multiple NUMA nodes, a shared fault that
2405 * occurs wholly within the set of nodes that the workload is
2406 * actively using should be counted as local. This allows the
2407 * scan rate to slow down when a workload has settled down.
2410 if (!priv && !local && ng && ng->active_nodes > 1 &&
2411 numa_is_active_node(cpu_node, ng) &&
2412 numa_is_active_node(mem_node, ng))
2416 * Retry to migrate task to preferred node periodically, in case it
2417 * previously failed, or the scheduler moved us.
2419 if (time_after(jiffies, p->numa_migrate_retry)) {
2420 task_numa_placement(p);
2421 numa_migrate_preferred(p);
2425 p->numa_pages_migrated += pages;
2426 if (flags & TNF_MIGRATE_FAIL)
2427 p->numa_faults_locality[2] += pages;
2429 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2430 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2431 p->numa_faults_locality[local] += pages;
2434 static void reset_ptenuma_scan(struct task_struct *p)
2437 * We only did a read acquisition of the mmap sem, so
2438 * p->mm->numa_scan_seq is written to without exclusive access
2439 * and the update is not guaranteed to be atomic. That's not
2440 * much of an issue though, since this is just used for
2441 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2442 * expensive, to avoid any form of compiler optimizations:
2444 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2445 p->mm->numa_scan_offset = 0;
2449 * The expensive part of numa migration is done from task_work context.
2450 * Triggered from task_tick_numa().
2452 void task_numa_work(struct callback_head *work)
2454 unsigned long migrate, next_scan, now = jiffies;
2455 struct task_struct *p = current;
2456 struct mm_struct *mm = p->mm;
2457 u64 runtime = p->se.sum_exec_runtime;
2458 struct vm_area_struct *vma;
2459 unsigned long start, end;
2460 unsigned long nr_pte_updates = 0;
2461 long pages, virtpages;
2463 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2465 work->next = work; /* protect against double add */
2467 * Who cares about NUMA placement when they're dying.
2469 * NOTE: make sure not to dereference p->mm before this check,
2470 * exit_task_work() happens _after_ exit_mm() so we could be called
2471 * without p->mm even though we still had it when we enqueued this
2474 if (p->flags & PF_EXITING)
2477 if (!mm->numa_next_scan) {
2478 mm->numa_next_scan = now +
2479 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2483 * Enforce maximal scan/migration frequency..
2485 migrate = mm->numa_next_scan;
2486 if (time_before(now, migrate))
2489 if (p->numa_scan_period == 0) {
2490 p->numa_scan_period_max = task_scan_max(p);
2491 p->numa_scan_period = task_scan_start(p);
2494 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2495 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2499 * Delay this task enough that another task of this mm will likely win
2500 * the next time around.
2502 p->node_stamp += 2 * TICK_NSEC;
2504 start = mm->numa_scan_offset;
2505 pages = sysctl_numa_balancing_scan_size;
2506 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2507 virtpages = pages * 8; /* Scan up to this much virtual space */
2512 if (!down_read_trylock(&mm->mmap_sem))
2514 vma = find_vma(mm, start);
2516 reset_ptenuma_scan(p);
2520 for (; vma; vma = vma->vm_next) {
2521 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2522 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2527 * Shared library pages mapped by multiple processes are not
2528 * migrated as it is expected they are cache replicated. Avoid
2529 * hinting faults in read-only file-backed mappings or the vdso
2530 * as migrating the pages will be of marginal benefit.
2533 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2537 * Skip inaccessible VMAs to avoid any confusion between
2538 * PROT_NONE and NUMA hinting ptes
2540 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2544 start = max(start, vma->vm_start);
2545 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2546 end = min(end, vma->vm_end);
2547 nr_pte_updates = change_prot_numa(vma, start, end);
2550 * Try to scan sysctl_numa_balancing_size worth of
2551 * hpages that have at least one present PTE that
2552 * is not already pte-numa. If the VMA contains
2553 * areas that are unused or already full of prot_numa
2554 * PTEs, scan up to virtpages, to skip through those
2558 pages -= (end - start) >> PAGE_SHIFT;
2559 virtpages -= (end - start) >> PAGE_SHIFT;
2562 if (pages <= 0 || virtpages <= 0)
2566 } while (end != vma->vm_end);
2571 * It is possible to reach the end of the VMA list but the last few
2572 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2573 * would find the !migratable VMA on the next scan but not reset the
2574 * scanner to the start so check it now.
2577 mm->numa_scan_offset = start;
2579 reset_ptenuma_scan(p);
2580 up_read(&mm->mmap_sem);
2583 * Make sure tasks use at least 32x as much time to run other code
2584 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2585 * Usually update_task_scan_period slows down scanning enough; on an
2586 * overloaded system we need to limit overhead on a per task basis.
2588 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2589 u64 diff = p->se.sum_exec_runtime - runtime;
2590 p->node_stamp += 32 * diff;
2595 * Drive the periodic memory faults..
2597 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2599 struct callback_head *work = &curr->numa_work;
2603 * We don't care about NUMA placement if we don't have memory.
2605 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2609 * Using runtime rather than walltime has the dual advantage that
2610 * we (mostly) drive the selection from busy threads and that the
2611 * task needs to have done some actual work before we bother with
2614 now = curr->se.sum_exec_runtime;
2615 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2617 if (now > curr->node_stamp + period) {
2618 if (!curr->node_stamp)
2619 curr->numa_scan_period = task_scan_start(curr);
2620 curr->node_stamp += period;
2622 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2623 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2624 task_work_add(curr, work, true);
2629 static void update_scan_period(struct task_struct *p, int new_cpu)
2631 int src_nid = cpu_to_node(task_cpu(p));
2632 int dst_nid = cpu_to_node(new_cpu);
2634 if (!static_branch_likely(&sched_numa_balancing))
2637 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2640 if (src_nid == dst_nid)
2644 * Allow resets if faults have been trapped before one scan
2645 * has completed. This is most likely due to a new task that
2646 * is pulled cross-node due to wakeups or load balancing.
2648 if (p->numa_scan_seq) {
2650 * Avoid scan adjustments if moving to the preferred
2651 * node or if the task was not previously running on
2652 * the preferred node.
2654 if (dst_nid == p->numa_preferred_nid ||
2655 (p->numa_preferred_nid != -1 && src_nid != p->numa_preferred_nid))
2659 p->numa_scan_period = task_scan_start(p);
2663 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2667 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2671 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2675 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2679 #endif /* CONFIG_NUMA_BALANCING */
2682 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2684 update_load_add(&cfs_rq->load, se->load.weight);
2685 if (!parent_entity(se))
2686 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2688 if (entity_is_task(se)) {
2689 struct rq *rq = rq_of(cfs_rq);
2691 account_numa_enqueue(rq, task_of(se));
2692 list_add(&se->group_node, &rq->cfs_tasks);
2695 cfs_rq->nr_running++;
2699 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2701 update_load_sub(&cfs_rq->load, se->load.weight);
2702 if (!parent_entity(se))
2703 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2705 if (entity_is_task(se)) {
2706 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2707 list_del_init(&se->group_node);
2710 cfs_rq->nr_running--;
2714 * Signed add and clamp on underflow.
2716 * Explicitly do a load-store to ensure the intermediate value never hits
2717 * memory. This allows lockless observations without ever seeing the negative
2720 #define add_positive(_ptr, _val) do { \
2721 typeof(_ptr) ptr = (_ptr); \
2722 typeof(_val) val = (_val); \
2723 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2727 if (val < 0 && res > var) \
2730 WRITE_ONCE(*ptr, res); \
2734 * Unsigned subtract and clamp on underflow.
2736 * Explicitly do a load-store to ensure the intermediate value never hits
2737 * memory. This allows lockless observations without ever seeing the negative
2740 #define sub_positive(_ptr, _val) do { \
2741 typeof(_ptr) ptr = (_ptr); \
2742 typeof(*ptr) val = (_val); \
2743 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2747 WRITE_ONCE(*ptr, res); \
2751 * Remove and clamp on negative, from a local variable.
2753 * A variant of sub_positive(), which does not use explicit load-store
2754 * and is thus optimized for local variable updates.
2756 #define lsub_positive(_ptr, _val) do { \
2757 typeof(_ptr) ptr = (_ptr); \
2758 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
2763 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2765 cfs_rq->runnable_weight += se->runnable_weight;
2767 cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2768 cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2772 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2774 cfs_rq->runnable_weight -= se->runnable_weight;
2776 sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2777 sub_positive(&cfs_rq->avg.runnable_load_sum,
2778 se_runnable(se) * se->avg.runnable_load_sum);
2782 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2784 cfs_rq->avg.load_avg += se->avg.load_avg;
2785 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2789 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2791 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2792 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2796 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2798 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2800 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2802 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2805 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2806 unsigned long weight, unsigned long runnable)
2809 /* commit outstanding execution time */
2810 if (cfs_rq->curr == se)
2811 update_curr(cfs_rq);
2812 account_entity_dequeue(cfs_rq, se);
2813 dequeue_runnable_load_avg(cfs_rq, se);
2815 dequeue_load_avg(cfs_rq, se);
2817 se->runnable_weight = runnable;
2818 update_load_set(&se->load, weight);
2822 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2824 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2825 se->avg.runnable_load_avg =
2826 div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2830 enqueue_load_avg(cfs_rq, se);
2832 account_entity_enqueue(cfs_rq, se);
2833 enqueue_runnable_load_avg(cfs_rq, se);
2837 void reweight_task(struct task_struct *p, int prio)
2839 struct sched_entity *se = &p->se;
2840 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2841 struct load_weight *load = &se->load;
2842 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2844 reweight_entity(cfs_rq, se, weight, weight);
2845 load->inv_weight = sched_prio_to_wmult[prio];
2848 #ifdef CONFIG_FAIR_GROUP_SCHED
2851 * All this does is approximate the hierarchical proportion which includes that
2852 * global sum we all love to hate.
2854 * That is, the weight of a group entity, is the proportional share of the
2855 * group weight based on the group runqueue weights. That is:
2857 * tg->weight * grq->load.weight
2858 * ge->load.weight = ----------------------------- (1)
2859 * \Sum grq->load.weight
2861 * Now, because computing that sum is prohibitively expensive to compute (been
2862 * there, done that) we approximate it with this average stuff. The average
2863 * moves slower and therefore the approximation is cheaper and more stable.
2865 * So instead of the above, we substitute:
2867 * grq->load.weight -> grq->avg.load_avg (2)
2869 * which yields the following:
2871 * tg->weight * grq->avg.load_avg
2872 * ge->load.weight = ------------------------------ (3)
2875 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2877 * That is shares_avg, and it is right (given the approximation (2)).
2879 * The problem with it is that because the average is slow -- it was designed
2880 * to be exactly that of course -- this leads to transients in boundary
2881 * conditions. In specific, the case where the group was idle and we start the
2882 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2883 * yielding bad latency etc..
2885 * Now, in that special case (1) reduces to:
2887 * tg->weight * grq->load.weight
2888 * ge->load.weight = ----------------------------- = tg->weight (4)
2891 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2893 * So what we do is modify our approximation (3) to approach (4) in the (near)
2898 * tg->weight * grq->load.weight
2899 * --------------------------------------------------- (5)
2900 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2902 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2903 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2906 * tg->weight * grq->load.weight
2907 * ge->load.weight = ----------------------------- (6)
2912 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2913 * max(grq->load.weight, grq->avg.load_avg)
2915 * And that is shares_weight and is icky. In the (near) UP case it approaches
2916 * (4) while in the normal case it approaches (3). It consistently
2917 * overestimates the ge->load.weight and therefore:
2919 * \Sum ge->load.weight >= tg->weight
2923 static long calc_group_shares(struct cfs_rq *cfs_rq)
2925 long tg_weight, tg_shares, load, shares;
2926 struct task_group *tg = cfs_rq->tg;
2928 tg_shares = READ_ONCE(tg->shares);
2930 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
2932 tg_weight = atomic_long_read(&tg->load_avg);
2934 /* Ensure tg_weight >= load */
2935 tg_weight -= cfs_rq->tg_load_avg_contrib;
2938 shares = (tg_shares * load);
2940 shares /= tg_weight;
2943 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2944 * of a group with small tg->shares value. It is a floor value which is
2945 * assigned as a minimum load.weight to the sched_entity representing
2946 * the group on a CPU.
2948 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2949 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2950 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2951 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2954 return clamp_t(long, shares, MIN_SHARES, tg_shares);
2958 * This calculates the effective runnable weight for a group entity based on
2959 * the group entity weight calculated above.
2961 * Because of the above approximation (2), our group entity weight is
2962 * an load_avg based ratio (3). This means that it includes blocked load and
2963 * does not represent the runnable weight.
2965 * Approximate the group entity's runnable weight per ratio from the group
2968 * grq->avg.runnable_load_avg
2969 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
2972 * However, analogous to above, since the avg numbers are slow, this leads to
2973 * transients in the from-idle case. Instead we use:
2975 * ge->runnable_weight = ge->load.weight *
2977 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
2978 * ----------------------------------------------------- (8)
2979 * max(grq->avg.load_avg, grq->load.weight)
2981 * Where these max() serve both to use the 'instant' values to fix the slow
2982 * from-idle and avoid the /0 on to-idle, similar to (6).
2984 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
2986 long runnable, load_avg;
2988 load_avg = max(cfs_rq->avg.load_avg,
2989 scale_load_down(cfs_rq->load.weight));
2991 runnable = max(cfs_rq->avg.runnable_load_avg,
2992 scale_load_down(cfs_rq->runnable_weight));
2996 runnable /= load_avg;
2998 return clamp_t(long, runnable, MIN_SHARES, shares);
3000 #endif /* CONFIG_SMP */
3002 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3005 * Recomputes the group entity based on the current state of its group
3008 static void update_cfs_group(struct sched_entity *se)
3010 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3011 long shares, runnable;
3016 if (throttled_hierarchy(gcfs_rq))
3020 runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
3022 if (likely(se->load.weight == shares))
3025 shares = calc_group_shares(gcfs_rq);
3026 runnable = calc_group_runnable(gcfs_rq, shares);
3029 reweight_entity(cfs_rq_of(se), se, shares, runnable);
3032 #else /* CONFIG_FAIR_GROUP_SCHED */
3033 static inline void update_cfs_group(struct sched_entity *se)
3036 #endif /* CONFIG_FAIR_GROUP_SCHED */
3038 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3040 struct rq *rq = rq_of(cfs_rq);
3042 if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) {
3044 * There are a few boundary cases this might miss but it should
3045 * get called often enough that that should (hopefully) not be
3048 * It will not get called when we go idle, because the idle
3049 * thread is a different class (!fair), nor will the utilization
3050 * number include things like RT tasks.
3052 * As is, the util number is not freq-invariant (we'd have to
3053 * implement arch_scale_freq_capacity() for that).
3057 cpufreq_update_util(rq, flags);
3062 #ifdef CONFIG_FAIR_GROUP_SCHED
3064 * update_tg_load_avg - update the tg's load avg
3065 * @cfs_rq: the cfs_rq whose avg changed
3066 * @force: update regardless of how small the difference
3068 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3069 * However, because tg->load_avg is a global value there are performance
3072 * In order to avoid having to look at the other cfs_rq's, we use a
3073 * differential update where we store the last value we propagated. This in
3074 * turn allows skipping updates if the differential is 'small'.
3076 * Updating tg's load_avg is necessary before update_cfs_share().
3078 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3080 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3083 * No need to update load_avg for root_task_group as it is not used.
3085 if (cfs_rq->tg == &root_task_group)
3088 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3089 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3090 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3095 * Called within set_task_rq() right before setting a task's CPU. The
3096 * caller only guarantees p->pi_lock is held; no other assumptions,
3097 * including the state of rq->lock, should be made.
3099 void set_task_rq_fair(struct sched_entity *se,
3100 struct cfs_rq *prev, struct cfs_rq *next)
3102 u64 p_last_update_time;
3103 u64 n_last_update_time;
3105 if (!sched_feat(ATTACH_AGE_LOAD))
3109 * We are supposed to update the task to "current" time, then its up to
3110 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3111 * getting what current time is, so simply throw away the out-of-date
3112 * time. This will result in the wakee task is less decayed, but giving
3113 * the wakee more load sounds not bad.
3115 if (!(se->avg.last_update_time && prev))
3118 #ifndef CONFIG_64BIT
3120 u64 p_last_update_time_copy;
3121 u64 n_last_update_time_copy;
3124 p_last_update_time_copy = prev->load_last_update_time_copy;
3125 n_last_update_time_copy = next->load_last_update_time_copy;
3129 p_last_update_time = prev->avg.last_update_time;
3130 n_last_update_time = next->avg.last_update_time;
3132 } while (p_last_update_time != p_last_update_time_copy ||
3133 n_last_update_time != n_last_update_time_copy);
3136 p_last_update_time = prev->avg.last_update_time;
3137 n_last_update_time = next->avg.last_update_time;
3139 __update_load_avg_blocked_se(p_last_update_time, se);
3140 se->avg.last_update_time = n_last_update_time;
3145 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3146 * propagate its contribution. The key to this propagation is the invariant
3147 * that for each group:
3149 * ge->avg == grq->avg (1)
3151 * _IFF_ we look at the pure running and runnable sums. Because they
3152 * represent the very same entity, just at different points in the hierarchy.
3154 * Per the above update_tg_cfs_util() is trivial and simply copies the running
3155 * sum over (but still wrong, because the group entity and group rq do not have
3156 * their PELT windows aligned).
3158 * However, update_tg_cfs_runnable() is more complex. So we have:
3160 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3162 * And since, like util, the runnable part should be directly transferable,
3163 * the following would _appear_ to be the straight forward approach:
3165 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3167 * And per (1) we have:
3169 * ge->avg.runnable_avg == grq->avg.runnable_avg
3173 * ge->load.weight * grq->avg.load_avg
3174 * ge->avg.load_avg = ----------------------------------- (4)
3177 * Except that is wrong!
3179 * Because while for entities historical weight is not important and we
3180 * really only care about our future and therefore can consider a pure
3181 * runnable sum, runqueues can NOT do this.
3183 * We specifically want runqueues to have a load_avg that includes
3184 * historical weights. Those represent the blocked load, the load we expect
3185 * to (shortly) return to us. This only works by keeping the weights as
3186 * integral part of the sum. We therefore cannot decompose as per (3).
3188 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3189 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3190 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3191 * runnable section of these tasks overlap (or not). If they were to perfectly
3192 * align the rq as a whole would be runnable 2/3 of the time. If however we
3193 * always have at least 1 runnable task, the rq as a whole is always runnable.
3195 * So we'll have to approximate.. :/
3197 * Given the constraint:
3199 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3201 * We can construct a rule that adds runnable to a rq by assuming minimal
3204 * On removal, we'll assume each task is equally runnable; which yields:
3206 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3208 * XXX: only do this for the part of runnable > running ?
3213 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3215 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3217 /* Nothing to update */
3222 * The relation between sum and avg is:
3224 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3226 * however, the PELT windows are not aligned between grq and gse.
3229 /* Set new sched_entity's utilization */
3230 se->avg.util_avg = gcfs_rq->avg.util_avg;
3231 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3233 /* Update parent cfs_rq utilization */
3234 add_positive(&cfs_rq->avg.util_avg, delta);
3235 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3239 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3241 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3242 unsigned long runnable_load_avg, load_avg;
3243 u64 runnable_load_sum, load_sum = 0;
3249 gcfs_rq->prop_runnable_sum = 0;
3251 if (runnable_sum >= 0) {
3253 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3254 * the CPU is saturated running == runnable.
3256 runnable_sum += se->avg.load_sum;
3257 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3260 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3261 * assuming all tasks are equally runnable.
3263 if (scale_load_down(gcfs_rq->load.weight)) {
3264 load_sum = div_s64(gcfs_rq->avg.load_sum,
3265 scale_load_down(gcfs_rq->load.weight));
3268 /* But make sure to not inflate se's runnable */
3269 runnable_sum = min(se->avg.load_sum, load_sum);
3273 * runnable_sum can't be lower than running_sum
3274 * Rescale running sum to be in the same range as runnable sum
3275 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3276 * runnable_sum is in [0 : LOAD_AVG_MAX]
3278 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3279 runnable_sum = max(runnable_sum, running_sum);
3281 load_sum = (s64)se_weight(se) * runnable_sum;
3282 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3284 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3285 delta_avg = load_avg - se->avg.load_avg;
3287 se->avg.load_sum = runnable_sum;
3288 se->avg.load_avg = load_avg;
3289 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3290 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3292 runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3293 runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3294 delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3295 delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3297 se->avg.runnable_load_sum = runnable_sum;
3298 se->avg.runnable_load_avg = runnable_load_avg;
3301 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3302 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3306 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3308 cfs_rq->propagate = 1;
3309 cfs_rq->prop_runnable_sum += runnable_sum;
3312 /* Update task and its cfs_rq load average */
3313 static inline int propagate_entity_load_avg(struct sched_entity *se)
3315 struct cfs_rq *cfs_rq, *gcfs_rq;
3317 if (entity_is_task(se))
3320 gcfs_rq = group_cfs_rq(se);
3321 if (!gcfs_rq->propagate)
3324 gcfs_rq->propagate = 0;
3326 cfs_rq = cfs_rq_of(se);
3328 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3330 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3331 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3337 * Check if we need to update the load and the utilization of a blocked
3340 static inline bool skip_blocked_update(struct sched_entity *se)
3342 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3345 * If sched_entity still have not zero load or utilization, we have to
3348 if (se->avg.load_avg || se->avg.util_avg)
3352 * If there is a pending propagation, we have to update the load and
3353 * the utilization of the sched_entity:
3355 if (gcfs_rq->propagate)
3359 * Otherwise, the load and the utilization of the sched_entity is
3360 * already zero and there is no pending propagation, so it will be a
3361 * waste of time to try to decay it:
3366 #else /* CONFIG_FAIR_GROUP_SCHED */
3368 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3370 static inline int propagate_entity_load_avg(struct sched_entity *se)
3375 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3377 #endif /* CONFIG_FAIR_GROUP_SCHED */
3380 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3381 * @now: current time, as per cfs_rq_clock_pelt()
3382 * @cfs_rq: cfs_rq to update
3384 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3385 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3386 * post_init_entity_util_avg().
3388 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3390 * Returns true if the load decayed or we removed load.
3392 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3393 * call update_tg_load_avg() when this function returns true.
3396 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3398 unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3399 struct sched_avg *sa = &cfs_rq->avg;
3402 if (cfs_rq->removed.nr) {
3404 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3406 raw_spin_lock(&cfs_rq->removed.lock);
3407 swap(cfs_rq->removed.util_avg, removed_util);
3408 swap(cfs_rq->removed.load_avg, removed_load);
3409 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3410 cfs_rq->removed.nr = 0;
3411 raw_spin_unlock(&cfs_rq->removed.lock);
3414 sub_positive(&sa->load_avg, r);
3415 sub_positive(&sa->load_sum, r * divider);
3418 sub_positive(&sa->util_avg, r);
3419 sub_positive(&sa->util_sum, r * divider);
3421 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3426 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
3428 #ifndef CONFIG_64BIT
3430 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3434 cfs_rq_util_change(cfs_rq, 0);
3440 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3441 * @cfs_rq: cfs_rq to attach to
3442 * @se: sched_entity to attach
3443 * @flags: migration hints
3445 * Must call update_cfs_rq_load_avg() before this, since we rely on
3446 * cfs_rq->avg.last_update_time being current.
3448 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3450 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3453 * When we attach the @se to the @cfs_rq, we must align the decay
3454 * window because without that, really weird and wonderful things can
3459 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3460 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3463 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3464 * period_contrib. This isn't strictly correct, but since we're
3465 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3468 se->avg.util_sum = se->avg.util_avg * divider;
3470 se->avg.load_sum = divider;
3471 if (se_weight(se)) {
3473 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3476 se->avg.runnable_load_sum = se->avg.load_sum;
3478 enqueue_load_avg(cfs_rq, se);
3479 cfs_rq->avg.util_avg += se->avg.util_avg;
3480 cfs_rq->avg.util_sum += se->avg.util_sum;
3482 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3484 cfs_rq_util_change(cfs_rq, flags);
3488 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3489 * @cfs_rq: cfs_rq to detach from
3490 * @se: sched_entity to detach
3492 * Must call update_cfs_rq_load_avg() before this, since we rely on
3493 * cfs_rq->avg.last_update_time being current.
3495 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3497 dequeue_load_avg(cfs_rq, se);
3498 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3499 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3501 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3503 cfs_rq_util_change(cfs_rq, 0);
3507 * Optional action to be done while updating the load average
3509 #define UPDATE_TG 0x1
3510 #define SKIP_AGE_LOAD 0x2
3511 #define DO_ATTACH 0x4
3513 /* Update task and its cfs_rq load average */
3514 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3516 u64 now = cfs_rq_clock_pelt(cfs_rq);
3520 * Track task load average for carrying it to new CPU after migrated, and
3521 * track group sched_entity load average for task_h_load calc in migration
3523 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3524 __update_load_avg_se(now, cfs_rq, se);
3526 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3527 decayed |= propagate_entity_load_avg(se);
3529 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3532 * DO_ATTACH means we're here from enqueue_entity().
3533 * !last_update_time means we've passed through
3534 * migrate_task_rq_fair() indicating we migrated.
3536 * IOW we're enqueueing a task on a new CPU.
3538 attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
3539 update_tg_load_avg(cfs_rq, 0);
3541 } else if (decayed && (flags & UPDATE_TG))
3542 update_tg_load_avg(cfs_rq, 0);
3545 #ifndef CONFIG_64BIT
3546 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3548 u64 last_update_time_copy;
3549 u64 last_update_time;
3552 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3554 last_update_time = cfs_rq->avg.last_update_time;
3555 } while (last_update_time != last_update_time_copy);
3557 return last_update_time;
3560 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3562 return cfs_rq->avg.last_update_time;
3567 * Synchronize entity load avg of dequeued entity without locking
3570 void sync_entity_load_avg(struct sched_entity *se)
3572 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3573 u64 last_update_time;
3575 last_update_time = cfs_rq_last_update_time(cfs_rq);
3576 __update_load_avg_blocked_se(last_update_time, se);
3580 * Task first catches up with cfs_rq, and then subtract
3581 * itself from the cfs_rq (task must be off the queue now).
3583 void remove_entity_load_avg(struct sched_entity *se)
3585 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3586 unsigned long flags;
3589 * tasks cannot exit without having gone through wake_up_new_task() ->
3590 * post_init_entity_util_avg() which will have added things to the
3591 * cfs_rq, so we can remove unconditionally.
3593 * Similarly for groups, they will have passed through
3594 * post_init_entity_util_avg() before unregister_sched_fair_group()
3598 sync_entity_load_avg(se);
3600 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3601 ++cfs_rq->removed.nr;
3602 cfs_rq->removed.util_avg += se->avg.util_avg;
3603 cfs_rq->removed.load_avg += se->avg.load_avg;
3604 cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */
3605 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3608 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3610 return cfs_rq->avg.runnable_load_avg;
3613 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3615 return cfs_rq->avg.load_avg;
3618 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3620 static inline unsigned long task_util(struct task_struct *p)
3622 return READ_ONCE(p->se.avg.util_avg);
3625 static inline unsigned long _task_util_est(struct task_struct *p)
3627 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3629 return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED);
3632 static inline unsigned long task_util_est(struct task_struct *p)
3634 return max(task_util(p), _task_util_est(p));
3637 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3638 struct task_struct *p)
3640 unsigned int enqueued;
3642 if (!sched_feat(UTIL_EST))
3645 /* Update root cfs_rq's estimated utilization */
3646 enqueued = cfs_rq->avg.util_est.enqueued;
3647 enqueued += _task_util_est(p);
3648 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3652 * Check if a (signed) value is within a specified (unsigned) margin,
3653 * based on the observation that:
3655 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3657 * NOTE: this only works when value + maring < INT_MAX.
3659 static inline bool within_margin(int value, int margin)
3661 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3665 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3667 long last_ewma_diff;
3671 if (!sched_feat(UTIL_EST))
3674 /* Update root cfs_rq's estimated utilization */
3675 ue.enqueued = cfs_rq->avg.util_est.enqueued;
3676 ue.enqueued -= min_t(unsigned int, ue.enqueued, _task_util_est(p));
3677 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3680 * Skip update of task's estimated utilization when the task has not
3681 * yet completed an activation, e.g. being migrated.
3687 * If the PELT values haven't changed since enqueue time,
3688 * skip the util_est update.
3690 ue = p->se.avg.util_est;
3691 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3695 * Skip update of task's estimated utilization when its EWMA is
3696 * already ~1% close to its last activation value.
3698 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3699 last_ewma_diff = ue.enqueued - ue.ewma;
3700 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3704 * To avoid overestimation of actual task utilization, skip updates if
3705 * we cannot grant there is idle time in this CPU.
3707 cpu = cpu_of(rq_of(cfs_rq));
3708 if (task_util(p) > capacity_orig_of(cpu))
3712 * Update Task's estimated utilization
3714 * When *p completes an activation we can consolidate another sample
3715 * of the task size. This is done by storing the current PELT value
3716 * as ue.enqueued and by using this value to update the Exponential
3717 * Weighted Moving Average (EWMA):
3719 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
3720 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
3721 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
3722 * = w * ( last_ewma_diff ) + ewma(t-1)
3723 * = w * (last_ewma_diff + ewma(t-1) / w)
3725 * Where 'w' is the weight of new samples, which is configured to be
3726 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3728 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
3729 ue.ewma += last_ewma_diff;
3730 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
3731 WRITE_ONCE(p->se.avg.util_est, ue);
3734 static inline int task_fits_capacity(struct task_struct *p, long capacity)
3736 return capacity * 1024 > task_util_est(p) * capacity_margin;
3739 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
3741 if (!static_branch_unlikely(&sched_asym_cpucapacity))
3745 rq->misfit_task_load = 0;
3749 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
3750 rq->misfit_task_load = 0;
3754 rq->misfit_task_load = task_h_load(p);
3757 #else /* CONFIG_SMP */
3759 #define UPDATE_TG 0x0
3760 #define SKIP_AGE_LOAD 0x0
3761 #define DO_ATTACH 0x0
3763 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3765 cfs_rq_util_change(cfs_rq, 0);
3768 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3771 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
3773 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3775 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3781 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
3784 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
3786 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
3788 #endif /* CONFIG_SMP */
3790 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3792 #ifdef CONFIG_SCHED_DEBUG
3793 s64 d = se->vruntime - cfs_rq->min_vruntime;
3798 if (d > 3*sysctl_sched_latency)
3799 schedstat_inc(cfs_rq->nr_spread_over);
3804 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3806 u64 vruntime = cfs_rq->min_vruntime;
3809 * The 'current' period is already promised to the current tasks,
3810 * however the extra weight of the new task will slow them down a
3811 * little, place the new task so that it fits in the slot that
3812 * stays open at the end.
3814 if (initial && sched_feat(START_DEBIT))
3815 vruntime += sched_vslice(cfs_rq, se);
3817 /* sleeps up to a single latency don't count. */
3819 unsigned long thresh = sysctl_sched_latency;
3822 * Halve their sleep time's effect, to allow
3823 * for a gentler effect of sleepers:
3825 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3831 /* ensure we never gain time by being placed backwards. */
3832 se->vruntime = max_vruntime(se->vruntime, vruntime);
3835 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3837 static inline void check_schedstat_required(void)
3839 #ifdef CONFIG_SCHEDSTATS
3840 if (schedstat_enabled())
3843 /* Force schedstat enabled if a dependent tracepoint is active */
3844 if (trace_sched_stat_wait_enabled() ||
3845 trace_sched_stat_sleep_enabled() ||
3846 trace_sched_stat_iowait_enabled() ||
3847 trace_sched_stat_blocked_enabled() ||
3848 trace_sched_stat_runtime_enabled()) {
3849 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3850 "stat_blocked and stat_runtime require the "
3851 "kernel parameter schedstats=enable or "
3852 "kernel.sched_schedstats=1\n");
3863 * update_min_vruntime()
3864 * vruntime -= min_vruntime
3868 * update_min_vruntime()
3869 * vruntime += min_vruntime
3871 * this way the vruntime transition between RQs is done when both
3872 * min_vruntime are up-to-date.
3876 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3877 * vruntime -= min_vruntime
3881 * update_min_vruntime()
3882 * vruntime += min_vruntime
3884 * this way we don't have the most up-to-date min_vruntime on the originating
3885 * CPU and an up-to-date min_vruntime on the destination CPU.
3889 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3891 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3892 bool curr = cfs_rq->curr == se;
3895 * If we're the current task, we must renormalise before calling
3899 se->vruntime += cfs_rq->min_vruntime;
3901 update_curr(cfs_rq);
3904 * Otherwise, renormalise after, such that we're placed at the current
3905 * moment in time, instead of some random moment in the past. Being
3906 * placed in the past could significantly boost this task to the
3907 * fairness detriment of existing tasks.
3909 if (renorm && !curr)
3910 se->vruntime += cfs_rq->min_vruntime;
3913 * When enqueuing a sched_entity, we must:
3914 * - Update loads to have both entity and cfs_rq synced with now.
3915 * - Add its load to cfs_rq->runnable_avg
3916 * - For group_entity, update its weight to reflect the new share of
3918 * - Add its new weight to cfs_rq->load.weight
3920 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
3921 update_cfs_group(se);
3922 enqueue_runnable_load_avg(cfs_rq, se);
3923 account_entity_enqueue(cfs_rq, se);
3925 if (flags & ENQUEUE_WAKEUP)
3926 place_entity(cfs_rq, se, 0);
3928 check_schedstat_required();
3929 update_stats_enqueue(cfs_rq, se, flags);
3930 check_spread(cfs_rq, se);
3932 __enqueue_entity(cfs_rq, se);
3935 if (cfs_rq->nr_running == 1) {
3936 list_add_leaf_cfs_rq(cfs_rq);
3937 check_enqueue_throttle(cfs_rq);
3941 static void __clear_buddies_last(struct sched_entity *se)
3943 for_each_sched_entity(se) {
3944 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3945 if (cfs_rq->last != se)
3948 cfs_rq->last = NULL;
3952 static void __clear_buddies_next(struct sched_entity *se)
3954 for_each_sched_entity(se) {
3955 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3956 if (cfs_rq->next != se)
3959 cfs_rq->next = NULL;
3963 static void __clear_buddies_skip(struct sched_entity *se)
3965 for_each_sched_entity(se) {
3966 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3967 if (cfs_rq->skip != se)
3970 cfs_rq->skip = NULL;
3974 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3976 if (cfs_rq->last == se)
3977 __clear_buddies_last(se);
3979 if (cfs_rq->next == se)
3980 __clear_buddies_next(se);
3982 if (cfs_rq->skip == se)
3983 __clear_buddies_skip(se);
3986 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3989 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3992 * Update run-time statistics of the 'current'.
3994 update_curr(cfs_rq);
3997 * When dequeuing a sched_entity, we must:
3998 * - Update loads to have both entity and cfs_rq synced with now.
3999 * - Subtract its load from the cfs_rq->runnable_avg.
4000 * - Subtract its previous weight from cfs_rq->load.weight.
4001 * - For group entity, update its weight to reflect the new share
4002 * of its group cfs_rq.
4004 update_load_avg(cfs_rq, se, UPDATE_TG);
4005 dequeue_runnable_load_avg(cfs_rq, se);
4007 update_stats_dequeue(cfs_rq, se, flags);
4009 clear_buddies(cfs_rq, se);
4011 if (se != cfs_rq->curr)
4012 __dequeue_entity(cfs_rq, se);
4014 account_entity_dequeue(cfs_rq, se);
4017 * Normalize after update_curr(); which will also have moved
4018 * min_vruntime if @se is the one holding it back. But before doing
4019 * update_min_vruntime() again, which will discount @se's position and
4020 * can move min_vruntime forward still more.
4022 if (!(flags & DEQUEUE_SLEEP))
4023 se->vruntime -= cfs_rq->min_vruntime;
4025 /* return excess runtime on last dequeue */
4026 return_cfs_rq_runtime(cfs_rq);
4028 update_cfs_group(se);
4031 * Now advance min_vruntime if @se was the entity holding it back,
4032 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4033 * put back on, and if we advance min_vruntime, we'll be placed back
4034 * further than we started -- ie. we'll be penalized.
4036 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4037 update_min_vruntime(cfs_rq);
4041 * Preempt the current task with a newly woken task if needed:
4044 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4046 unsigned long ideal_runtime, delta_exec;
4047 struct sched_entity *se;
4050 ideal_runtime = sched_slice(cfs_rq, curr);
4051 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4052 if (delta_exec > ideal_runtime) {
4053 resched_curr(rq_of(cfs_rq));
4055 * The current task ran long enough, ensure it doesn't get
4056 * re-elected due to buddy favours.
4058 clear_buddies(cfs_rq, curr);
4063 * Ensure that a task that missed wakeup preemption by a
4064 * narrow margin doesn't have to wait for a full slice.
4065 * This also mitigates buddy induced latencies under load.
4067 if (delta_exec < sysctl_sched_min_granularity)
4070 se = __pick_first_entity(cfs_rq);
4071 delta = curr->vruntime - se->vruntime;
4076 if (delta > ideal_runtime)
4077 resched_curr(rq_of(cfs_rq));
4081 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4083 /* 'current' is not kept within the tree. */
4086 * Any task has to be enqueued before it get to execute on
4087 * a CPU. So account for the time it spent waiting on the
4090 update_stats_wait_end(cfs_rq, se);
4091 __dequeue_entity(cfs_rq, se);
4092 update_load_avg(cfs_rq, se, UPDATE_TG);
4095 update_stats_curr_start(cfs_rq, se);
4099 * Track our maximum slice length, if the CPU's load is at
4100 * least twice that of our own weight (i.e. dont track it
4101 * when there are only lesser-weight tasks around):
4103 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
4104 schedstat_set(se->statistics.slice_max,
4105 max((u64)schedstat_val(se->statistics.slice_max),
4106 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4109 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4113 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4116 * Pick the next process, keeping these things in mind, in this order:
4117 * 1) keep things fair between processes/task groups
4118 * 2) pick the "next" process, since someone really wants that to run
4119 * 3) pick the "last" process, for cache locality
4120 * 4) do not run the "skip" process, if something else is available
4122 static struct sched_entity *
4123 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4125 struct sched_entity *left = __pick_first_entity(cfs_rq);
4126 struct sched_entity *se;
4129 * If curr is set we have to see if its left of the leftmost entity
4130 * still in the tree, provided there was anything in the tree at all.
4132 if (!left || (curr && entity_before(curr, left)))
4135 se = left; /* ideally we run the leftmost entity */
4138 * Avoid running the skip buddy, if running something else can
4139 * be done without getting too unfair.
4141 if (cfs_rq->skip == se) {
4142 struct sched_entity *second;
4145 second = __pick_first_entity(cfs_rq);
4147 second = __pick_next_entity(se);
4148 if (!second || (curr && entity_before(curr, second)))
4152 if (second && wakeup_preempt_entity(second, left) < 1)
4157 * Prefer last buddy, try to return the CPU to a preempted task.
4159 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4163 * Someone really wants this to run. If it's not unfair, run it.
4165 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4168 clear_buddies(cfs_rq, se);
4173 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4175 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4178 * If still on the runqueue then deactivate_task()
4179 * was not called and update_curr() has to be done:
4182 update_curr(cfs_rq);
4184 /* throttle cfs_rqs exceeding runtime */
4185 check_cfs_rq_runtime(cfs_rq);
4187 check_spread(cfs_rq, prev);
4190 update_stats_wait_start(cfs_rq, prev);
4191 /* Put 'current' back into the tree. */
4192 __enqueue_entity(cfs_rq, prev);
4193 /* in !on_rq case, update occurred at dequeue */
4194 update_load_avg(cfs_rq, prev, 0);
4196 cfs_rq->curr = NULL;
4200 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4203 * Update run-time statistics of the 'current'.
4205 update_curr(cfs_rq);
4208 * Ensure that runnable average is periodically updated.
4210 update_load_avg(cfs_rq, curr, UPDATE_TG);
4211 update_cfs_group(curr);
4213 #ifdef CONFIG_SCHED_HRTICK
4215 * queued ticks are scheduled to match the slice, so don't bother
4216 * validating it and just reschedule.
4219 resched_curr(rq_of(cfs_rq));
4223 * don't let the period tick interfere with the hrtick preemption
4225 if (!sched_feat(DOUBLE_TICK) &&
4226 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4230 if (cfs_rq->nr_running > 1)
4231 check_preempt_tick(cfs_rq, curr);
4235 /**************************************************
4236 * CFS bandwidth control machinery
4239 #ifdef CONFIG_CFS_BANDWIDTH
4241 #ifdef CONFIG_JUMP_LABEL
4242 static struct static_key __cfs_bandwidth_used;
4244 static inline bool cfs_bandwidth_used(void)
4246 return static_key_false(&__cfs_bandwidth_used);
4249 void cfs_bandwidth_usage_inc(void)
4251 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4254 void cfs_bandwidth_usage_dec(void)
4256 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4258 #else /* CONFIG_JUMP_LABEL */
4259 static bool cfs_bandwidth_used(void)
4264 void cfs_bandwidth_usage_inc(void) {}
4265 void cfs_bandwidth_usage_dec(void) {}
4266 #endif /* CONFIG_JUMP_LABEL */
4269 * default period for cfs group bandwidth.
4270 * default: 0.1s, units: nanoseconds
4272 static inline u64 default_cfs_period(void)
4274 return 100000000ULL;
4277 static inline u64 sched_cfs_bandwidth_slice(void)
4279 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4283 * Replenish runtime according to assigned quota and update expiration time.
4284 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4285 * additional synchronization around rq->lock.
4287 * requires cfs_b->lock
4289 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4293 if (cfs_b->quota == RUNTIME_INF)
4296 now = sched_clock_cpu(smp_processor_id());
4297 cfs_b->runtime = cfs_b->quota;
4298 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4299 cfs_b->expires_seq++;
4302 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4304 return &tg->cfs_bandwidth;
4307 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4308 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4310 if (unlikely(cfs_rq->throttle_count))
4311 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4313 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4316 /* returns 0 on failure to allocate runtime */
4317 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4319 struct task_group *tg = cfs_rq->tg;
4320 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4321 u64 amount = 0, min_amount, expires;
4324 /* note: this is a positive sum as runtime_remaining <= 0 */
4325 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4327 raw_spin_lock(&cfs_b->lock);
4328 if (cfs_b->quota == RUNTIME_INF)
4329 amount = min_amount;
4331 start_cfs_bandwidth(cfs_b);
4333 if (cfs_b->runtime > 0) {
4334 amount = min(cfs_b->runtime, min_amount);
4335 cfs_b->runtime -= amount;
4339 expires_seq = cfs_b->expires_seq;
4340 expires = cfs_b->runtime_expires;
4341 raw_spin_unlock(&cfs_b->lock);
4343 cfs_rq->runtime_remaining += amount;
4345 * we may have advanced our local expiration to account for allowed
4346 * spread between our sched_clock and the one on which runtime was
4349 if (cfs_rq->expires_seq != expires_seq) {
4350 cfs_rq->expires_seq = expires_seq;
4351 cfs_rq->runtime_expires = expires;
4354 return cfs_rq->runtime_remaining > 0;
4358 * Note: This depends on the synchronization provided by sched_clock and the
4359 * fact that rq->clock snapshots this value.
4361 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4363 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4365 /* if the deadline is ahead of our clock, nothing to do */
4366 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4369 if (cfs_rq->runtime_remaining < 0)
4373 * If the local deadline has passed we have to consider the
4374 * possibility that our sched_clock is 'fast' and the global deadline
4375 * has not truly expired.
4377 * Fortunately we can check determine whether this the case by checking
4378 * whether the global deadline(cfs_b->expires_seq) has advanced.
4380 if (cfs_rq->expires_seq == cfs_b->expires_seq) {
4381 /* extend local deadline, drift is bounded above by 2 ticks */
4382 cfs_rq->runtime_expires += TICK_NSEC;
4384 /* global deadline is ahead, expiration has passed */
4385 cfs_rq->runtime_remaining = 0;
4389 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4391 /* dock delta_exec before expiring quota (as it could span periods) */
4392 cfs_rq->runtime_remaining -= delta_exec;
4393 expire_cfs_rq_runtime(cfs_rq);
4395 if (likely(cfs_rq->runtime_remaining > 0))
4399 * if we're unable to extend our runtime we resched so that the active
4400 * hierarchy can be throttled
4402 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4403 resched_curr(rq_of(cfs_rq));
4406 static __always_inline
4407 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4409 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4412 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4415 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4417 return cfs_bandwidth_used() && cfs_rq->throttled;
4420 /* check whether cfs_rq, or any parent, is throttled */
4421 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4423 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4427 * Ensure that neither of the group entities corresponding to src_cpu or
4428 * dest_cpu are members of a throttled hierarchy when performing group
4429 * load-balance operations.
4431 static inline int throttled_lb_pair(struct task_group *tg,
4432 int src_cpu, int dest_cpu)
4434 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4436 src_cfs_rq = tg->cfs_rq[src_cpu];
4437 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4439 return throttled_hierarchy(src_cfs_rq) ||
4440 throttled_hierarchy(dest_cfs_rq);
4443 static int tg_unthrottle_up(struct task_group *tg, void *data)
4445 struct rq *rq = data;
4446 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4448 cfs_rq->throttle_count--;
4449 if (!cfs_rq->throttle_count) {
4450 /* adjust cfs_rq_clock_task() */
4451 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4452 cfs_rq->throttled_clock_task;
4454 /* Add cfs_rq with already running entity in the list */
4455 if (cfs_rq->nr_running >= 1)
4456 list_add_leaf_cfs_rq(cfs_rq);
4462 static int tg_throttle_down(struct task_group *tg, void *data)
4464 struct rq *rq = data;
4465 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4467 /* group is entering throttled state, stop time */
4468 if (!cfs_rq->throttle_count) {
4469 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4470 list_del_leaf_cfs_rq(cfs_rq);
4472 cfs_rq->throttle_count++;
4477 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4479 struct rq *rq = rq_of(cfs_rq);
4480 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4481 struct sched_entity *se;
4482 long task_delta, dequeue = 1;
4485 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4487 /* freeze hierarchy runnable averages while throttled */
4489 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4492 task_delta = cfs_rq->h_nr_running;
4493 for_each_sched_entity(se) {
4494 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4495 /* throttled entity or throttle-on-deactivate */
4500 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4501 qcfs_rq->h_nr_running -= task_delta;
4503 if (qcfs_rq->load.weight)
4508 sub_nr_running(rq, task_delta);
4510 cfs_rq->throttled = 1;
4511 cfs_rq->throttled_clock = rq_clock(rq);
4512 raw_spin_lock(&cfs_b->lock);
4513 empty = list_empty(&cfs_b->throttled_cfs_rq);
4516 * Add to the _head_ of the list, so that an already-started
4517 * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
4518 * not running add to the tail so that later runqueues don't get starved.
4520 if (cfs_b->distribute_running)
4521 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4523 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4526 * If we're the first throttled task, make sure the bandwidth
4530 start_cfs_bandwidth(cfs_b);
4532 raw_spin_unlock(&cfs_b->lock);
4535 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4537 struct rq *rq = rq_of(cfs_rq);
4538 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4539 struct sched_entity *se;
4543 se = cfs_rq->tg->se[cpu_of(rq)];
4545 cfs_rq->throttled = 0;
4547 update_rq_clock(rq);
4549 raw_spin_lock(&cfs_b->lock);
4550 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4551 list_del_rcu(&cfs_rq->throttled_list);
4552 raw_spin_unlock(&cfs_b->lock);
4554 /* update hierarchical throttle state */
4555 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4557 if (!cfs_rq->load.weight)
4560 task_delta = cfs_rq->h_nr_running;
4561 for_each_sched_entity(se) {
4565 cfs_rq = cfs_rq_of(se);
4567 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4568 cfs_rq->h_nr_running += task_delta;
4570 if (cfs_rq_throttled(cfs_rq))
4574 assert_list_leaf_cfs_rq(rq);
4577 add_nr_running(rq, task_delta);
4579 /* Determine whether we need to wake up potentially idle CPU: */
4580 if (rq->curr == rq->idle && rq->cfs.nr_running)
4584 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4585 u64 remaining, u64 expires)
4587 struct cfs_rq *cfs_rq;
4589 u64 starting_runtime = remaining;
4592 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4594 struct rq *rq = rq_of(cfs_rq);
4597 rq_lock_irqsave(rq, &rf);
4598 if (!cfs_rq_throttled(cfs_rq))
4601 runtime = -cfs_rq->runtime_remaining + 1;
4602 if (runtime > remaining)
4603 runtime = remaining;
4604 remaining -= runtime;
4606 cfs_rq->runtime_remaining += runtime;
4607 cfs_rq->runtime_expires = expires;
4609 /* we check whether we're throttled above */
4610 if (cfs_rq->runtime_remaining > 0)
4611 unthrottle_cfs_rq(cfs_rq);
4614 rq_unlock_irqrestore(rq, &rf);
4621 return starting_runtime - remaining;
4625 * Responsible for refilling a task_group's bandwidth and unthrottling its
4626 * cfs_rqs as appropriate. If there has been no activity within the last
4627 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4628 * used to track this state.
4630 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
4632 u64 runtime, runtime_expires;
4635 /* no need to continue the timer with no bandwidth constraint */
4636 if (cfs_b->quota == RUNTIME_INF)
4637 goto out_deactivate;
4639 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4640 cfs_b->nr_periods += overrun;
4643 * idle depends on !throttled (for the case of a large deficit), and if
4644 * we're going inactive then everything else can be deferred
4646 if (cfs_b->idle && !throttled)
4647 goto out_deactivate;
4649 __refill_cfs_bandwidth_runtime(cfs_b);
4652 /* mark as potentially idle for the upcoming period */
4657 /* account preceding periods in which throttling occurred */
4658 cfs_b->nr_throttled += overrun;
4660 runtime_expires = cfs_b->runtime_expires;
4663 * This check is repeated as we are holding onto the new bandwidth while
4664 * we unthrottle. This can potentially race with an unthrottled group
4665 * trying to acquire new bandwidth from the global pool. This can result
4666 * in us over-using our runtime if it is all used during this loop, but
4667 * only by limited amounts in that extreme case.
4669 while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
4670 runtime = cfs_b->runtime;
4671 cfs_b->distribute_running = 1;
4672 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4673 /* we can't nest cfs_b->lock while distributing bandwidth */
4674 runtime = distribute_cfs_runtime(cfs_b, runtime,
4676 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4678 cfs_b->distribute_running = 0;
4679 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4681 lsub_positive(&cfs_b->runtime, runtime);
4685 * While we are ensured activity in the period following an
4686 * unthrottle, this also covers the case in which the new bandwidth is
4687 * insufficient to cover the existing bandwidth deficit. (Forcing the
4688 * timer to remain active while there are any throttled entities.)
4698 /* a cfs_rq won't donate quota below this amount */
4699 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4700 /* minimum remaining period time to redistribute slack quota */
4701 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4702 /* how long we wait to gather additional slack before distributing */
4703 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4706 * Are we near the end of the current quota period?
4708 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4709 * hrtimer base being cleared by hrtimer_start. In the case of
4710 * migrate_hrtimers, base is never cleared, so we are fine.
4712 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4714 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4717 /* if the call-back is running a quota refresh is already occurring */
4718 if (hrtimer_callback_running(refresh_timer))
4721 /* is a quota refresh about to occur? */
4722 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4723 if (remaining < min_expire)
4729 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4731 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4733 /* if there's a quota refresh soon don't bother with slack */
4734 if (runtime_refresh_within(cfs_b, min_left))
4737 hrtimer_start(&cfs_b->slack_timer,
4738 ns_to_ktime(cfs_bandwidth_slack_period),
4742 /* we know any runtime found here is valid as update_curr() precedes return */
4743 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4745 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4746 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4748 if (slack_runtime <= 0)
4751 raw_spin_lock(&cfs_b->lock);
4752 if (cfs_b->quota != RUNTIME_INF &&
4753 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4754 cfs_b->runtime += slack_runtime;
4756 /* we are under rq->lock, defer unthrottling using a timer */
4757 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4758 !list_empty(&cfs_b->throttled_cfs_rq))
4759 start_cfs_slack_bandwidth(cfs_b);
4761 raw_spin_unlock(&cfs_b->lock);
4763 /* even if it's not valid for return we don't want to try again */
4764 cfs_rq->runtime_remaining -= slack_runtime;
4767 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4769 if (!cfs_bandwidth_used())
4772 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4775 __return_cfs_rq_runtime(cfs_rq);
4779 * This is done with a timer (instead of inline with bandwidth return) since
4780 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4782 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4784 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4785 unsigned long flags;
4788 /* confirm we're still not at a refresh boundary */
4789 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4790 if (cfs_b->distribute_running) {
4791 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4795 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4796 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4800 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4801 runtime = cfs_b->runtime;
4803 expires = cfs_b->runtime_expires;
4805 cfs_b->distribute_running = 1;
4807 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4812 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4814 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4815 if (expires == cfs_b->runtime_expires)
4816 lsub_positive(&cfs_b->runtime, runtime);
4817 cfs_b->distribute_running = 0;
4818 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4822 * When a group wakes up we want to make sure that its quota is not already
4823 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4824 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4826 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4828 if (!cfs_bandwidth_used())
4831 /* an active group must be handled by the update_curr()->put() path */
4832 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4835 /* ensure the group is not already throttled */
4836 if (cfs_rq_throttled(cfs_rq))
4839 /* update runtime allocation */
4840 account_cfs_rq_runtime(cfs_rq, 0);
4841 if (cfs_rq->runtime_remaining <= 0)
4842 throttle_cfs_rq(cfs_rq);
4845 static void sync_throttle(struct task_group *tg, int cpu)
4847 struct cfs_rq *pcfs_rq, *cfs_rq;
4849 if (!cfs_bandwidth_used())
4855 cfs_rq = tg->cfs_rq[cpu];
4856 pcfs_rq = tg->parent->cfs_rq[cpu];
4858 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4859 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4862 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4863 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4865 if (!cfs_bandwidth_used())
4868 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4872 * it's possible for a throttled entity to be forced into a running
4873 * state (e.g. set_curr_task), in this case we're finished.
4875 if (cfs_rq_throttled(cfs_rq))
4878 throttle_cfs_rq(cfs_rq);
4882 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4884 struct cfs_bandwidth *cfs_b =
4885 container_of(timer, struct cfs_bandwidth, slack_timer);
4887 do_sched_cfs_slack_timer(cfs_b);
4889 return HRTIMER_NORESTART;
4892 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4894 struct cfs_bandwidth *cfs_b =
4895 container_of(timer, struct cfs_bandwidth, period_timer);
4896 unsigned long flags;
4900 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4902 overrun = hrtimer_forward_now(timer, cfs_b->period);
4906 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
4909 cfs_b->period_active = 0;
4910 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4912 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4915 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4917 raw_spin_lock_init(&cfs_b->lock);
4919 cfs_b->quota = RUNTIME_INF;
4920 cfs_b->period = ns_to_ktime(default_cfs_period());
4922 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4923 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4924 cfs_b->period_timer.function = sched_cfs_period_timer;
4925 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4926 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4927 cfs_b->distribute_running = 0;
4930 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4932 cfs_rq->runtime_enabled = 0;
4933 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4936 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4940 lockdep_assert_held(&cfs_b->lock);
4942 if (cfs_b->period_active)
4945 cfs_b->period_active = 1;
4946 overrun = hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4947 cfs_b->runtime_expires += (overrun + 1) * ktime_to_ns(cfs_b->period);
4948 cfs_b->expires_seq++;
4949 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4952 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4954 /* init_cfs_bandwidth() was not called */
4955 if (!cfs_b->throttled_cfs_rq.next)
4958 hrtimer_cancel(&cfs_b->period_timer);
4959 hrtimer_cancel(&cfs_b->slack_timer);
4963 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
4965 * The race is harmless, since modifying bandwidth settings of unhooked group
4966 * bits doesn't do much.
4969 /* cpu online calback */
4970 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4972 struct task_group *tg;
4974 lockdep_assert_held(&rq->lock);
4977 list_for_each_entry_rcu(tg, &task_groups, list) {
4978 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
4979 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4981 raw_spin_lock(&cfs_b->lock);
4982 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4983 raw_spin_unlock(&cfs_b->lock);
4988 /* cpu offline callback */
4989 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4991 struct task_group *tg;
4993 lockdep_assert_held(&rq->lock);
4996 list_for_each_entry_rcu(tg, &task_groups, list) {
4997 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4999 if (!cfs_rq->runtime_enabled)
5003 * clock_task is not advancing so we just need to make sure
5004 * there's some valid quota amount
5006 cfs_rq->runtime_remaining = 1;
5008 * Offline rq is schedulable till CPU is completely disabled
5009 * in take_cpu_down(), so we prevent new cfs throttling here.
5011 cfs_rq->runtime_enabled = 0;
5013 if (cfs_rq_throttled(cfs_rq))
5014 unthrottle_cfs_rq(cfs_rq);
5019 #else /* CONFIG_CFS_BANDWIDTH */
5021 static inline bool cfs_bandwidth_used(void)
5026 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
5028 return rq_clock_task(rq_of(cfs_rq));
5031 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5032 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5033 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5034 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5035 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5037 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5042 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5047 static inline int throttled_lb_pair(struct task_group *tg,
5048 int src_cpu, int dest_cpu)
5053 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5055 #ifdef CONFIG_FAIR_GROUP_SCHED
5056 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5059 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5063 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5064 static inline void update_runtime_enabled(struct rq *rq) {}
5065 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5067 #endif /* CONFIG_CFS_BANDWIDTH */
5069 /**************************************************
5070 * CFS operations on tasks:
5073 #ifdef CONFIG_SCHED_HRTICK
5074 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5076 struct sched_entity *se = &p->se;
5077 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5079 SCHED_WARN_ON(task_rq(p) != rq);
5081 if (rq->cfs.h_nr_running > 1) {
5082 u64 slice = sched_slice(cfs_rq, se);
5083 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5084 s64 delta = slice - ran;
5091 hrtick_start(rq, delta);
5096 * called from enqueue/dequeue and updates the hrtick when the
5097 * current task is from our class and nr_running is low enough
5100 static void hrtick_update(struct rq *rq)
5102 struct task_struct *curr = rq->curr;
5104 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5107 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5108 hrtick_start_fair(rq, curr);
5110 #else /* !CONFIG_SCHED_HRTICK */
5112 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5116 static inline void hrtick_update(struct rq *rq)
5122 static inline unsigned long cpu_util(int cpu);
5123 static unsigned long capacity_of(int cpu);
5125 static inline bool cpu_overutilized(int cpu)
5127 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5130 static inline void update_overutilized_status(struct rq *rq)
5132 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu))
5133 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5136 static inline void update_overutilized_status(struct rq *rq) { }
5140 * The enqueue_task method is called before nr_running is
5141 * increased. Here we update the fair scheduling stats and
5142 * then put the task into the rbtree:
5145 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5147 struct cfs_rq *cfs_rq;
5148 struct sched_entity *se = &p->se;
5151 * The code below (indirectly) updates schedutil which looks at
5152 * the cfs_rq utilization to select a frequency.
5153 * Let's add the task's estimated utilization to the cfs_rq's
5154 * estimated utilization, before we update schedutil.
5156 util_est_enqueue(&rq->cfs, p);
5159 * If in_iowait is set, the code below may not trigger any cpufreq
5160 * utilization updates, so do it here explicitly with the IOWAIT flag
5164 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5166 for_each_sched_entity(se) {
5169 cfs_rq = cfs_rq_of(se);
5170 enqueue_entity(cfs_rq, se, flags);
5173 * end evaluation on encountering a throttled cfs_rq
5175 * note: in the case of encountering a throttled cfs_rq we will
5176 * post the final h_nr_running increment below.
5178 if (cfs_rq_throttled(cfs_rq))
5180 cfs_rq->h_nr_running++;
5182 flags = ENQUEUE_WAKEUP;
5185 for_each_sched_entity(se) {
5186 cfs_rq = cfs_rq_of(se);
5187 cfs_rq->h_nr_running++;
5189 if (cfs_rq_throttled(cfs_rq))
5192 update_load_avg(cfs_rq, se, UPDATE_TG);
5193 update_cfs_group(se);
5197 add_nr_running(rq, 1);
5199 * Since new tasks are assigned an initial util_avg equal to
5200 * half of the spare capacity of their CPU, tiny tasks have the
5201 * ability to cross the overutilized threshold, which will
5202 * result in the load balancer ruining all the task placement
5203 * done by EAS. As a way to mitigate that effect, do not account
5204 * for the first enqueue operation of new tasks during the
5205 * overutilized flag detection.
5207 * A better way of solving this problem would be to wait for
5208 * the PELT signals of tasks to converge before taking them
5209 * into account, but that is not straightforward to implement,
5210 * and the following generally works well enough in practice.
5212 if (flags & ENQUEUE_WAKEUP)
5213 update_overutilized_status(rq);
5217 if (cfs_bandwidth_used()) {
5219 * When bandwidth control is enabled; the cfs_rq_throttled()
5220 * breaks in the above iteration can result in incomplete
5221 * leaf list maintenance, resulting in triggering the assertion
5224 for_each_sched_entity(se) {
5225 cfs_rq = cfs_rq_of(se);
5227 if (list_add_leaf_cfs_rq(cfs_rq))
5232 assert_list_leaf_cfs_rq(rq);
5237 static void set_next_buddy(struct sched_entity *se);
5240 * The dequeue_task method is called before nr_running is
5241 * decreased. We remove the task from the rbtree and
5242 * update the fair scheduling stats:
5244 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5246 struct cfs_rq *cfs_rq;
5247 struct sched_entity *se = &p->se;
5248 int task_sleep = flags & DEQUEUE_SLEEP;
5250 for_each_sched_entity(se) {
5251 cfs_rq = cfs_rq_of(se);
5252 dequeue_entity(cfs_rq, se, flags);
5255 * end evaluation on encountering a throttled cfs_rq
5257 * note: in the case of encountering a throttled cfs_rq we will
5258 * post the final h_nr_running decrement below.
5260 if (cfs_rq_throttled(cfs_rq))
5262 cfs_rq->h_nr_running--;
5264 /* Don't dequeue parent if it has other entities besides us */
5265 if (cfs_rq->load.weight) {
5266 /* Avoid re-evaluating load for this entity: */
5267 se = parent_entity(se);
5269 * Bias pick_next to pick a task from this cfs_rq, as
5270 * p is sleeping when it is within its sched_slice.
5272 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5276 flags |= DEQUEUE_SLEEP;
5279 for_each_sched_entity(se) {
5280 cfs_rq = cfs_rq_of(se);
5281 cfs_rq->h_nr_running--;
5283 if (cfs_rq_throttled(cfs_rq))
5286 update_load_avg(cfs_rq, se, UPDATE_TG);
5287 update_cfs_group(se);
5291 sub_nr_running(rq, 1);
5293 util_est_dequeue(&rq->cfs, p, task_sleep);
5299 /* Working cpumask for: load_balance, load_balance_newidle. */
5300 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5301 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5303 #ifdef CONFIG_NO_HZ_COMMON
5305 * per rq 'load' arrray crap; XXX kill this.
5309 * The exact cpuload calculated at every tick would be:
5311 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5313 * If a CPU misses updates for n ticks (as it was idle) and update gets
5314 * called on the n+1-th tick when CPU may be busy, then we have:
5316 * load_n = (1 - 1/2^i)^n * load_0
5317 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
5319 * decay_load_missed() below does efficient calculation of
5321 * load' = (1 - 1/2^i)^n * load
5323 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5324 * This allows us to precompute the above in said factors, thereby allowing the
5325 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5326 * fixed_power_int())
5328 * The calculation is approximated on a 128 point scale.
5330 #define DEGRADE_SHIFT 7
5332 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5333 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5334 { 0, 0, 0, 0, 0, 0, 0, 0 },
5335 { 64, 32, 8, 0, 0, 0, 0, 0 },
5336 { 96, 72, 40, 12, 1, 0, 0, 0 },
5337 { 112, 98, 75, 43, 15, 1, 0, 0 },
5338 { 120, 112, 98, 76, 45, 16, 2, 0 }
5342 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5343 * would be when CPU is idle and so we just decay the old load without
5344 * adding any new load.
5346 static unsigned long
5347 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5351 if (!missed_updates)
5354 if (missed_updates >= degrade_zero_ticks[idx])
5358 return load >> missed_updates;
5360 while (missed_updates) {
5361 if (missed_updates % 2)
5362 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5364 missed_updates >>= 1;
5371 cpumask_var_t idle_cpus_mask;
5373 int has_blocked; /* Idle CPUS has blocked load */
5374 unsigned long next_balance; /* in jiffy units */
5375 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5376 } nohz ____cacheline_aligned;
5378 #endif /* CONFIG_NO_HZ_COMMON */
5381 * __cpu_load_update - update the rq->cpu_load[] statistics
5382 * @this_rq: The rq to update statistics for
5383 * @this_load: The current load
5384 * @pending_updates: The number of missed updates
5386 * Update rq->cpu_load[] statistics. This function is usually called every
5387 * scheduler tick (TICK_NSEC).
5389 * This function computes a decaying average:
5391 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5393 * Because of NOHZ it might not get called on every tick which gives need for
5394 * the @pending_updates argument.
5396 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5397 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5398 * = A * (A * load[i]_n-2 + B) + B
5399 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5400 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5401 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5402 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5403 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5405 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5406 * any change in load would have resulted in the tick being turned back on.
5408 * For regular NOHZ, this reduces to:
5410 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5412 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5415 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5416 unsigned long pending_updates)
5418 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5421 this_rq->nr_load_updates++;
5423 /* Update our load: */
5424 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5425 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5426 unsigned long old_load, new_load;
5428 /* scale is effectively 1 << i now, and >> i divides by scale */
5430 old_load = this_rq->cpu_load[i];
5431 #ifdef CONFIG_NO_HZ_COMMON
5432 old_load = decay_load_missed(old_load, pending_updates - 1, i);
5433 if (tickless_load) {
5434 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5436 * old_load can never be a negative value because a
5437 * decayed tickless_load cannot be greater than the
5438 * original tickless_load.
5440 old_load += tickless_load;
5443 new_load = this_load;
5445 * Round up the averaging division if load is increasing. This
5446 * prevents us from getting stuck on 9 if the load is 10, for
5449 if (new_load > old_load)
5450 new_load += scale - 1;
5452 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5456 /* Used instead of source_load when we know the type == 0 */
5457 static unsigned long weighted_cpuload(struct rq *rq)
5459 return cfs_rq_runnable_load_avg(&rq->cfs);
5462 #ifdef CONFIG_NO_HZ_COMMON
5464 * There is no sane way to deal with nohz on smp when using jiffies because the
5465 * CPU doing the jiffies update might drift wrt the CPU doing the jiffy reading
5466 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5468 * Therefore we need to avoid the delta approach from the regular tick when
5469 * possible since that would seriously skew the load calculation. This is why we
5470 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5471 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5472 * loop exit, nohz_idle_balance, nohz full exit...)
5474 * This means we might still be one tick off for nohz periods.
5477 static void cpu_load_update_nohz(struct rq *this_rq,
5478 unsigned long curr_jiffies,
5481 unsigned long pending_updates;
5483 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5484 if (pending_updates) {
5485 this_rq->last_load_update_tick = curr_jiffies;
5487 * In the regular NOHZ case, we were idle, this means load 0.
5488 * In the NOHZ_FULL case, we were non-idle, we should consider
5489 * its weighted load.
5491 cpu_load_update(this_rq, load, pending_updates);
5496 * Called from nohz_idle_balance() to update the load ratings before doing the
5499 static void cpu_load_update_idle(struct rq *this_rq)
5502 * bail if there's load or we're actually up-to-date.
5504 if (weighted_cpuload(this_rq))
5507 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5511 * Record CPU load on nohz entry so we know the tickless load to account
5512 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5513 * than other cpu_load[idx] but it should be fine as cpu_load readers
5514 * shouldn't rely into synchronized cpu_load[*] updates.
5516 void cpu_load_update_nohz_start(void)
5518 struct rq *this_rq = this_rq();
5521 * This is all lockless but should be fine. If weighted_cpuload changes
5522 * concurrently we'll exit nohz. And cpu_load write can race with
5523 * cpu_load_update_idle() but both updater would be writing the same.
5525 this_rq->cpu_load[0] = weighted_cpuload(this_rq);
5529 * Account the tickless load in the end of a nohz frame.
5531 void cpu_load_update_nohz_stop(void)
5533 unsigned long curr_jiffies = READ_ONCE(jiffies);
5534 struct rq *this_rq = this_rq();
5538 if (curr_jiffies == this_rq->last_load_update_tick)
5541 load = weighted_cpuload(this_rq);
5542 rq_lock(this_rq, &rf);
5543 update_rq_clock(this_rq);
5544 cpu_load_update_nohz(this_rq, curr_jiffies, load);
5545 rq_unlock(this_rq, &rf);
5547 #else /* !CONFIG_NO_HZ_COMMON */
5548 static inline void cpu_load_update_nohz(struct rq *this_rq,
5549 unsigned long curr_jiffies,
5550 unsigned long load) { }
5551 #endif /* CONFIG_NO_HZ_COMMON */
5553 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5555 #ifdef CONFIG_NO_HZ_COMMON
5556 /* See the mess around cpu_load_update_nohz(). */
5557 this_rq->last_load_update_tick = READ_ONCE(jiffies);
5559 cpu_load_update(this_rq, load, 1);
5563 * Called from scheduler_tick()
5565 void cpu_load_update_active(struct rq *this_rq)
5567 unsigned long load = weighted_cpuload(this_rq);
5569 if (tick_nohz_tick_stopped())
5570 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5572 cpu_load_update_periodic(this_rq, load);
5576 * Return a low guess at the load of a migration-source CPU weighted
5577 * according to the scheduling class and "nice" value.
5579 * We want to under-estimate the load of migration sources, to
5580 * balance conservatively.
5582 static unsigned long source_load(int cpu, int type)
5584 struct rq *rq = cpu_rq(cpu);
5585 unsigned long total = weighted_cpuload(rq);
5587 if (type == 0 || !sched_feat(LB_BIAS))
5590 return min(rq->cpu_load[type-1], total);
5594 * Return a high guess at the load of a migration-target CPU weighted
5595 * according to the scheduling class and "nice" value.
5597 static unsigned long target_load(int cpu, int type)
5599 struct rq *rq = cpu_rq(cpu);
5600 unsigned long total = weighted_cpuload(rq);
5602 if (type == 0 || !sched_feat(LB_BIAS))
5605 return max(rq->cpu_load[type-1], total);
5608 static unsigned long capacity_of(int cpu)
5610 return cpu_rq(cpu)->cpu_capacity;
5613 static unsigned long cpu_avg_load_per_task(int cpu)
5615 struct rq *rq = cpu_rq(cpu);
5616 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5617 unsigned long load_avg = weighted_cpuload(rq);
5620 return load_avg / nr_running;
5625 static void record_wakee(struct task_struct *p)
5628 * Only decay a single time; tasks that have less then 1 wakeup per
5629 * jiffy will not have built up many flips.
5631 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5632 current->wakee_flips >>= 1;
5633 current->wakee_flip_decay_ts = jiffies;
5636 if (current->last_wakee != p) {
5637 current->last_wakee = p;
5638 current->wakee_flips++;
5643 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5645 * A waker of many should wake a different task than the one last awakened
5646 * at a frequency roughly N times higher than one of its wakees.
5648 * In order to determine whether we should let the load spread vs consolidating
5649 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5650 * partner, and a factor of lls_size higher frequency in the other.
5652 * With both conditions met, we can be relatively sure that the relationship is
5653 * non-monogamous, with partner count exceeding socket size.
5655 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5656 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5659 static int wake_wide(struct task_struct *p)
5661 unsigned int master = current->wakee_flips;
5662 unsigned int slave = p->wakee_flips;
5663 int factor = this_cpu_read(sd_llc_size);
5666 swap(master, slave);
5667 if (slave < factor || master < slave * factor)
5673 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5674 * soonest. For the purpose of speed we only consider the waking and previous
5677 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5678 * cache-affine and is (or will be) idle.
5680 * wake_affine_weight() - considers the weight to reflect the average
5681 * scheduling latency of the CPUs. This seems to work
5682 * for the overloaded case.
5685 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5688 * If this_cpu is idle, it implies the wakeup is from interrupt
5689 * context. Only allow the move if cache is shared. Otherwise an
5690 * interrupt intensive workload could force all tasks onto one
5691 * node depending on the IO topology or IRQ affinity settings.
5693 * If the prev_cpu is idle and cache affine then avoid a migration.
5694 * There is no guarantee that the cache hot data from an interrupt
5695 * is more important than cache hot data on the prev_cpu and from
5696 * a cpufreq perspective, it's better to have higher utilisation
5699 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5700 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5702 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5705 return nr_cpumask_bits;
5709 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5710 int this_cpu, int prev_cpu, int sync)
5712 s64 this_eff_load, prev_eff_load;
5713 unsigned long task_load;
5715 this_eff_load = target_load(this_cpu, sd->wake_idx);
5718 unsigned long current_load = task_h_load(current);
5720 if (current_load > this_eff_load)
5723 this_eff_load -= current_load;
5726 task_load = task_h_load(p);
5728 this_eff_load += task_load;
5729 if (sched_feat(WA_BIAS))
5730 this_eff_load *= 100;
5731 this_eff_load *= capacity_of(prev_cpu);
5733 prev_eff_load = source_load(prev_cpu, sd->wake_idx);
5734 prev_eff_load -= task_load;
5735 if (sched_feat(WA_BIAS))
5736 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5737 prev_eff_load *= capacity_of(this_cpu);
5740 * If sync, adjust the weight of prev_eff_load such that if
5741 * prev_eff == this_eff that select_idle_sibling() will consider
5742 * stacking the wakee on top of the waker if no other CPU is
5748 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5751 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5752 int this_cpu, int prev_cpu, int sync)
5754 int target = nr_cpumask_bits;
5756 if (sched_feat(WA_IDLE))
5757 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5759 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5760 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5762 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5763 if (target == nr_cpumask_bits)
5766 schedstat_inc(sd->ttwu_move_affine);
5767 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5771 static unsigned long cpu_util_without(int cpu, struct task_struct *p);
5773 static unsigned long capacity_spare_without(int cpu, struct task_struct *p)
5775 return max_t(long, capacity_of(cpu) - cpu_util_without(cpu, p), 0);
5779 * find_idlest_group finds and returns the least busy CPU group within the
5782 * Assumes p is allowed on at least one CPU in sd.
5784 static struct sched_group *
5785 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5786 int this_cpu, int sd_flag)
5788 struct sched_group *idlest = NULL, *group = sd->groups;
5789 struct sched_group *most_spare_sg = NULL;
5790 unsigned long min_runnable_load = ULONG_MAX;
5791 unsigned long this_runnable_load = ULONG_MAX;
5792 unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5793 unsigned long most_spare = 0, this_spare = 0;
5794 int load_idx = sd->forkexec_idx;
5795 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5796 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5797 (sd->imbalance_pct-100) / 100;
5799 if (sd_flag & SD_BALANCE_WAKE)
5800 load_idx = sd->wake_idx;
5803 unsigned long load, avg_load, runnable_load;
5804 unsigned long spare_cap, max_spare_cap;
5808 /* Skip over this group if it has no CPUs allowed */
5809 if (!cpumask_intersects(sched_group_span(group),
5813 local_group = cpumask_test_cpu(this_cpu,
5814 sched_group_span(group));
5817 * Tally up the load of all CPUs in the group and find
5818 * the group containing the CPU with most spare capacity.
5824 for_each_cpu(i, sched_group_span(group)) {
5825 /* Bias balancing toward CPUs of our domain */
5827 load = source_load(i, load_idx);
5829 load = target_load(i, load_idx);
5831 runnable_load += load;
5833 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5835 spare_cap = capacity_spare_without(i, p);
5837 if (spare_cap > max_spare_cap)
5838 max_spare_cap = spare_cap;
5841 /* Adjust by relative CPU capacity of the group */
5842 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5843 group->sgc->capacity;
5844 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5845 group->sgc->capacity;
5848 this_runnable_load = runnable_load;
5849 this_avg_load = avg_load;
5850 this_spare = max_spare_cap;
5852 if (min_runnable_load > (runnable_load + imbalance)) {
5854 * The runnable load is significantly smaller
5855 * so we can pick this new CPU:
5857 min_runnable_load = runnable_load;
5858 min_avg_load = avg_load;
5860 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5861 (100*min_avg_load > imbalance_scale*avg_load)) {
5863 * The runnable loads are close so take the
5864 * blocked load into account through avg_load:
5866 min_avg_load = avg_load;
5870 if (most_spare < max_spare_cap) {
5871 most_spare = max_spare_cap;
5872 most_spare_sg = group;
5875 } while (group = group->next, group != sd->groups);
5878 * The cross-over point between using spare capacity or least load
5879 * is too conservative for high utilization tasks on partially
5880 * utilized systems if we require spare_capacity > task_util(p),
5881 * so we allow for some task stuffing by using
5882 * spare_capacity > task_util(p)/2.
5884 * Spare capacity can't be used for fork because the utilization has
5885 * not been set yet, we must first select a rq to compute the initial
5888 if (sd_flag & SD_BALANCE_FORK)
5891 if (this_spare > task_util(p) / 2 &&
5892 imbalance_scale*this_spare > 100*most_spare)
5895 if (most_spare > task_util(p) / 2)
5896 return most_spare_sg;
5903 * When comparing groups across NUMA domains, it's possible for the
5904 * local domain to be very lightly loaded relative to the remote
5905 * domains but "imbalance" skews the comparison making remote CPUs
5906 * look much more favourable. When considering cross-domain, add
5907 * imbalance to the runnable load on the remote node and consider
5910 if ((sd->flags & SD_NUMA) &&
5911 min_runnable_load + imbalance >= this_runnable_load)
5914 if (min_runnable_load > (this_runnable_load + imbalance))
5917 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5918 (100*this_avg_load < imbalance_scale*min_avg_load))
5925 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5928 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5930 unsigned long load, min_load = ULONG_MAX;
5931 unsigned int min_exit_latency = UINT_MAX;
5932 u64 latest_idle_timestamp = 0;
5933 int least_loaded_cpu = this_cpu;
5934 int shallowest_idle_cpu = -1;
5937 /* Check if we have any choice: */
5938 if (group->group_weight == 1)
5939 return cpumask_first(sched_group_span(group));
5941 /* Traverse only the allowed CPUs */
5942 for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
5943 if (available_idle_cpu(i)) {
5944 struct rq *rq = cpu_rq(i);
5945 struct cpuidle_state *idle = idle_get_state(rq);
5946 if (idle && idle->exit_latency < min_exit_latency) {
5948 * We give priority to a CPU whose idle state
5949 * has the smallest exit latency irrespective
5950 * of any idle timestamp.
5952 min_exit_latency = idle->exit_latency;
5953 latest_idle_timestamp = rq->idle_stamp;
5954 shallowest_idle_cpu = i;
5955 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5956 rq->idle_stamp > latest_idle_timestamp) {
5958 * If equal or no active idle state, then
5959 * the most recently idled CPU might have
5962 latest_idle_timestamp = rq->idle_stamp;
5963 shallowest_idle_cpu = i;
5965 } else if (shallowest_idle_cpu == -1) {
5966 load = weighted_cpuload(cpu_rq(i));
5967 if (load < min_load) {
5969 least_loaded_cpu = i;
5974 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5977 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5978 int cpu, int prev_cpu, int sd_flag)
5982 if (!cpumask_intersects(sched_domain_span(sd), &p->cpus_allowed))
5986 * We need task's util for capacity_spare_without, sync it up to
5987 * prev_cpu's last_update_time.
5989 if (!(sd_flag & SD_BALANCE_FORK))
5990 sync_entity_load_avg(&p->se);
5993 struct sched_group *group;
5994 struct sched_domain *tmp;
5997 if (!(sd->flags & sd_flag)) {
6002 group = find_idlest_group(sd, p, cpu, sd_flag);
6008 new_cpu = find_idlest_group_cpu(group, p, cpu);
6009 if (new_cpu == cpu) {
6010 /* Now try balancing at a lower domain level of 'cpu': */
6015 /* Now try balancing at a lower domain level of 'new_cpu': */
6017 weight = sd->span_weight;
6019 for_each_domain(cpu, tmp) {
6020 if (weight <= tmp->span_weight)
6022 if (tmp->flags & sd_flag)
6030 #ifdef CONFIG_SCHED_SMT
6031 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
6032 EXPORT_SYMBOL_GPL(sched_smt_present);
6034 static inline void set_idle_cores(int cpu, int val)
6036 struct sched_domain_shared *sds;
6038 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6040 WRITE_ONCE(sds->has_idle_cores, val);
6043 static inline bool test_idle_cores(int cpu, bool def)
6045 struct sched_domain_shared *sds;
6047 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6049 return READ_ONCE(sds->has_idle_cores);
6055 * Scans the local SMT mask to see if the entire core is idle, and records this
6056 * information in sd_llc_shared->has_idle_cores.
6058 * Since SMT siblings share all cache levels, inspecting this limited remote
6059 * state should be fairly cheap.
6061 void __update_idle_core(struct rq *rq)
6063 int core = cpu_of(rq);
6067 if (test_idle_cores(core, true))
6070 for_each_cpu(cpu, cpu_smt_mask(core)) {
6074 if (!available_idle_cpu(cpu))
6078 set_idle_cores(core, 1);
6084 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6085 * there are no idle cores left in the system; tracked through
6086 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6088 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6090 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6093 if (!static_branch_likely(&sched_smt_present))
6096 if (!test_idle_cores(target, false))
6099 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
6101 for_each_cpu_wrap(core, cpus, target) {
6104 for_each_cpu(cpu, cpu_smt_mask(core)) {
6105 cpumask_clear_cpu(cpu, cpus);
6106 if (!available_idle_cpu(cpu))
6115 * Failed to find an idle core; stop looking for one.
6117 set_idle_cores(target, 0);
6123 * Scan the local SMT mask for idle CPUs.
6125 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6129 if (!static_branch_likely(&sched_smt_present))
6132 for_each_cpu(cpu, cpu_smt_mask(target)) {
6133 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6135 if (available_idle_cpu(cpu))
6142 #else /* CONFIG_SCHED_SMT */
6144 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6149 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6154 #endif /* CONFIG_SCHED_SMT */
6157 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6158 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6159 * average idle time for this rq (as found in rq->avg_idle).
6161 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
6163 struct sched_domain *this_sd;
6164 u64 avg_cost, avg_idle;
6167 int cpu, nr = INT_MAX;
6169 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6174 * Due to large variance we need a large fuzz factor; hackbench in
6175 * particularly is sensitive here.
6177 avg_idle = this_rq()->avg_idle / 512;
6178 avg_cost = this_sd->avg_scan_cost + 1;
6180 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
6183 if (sched_feat(SIS_PROP)) {
6184 u64 span_avg = sd->span_weight * avg_idle;
6185 if (span_avg > 4*avg_cost)
6186 nr = div_u64(span_avg, avg_cost);
6191 time = local_clock();
6193 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
6196 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6198 if (available_idle_cpu(cpu))
6202 time = local_clock() - time;
6203 cost = this_sd->avg_scan_cost;
6204 delta = (s64)(time - cost) / 8;
6205 this_sd->avg_scan_cost += delta;
6211 * Try and locate an idle core/thread in the LLC cache domain.
6213 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6215 struct sched_domain *sd;
6216 int i, recent_used_cpu;
6218 if (available_idle_cpu(target))
6222 * If the previous CPU is cache affine and idle, don't be stupid:
6224 if (prev != target && cpus_share_cache(prev, target) && available_idle_cpu(prev))
6227 /* Check a recently used CPU as a potential idle candidate: */
6228 recent_used_cpu = p->recent_used_cpu;
6229 if (recent_used_cpu != prev &&
6230 recent_used_cpu != target &&
6231 cpus_share_cache(recent_used_cpu, target) &&
6232 available_idle_cpu(recent_used_cpu) &&
6233 cpumask_test_cpu(p->recent_used_cpu, &p->cpus_allowed)) {
6235 * Replace recent_used_cpu with prev as it is a potential
6236 * candidate for the next wake:
6238 p->recent_used_cpu = prev;
6239 return recent_used_cpu;
6242 sd = rcu_dereference(per_cpu(sd_llc, target));
6246 i = select_idle_core(p, sd, target);
6247 if ((unsigned)i < nr_cpumask_bits)
6250 i = select_idle_cpu(p, sd, target);
6251 if ((unsigned)i < nr_cpumask_bits)
6254 i = select_idle_smt(p, sd, target);
6255 if ((unsigned)i < nr_cpumask_bits)
6262 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6263 * @cpu: the CPU to get the utilization of
6265 * The unit of the return value must be the one of capacity so we can compare
6266 * the utilization with the capacity of the CPU that is available for CFS task
6267 * (ie cpu_capacity).
6269 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6270 * recent utilization of currently non-runnable tasks on a CPU. It represents
6271 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6272 * capacity_orig is the cpu_capacity available at the highest frequency
6273 * (arch_scale_freq_capacity()).
6274 * The utilization of a CPU converges towards a sum equal to or less than the
6275 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6276 * the running time on this CPU scaled by capacity_curr.
6278 * The estimated utilization of a CPU is defined to be the maximum between its
6279 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6280 * currently RUNNABLE on that CPU.
6281 * This allows to properly represent the expected utilization of a CPU which
6282 * has just got a big task running since a long sleep period. At the same time
6283 * however it preserves the benefits of the "blocked utilization" in
6284 * describing the potential for other tasks waking up on the same CPU.
6286 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6287 * higher than capacity_orig because of unfortunate rounding in
6288 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6289 * the average stabilizes with the new running time. We need to check that the
6290 * utilization stays within the range of [0..capacity_orig] and cap it if
6291 * necessary. Without utilization capping, a group could be seen as overloaded
6292 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6293 * available capacity. We allow utilization to overshoot capacity_curr (but not
6294 * capacity_orig) as it useful for predicting the capacity required after task
6295 * migrations (scheduler-driven DVFS).
6297 * Return: the (estimated) utilization for the specified CPU
6299 static inline unsigned long cpu_util(int cpu)
6301 struct cfs_rq *cfs_rq;
6304 cfs_rq = &cpu_rq(cpu)->cfs;
6305 util = READ_ONCE(cfs_rq->avg.util_avg);
6307 if (sched_feat(UTIL_EST))
6308 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6310 return min_t(unsigned long, util, capacity_orig_of(cpu));
6314 * cpu_util_without: compute cpu utilization without any contributions from *p
6315 * @cpu: the CPU which utilization is requested
6316 * @p: the task which utilization should be discounted
6318 * The utilization of a CPU is defined by the utilization of tasks currently
6319 * enqueued on that CPU as well as tasks which are currently sleeping after an
6320 * execution on that CPU.
6322 * This method returns the utilization of the specified CPU by discounting the
6323 * utilization of the specified task, whenever the task is currently
6324 * contributing to the CPU utilization.
6326 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6328 struct cfs_rq *cfs_rq;
6331 /* Task has no contribution or is new */
6332 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6333 return cpu_util(cpu);
6335 cfs_rq = &cpu_rq(cpu)->cfs;
6336 util = READ_ONCE(cfs_rq->avg.util_avg);
6338 /* Discount task's util from CPU's util */
6339 lsub_positive(&util, task_util(p));
6344 * a) if *p is the only task sleeping on this CPU, then:
6345 * cpu_util (== task_util) > util_est (== 0)
6346 * and thus we return:
6347 * cpu_util_without = (cpu_util - task_util) = 0
6349 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6351 * cpu_util >= task_util
6352 * cpu_util > util_est (== 0)
6353 * and thus we discount *p's blocked utilization to return:
6354 * cpu_util_without = (cpu_util - task_util) >= 0
6356 * c) if other tasks are RUNNABLE on that CPU and
6357 * util_est > cpu_util
6358 * then we use util_est since it returns a more restrictive
6359 * estimation of the spare capacity on that CPU, by just
6360 * considering the expected utilization of tasks already
6361 * runnable on that CPU.
6363 * Cases a) and b) are covered by the above code, while case c) is
6364 * covered by the following code when estimated utilization is
6367 if (sched_feat(UTIL_EST)) {
6368 unsigned int estimated =
6369 READ_ONCE(cfs_rq->avg.util_est.enqueued);
6372 * Despite the following checks we still have a small window
6373 * for a possible race, when an execl's select_task_rq_fair()
6374 * races with LB's detach_task():
6377 * p->on_rq = TASK_ON_RQ_MIGRATING;
6378 * ---------------------------------- A
6379 * deactivate_task() \
6380 * dequeue_task() + RaceTime
6381 * util_est_dequeue() /
6382 * ---------------------------------- B
6384 * The additional check on "current == p" it's required to
6385 * properly fix the execl regression and it helps in further
6386 * reducing the chances for the above race.
6388 if (unlikely(task_on_rq_queued(p) || current == p))
6389 lsub_positive(&estimated, _task_util_est(p));
6391 util = max(util, estimated);
6395 * Utilization (estimated) can exceed the CPU capacity, thus let's
6396 * clamp to the maximum CPU capacity to ensure consistency with
6397 * the cpu_util call.
6399 return min_t(unsigned long, util, capacity_orig_of(cpu));
6403 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6404 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6406 * In that case WAKE_AFFINE doesn't make sense and we'll let
6407 * BALANCE_WAKE sort things out.
6409 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6411 long min_cap, max_cap;
6413 if (!static_branch_unlikely(&sched_asym_cpucapacity))
6416 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6417 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6419 /* Minimum capacity is close to max, no need to abort wake_affine */
6420 if (max_cap - min_cap < max_cap >> 3)
6423 /* Bring task utilization in sync with prev_cpu */
6424 sync_entity_load_avg(&p->se);
6426 return !task_fits_capacity(p, min_cap);
6430 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6433 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6435 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6436 unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
6439 * If @p migrates from @cpu to another, remove its contribution. Or,
6440 * if @p migrates from another CPU to @cpu, add its contribution. In
6441 * the other cases, @cpu is not impacted by the migration, so the
6442 * util_avg should already be correct.
6444 if (task_cpu(p) == cpu && dst_cpu != cpu)
6445 sub_positive(&util, task_util(p));
6446 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6447 util += task_util(p);
6449 if (sched_feat(UTIL_EST)) {
6450 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6453 * During wake-up, the task isn't enqueued yet and doesn't
6454 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6455 * so just add it (if needed) to "simulate" what will be
6456 * cpu_util() after the task has been enqueued.
6459 util_est += _task_util_est(p);
6461 util = max(util, util_est);
6464 return min(util, capacity_orig_of(cpu));
6468 * compute_energy(): Estimates the energy that would be consumed if @p was
6469 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6470 * landscape of the * CPUs after the task migration, and uses the Energy Model
6471 * to compute what would be the energy if we decided to actually migrate that
6475 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
6477 long util, max_util, sum_util, energy = 0;
6480 for (; pd; pd = pd->next) {
6481 max_util = sum_util = 0;
6483 * The capacity state of CPUs of the current rd can be driven by
6484 * CPUs of another rd if they belong to the same performance
6485 * domain. So, account for the utilization of these CPUs too
6486 * by masking pd with cpu_online_mask instead of the rd span.
6488 * If an entire performance domain is outside of the current rd,
6489 * it will not appear in its pd list and will not be accounted
6490 * by compute_energy().
6492 for_each_cpu_and(cpu, perf_domain_span(pd), cpu_online_mask) {
6493 util = cpu_util_next(cpu, p, dst_cpu);
6494 util = schedutil_energy_util(cpu, util);
6495 max_util = max(util, max_util);
6499 energy += em_pd_energy(pd->em_pd, max_util, sum_util);
6506 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6507 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6508 * spare capacity in each performance domain and uses it as a potential
6509 * candidate to execute the task. Then, it uses the Energy Model to figure
6510 * out which of the CPU candidates is the most energy-efficient.
6512 * The rationale for this heuristic is as follows. In a performance domain,
6513 * all the most energy efficient CPU candidates (according to the Energy
6514 * Model) are those for which we'll request a low frequency. When there are
6515 * several CPUs for which the frequency request will be the same, we don't
6516 * have enough data to break the tie between them, because the Energy Model
6517 * only includes active power costs. With this model, if we assume that
6518 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6519 * the maximum spare capacity in a performance domain is guaranteed to be among
6520 * the best candidates of the performance domain.
6522 * In practice, it could be preferable from an energy standpoint to pack
6523 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6524 * but that could also hurt our chances to go cluster idle, and we have no
6525 * ways to tell with the current Energy Model if this is actually a good
6526 * idea or not. So, find_energy_efficient_cpu() basically favors
6527 * cluster-packing, and spreading inside a cluster. That should at least be
6528 * a good thing for latency, and this is consistent with the idea that most
6529 * of the energy savings of EAS come from the asymmetry of the system, and
6530 * not so much from breaking the tie between identical CPUs. That's also the
6531 * reason why EAS is enabled in the topology code only for systems where
6532 * SD_ASYM_CPUCAPACITY is set.
6534 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6535 * they don't have any useful utilization data yet and it's not possible to
6536 * forecast their impact on energy consumption. Consequently, they will be
6537 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6538 * to be energy-inefficient in some use-cases. The alternative would be to
6539 * bias new tasks towards specific types of CPUs first, or to try to infer
6540 * their util_avg from the parent task, but those heuristics could hurt
6541 * other use-cases too. So, until someone finds a better way to solve this,
6542 * let's keep things simple by re-using the existing slow path.
6545 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6547 unsigned long prev_energy = ULONG_MAX, best_energy = ULONG_MAX;
6548 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6549 int cpu, best_energy_cpu = prev_cpu;
6550 struct perf_domain *head, *pd;
6551 unsigned long cpu_cap, util;
6552 struct sched_domain *sd;
6555 pd = rcu_dereference(rd->pd);
6556 if (!pd || READ_ONCE(rd->overutilized))
6561 * Energy-aware wake-up happens on the lowest sched_domain starting
6562 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6564 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6565 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6570 sync_entity_load_avg(&p->se);
6571 if (!task_util_est(p))
6574 for (; pd; pd = pd->next) {
6575 unsigned long cur_energy, spare_cap, max_spare_cap = 0;
6576 int max_spare_cap_cpu = -1;
6578 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
6579 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6582 /* Skip CPUs that will be overutilized. */
6583 util = cpu_util_next(cpu, p, cpu);
6584 cpu_cap = capacity_of(cpu);
6585 if (cpu_cap * 1024 < util * capacity_margin)
6588 /* Always use prev_cpu as a candidate. */
6589 if (cpu == prev_cpu) {
6590 prev_energy = compute_energy(p, prev_cpu, head);
6591 best_energy = min(best_energy, prev_energy);
6596 * Find the CPU with the maximum spare capacity in
6597 * the performance domain
6599 spare_cap = cpu_cap - util;
6600 if (spare_cap > max_spare_cap) {
6601 max_spare_cap = spare_cap;
6602 max_spare_cap_cpu = cpu;
6606 /* Evaluate the energy impact of using this CPU. */
6607 if (max_spare_cap_cpu >= 0) {
6608 cur_energy = compute_energy(p, max_spare_cap_cpu, head);
6609 if (cur_energy < best_energy) {
6610 best_energy = cur_energy;
6611 best_energy_cpu = max_spare_cap_cpu;
6619 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6620 * least 6% of the energy used by prev_cpu.
6622 if (prev_energy == ULONG_MAX)
6623 return best_energy_cpu;
6625 if ((prev_energy - best_energy) > (prev_energy >> 4))
6626 return best_energy_cpu;
6637 * select_task_rq_fair: Select target runqueue for the waking task in domains
6638 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6639 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6641 * Balances load by selecting the idlest CPU in the idlest group, or under
6642 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6644 * Returns the target CPU number.
6646 * preempt must be disabled.
6649 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6651 struct sched_domain *tmp, *sd = NULL;
6652 int cpu = smp_processor_id();
6653 int new_cpu = prev_cpu;
6654 int want_affine = 0;
6655 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6657 if (sd_flag & SD_BALANCE_WAKE) {
6660 if (sched_energy_enabled()) {
6661 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
6667 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu) &&
6668 cpumask_test_cpu(cpu, &p->cpus_allowed);
6672 for_each_domain(cpu, tmp) {
6673 if (!(tmp->flags & SD_LOAD_BALANCE))
6677 * If both 'cpu' and 'prev_cpu' are part of this domain,
6678 * cpu is a valid SD_WAKE_AFFINE target.
6680 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6681 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6682 if (cpu != prev_cpu)
6683 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6685 sd = NULL; /* Prefer wake_affine over balance flags */
6689 if (tmp->flags & sd_flag)
6691 else if (!want_affine)
6697 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6698 } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6701 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6704 current->recent_used_cpu = cpu;
6711 static void detach_entity_cfs_rq(struct sched_entity *se);
6714 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6715 * cfs_rq_of(p) references at time of call are still valid and identify the
6716 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6718 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6721 * As blocked tasks retain absolute vruntime the migration needs to
6722 * deal with this by subtracting the old and adding the new
6723 * min_vruntime -- the latter is done by enqueue_entity() when placing
6724 * the task on the new runqueue.
6726 if (p->state == TASK_WAKING) {
6727 struct sched_entity *se = &p->se;
6728 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6731 #ifndef CONFIG_64BIT
6732 u64 min_vruntime_copy;
6735 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6737 min_vruntime = cfs_rq->min_vruntime;
6738 } while (min_vruntime != min_vruntime_copy);
6740 min_vruntime = cfs_rq->min_vruntime;
6743 se->vruntime -= min_vruntime;
6746 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6748 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6749 * rq->lock and can modify state directly.
6751 lockdep_assert_held(&task_rq(p)->lock);
6752 detach_entity_cfs_rq(&p->se);
6756 * We are supposed to update the task to "current" time, then
6757 * its up to date and ready to go to new CPU/cfs_rq. But we
6758 * have difficulty in getting what current time is, so simply
6759 * throw away the out-of-date time. This will result in the
6760 * wakee task is less decayed, but giving the wakee more load
6763 remove_entity_load_avg(&p->se);
6766 /* Tell new CPU we are migrated */
6767 p->se.avg.last_update_time = 0;
6769 /* We have migrated, no longer consider this task hot */
6770 p->se.exec_start = 0;
6772 update_scan_period(p, new_cpu);
6775 static void task_dead_fair(struct task_struct *p)
6777 remove_entity_load_avg(&p->se);
6779 #endif /* CONFIG_SMP */
6781 static unsigned long wakeup_gran(struct sched_entity *se)
6783 unsigned long gran = sysctl_sched_wakeup_granularity;
6786 * Since its curr running now, convert the gran from real-time
6787 * to virtual-time in his units.
6789 * By using 'se' instead of 'curr' we penalize light tasks, so
6790 * they get preempted easier. That is, if 'se' < 'curr' then
6791 * the resulting gran will be larger, therefore penalizing the
6792 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6793 * be smaller, again penalizing the lighter task.
6795 * This is especially important for buddies when the leftmost
6796 * task is higher priority than the buddy.
6798 return calc_delta_fair(gran, se);
6802 * Should 'se' preempt 'curr'.
6816 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6818 s64 gran, vdiff = curr->vruntime - se->vruntime;
6823 gran = wakeup_gran(se);
6830 static void set_last_buddy(struct sched_entity *se)
6832 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6835 for_each_sched_entity(se) {
6836 if (SCHED_WARN_ON(!se->on_rq))
6838 cfs_rq_of(se)->last = se;
6842 static void set_next_buddy(struct sched_entity *se)
6844 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6847 for_each_sched_entity(se) {
6848 if (SCHED_WARN_ON(!se->on_rq))
6850 cfs_rq_of(se)->next = se;
6854 static void set_skip_buddy(struct sched_entity *se)
6856 for_each_sched_entity(se)
6857 cfs_rq_of(se)->skip = se;
6861 * Preempt the current task with a newly woken task if needed:
6863 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6865 struct task_struct *curr = rq->curr;
6866 struct sched_entity *se = &curr->se, *pse = &p->se;
6867 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6868 int scale = cfs_rq->nr_running >= sched_nr_latency;
6869 int next_buddy_marked = 0;
6871 if (unlikely(se == pse))
6875 * This is possible from callers such as attach_tasks(), in which we
6876 * unconditionally check_prempt_curr() after an enqueue (which may have
6877 * lead to a throttle). This both saves work and prevents false
6878 * next-buddy nomination below.
6880 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6883 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6884 set_next_buddy(pse);
6885 next_buddy_marked = 1;
6889 * We can come here with TIF_NEED_RESCHED already set from new task
6892 * Note: this also catches the edge-case of curr being in a throttled
6893 * group (e.g. via set_curr_task), since update_curr() (in the
6894 * enqueue of curr) will have resulted in resched being set. This
6895 * prevents us from potentially nominating it as a false LAST_BUDDY
6898 if (test_tsk_need_resched(curr))
6901 /* Idle tasks are by definition preempted by non-idle tasks. */
6902 if (unlikely(task_has_idle_policy(curr)) &&
6903 likely(!task_has_idle_policy(p)))
6907 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6908 * is driven by the tick):
6910 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6913 find_matching_se(&se, &pse);
6914 update_curr(cfs_rq_of(se));
6916 if (wakeup_preempt_entity(se, pse) == 1) {
6918 * Bias pick_next to pick the sched entity that is
6919 * triggering this preemption.
6921 if (!next_buddy_marked)
6922 set_next_buddy(pse);
6931 * Only set the backward buddy when the current task is still
6932 * on the rq. This can happen when a wakeup gets interleaved
6933 * with schedule on the ->pre_schedule() or idle_balance()
6934 * point, either of which can * drop the rq lock.
6936 * Also, during early boot the idle thread is in the fair class,
6937 * for obvious reasons its a bad idea to schedule back to it.
6939 if (unlikely(!se->on_rq || curr == rq->idle))
6942 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6946 static struct task_struct *
6947 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6949 struct cfs_rq *cfs_rq = &rq->cfs;
6950 struct sched_entity *se;
6951 struct task_struct *p;
6955 if (!cfs_rq->nr_running)
6958 #ifdef CONFIG_FAIR_GROUP_SCHED
6959 if (prev->sched_class != &fair_sched_class)
6963 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6964 * likely that a next task is from the same cgroup as the current.
6966 * Therefore attempt to avoid putting and setting the entire cgroup
6967 * hierarchy, only change the part that actually changes.
6971 struct sched_entity *curr = cfs_rq->curr;
6974 * Since we got here without doing put_prev_entity() we also
6975 * have to consider cfs_rq->curr. If it is still a runnable
6976 * entity, update_curr() will update its vruntime, otherwise
6977 * forget we've ever seen it.
6981 update_curr(cfs_rq);
6986 * This call to check_cfs_rq_runtime() will do the
6987 * throttle and dequeue its entity in the parent(s).
6988 * Therefore the nr_running test will indeed
6991 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6994 if (!cfs_rq->nr_running)
7001 se = pick_next_entity(cfs_rq, curr);
7002 cfs_rq = group_cfs_rq(se);
7008 * Since we haven't yet done put_prev_entity and if the selected task
7009 * is a different task than we started out with, try and touch the
7010 * least amount of cfs_rqs.
7013 struct sched_entity *pse = &prev->se;
7015 while (!(cfs_rq = is_same_group(se, pse))) {
7016 int se_depth = se->depth;
7017 int pse_depth = pse->depth;
7019 if (se_depth <= pse_depth) {
7020 put_prev_entity(cfs_rq_of(pse), pse);
7021 pse = parent_entity(pse);
7023 if (se_depth >= pse_depth) {
7024 set_next_entity(cfs_rq_of(se), se);
7025 se = parent_entity(se);
7029 put_prev_entity(cfs_rq, pse);
7030 set_next_entity(cfs_rq, se);
7037 put_prev_task(rq, prev);
7040 se = pick_next_entity(cfs_rq, NULL);
7041 set_next_entity(cfs_rq, se);
7042 cfs_rq = group_cfs_rq(se);
7047 done: __maybe_unused;
7050 * Move the next running task to the front of
7051 * the list, so our cfs_tasks list becomes MRU
7054 list_move(&p->se.group_node, &rq->cfs_tasks);
7057 if (hrtick_enabled(rq))
7058 hrtick_start_fair(rq, p);
7060 update_misfit_status(p, rq);
7065 update_misfit_status(NULL, rq);
7066 new_tasks = idle_balance(rq, rf);
7069 * Because idle_balance() releases (and re-acquires) rq->lock, it is
7070 * possible for any higher priority task to appear. In that case we
7071 * must re-start the pick_next_entity() loop.
7080 * rq is about to be idle, check if we need to update the
7081 * lost_idle_time of clock_pelt
7083 update_idle_rq_clock_pelt(rq);
7089 * Account for a descheduled task:
7091 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
7093 struct sched_entity *se = &prev->se;
7094 struct cfs_rq *cfs_rq;
7096 for_each_sched_entity(se) {
7097 cfs_rq = cfs_rq_of(se);
7098 put_prev_entity(cfs_rq, se);
7103 * sched_yield() is very simple
7105 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7107 static void yield_task_fair(struct rq *rq)
7109 struct task_struct *curr = rq->curr;
7110 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7111 struct sched_entity *se = &curr->se;
7114 * Are we the only task in the tree?
7116 if (unlikely(rq->nr_running == 1))
7119 clear_buddies(cfs_rq, se);
7121 if (curr->policy != SCHED_BATCH) {
7122 update_rq_clock(rq);
7124 * Update run-time statistics of the 'current'.
7126 update_curr(cfs_rq);
7128 * Tell update_rq_clock() that we've just updated,
7129 * so we don't do microscopic update in schedule()
7130 * and double the fastpath cost.
7132 rq_clock_skip_update(rq);
7138 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
7140 struct sched_entity *se = &p->se;
7142 /* throttled hierarchies are not runnable */
7143 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
7146 /* Tell the scheduler that we'd really like pse to run next. */
7149 yield_task_fair(rq);
7155 /**************************************************
7156 * Fair scheduling class load-balancing methods.
7160 * The purpose of load-balancing is to achieve the same basic fairness the
7161 * per-CPU scheduler provides, namely provide a proportional amount of compute
7162 * time to each task. This is expressed in the following equation:
7164 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7166 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7167 * W_i,0 is defined as:
7169 * W_i,0 = \Sum_j w_i,j (2)
7171 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7172 * is derived from the nice value as per sched_prio_to_weight[].
7174 * The weight average is an exponential decay average of the instantaneous
7177 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7179 * C_i is the compute capacity of CPU i, typically it is the
7180 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7181 * can also include other factors [XXX].
7183 * To achieve this balance we define a measure of imbalance which follows
7184 * directly from (1):
7186 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7188 * We them move tasks around to minimize the imbalance. In the continuous
7189 * function space it is obvious this converges, in the discrete case we get
7190 * a few fun cases generally called infeasible weight scenarios.
7193 * - infeasible weights;
7194 * - local vs global optima in the discrete case. ]
7199 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7200 * for all i,j solution, we create a tree of CPUs that follows the hardware
7201 * topology where each level pairs two lower groups (or better). This results
7202 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7203 * tree to only the first of the previous level and we decrease the frequency
7204 * of load-balance at each level inv. proportional to the number of CPUs in
7210 * \Sum { --- * --- * 2^i } = O(n) (5)
7212 * `- size of each group
7213 * | | `- number of CPUs doing load-balance
7215 * `- sum over all levels
7217 * Coupled with a limit on how many tasks we can migrate every balance pass,
7218 * this makes (5) the runtime complexity of the balancer.
7220 * An important property here is that each CPU is still (indirectly) connected
7221 * to every other CPU in at most O(log n) steps:
7223 * The adjacency matrix of the resulting graph is given by:
7226 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7229 * And you'll find that:
7231 * A^(log_2 n)_i,j != 0 for all i,j (7)
7233 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7234 * The task movement gives a factor of O(m), giving a convergence complexity
7237 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7242 * In order to avoid CPUs going idle while there's still work to do, new idle
7243 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7244 * tree itself instead of relying on other CPUs to bring it work.
7246 * This adds some complexity to both (5) and (8) but it reduces the total idle
7254 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7257 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7262 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7264 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7266 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7269 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7270 * rewrite all of this once again.]
7273 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7275 enum fbq_type { regular, remote, all };
7284 #define LBF_ALL_PINNED 0x01
7285 #define LBF_NEED_BREAK 0x02
7286 #define LBF_DST_PINNED 0x04
7287 #define LBF_SOME_PINNED 0x08
7288 #define LBF_NOHZ_STATS 0x10
7289 #define LBF_NOHZ_AGAIN 0x20
7292 struct sched_domain *sd;
7300 struct cpumask *dst_grpmask;
7302 enum cpu_idle_type idle;
7304 /* The set of CPUs under consideration for load-balancing */
7305 struct cpumask *cpus;
7310 unsigned int loop_break;
7311 unsigned int loop_max;
7313 enum fbq_type fbq_type;
7314 enum group_type src_grp_type;
7315 struct list_head tasks;
7319 * Is this task likely cache-hot:
7321 static int task_hot(struct task_struct *p, struct lb_env *env)
7325 lockdep_assert_held(&env->src_rq->lock);
7327 if (p->sched_class != &fair_sched_class)
7330 if (unlikely(task_has_idle_policy(p)))
7334 * Buddy candidates are cache hot:
7336 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7337 (&p->se == cfs_rq_of(&p->se)->next ||
7338 &p->se == cfs_rq_of(&p->se)->last))
7341 if (sysctl_sched_migration_cost == -1)
7343 if (sysctl_sched_migration_cost == 0)
7346 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7348 return delta < (s64)sysctl_sched_migration_cost;
7351 #ifdef CONFIG_NUMA_BALANCING
7353 * Returns 1, if task migration degrades locality
7354 * Returns 0, if task migration improves locality i.e migration preferred.
7355 * Returns -1, if task migration is not affected by locality.
7357 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7359 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7360 unsigned long src_weight, dst_weight;
7361 int src_nid, dst_nid, dist;
7363 if (!static_branch_likely(&sched_numa_balancing))
7366 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7369 src_nid = cpu_to_node(env->src_cpu);
7370 dst_nid = cpu_to_node(env->dst_cpu);
7372 if (src_nid == dst_nid)
7375 /* Migrating away from the preferred node is always bad. */
7376 if (src_nid == p->numa_preferred_nid) {
7377 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7383 /* Encourage migration to the preferred node. */
7384 if (dst_nid == p->numa_preferred_nid)
7387 /* Leaving a core idle is often worse than degrading locality. */
7388 if (env->idle == CPU_IDLE)
7391 dist = node_distance(src_nid, dst_nid);
7393 src_weight = group_weight(p, src_nid, dist);
7394 dst_weight = group_weight(p, dst_nid, dist);
7396 src_weight = task_weight(p, src_nid, dist);
7397 dst_weight = task_weight(p, dst_nid, dist);
7400 return dst_weight < src_weight;
7404 static inline int migrate_degrades_locality(struct task_struct *p,
7412 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7415 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7419 lockdep_assert_held(&env->src_rq->lock);
7422 * We do not migrate tasks that are:
7423 * 1) throttled_lb_pair, or
7424 * 2) cannot be migrated to this CPU due to cpus_allowed, or
7425 * 3) running (obviously), or
7426 * 4) are cache-hot on their current CPU.
7428 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7431 if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) {
7434 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7436 env->flags |= LBF_SOME_PINNED;
7439 * Remember if this task can be migrated to any other CPU in
7440 * our sched_group. We may want to revisit it if we couldn't
7441 * meet load balance goals by pulling other tasks on src_cpu.
7443 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7444 * already computed one in current iteration.
7446 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7449 /* Prevent to re-select dst_cpu via env's CPUs: */
7450 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7451 if (cpumask_test_cpu(cpu, &p->cpus_allowed)) {
7452 env->flags |= LBF_DST_PINNED;
7453 env->new_dst_cpu = cpu;
7461 /* Record that we found atleast one task that could run on dst_cpu */
7462 env->flags &= ~LBF_ALL_PINNED;
7464 if (task_running(env->src_rq, p)) {
7465 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7470 * Aggressive migration if:
7471 * 1) destination numa is preferred
7472 * 2) task is cache cold, or
7473 * 3) too many balance attempts have failed.
7475 tsk_cache_hot = migrate_degrades_locality(p, env);
7476 if (tsk_cache_hot == -1)
7477 tsk_cache_hot = task_hot(p, env);
7479 if (tsk_cache_hot <= 0 ||
7480 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7481 if (tsk_cache_hot == 1) {
7482 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7483 schedstat_inc(p->se.statistics.nr_forced_migrations);
7488 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7493 * detach_task() -- detach the task for the migration specified in env
7495 static void detach_task(struct task_struct *p, struct lb_env *env)
7497 lockdep_assert_held(&env->src_rq->lock);
7499 p->on_rq = TASK_ON_RQ_MIGRATING;
7500 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7501 set_task_cpu(p, env->dst_cpu);
7505 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7506 * part of active balancing operations within "domain".
7508 * Returns a task if successful and NULL otherwise.
7510 static struct task_struct *detach_one_task(struct lb_env *env)
7512 struct task_struct *p;
7514 lockdep_assert_held(&env->src_rq->lock);
7516 list_for_each_entry_reverse(p,
7517 &env->src_rq->cfs_tasks, se.group_node) {
7518 if (!can_migrate_task(p, env))
7521 detach_task(p, env);
7524 * Right now, this is only the second place where
7525 * lb_gained[env->idle] is updated (other is detach_tasks)
7526 * so we can safely collect stats here rather than
7527 * inside detach_tasks().
7529 schedstat_inc(env->sd->lb_gained[env->idle]);
7535 static const unsigned int sched_nr_migrate_break = 32;
7538 * detach_tasks() -- tries to detach up to imbalance weighted load from
7539 * busiest_rq, as part of a balancing operation within domain "sd".
7541 * Returns number of detached tasks if successful and 0 otherwise.
7543 static int detach_tasks(struct lb_env *env)
7545 struct list_head *tasks = &env->src_rq->cfs_tasks;
7546 struct task_struct *p;
7550 lockdep_assert_held(&env->src_rq->lock);
7552 if (env->imbalance <= 0)
7555 while (!list_empty(tasks)) {
7557 * We don't want to steal all, otherwise we may be treated likewise,
7558 * which could at worst lead to a livelock crash.
7560 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7563 p = list_last_entry(tasks, struct task_struct, se.group_node);
7566 /* We've more or less seen every task there is, call it quits */
7567 if (env->loop > env->loop_max)
7570 /* take a breather every nr_migrate tasks */
7571 if (env->loop > env->loop_break) {
7572 env->loop_break += sched_nr_migrate_break;
7573 env->flags |= LBF_NEED_BREAK;
7577 if (!can_migrate_task(p, env))
7580 load = task_h_load(p);
7582 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7585 if ((load / 2) > env->imbalance)
7588 detach_task(p, env);
7589 list_add(&p->se.group_node, &env->tasks);
7592 env->imbalance -= load;
7594 #ifdef CONFIG_PREEMPT
7596 * NEWIDLE balancing is a source of latency, so preemptible
7597 * kernels will stop after the first task is detached to minimize
7598 * the critical section.
7600 if (env->idle == CPU_NEWLY_IDLE)
7605 * We only want to steal up to the prescribed amount of
7608 if (env->imbalance <= 0)
7613 list_move(&p->se.group_node, tasks);
7617 * Right now, this is one of only two places we collect this stat
7618 * so we can safely collect detach_one_task() stats here rather
7619 * than inside detach_one_task().
7621 schedstat_add(env->sd->lb_gained[env->idle], detached);
7627 * attach_task() -- attach the task detached by detach_task() to its new rq.
7629 static void attach_task(struct rq *rq, struct task_struct *p)
7631 lockdep_assert_held(&rq->lock);
7633 BUG_ON(task_rq(p) != rq);
7634 activate_task(rq, p, ENQUEUE_NOCLOCK);
7635 p->on_rq = TASK_ON_RQ_QUEUED;
7636 check_preempt_curr(rq, p, 0);
7640 * attach_one_task() -- attaches the task returned from detach_one_task() to
7643 static void attach_one_task(struct rq *rq, struct task_struct *p)
7648 update_rq_clock(rq);
7654 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7657 static void attach_tasks(struct lb_env *env)
7659 struct list_head *tasks = &env->tasks;
7660 struct task_struct *p;
7663 rq_lock(env->dst_rq, &rf);
7664 update_rq_clock(env->dst_rq);
7666 while (!list_empty(tasks)) {
7667 p = list_first_entry(tasks, struct task_struct, se.group_node);
7668 list_del_init(&p->se.group_node);
7670 attach_task(env->dst_rq, p);
7673 rq_unlock(env->dst_rq, &rf);
7676 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7678 if (cfs_rq->avg.load_avg)
7681 if (cfs_rq->avg.util_avg)
7687 static inline bool others_have_blocked(struct rq *rq)
7689 if (READ_ONCE(rq->avg_rt.util_avg))
7692 if (READ_ONCE(rq->avg_dl.util_avg))
7695 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7696 if (READ_ONCE(rq->avg_irq.util_avg))
7703 #ifdef CONFIG_FAIR_GROUP_SCHED
7705 static void update_blocked_averages(int cpu)
7707 struct rq *rq = cpu_rq(cpu);
7708 struct cfs_rq *cfs_rq;
7709 const struct sched_class *curr_class;
7713 rq_lock_irqsave(rq, &rf);
7714 update_rq_clock(rq);
7717 * Iterates the task_group tree in a bottom up fashion, see
7718 * list_add_leaf_cfs_rq() for details.
7720 for_each_leaf_cfs_rq(rq, cfs_rq) {
7721 struct sched_entity *se;
7723 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq))
7724 update_tg_load_avg(cfs_rq, 0);
7726 /* Propagate pending load changes to the parent, if any: */
7727 se = cfs_rq->tg->se[cpu];
7728 if (se && !skip_blocked_update(se))
7729 update_load_avg(cfs_rq_of(se), se, 0);
7731 /* Don't need periodic decay once load/util_avg are null */
7732 if (cfs_rq_has_blocked(cfs_rq))
7736 curr_class = rq->curr->sched_class;
7737 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class);
7738 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class);
7739 update_irq_load_avg(rq, 0);
7740 /* Don't need periodic decay once load/util_avg are null */
7741 if (others_have_blocked(rq))
7744 #ifdef CONFIG_NO_HZ_COMMON
7745 rq->last_blocked_load_update_tick = jiffies;
7747 rq->has_blocked_load = 0;
7749 rq_unlock_irqrestore(rq, &rf);
7753 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7754 * This needs to be done in a top-down fashion because the load of a child
7755 * group is a fraction of its parents load.
7757 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7759 struct rq *rq = rq_of(cfs_rq);
7760 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7761 unsigned long now = jiffies;
7764 if (cfs_rq->last_h_load_update == now)
7767 cfs_rq->h_load_next = NULL;
7768 for_each_sched_entity(se) {
7769 cfs_rq = cfs_rq_of(se);
7770 cfs_rq->h_load_next = se;
7771 if (cfs_rq->last_h_load_update == now)
7776 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7777 cfs_rq->last_h_load_update = now;
7780 while ((se = cfs_rq->h_load_next) != NULL) {
7781 load = cfs_rq->h_load;
7782 load = div64_ul(load * se->avg.load_avg,
7783 cfs_rq_load_avg(cfs_rq) + 1);
7784 cfs_rq = group_cfs_rq(se);
7785 cfs_rq->h_load = load;
7786 cfs_rq->last_h_load_update = now;
7790 static unsigned long task_h_load(struct task_struct *p)
7792 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7794 update_cfs_rq_h_load(cfs_rq);
7795 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7796 cfs_rq_load_avg(cfs_rq) + 1);
7799 static inline void update_blocked_averages(int cpu)
7801 struct rq *rq = cpu_rq(cpu);
7802 struct cfs_rq *cfs_rq = &rq->cfs;
7803 const struct sched_class *curr_class;
7806 rq_lock_irqsave(rq, &rf);
7807 update_rq_clock(rq);
7808 update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
7810 curr_class = rq->curr->sched_class;
7811 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class);
7812 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class);
7813 update_irq_load_avg(rq, 0);
7814 #ifdef CONFIG_NO_HZ_COMMON
7815 rq->last_blocked_load_update_tick = jiffies;
7816 if (!cfs_rq_has_blocked(cfs_rq) && !others_have_blocked(rq))
7817 rq->has_blocked_load = 0;
7819 rq_unlock_irqrestore(rq, &rf);
7822 static unsigned long task_h_load(struct task_struct *p)
7824 return p->se.avg.load_avg;
7828 /********** Helpers for find_busiest_group ************************/
7831 * sg_lb_stats - stats of a sched_group required for load_balancing
7833 struct sg_lb_stats {
7834 unsigned long avg_load; /*Avg load across the CPUs of the group */
7835 unsigned long group_load; /* Total load over the CPUs of the group */
7836 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7837 unsigned long load_per_task;
7838 unsigned long group_capacity;
7839 unsigned long group_util; /* Total utilization of the group */
7840 unsigned int sum_nr_running; /* Nr tasks running in the group */
7841 unsigned int idle_cpus;
7842 unsigned int group_weight;
7843 enum group_type group_type;
7844 int group_no_capacity;
7845 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
7846 #ifdef CONFIG_NUMA_BALANCING
7847 unsigned int nr_numa_running;
7848 unsigned int nr_preferred_running;
7853 * sd_lb_stats - Structure to store the statistics of a sched_domain
7854 * during load balancing.
7856 struct sd_lb_stats {
7857 struct sched_group *busiest; /* Busiest group in this sd */
7858 struct sched_group *local; /* Local group in this sd */
7859 unsigned long total_running;
7860 unsigned long total_load; /* Total load of all groups in sd */
7861 unsigned long total_capacity; /* Total capacity of all groups in sd */
7862 unsigned long avg_load; /* Average load across all groups in sd */
7864 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7865 struct sg_lb_stats local_stat; /* Statistics of the local group */
7868 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7871 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7872 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7873 * We must however clear busiest_stat::avg_load because
7874 * update_sd_pick_busiest() reads this before assignment.
7876 *sds = (struct sd_lb_stats){
7879 .total_running = 0UL,
7881 .total_capacity = 0UL,
7884 .sum_nr_running = 0,
7885 .group_type = group_other,
7891 * get_sd_load_idx - Obtain the load index for a given sched domain.
7892 * @sd: The sched_domain whose load_idx is to be obtained.
7893 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7895 * Return: The load index.
7897 static inline int get_sd_load_idx(struct sched_domain *sd,
7898 enum cpu_idle_type idle)
7904 load_idx = sd->busy_idx;
7907 case CPU_NEWLY_IDLE:
7908 load_idx = sd->newidle_idx;
7911 load_idx = sd->idle_idx;
7918 static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu)
7920 struct rq *rq = cpu_rq(cpu);
7921 unsigned long max = arch_scale_cpu_capacity(sd, cpu);
7922 unsigned long used, free;
7925 irq = cpu_util_irq(rq);
7927 if (unlikely(irq >= max))
7930 used = READ_ONCE(rq->avg_rt.util_avg);
7931 used += READ_ONCE(rq->avg_dl.util_avg);
7933 if (unlikely(used >= max))
7938 return scale_irq_capacity(free, irq, max);
7941 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7943 unsigned long capacity = scale_rt_capacity(sd, cpu);
7944 struct sched_group *sdg = sd->groups;
7946 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(sd, cpu);
7951 cpu_rq(cpu)->cpu_capacity = capacity;
7952 sdg->sgc->capacity = capacity;
7953 sdg->sgc->min_capacity = capacity;
7954 sdg->sgc->max_capacity = capacity;
7957 void update_group_capacity(struct sched_domain *sd, int cpu)
7959 struct sched_domain *child = sd->child;
7960 struct sched_group *group, *sdg = sd->groups;
7961 unsigned long capacity, min_capacity, max_capacity;
7962 unsigned long interval;
7964 interval = msecs_to_jiffies(sd->balance_interval);
7965 interval = clamp(interval, 1UL, max_load_balance_interval);
7966 sdg->sgc->next_update = jiffies + interval;
7969 update_cpu_capacity(sd, cpu);
7974 min_capacity = ULONG_MAX;
7977 if (child->flags & SD_OVERLAP) {
7979 * SD_OVERLAP domains cannot assume that child groups
7980 * span the current group.
7983 for_each_cpu(cpu, sched_group_span(sdg)) {
7984 struct sched_group_capacity *sgc;
7985 struct rq *rq = cpu_rq(cpu);
7988 * build_sched_domains() -> init_sched_groups_capacity()
7989 * gets here before we've attached the domains to the
7992 * Use capacity_of(), which is set irrespective of domains
7993 * in update_cpu_capacity().
7995 * This avoids capacity from being 0 and
7996 * causing divide-by-zero issues on boot.
7998 if (unlikely(!rq->sd)) {
7999 capacity += capacity_of(cpu);
8001 sgc = rq->sd->groups->sgc;
8002 capacity += sgc->capacity;
8005 min_capacity = min(capacity, min_capacity);
8006 max_capacity = max(capacity, max_capacity);
8010 * !SD_OVERLAP domains can assume that child groups
8011 * span the current group.
8014 group = child->groups;
8016 struct sched_group_capacity *sgc = group->sgc;
8018 capacity += sgc->capacity;
8019 min_capacity = min(sgc->min_capacity, min_capacity);
8020 max_capacity = max(sgc->max_capacity, max_capacity);
8021 group = group->next;
8022 } while (group != child->groups);
8025 sdg->sgc->capacity = capacity;
8026 sdg->sgc->min_capacity = min_capacity;
8027 sdg->sgc->max_capacity = max_capacity;
8031 * Check whether the capacity of the rq has been noticeably reduced by side
8032 * activity. The imbalance_pct is used for the threshold.
8033 * Return true is the capacity is reduced
8036 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
8038 return ((rq->cpu_capacity * sd->imbalance_pct) <
8039 (rq->cpu_capacity_orig * 100));
8043 * Group imbalance indicates (and tries to solve) the problem where balancing
8044 * groups is inadequate due to ->cpus_allowed constraints.
8046 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
8047 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
8050 * { 0 1 2 3 } { 4 5 6 7 }
8053 * If we were to balance group-wise we'd place two tasks in the first group and
8054 * two tasks in the second group. Clearly this is undesired as it will overload
8055 * cpu 3 and leave one of the CPUs in the second group unused.
8057 * The current solution to this issue is detecting the skew in the first group
8058 * by noticing the lower domain failed to reach balance and had difficulty
8059 * moving tasks due to affinity constraints.
8061 * When this is so detected; this group becomes a candidate for busiest; see
8062 * update_sd_pick_busiest(). And calculate_imbalance() and
8063 * find_busiest_group() avoid some of the usual balance conditions to allow it
8064 * to create an effective group imbalance.
8066 * This is a somewhat tricky proposition since the next run might not find the
8067 * group imbalance and decide the groups need to be balanced again. A most
8068 * subtle and fragile situation.
8071 static inline int sg_imbalanced(struct sched_group *group)
8073 return group->sgc->imbalance;
8077 * group_has_capacity returns true if the group has spare capacity that could
8078 * be used by some tasks.
8079 * We consider that a group has spare capacity if the * number of task is
8080 * smaller than the number of CPUs or if the utilization is lower than the
8081 * available capacity for CFS tasks.
8082 * For the latter, we use a threshold to stabilize the state, to take into
8083 * account the variance of the tasks' load and to return true if the available
8084 * capacity in meaningful for the load balancer.
8085 * As an example, an available capacity of 1% can appear but it doesn't make
8086 * any benefit for the load balance.
8089 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
8091 if (sgs->sum_nr_running < sgs->group_weight)
8094 if ((sgs->group_capacity * 100) >
8095 (sgs->group_util * env->sd->imbalance_pct))
8102 * group_is_overloaded returns true if the group has more tasks than it can
8104 * group_is_overloaded is not equals to !group_has_capacity because a group
8105 * with the exact right number of tasks, has no more spare capacity but is not
8106 * overloaded so both group_has_capacity and group_is_overloaded return
8110 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
8112 if (sgs->sum_nr_running <= sgs->group_weight)
8115 if ((sgs->group_capacity * 100) <
8116 (sgs->group_util * env->sd->imbalance_pct))
8123 * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller
8124 * per-CPU capacity than sched_group ref.
8127 group_smaller_min_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
8129 return sg->sgc->min_capacity * capacity_margin <
8130 ref->sgc->min_capacity * 1024;
8134 * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller
8135 * per-CPU capacity_orig than sched_group ref.
8138 group_smaller_max_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
8140 return sg->sgc->max_capacity * capacity_margin <
8141 ref->sgc->max_capacity * 1024;
8145 group_type group_classify(struct sched_group *group,
8146 struct sg_lb_stats *sgs)
8148 if (sgs->group_no_capacity)
8149 return group_overloaded;
8151 if (sg_imbalanced(group))
8152 return group_imbalanced;
8154 if (sgs->group_misfit_task_load)
8155 return group_misfit_task;
8160 static bool update_nohz_stats(struct rq *rq, bool force)
8162 #ifdef CONFIG_NO_HZ_COMMON
8163 unsigned int cpu = rq->cpu;
8165 if (!rq->has_blocked_load)
8168 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
8171 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
8174 update_blocked_averages(cpu);
8176 return rq->has_blocked_load;
8183 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8184 * @env: The load balancing environment.
8185 * @group: sched_group whose statistics are to be updated.
8186 * @sgs: variable to hold the statistics for this group.
8187 * @sg_status: Holds flag indicating the status of the sched_group
8189 static inline void update_sg_lb_stats(struct lb_env *env,
8190 struct sched_group *group,
8191 struct sg_lb_stats *sgs,
8194 int local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(group));
8195 int load_idx = get_sd_load_idx(env->sd, env->idle);
8199 memset(sgs, 0, sizeof(*sgs));
8201 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8202 struct rq *rq = cpu_rq(i);
8204 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
8205 env->flags |= LBF_NOHZ_AGAIN;
8207 /* Bias balancing toward CPUs of our domain: */
8209 load = target_load(i, load_idx);
8211 load = source_load(i, load_idx);
8213 sgs->group_load += load;
8214 sgs->group_util += cpu_util(i);
8215 sgs->sum_nr_running += rq->cfs.h_nr_running;
8217 nr_running = rq->nr_running;
8219 *sg_status |= SG_OVERLOAD;
8221 if (cpu_overutilized(i))
8222 *sg_status |= SG_OVERUTILIZED;
8224 #ifdef CONFIG_NUMA_BALANCING
8225 sgs->nr_numa_running += rq->nr_numa_running;
8226 sgs->nr_preferred_running += rq->nr_preferred_running;
8228 sgs->sum_weighted_load += weighted_cpuload(rq);
8230 * No need to call idle_cpu() if nr_running is not 0
8232 if (!nr_running && idle_cpu(i))
8235 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8236 sgs->group_misfit_task_load < rq->misfit_task_load) {
8237 sgs->group_misfit_task_load = rq->misfit_task_load;
8238 *sg_status |= SG_OVERLOAD;
8242 /* Adjust by relative CPU capacity of the group */
8243 sgs->group_capacity = group->sgc->capacity;
8244 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
8246 if (sgs->sum_nr_running)
8247 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
8249 sgs->group_weight = group->group_weight;
8251 sgs->group_no_capacity = group_is_overloaded(env, sgs);
8252 sgs->group_type = group_classify(group, sgs);
8256 * update_sd_pick_busiest - return 1 on busiest group
8257 * @env: The load balancing environment.
8258 * @sds: sched_domain statistics
8259 * @sg: sched_group candidate to be checked for being the busiest
8260 * @sgs: sched_group statistics
8262 * Determine if @sg is a busier group than the previously selected
8265 * Return: %true if @sg is a busier group than the previously selected
8266 * busiest group. %false otherwise.
8268 static bool update_sd_pick_busiest(struct lb_env *env,
8269 struct sd_lb_stats *sds,
8270 struct sched_group *sg,
8271 struct sg_lb_stats *sgs)
8273 struct sg_lb_stats *busiest = &sds->busiest_stat;
8276 * Don't try to pull misfit tasks we can't help.
8277 * We can use max_capacity here as reduction in capacity on some
8278 * CPUs in the group should either be possible to resolve
8279 * internally or be covered by avg_load imbalance (eventually).
8281 if (sgs->group_type == group_misfit_task &&
8282 (!group_smaller_max_cpu_capacity(sg, sds->local) ||
8283 !group_has_capacity(env, &sds->local_stat)))
8286 if (sgs->group_type > busiest->group_type)
8289 if (sgs->group_type < busiest->group_type)
8292 if (sgs->avg_load <= busiest->avg_load)
8295 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
8299 * Candidate sg has no more than one task per CPU and
8300 * has higher per-CPU capacity. Migrating tasks to less
8301 * capable CPUs may harm throughput. Maximize throughput,
8302 * power/energy consequences are not considered.
8304 if (sgs->sum_nr_running <= sgs->group_weight &&
8305 group_smaller_min_cpu_capacity(sds->local, sg))
8309 * If we have more than one misfit sg go with the biggest misfit.
8311 if (sgs->group_type == group_misfit_task &&
8312 sgs->group_misfit_task_load < busiest->group_misfit_task_load)
8316 /* This is the busiest node in its class. */
8317 if (!(env->sd->flags & SD_ASYM_PACKING))
8320 /* No ASYM_PACKING if target CPU is already busy */
8321 if (env->idle == CPU_NOT_IDLE)
8324 * ASYM_PACKING needs to move all the work to the highest
8325 * prority CPUs in the group, therefore mark all groups
8326 * of lower priority than ourself as busy.
8328 if (sgs->sum_nr_running &&
8329 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
8333 /* Prefer to move from lowest priority CPU's work */
8334 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
8335 sg->asym_prefer_cpu))
8342 #ifdef CONFIG_NUMA_BALANCING
8343 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8345 if (sgs->sum_nr_running > sgs->nr_numa_running)
8347 if (sgs->sum_nr_running > sgs->nr_preferred_running)
8352 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8354 if (rq->nr_running > rq->nr_numa_running)
8356 if (rq->nr_running > rq->nr_preferred_running)
8361 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8366 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8370 #endif /* CONFIG_NUMA_BALANCING */
8373 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8374 * @env: The load balancing environment.
8375 * @sds: variable to hold the statistics for this sched_domain.
8377 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8379 struct sched_domain *child = env->sd->child;
8380 struct sched_group *sg = env->sd->groups;
8381 struct sg_lb_stats *local = &sds->local_stat;
8382 struct sg_lb_stats tmp_sgs;
8383 bool prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
8386 #ifdef CONFIG_NO_HZ_COMMON
8387 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
8388 env->flags |= LBF_NOHZ_STATS;
8392 struct sg_lb_stats *sgs = &tmp_sgs;
8395 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
8400 if (env->idle != CPU_NEWLY_IDLE ||
8401 time_after_eq(jiffies, sg->sgc->next_update))
8402 update_group_capacity(env->sd, env->dst_cpu);
8405 update_sg_lb_stats(env, sg, sgs, &sg_status);
8411 * In case the child domain prefers tasks go to siblings
8412 * first, lower the sg capacity so that we'll try
8413 * and move all the excess tasks away. We lower the capacity
8414 * of a group only if the local group has the capacity to fit
8415 * these excess tasks. The extra check prevents the case where
8416 * you always pull from the heaviest group when it is already
8417 * under-utilized (possible with a large weight task outweighs
8418 * the tasks on the system).
8420 if (prefer_sibling && sds->local &&
8421 group_has_capacity(env, local) &&
8422 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
8423 sgs->group_no_capacity = 1;
8424 sgs->group_type = group_classify(sg, sgs);
8427 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8429 sds->busiest_stat = *sgs;
8433 /* Now, start updating sd_lb_stats */
8434 sds->total_running += sgs->sum_nr_running;
8435 sds->total_load += sgs->group_load;
8436 sds->total_capacity += sgs->group_capacity;
8439 } while (sg != env->sd->groups);
8441 #ifdef CONFIG_NO_HZ_COMMON
8442 if ((env->flags & LBF_NOHZ_AGAIN) &&
8443 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8445 WRITE_ONCE(nohz.next_blocked,
8446 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8450 if (env->sd->flags & SD_NUMA)
8451 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8453 if (!env->sd->parent) {
8454 struct root_domain *rd = env->dst_rq->rd;
8456 /* update overload indicator if we are at root domain */
8457 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
8459 /* Update over-utilization (tipping point, U >= 0) indicator */
8460 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
8461 } else if (sg_status & SG_OVERUTILIZED) {
8462 WRITE_ONCE(env->dst_rq->rd->overutilized, SG_OVERUTILIZED);
8467 * check_asym_packing - Check to see if the group is packed into the
8470 * This is primarily intended to used at the sibling level. Some
8471 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8472 * case of POWER7, it can move to lower SMT modes only when higher
8473 * threads are idle. When in lower SMT modes, the threads will
8474 * perform better since they share less core resources. Hence when we
8475 * have idle threads, we want them to be the higher ones.
8477 * This packing function is run on idle threads. It checks to see if
8478 * the busiest CPU in this domain (core in the P7 case) has a higher
8479 * CPU number than the packing function is being run on. Here we are
8480 * assuming lower CPU number will be equivalent to lower a SMT thread
8483 * Return: 1 when packing is required and a task should be moved to
8484 * this CPU. The amount of the imbalance is returned in env->imbalance.
8486 * @env: The load balancing environment.
8487 * @sds: Statistics of the sched_domain which is to be packed
8489 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8493 if (!(env->sd->flags & SD_ASYM_PACKING))
8496 if (env->idle == CPU_NOT_IDLE)
8502 busiest_cpu = sds->busiest->asym_prefer_cpu;
8503 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
8506 env->imbalance = sds->busiest_stat.group_load;
8512 * fix_small_imbalance - Calculate the minor imbalance that exists
8513 * amongst the groups of a sched_domain, during
8515 * @env: The load balancing environment.
8516 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8519 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8521 unsigned long tmp, capa_now = 0, capa_move = 0;
8522 unsigned int imbn = 2;
8523 unsigned long scaled_busy_load_per_task;
8524 struct sg_lb_stats *local, *busiest;
8526 local = &sds->local_stat;
8527 busiest = &sds->busiest_stat;
8529 if (!local->sum_nr_running)
8530 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8531 else if (busiest->load_per_task > local->load_per_task)
8534 scaled_busy_load_per_task =
8535 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8536 busiest->group_capacity;
8538 if (busiest->avg_load + scaled_busy_load_per_task >=
8539 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8540 env->imbalance = busiest->load_per_task;
8545 * OK, we don't have enough imbalance to justify moving tasks,
8546 * however we may be able to increase total CPU capacity used by
8550 capa_now += busiest->group_capacity *
8551 min(busiest->load_per_task, busiest->avg_load);
8552 capa_now += local->group_capacity *
8553 min(local->load_per_task, local->avg_load);
8554 capa_now /= SCHED_CAPACITY_SCALE;
8556 /* Amount of load we'd subtract */
8557 if (busiest->avg_load > scaled_busy_load_per_task) {
8558 capa_move += busiest->group_capacity *
8559 min(busiest->load_per_task,
8560 busiest->avg_load - scaled_busy_load_per_task);
8563 /* Amount of load we'd add */
8564 if (busiest->avg_load * busiest->group_capacity <
8565 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8566 tmp = (busiest->avg_load * busiest->group_capacity) /
8567 local->group_capacity;
8569 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8570 local->group_capacity;
8572 capa_move += local->group_capacity *
8573 min(local->load_per_task, local->avg_load + tmp);
8574 capa_move /= SCHED_CAPACITY_SCALE;
8576 /* Move if we gain throughput */
8577 if (capa_move > capa_now)
8578 env->imbalance = busiest->load_per_task;
8582 * calculate_imbalance - Calculate the amount of imbalance present within the
8583 * groups of a given sched_domain during load balance.
8584 * @env: load balance environment
8585 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8587 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8589 unsigned long max_pull, load_above_capacity = ~0UL;
8590 struct sg_lb_stats *local, *busiest;
8592 local = &sds->local_stat;
8593 busiest = &sds->busiest_stat;
8595 if (busiest->group_type == group_imbalanced) {
8597 * In the group_imb case we cannot rely on group-wide averages
8598 * to ensure CPU-load equilibrium, look at wider averages. XXX
8600 busiest->load_per_task =
8601 min(busiest->load_per_task, sds->avg_load);
8605 * Avg load of busiest sg can be less and avg load of local sg can
8606 * be greater than avg load across all sgs of sd because avg load
8607 * factors in sg capacity and sgs with smaller group_type are
8608 * skipped when updating the busiest sg:
8610 if (busiest->group_type != group_misfit_task &&
8611 (busiest->avg_load <= sds->avg_load ||
8612 local->avg_load >= sds->avg_load)) {
8614 return fix_small_imbalance(env, sds);
8618 * If there aren't any idle CPUs, avoid creating some.
8620 if (busiest->group_type == group_overloaded &&
8621 local->group_type == group_overloaded) {
8622 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
8623 if (load_above_capacity > busiest->group_capacity) {
8624 load_above_capacity -= busiest->group_capacity;
8625 load_above_capacity *= scale_load_down(NICE_0_LOAD);
8626 load_above_capacity /= busiest->group_capacity;
8628 load_above_capacity = ~0UL;
8632 * We're trying to get all the CPUs to the average_load, so we don't
8633 * want to push ourselves above the average load, nor do we wish to
8634 * reduce the max loaded CPU below the average load. At the same time,
8635 * we also don't want to reduce the group load below the group
8636 * capacity. Thus we look for the minimum possible imbalance.
8638 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8640 /* How much load to actually move to equalise the imbalance */
8641 env->imbalance = min(
8642 max_pull * busiest->group_capacity,
8643 (sds->avg_load - local->avg_load) * local->group_capacity
8644 ) / SCHED_CAPACITY_SCALE;
8646 /* Boost imbalance to allow misfit task to be balanced. */
8647 if (busiest->group_type == group_misfit_task) {
8648 env->imbalance = max_t(long, env->imbalance,
8649 busiest->group_misfit_task_load);
8653 * if *imbalance is less than the average load per runnable task
8654 * there is no guarantee that any tasks will be moved so we'll have
8655 * a think about bumping its value to force at least one task to be
8658 if (env->imbalance < busiest->load_per_task)
8659 return fix_small_imbalance(env, sds);
8662 /******* find_busiest_group() helpers end here *********************/
8665 * find_busiest_group - Returns the busiest group within the sched_domain
8666 * if there is an imbalance.
8668 * Also calculates the amount of weighted load which should be moved
8669 * to restore balance.
8671 * @env: The load balancing environment.
8673 * Return: - The busiest group if imbalance exists.
8675 static struct sched_group *find_busiest_group(struct lb_env *env)
8677 struct sg_lb_stats *local, *busiest;
8678 struct sd_lb_stats sds;
8680 init_sd_lb_stats(&sds);
8683 * Compute the various statistics relavent for load balancing at
8686 update_sd_lb_stats(env, &sds);
8688 if (sched_energy_enabled()) {
8689 struct root_domain *rd = env->dst_rq->rd;
8691 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
8695 local = &sds.local_stat;
8696 busiest = &sds.busiest_stat;
8698 /* ASYM feature bypasses nice load balance check */
8699 if (check_asym_packing(env, &sds))
8702 /* There is no busy sibling group to pull tasks from */
8703 if (!sds.busiest || busiest->sum_nr_running == 0)
8706 /* XXX broken for overlapping NUMA groups */
8707 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8708 / sds.total_capacity;
8711 * If the busiest group is imbalanced the below checks don't
8712 * work because they assume all things are equal, which typically
8713 * isn't true due to cpus_allowed constraints and the like.
8715 if (busiest->group_type == group_imbalanced)
8719 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8720 * capacities from resulting in underutilization due to avg_load.
8722 if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8723 busiest->group_no_capacity)
8726 /* Misfit tasks should be dealt with regardless of the avg load */
8727 if (busiest->group_type == group_misfit_task)
8731 * If the local group is busier than the selected busiest group
8732 * don't try and pull any tasks.
8734 if (local->avg_load >= busiest->avg_load)
8738 * Don't pull any tasks if this group is already above the domain
8741 if (local->avg_load >= sds.avg_load)
8744 if (env->idle == CPU_IDLE) {
8746 * This CPU is idle. If the busiest group is not overloaded
8747 * and there is no imbalance between this and busiest group
8748 * wrt idle CPUs, it is balanced. The imbalance becomes
8749 * significant if the diff is greater than 1 otherwise we
8750 * might end up to just move the imbalance on another group
8752 if ((busiest->group_type != group_overloaded) &&
8753 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8757 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8758 * imbalance_pct to be conservative.
8760 if (100 * busiest->avg_load <=
8761 env->sd->imbalance_pct * local->avg_load)
8766 /* Looks like there is an imbalance. Compute it */
8767 env->src_grp_type = busiest->group_type;
8768 calculate_imbalance(env, &sds);
8769 return env->imbalance ? sds.busiest : NULL;
8777 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
8779 static struct rq *find_busiest_queue(struct lb_env *env,
8780 struct sched_group *group)
8782 struct rq *busiest = NULL, *rq;
8783 unsigned long busiest_load = 0, busiest_capacity = 1;
8786 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8787 unsigned long capacity, wl;
8791 rt = fbq_classify_rq(rq);
8794 * We classify groups/runqueues into three groups:
8795 * - regular: there are !numa tasks
8796 * - remote: there are numa tasks that run on the 'wrong' node
8797 * - all: there is no distinction
8799 * In order to avoid migrating ideally placed numa tasks,
8800 * ignore those when there's better options.
8802 * If we ignore the actual busiest queue to migrate another
8803 * task, the next balance pass can still reduce the busiest
8804 * queue by moving tasks around inside the node.
8806 * If we cannot move enough load due to this classification
8807 * the next pass will adjust the group classification and
8808 * allow migration of more tasks.
8810 * Both cases only affect the total convergence complexity.
8812 if (rt > env->fbq_type)
8816 * For ASYM_CPUCAPACITY domains with misfit tasks we simply
8817 * seek the "biggest" misfit task.
8819 if (env->src_grp_type == group_misfit_task) {
8820 if (rq->misfit_task_load > busiest_load) {
8821 busiest_load = rq->misfit_task_load;
8828 capacity = capacity_of(i);
8831 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
8832 * eventually lead to active_balancing high->low capacity.
8833 * Higher per-CPU capacity is considered better than balancing
8836 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8837 capacity_of(env->dst_cpu) < capacity &&
8838 rq->nr_running == 1)
8841 wl = weighted_cpuload(rq);
8844 * When comparing with imbalance, use weighted_cpuload()
8845 * which is not scaled with the CPU capacity.
8848 if (rq->nr_running == 1 && wl > env->imbalance &&
8849 !check_cpu_capacity(rq, env->sd))
8853 * For the load comparisons with the other CPU's, consider
8854 * the weighted_cpuload() scaled with the CPU capacity, so
8855 * that the load can be moved away from the CPU that is
8856 * potentially running at a lower capacity.
8858 * Thus we're looking for max(wl_i / capacity_i), crosswise
8859 * multiplication to rid ourselves of the division works out
8860 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8861 * our previous maximum.
8863 if (wl * busiest_capacity > busiest_load * capacity) {
8865 busiest_capacity = capacity;
8874 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8875 * so long as it is large enough.
8877 #define MAX_PINNED_INTERVAL 512
8880 asym_active_balance(struct lb_env *env)
8883 * ASYM_PACKING needs to force migrate tasks from busy but
8884 * lower priority CPUs in order to pack all tasks in the
8885 * highest priority CPUs.
8887 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
8888 sched_asym_prefer(env->dst_cpu, env->src_cpu);
8892 voluntary_active_balance(struct lb_env *env)
8894 struct sched_domain *sd = env->sd;
8896 if (asym_active_balance(env))
8900 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8901 * It's worth migrating the task if the src_cpu's capacity is reduced
8902 * because of other sched_class or IRQs if more capacity stays
8903 * available on dst_cpu.
8905 if ((env->idle != CPU_NOT_IDLE) &&
8906 (env->src_rq->cfs.h_nr_running == 1)) {
8907 if ((check_cpu_capacity(env->src_rq, sd)) &&
8908 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8912 if (env->src_grp_type == group_misfit_task)
8918 static int need_active_balance(struct lb_env *env)
8920 struct sched_domain *sd = env->sd;
8922 if (voluntary_active_balance(env))
8925 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8928 static int active_load_balance_cpu_stop(void *data);
8930 static int should_we_balance(struct lb_env *env)
8932 struct sched_group *sg = env->sd->groups;
8933 int cpu, balance_cpu = -1;
8936 * Ensure the balancing environment is consistent; can happen
8937 * when the softirq triggers 'during' hotplug.
8939 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8943 * In the newly idle case, we will allow all the CPUs
8944 * to do the newly idle load balance.
8946 if (env->idle == CPU_NEWLY_IDLE)
8949 /* Try to find first idle CPU */
8950 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8958 if (balance_cpu == -1)
8959 balance_cpu = group_balance_cpu(sg);
8962 * First idle CPU or the first CPU(busiest) in this sched group
8963 * is eligible for doing load balancing at this and above domains.
8965 return balance_cpu == env->dst_cpu;
8969 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8970 * tasks if there is an imbalance.
8972 static int load_balance(int this_cpu, struct rq *this_rq,
8973 struct sched_domain *sd, enum cpu_idle_type idle,
8974 int *continue_balancing)
8976 int ld_moved, cur_ld_moved, active_balance = 0;
8977 struct sched_domain *sd_parent = sd->parent;
8978 struct sched_group *group;
8981 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8983 struct lb_env env = {
8985 .dst_cpu = this_cpu,
8987 .dst_grpmask = sched_group_span(sd->groups),
8989 .loop_break = sched_nr_migrate_break,
8992 .tasks = LIST_HEAD_INIT(env.tasks),
8995 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8997 schedstat_inc(sd->lb_count[idle]);
9000 if (!should_we_balance(&env)) {
9001 *continue_balancing = 0;
9005 group = find_busiest_group(&env);
9007 schedstat_inc(sd->lb_nobusyg[idle]);
9011 busiest = find_busiest_queue(&env, group);
9013 schedstat_inc(sd->lb_nobusyq[idle]);
9017 BUG_ON(busiest == env.dst_rq);
9019 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
9021 env.src_cpu = busiest->cpu;
9022 env.src_rq = busiest;
9025 if (busiest->nr_running > 1) {
9027 * Attempt to move tasks. If find_busiest_group has found
9028 * an imbalance but busiest->nr_running <= 1, the group is
9029 * still unbalanced. ld_moved simply stays zero, so it is
9030 * correctly treated as an imbalance.
9032 env.flags |= LBF_ALL_PINNED;
9033 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
9036 rq_lock_irqsave(busiest, &rf);
9037 update_rq_clock(busiest);
9040 * cur_ld_moved - load moved in current iteration
9041 * ld_moved - cumulative load moved across iterations
9043 cur_ld_moved = detach_tasks(&env);
9046 * We've detached some tasks from busiest_rq. Every
9047 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
9048 * unlock busiest->lock, and we are able to be sure
9049 * that nobody can manipulate the tasks in parallel.
9050 * See task_rq_lock() family for the details.
9053 rq_unlock(busiest, &rf);
9057 ld_moved += cur_ld_moved;
9060 local_irq_restore(rf.flags);
9062 if (env.flags & LBF_NEED_BREAK) {
9063 env.flags &= ~LBF_NEED_BREAK;
9068 * Revisit (affine) tasks on src_cpu that couldn't be moved to
9069 * us and move them to an alternate dst_cpu in our sched_group
9070 * where they can run. The upper limit on how many times we
9071 * iterate on same src_cpu is dependent on number of CPUs in our
9074 * This changes load balance semantics a bit on who can move
9075 * load to a given_cpu. In addition to the given_cpu itself
9076 * (or a ilb_cpu acting on its behalf where given_cpu is
9077 * nohz-idle), we now have balance_cpu in a position to move
9078 * load to given_cpu. In rare situations, this may cause
9079 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
9080 * _independently_ and at _same_ time to move some load to
9081 * given_cpu) causing exceess load to be moved to given_cpu.
9082 * This however should not happen so much in practice and
9083 * moreover subsequent load balance cycles should correct the
9084 * excess load moved.
9086 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
9088 /* Prevent to re-select dst_cpu via env's CPUs */
9089 cpumask_clear_cpu(env.dst_cpu, env.cpus);
9091 env.dst_rq = cpu_rq(env.new_dst_cpu);
9092 env.dst_cpu = env.new_dst_cpu;
9093 env.flags &= ~LBF_DST_PINNED;
9095 env.loop_break = sched_nr_migrate_break;
9098 * Go back to "more_balance" rather than "redo" since we
9099 * need to continue with same src_cpu.
9105 * We failed to reach balance because of affinity.
9108 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9110 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
9111 *group_imbalance = 1;
9114 /* All tasks on this runqueue were pinned by CPU affinity */
9115 if (unlikely(env.flags & LBF_ALL_PINNED)) {
9116 cpumask_clear_cpu(cpu_of(busiest), cpus);
9118 * Attempting to continue load balancing at the current
9119 * sched_domain level only makes sense if there are
9120 * active CPUs remaining as possible busiest CPUs to
9121 * pull load from which are not contained within the
9122 * destination group that is receiving any migrated
9125 if (!cpumask_subset(cpus, env.dst_grpmask)) {
9127 env.loop_break = sched_nr_migrate_break;
9130 goto out_all_pinned;
9135 schedstat_inc(sd->lb_failed[idle]);
9137 * Increment the failure counter only on periodic balance.
9138 * We do not want newidle balance, which can be very
9139 * frequent, pollute the failure counter causing
9140 * excessive cache_hot migrations and active balances.
9142 if (idle != CPU_NEWLY_IDLE)
9143 sd->nr_balance_failed++;
9145 if (need_active_balance(&env)) {
9146 unsigned long flags;
9148 raw_spin_lock_irqsave(&busiest->lock, flags);
9151 * Don't kick the active_load_balance_cpu_stop,
9152 * if the curr task on busiest CPU can't be
9153 * moved to this_cpu:
9155 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
9156 raw_spin_unlock_irqrestore(&busiest->lock,
9158 env.flags |= LBF_ALL_PINNED;
9159 goto out_one_pinned;
9163 * ->active_balance synchronizes accesses to
9164 * ->active_balance_work. Once set, it's cleared
9165 * only after active load balance is finished.
9167 if (!busiest->active_balance) {
9168 busiest->active_balance = 1;
9169 busiest->push_cpu = this_cpu;
9172 raw_spin_unlock_irqrestore(&busiest->lock, flags);
9174 if (active_balance) {
9175 stop_one_cpu_nowait(cpu_of(busiest),
9176 active_load_balance_cpu_stop, busiest,
9177 &busiest->active_balance_work);
9180 /* We've kicked active balancing, force task migration. */
9181 sd->nr_balance_failed = sd->cache_nice_tries+1;
9184 sd->nr_balance_failed = 0;
9186 if (likely(!active_balance) || voluntary_active_balance(&env)) {
9187 /* We were unbalanced, so reset the balancing interval */
9188 sd->balance_interval = sd->min_interval;
9191 * If we've begun active balancing, start to back off. This
9192 * case may not be covered by the all_pinned logic if there
9193 * is only 1 task on the busy runqueue (because we don't call
9196 if (sd->balance_interval < sd->max_interval)
9197 sd->balance_interval *= 2;
9204 * We reach balance although we may have faced some affinity
9205 * constraints. Clear the imbalance flag if it was set.
9208 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9210 if (*group_imbalance)
9211 *group_imbalance = 0;
9216 * We reach balance because all tasks are pinned at this level so
9217 * we can't migrate them. Let the imbalance flag set so parent level
9218 * can try to migrate them.
9220 schedstat_inc(sd->lb_balanced[idle]);
9222 sd->nr_balance_failed = 0;
9228 * idle_balance() disregards balance intervals, so we could repeatedly
9229 * reach this code, which would lead to balance_interval skyrocketting
9230 * in a short amount of time. Skip the balance_interval increase logic
9233 if (env.idle == CPU_NEWLY_IDLE)
9236 /* tune up the balancing interval */
9237 if ((env.flags & LBF_ALL_PINNED &&
9238 sd->balance_interval < MAX_PINNED_INTERVAL) ||
9239 sd->balance_interval < sd->max_interval)
9240 sd->balance_interval *= 2;
9245 static inline unsigned long
9246 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
9248 unsigned long interval = sd->balance_interval;
9251 interval *= sd->busy_factor;
9253 /* scale ms to jiffies */
9254 interval = msecs_to_jiffies(interval);
9255 interval = clamp(interval, 1UL, max_load_balance_interval);
9261 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
9263 unsigned long interval, next;
9265 /* used by idle balance, so cpu_busy = 0 */
9266 interval = get_sd_balance_interval(sd, 0);
9267 next = sd->last_balance + interval;
9269 if (time_after(*next_balance, next))
9270 *next_balance = next;
9274 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9275 * running tasks off the busiest CPU onto idle CPUs. It requires at
9276 * least 1 task to be running on each physical CPU where possible, and
9277 * avoids physical / logical imbalances.
9279 static int active_load_balance_cpu_stop(void *data)
9281 struct rq *busiest_rq = data;
9282 int busiest_cpu = cpu_of(busiest_rq);
9283 int target_cpu = busiest_rq->push_cpu;
9284 struct rq *target_rq = cpu_rq(target_cpu);
9285 struct sched_domain *sd;
9286 struct task_struct *p = NULL;
9289 rq_lock_irq(busiest_rq, &rf);
9291 * Between queueing the stop-work and running it is a hole in which
9292 * CPUs can become inactive. We should not move tasks from or to
9295 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
9298 /* Make sure the requested CPU hasn't gone down in the meantime: */
9299 if (unlikely(busiest_cpu != smp_processor_id() ||
9300 !busiest_rq->active_balance))
9303 /* Is there any task to move? */
9304 if (busiest_rq->nr_running <= 1)
9308 * This condition is "impossible", if it occurs
9309 * we need to fix it. Originally reported by
9310 * Bjorn Helgaas on a 128-CPU setup.
9312 BUG_ON(busiest_rq == target_rq);
9314 /* Search for an sd spanning us and the target CPU. */
9316 for_each_domain(target_cpu, sd) {
9317 if ((sd->flags & SD_LOAD_BALANCE) &&
9318 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9323 struct lb_env env = {
9325 .dst_cpu = target_cpu,
9326 .dst_rq = target_rq,
9327 .src_cpu = busiest_rq->cpu,
9328 .src_rq = busiest_rq,
9331 * can_migrate_task() doesn't need to compute new_dst_cpu
9332 * for active balancing. Since we have CPU_IDLE, but no
9333 * @dst_grpmask we need to make that test go away with lying
9336 .flags = LBF_DST_PINNED,
9339 schedstat_inc(sd->alb_count);
9340 update_rq_clock(busiest_rq);
9342 p = detach_one_task(&env);
9344 schedstat_inc(sd->alb_pushed);
9345 /* Active balancing done, reset the failure counter. */
9346 sd->nr_balance_failed = 0;
9348 schedstat_inc(sd->alb_failed);
9353 busiest_rq->active_balance = 0;
9354 rq_unlock(busiest_rq, &rf);
9357 attach_one_task(target_rq, p);
9364 static DEFINE_SPINLOCK(balancing);
9367 * Scale the max load_balance interval with the number of CPUs in the system.
9368 * This trades load-balance latency on larger machines for less cross talk.
9370 void update_max_interval(void)
9372 max_load_balance_interval = HZ*num_online_cpus()/10;
9376 * It checks each scheduling domain to see if it is due to be balanced,
9377 * and initiates a balancing operation if so.
9379 * Balancing parameters are set up in init_sched_domains.
9381 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9383 int continue_balancing = 1;
9385 unsigned long interval;
9386 struct sched_domain *sd;
9387 /* Earliest time when we have to do rebalance again */
9388 unsigned long next_balance = jiffies + 60*HZ;
9389 int update_next_balance = 0;
9390 int need_serialize, need_decay = 0;
9394 for_each_domain(cpu, sd) {
9396 * Decay the newidle max times here because this is a regular
9397 * visit to all the domains. Decay ~1% per second.
9399 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9400 sd->max_newidle_lb_cost =
9401 (sd->max_newidle_lb_cost * 253) / 256;
9402 sd->next_decay_max_lb_cost = jiffies + HZ;
9405 max_cost += sd->max_newidle_lb_cost;
9407 if (!(sd->flags & SD_LOAD_BALANCE))
9411 * Stop the load balance at this level. There is another
9412 * CPU in our sched group which is doing load balancing more
9415 if (!continue_balancing) {
9421 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9423 need_serialize = sd->flags & SD_SERIALIZE;
9424 if (need_serialize) {
9425 if (!spin_trylock(&balancing))
9429 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9430 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9432 * The LBF_DST_PINNED logic could have changed
9433 * env->dst_cpu, so we can't know our idle
9434 * state even if we migrated tasks. Update it.
9436 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9438 sd->last_balance = jiffies;
9439 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9442 spin_unlock(&balancing);
9444 if (time_after(next_balance, sd->last_balance + interval)) {
9445 next_balance = sd->last_balance + interval;
9446 update_next_balance = 1;
9451 * Ensure the rq-wide value also decays but keep it at a
9452 * reasonable floor to avoid funnies with rq->avg_idle.
9454 rq->max_idle_balance_cost =
9455 max((u64)sysctl_sched_migration_cost, max_cost);
9460 * next_balance will be updated only when there is a need.
9461 * When the cpu is attached to null domain for ex, it will not be
9464 if (likely(update_next_balance)) {
9465 rq->next_balance = next_balance;
9467 #ifdef CONFIG_NO_HZ_COMMON
9469 * If this CPU has been elected to perform the nohz idle
9470 * balance. Other idle CPUs have already rebalanced with
9471 * nohz_idle_balance() and nohz.next_balance has been
9472 * updated accordingly. This CPU is now running the idle load
9473 * balance for itself and we need to update the
9474 * nohz.next_balance accordingly.
9476 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9477 nohz.next_balance = rq->next_balance;
9482 static inline int on_null_domain(struct rq *rq)
9484 return unlikely(!rcu_dereference_sched(rq->sd));
9487 #ifdef CONFIG_NO_HZ_COMMON
9489 * idle load balancing details
9490 * - When one of the busy CPUs notice that there may be an idle rebalancing
9491 * needed, they will kick the idle load balancer, which then does idle
9492 * load balancing for all the idle CPUs.
9495 static inline int find_new_ilb(void)
9497 int ilb = cpumask_first(nohz.idle_cpus_mask);
9499 if (ilb < nr_cpu_ids && idle_cpu(ilb))
9506 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
9507 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
9508 * CPU (if there is one).
9510 static void kick_ilb(unsigned int flags)
9514 nohz.next_balance++;
9516 ilb_cpu = find_new_ilb();
9518 if (ilb_cpu >= nr_cpu_ids)
9521 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
9522 if (flags & NOHZ_KICK_MASK)
9526 * Use smp_send_reschedule() instead of resched_cpu().
9527 * This way we generate a sched IPI on the target CPU which
9528 * is idle. And the softirq performing nohz idle load balance
9529 * will be run before returning from the IPI.
9531 smp_send_reschedule(ilb_cpu);
9535 * Current heuristic for kicking the idle load balancer in the presence
9536 * of an idle cpu in the system.
9537 * - This rq has more than one task.
9538 * - This rq has at least one CFS task and the capacity of the CPU is
9539 * significantly reduced because of RT tasks or IRQs.
9540 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9541 * multiple busy cpu.
9542 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9543 * domain span are idle.
9545 static void nohz_balancer_kick(struct rq *rq)
9547 unsigned long now = jiffies;
9548 struct sched_domain_shared *sds;
9549 struct sched_domain *sd;
9550 int nr_busy, i, cpu = rq->cpu;
9551 unsigned int flags = 0;
9553 if (unlikely(rq->idle_balance))
9557 * We may be recently in ticked or tickless idle mode. At the first
9558 * busy tick after returning from idle, we will update the busy stats.
9560 nohz_balance_exit_idle(rq);
9563 * None are in tickless mode and hence no need for NOHZ idle load
9566 if (likely(!atomic_read(&nohz.nr_cpus)))
9569 if (READ_ONCE(nohz.has_blocked) &&
9570 time_after(now, READ_ONCE(nohz.next_blocked)))
9571 flags = NOHZ_STATS_KICK;
9573 if (time_before(now, nohz.next_balance))
9576 if (rq->nr_running >= 2 || rq->misfit_task_load) {
9577 flags = NOHZ_KICK_MASK;
9582 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9585 * XXX: write a coherent comment on why we do this.
9586 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
9588 nr_busy = atomic_read(&sds->nr_busy_cpus);
9590 flags = NOHZ_KICK_MASK;
9596 sd = rcu_dereference(rq->sd);
9598 if ((rq->cfs.h_nr_running >= 1) &&
9599 check_cpu_capacity(rq, sd)) {
9600 flags = NOHZ_KICK_MASK;
9605 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
9607 for_each_cpu(i, sched_domain_span(sd)) {
9609 !cpumask_test_cpu(i, nohz.idle_cpus_mask))
9612 if (sched_asym_prefer(i, cpu)) {
9613 flags = NOHZ_KICK_MASK;
9625 static void set_cpu_sd_state_busy(int cpu)
9627 struct sched_domain *sd;
9630 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9632 if (!sd || !sd->nohz_idle)
9636 atomic_inc(&sd->shared->nr_busy_cpus);
9641 void nohz_balance_exit_idle(struct rq *rq)
9643 SCHED_WARN_ON(rq != this_rq());
9645 if (likely(!rq->nohz_tick_stopped))
9648 rq->nohz_tick_stopped = 0;
9649 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
9650 atomic_dec(&nohz.nr_cpus);
9652 set_cpu_sd_state_busy(rq->cpu);
9655 static void set_cpu_sd_state_idle(int cpu)
9657 struct sched_domain *sd;
9660 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9662 if (!sd || sd->nohz_idle)
9666 atomic_dec(&sd->shared->nr_busy_cpus);
9672 * This routine will record that the CPU is going idle with tick stopped.
9673 * This info will be used in performing idle load balancing in the future.
9675 void nohz_balance_enter_idle(int cpu)
9677 struct rq *rq = cpu_rq(cpu);
9679 SCHED_WARN_ON(cpu != smp_processor_id());
9681 /* If this CPU is going down, then nothing needs to be done: */
9682 if (!cpu_active(cpu))
9685 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9686 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9690 * Can be set safely without rq->lock held
9691 * If a clear happens, it will have evaluated last additions because
9692 * rq->lock is held during the check and the clear
9694 rq->has_blocked_load = 1;
9697 * The tick is still stopped but load could have been added in the
9698 * meantime. We set the nohz.has_blocked flag to trig a check of the
9699 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
9700 * of nohz.has_blocked can only happen after checking the new load
9702 if (rq->nohz_tick_stopped)
9705 /* If we're a completely isolated CPU, we don't play: */
9706 if (on_null_domain(rq))
9709 rq->nohz_tick_stopped = 1;
9711 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9712 atomic_inc(&nohz.nr_cpus);
9715 * Ensures that if nohz_idle_balance() fails to observe our
9716 * @idle_cpus_mask store, it must observe the @has_blocked
9719 smp_mb__after_atomic();
9721 set_cpu_sd_state_idle(cpu);
9725 * Each time a cpu enter idle, we assume that it has blocked load and
9726 * enable the periodic update of the load of idle cpus
9728 WRITE_ONCE(nohz.has_blocked, 1);
9732 * Internal function that runs load balance for all idle cpus. The load balance
9733 * can be a simple update of blocked load or a complete load balance with
9734 * tasks movement depending of flags.
9735 * The function returns false if the loop has stopped before running
9736 * through all idle CPUs.
9738 static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
9739 enum cpu_idle_type idle)
9741 /* Earliest time when we have to do rebalance again */
9742 unsigned long now = jiffies;
9743 unsigned long next_balance = now + 60*HZ;
9744 bool has_blocked_load = false;
9745 int update_next_balance = 0;
9746 int this_cpu = this_rq->cpu;
9751 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
9754 * We assume there will be no idle load after this update and clear
9755 * the has_blocked flag. If a cpu enters idle in the mean time, it will
9756 * set the has_blocked flag and trig another update of idle load.
9757 * Because a cpu that becomes idle, is added to idle_cpus_mask before
9758 * setting the flag, we are sure to not clear the state and not
9759 * check the load of an idle cpu.
9761 WRITE_ONCE(nohz.has_blocked, 0);
9764 * Ensures that if we miss the CPU, we must see the has_blocked
9765 * store from nohz_balance_enter_idle().
9769 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9770 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9774 * If this CPU gets work to do, stop the load balancing
9775 * work being done for other CPUs. Next load
9776 * balancing owner will pick it up.
9778 if (need_resched()) {
9779 has_blocked_load = true;
9783 rq = cpu_rq(balance_cpu);
9785 has_blocked_load |= update_nohz_stats(rq, true);
9788 * If time for next balance is due,
9791 if (time_after_eq(jiffies, rq->next_balance)) {
9794 rq_lock_irqsave(rq, &rf);
9795 update_rq_clock(rq);
9796 cpu_load_update_idle(rq);
9797 rq_unlock_irqrestore(rq, &rf);
9799 if (flags & NOHZ_BALANCE_KICK)
9800 rebalance_domains(rq, CPU_IDLE);
9803 if (time_after(next_balance, rq->next_balance)) {
9804 next_balance = rq->next_balance;
9805 update_next_balance = 1;
9809 /* Newly idle CPU doesn't need an update */
9810 if (idle != CPU_NEWLY_IDLE) {
9811 update_blocked_averages(this_cpu);
9812 has_blocked_load |= this_rq->has_blocked_load;
9815 if (flags & NOHZ_BALANCE_KICK)
9816 rebalance_domains(this_rq, CPU_IDLE);
9818 WRITE_ONCE(nohz.next_blocked,
9819 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
9821 /* The full idle balance loop has been done */
9825 /* There is still blocked load, enable periodic update */
9826 if (has_blocked_load)
9827 WRITE_ONCE(nohz.has_blocked, 1);
9830 * next_balance will be updated only when there is a need.
9831 * When the CPU is attached to null domain for ex, it will not be
9834 if (likely(update_next_balance))
9835 nohz.next_balance = next_balance;
9841 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9842 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9844 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9846 int this_cpu = this_rq->cpu;
9849 if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
9852 if (idle != CPU_IDLE) {
9853 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9857 /* could be _relaxed() */
9858 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9859 if (!(flags & NOHZ_KICK_MASK))
9862 _nohz_idle_balance(this_rq, flags, idle);
9867 static void nohz_newidle_balance(struct rq *this_rq)
9869 int this_cpu = this_rq->cpu;
9872 * This CPU doesn't want to be disturbed by scheduler
9875 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
9878 /* Will wake up very soon. No time for doing anything else*/
9879 if (this_rq->avg_idle < sysctl_sched_migration_cost)
9882 /* Don't need to update blocked load of idle CPUs*/
9883 if (!READ_ONCE(nohz.has_blocked) ||
9884 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
9887 raw_spin_unlock(&this_rq->lock);
9889 * This CPU is going to be idle and blocked load of idle CPUs
9890 * need to be updated. Run the ilb locally as it is a good
9891 * candidate for ilb instead of waking up another idle CPU.
9892 * Kick an normal ilb if we failed to do the update.
9894 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
9895 kick_ilb(NOHZ_STATS_KICK);
9896 raw_spin_lock(&this_rq->lock);
9899 #else /* !CONFIG_NO_HZ_COMMON */
9900 static inline void nohz_balancer_kick(struct rq *rq) { }
9902 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9907 static inline void nohz_newidle_balance(struct rq *this_rq) { }
9908 #endif /* CONFIG_NO_HZ_COMMON */
9911 * idle_balance is called by schedule() if this_cpu is about to become
9912 * idle. Attempts to pull tasks from other CPUs.
9914 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
9916 unsigned long next_balance = jiffies + HZ;
9917 int this_cpu = this_rq->cpu;
9918 struct sched_domain *sd;
9919 int pulled_task = 0;
9923 * We must set idle_stamp _before_ calling idle_balance(), such that we
9924 * measure the duration of idle_balance() as idle time.
9926 this_rq->idle_stamp = rq_clock(this_rq);
9929 * Do not pull tasks towards !active CPUs...
9931 if (!cpu_active(this_cpu))
9935 * This is OK, because current is on_cpu, which avoids it being picked
9936 * for load-balance and preemption/IRQs are still disabled avoiding
9937 * further scheduler activity on it and we're being very careful to
9938 * re-start the picking loop.
9940 rq_unpin_lock(this_rq, rf);
9942 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
9943 !READ_ONCE(this_rq->rd->overload)) {
9946 sd = rcu_dereference_check_sched_domain(this_rq->sd);
9948 update_next_balance(sd, &next_balance);
9951 nohz_newidle_balance(this_rq);
9956 raw_spin_unlock(&this_rq->lock);
9958 update_blocked_averages(this_cpu);
9960 for_each_domain(this_cpu, sd) {
9961 int continue_balancing = 1;
9962 u64 t0, domain_cost;
9964 if (!(sd->flags & SD_LOAD_BALANCE))
9967 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
9968 update_next_balance(sd, &next_balance);
9972 if (sd->flags & SD_BALANCE_NEWIDLE) {
9973 t0 = sched_clock_cpu(this_cpu);
9975 pulled_task = load_balance(this_cpu, this_rq,
9977 &continue_balancing);
9979 domain_cost = sched_clock_cpu(this_cpu) - t0;
9980 if (domain_cost > sd->max_newidle_lb_cost)
9981 sd->max_newidle_lb_cost = domain_cost;
9983 curr_cost += domain_cost;
9986 update_next_balance(sd, &next_balance);
9989 * Stop searching for tasks to pull if there are
9990 * now runnable tasks on this rq.
9992 if (pulled_task || this_rq->nr_running > 0)
9997 raw_spin_lock(&this_rq->lock);
9999 if (curr_cost > this_rq->max_idle_balance_cost)
10000 this_rq->max_idle_balance_cost = curr_cost;
10004 * While browsing the domains, we released the rq lock, a task could
10005 * have been enqueued in the meantime. Since we're not going idle,
10006 * pretend we pulled a task.
10008 if (this_rq->cfs.h_nr_running && !pulled_task)
10011 /* Move the next balance forward */
10012 if (time_after(this_rq->next_balance, next_balance))
10013 this_rq->next_balance = next_balance;
10015 /* Is there a task of a high priority class? */
10016 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
10020 this_rq->idle_stamp = 0;
10022 rq_repin_lock(this_rq, rf);
10024 return pulled_task;
10028 * run_rebalance_domains is triggered when needed from the scheduler tick.
10029 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
10031 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
10033 struct rq *this_rq = this_rq();
10034 enum cpu_idle_type idle = this_rq->idle_balance ?
10035 CPU_IDLE : CPU_NOT_IDLE;
10038 * If this CPU has a pending nohz_balance_kick, then do the
10039 * balancing on behalf of the other idle CPUs whose ticks are
10040 * stopped. Do nohz_idle_balance *before* rebalance_domains to
10041 * give the idle CPUs a chance to load balance. Else we may
10042 * load balance only within the local sched_domain hierarchy
10043 * and abort nohz_idle_balance altogether if we pull some load.
10045 if (nohz_idle_balance(this_rq, idle))
10048 /* normal load balance */
10049 update_blocked_averages(this_rq->cpu);
10050 rebalance_domains(this_rq, idle);
10054 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
10056 void trigger_load_balance(struct rq *rq)
10058 /* Don't need to rebalance while attached to NULL domain */
10059 if (unlikely(on_null_domain(rq)))
10062 if (time_after_eq(jiffies, rq->next_balance))
10063 raise_softirq(SCHED_SOFTIRQ);
10065 nohz_balancer_kick(rq);
10068 static void rq_online_fair(struct rq *rq)
10072 update_runtime_enabled(rq);
10075 static void rq_offline_fair(struct rq *rq)
10079 /* Ensure any throttled groups are reachable by pick_next_task */
10080 unthrottle_offline_cfs_rqs(rq);
10083 #endif /* CONFIG_SMP */
10086 * scheduler tick hitting a task of our scheduling class.
10088 * NOTE: This function can be called remotely by the tick offload that
10089 * goes along full dynticks. Therefore no local assumption can be made
10090 * and everything must be accessed through the @rq and @curr passed in
10093 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
10095 struct cfs_rq *cfs_rq;
10096 struct sched_entity *se = &curr->se;
10098 for_each_sched_entity(se) {
10099 cfs_rq = cfs_rq_of(se);
10100 entity_tick(cfs_rq, se, queued);
10103 if (static_branch_unlikely(&sched_numa_balancing))
10104 task_tick_numa(rq, curr);
10106 update_misfit_status(curr, rq);
10107 update_overutilized_status(task_rq(curr));
10111 * called on fork with the child task as argument from the parent's context
10112 * - child not yet on the tasklist
10113 * - preemption disabled
10115 static void task_fork_fair(struct task_struct *p)
10117 struct cfs_rq *cfs_rq;
10118 struct sched_entity *se = &p->se, *curr;
10119 struct rq *rq = this_rq();
10120 struct rq_flags rf;
10123 update_rq_clock(rq);
10125 cfs_rq = task_cfs_rq(current);
10126 curr = cfs_rq->curr;
10128 update_curr(cfs_rq);
10129 se->vruntime = curr->vruntime;
10131 place_entity(cfs_rq, se, 1);
10133 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
10135 * Upon rescheduling, sched_class::put_prev_task() will place
10136 * 'current' within the tree based on its new key value.
10138 swap(curr->vruntime, se->vruntime);
10142 se->vruntime -= cfs_rq->min_vruntime;
10143 rq_unlock(rq, &rf);
10147 * Priority of the task has changed. Check to see if we preempt
10148 * the current task.
10151 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
10153 if (!task_on_rq_queued(p))
10157 * Reschedule if we are currently running on this runqueue and
10158 * our priority decreased, or if we are not currently running on
10159 * this runqueue and our priority is higher than the current's
10161 if (rq->curr == p) {
10162 if (p->prio > oldprio)
10165 check_preempt_curr(rq, p, 0);
10168 static inline bool vruntime_normalized(struct task_struct *p)
10170 struct sched_entity *se = &p->se;
10173 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
10174 * the dequeue_entity(.flags=0) will already have normalized the
10181 * When !on_rq, vruntime of the task has usually NOT been normalized.
10182 * But there are some cases where it has already been normalized:
10184 * - A forked child which is waiting for being woken up by
10185 * wake_up_new_task().
10186 * - A task which has been woken up by try_to_wake_up() and
10187 * waiting for actually being woken up by sched_ttwu_pending().
10189 if (!se->sum_exec_runtime ||
10190 (p->state == TASK_WAKING && p->sched_remote_wakeup))
10196 #ifdef CONFIG_FAIR_GROUP_SCHED
10198 * Propagate the changes of the sched_entity across the tg tree to make it
10199 * visible to the root
10201 static void propagate_entity_cfs_rq(struct sched_entity *se)
10203 struct cfs_rq *cfs_rq;
10205 /* Start to propagate at parent */
10208 for_each_sched_entity(se) {
10209 cfs_rq = cfs_rq_of(se);
10211 if (cfs_rq_throttled(cfs_rq))
10214 update_load_avg(cfs_rq, se, UPDATE_TG);
10218 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
10221 static void detach_entity_cfs_rq(struct sched_entity *se)
10223 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10225 /* Catch up with the cfs_rq and remove our load when we leave */
10226 update_load_avg(cfs_rq, se, 0);
10227 detach_entity_load_avg(cfs_rq, se);
10228 update_tg_load_avg(cfs_rq, false);
10229 propagate_entity_cfs_rq(se);
10232 static void attach_entity_cfs_rq(struct sched_entity *se)
10234 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10236 #ifdef CONFIG_FAIR_GROUP_SCHED
10238 * Since the real-depth could have been changed (only FAIR
10239 * class maintain depth value), reset depth properly.
10241 se->depth = se->parent ? se->parent->depth + 1 : 0;
10244 /* Synchronize entity with its cfs_rq */
10245 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
10246 attach_entity_load_avg(cfs_rq, se, 0);
10247 update_tg_load_avg(cfs_rq, false);
10248 propagate_entity_cfs_rq(se);
10251 static void detach_task_cfs_rq(struct task_struct *p)
10253 struct sched_entity *se = &p->se;
10254 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10256 if (!vruntime_normalized(p)) {
10258 * Fix up our vruntime so that the current sleep doesn't
10259 * cause 'unlimited' sleep bonus.
10261 place_entity(cfs_rq, se, 0);
10262 se->vruntime -= cfs_rq->min_vruntime;
10265 detach_entity_cfs_rq(se);
10268 static void attach_task_cfs_rq(struct task_struct *p)
10270 struct sched_entity *se = &p->se;
10271 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10273 attach_entity_cfs_rq(se);
10275 if (!vruntime_normalized(p))
10276 se->vruntime += cfs_rq->min_vruntime;
10279 static void switched_from_fair(struct rq *rq, struct task_struct *p)
10281 detach_task_cfs_rq(p);
10284 static void switched_to_fair(struct rq *rq, struct task_struct *p)
10286 attach_task_cfs_rq(p);
10288 if (task_on_rq_queued(p)) {
10290 * We were most likely switched from sched_rt, so
10291 * kick off the schedule if running, otherwise just see
10292 * if we can still preempt the current task.
10297 check_preempt_curr(rq, p, 0);
10301 /* Account for a task changing its policy or group.
10303 * This routine is mostly called to set cfs_rq->curr field when a task
10304 * migrates between groups/classes.
10306 static void set_curr_task_fair(struct rq *rq)
10308 struct sched_entity *se = &rq->curr->se;
10310 for_each_sched_entity(se) {
10311 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10313 set_next_entity(cfs_rq, se);
10314 /* ensure bandwidth has been allocated on our new cfs_rq */
10315 account_cfs_rq_runtime(cfs_rq, 0);
10319 void init_cfs_rq(struct cfs_rq *cfs_rq)
10321 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
10322 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
10323 #ifndef CONFIG_64BIT
10324 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
10327 raw_spin_lock_init(&cfs_rq->removed.lock);
10331 #ifdef CONFIG_FAIR_GROUP_SCHED
10332 static void task_set_group_fair(struct task_struct *p)
10334 struct sched_entity *se = &p->se;
10336 set_task_rq(p, task_cpu(p));
10337 se->depth = se->parent ? se->parent->depth + 1 : 0;
10340 static void task_move_group_fair(struct task_struct *p)
10342 detach_task_cfs_rq(p);
10343 set_task_rq(p, task_cpu(p));
10346 /* Tell se's cfs_rq has been changed -- migrated */
10347 p->se.avg.last_update_time = 0;
10349 attach_task_cfs_rq(p);
10352 static void task_change_group_fair(struct task_struct *p, int type)
10355 case TASK_SET_GROUP:
10356 task_set_group_fair(p);
10359 case TASK_MOVE_GROUP:
10360 task_move_group_fair(p);
10365 void free_fair_sched_group(struct task_group *tg)
10369 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
10371 for_each_possible_cpu(i) {
10373 kfree(tg->cfs_rq[i]);
10382 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10384 struct sched_entity *se;
10385 struct cfs_rq *cfs_rq;
10388 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
10391 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
10395 tg->shares = NICE_0_LOAD;
10397 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
10399 for_each_possible_cpu(i) {
10400 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
10401 GFP_KERNEL, cpu_to_node(i));
10405 se = kzalloc_node(sizeof(struct sched_entity),
10406 GFP_KERNEL, cpu_to_node(i));
10410 init_cfs_rq(cfs_rq);
10411 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
10412 init_entity_runnable_average(se);
10423 void online_fair_sched_group(struct task_group *tg)
10425 struct sched_entity *se;
10429 for_each_possible_cpu(i) {
10433 raw_spin_lock_irq(&rq->lock);
10434 update_rq_clock(rq);
10435 attach_entity_cfs_rq(se);
10436 sync_throttle(tg, i);
10437 raw_spin_unlock_irq(&rq->lock);
10441 void unregister_fair_sched_group(struct task_group *tg)
10443 unsigned long flags;
10447 for_each_possible_cpu(cpu) {
10449 remove_entity_load_avg(tg->se[cpu]);
10452 * Only empty task groups can be destroyed; so we can speculatively
10453 * check on_list without danger of it being re-added.
10455 if (!tg->cfs_rq[cpu]->on_list)
10460 raw_spin_lock_irqsave(&rq->lock, flags);
10461 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
10462 raw_spin_unlock_irqrestore(&rq->lock, flags);
10466 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
10467 struct sched_entity *se, int cpu,
10468 struct sched_entity *parent)
10470 struct rq *rq = cpu_rq(cpu);
10474 init_cfs_rq_runtime(cfs_rq);
10476 tg->cfs_rq[cpu] = cfs_rq;
10479 /* se could be NULL for root_task_group */
10484 se->cfs_rq = &rq->cfs;
10487 se->cfs_rq = parent->my_q;
10488 se->depth = parent->depth + 1;
10492 /* guarantee group entities always have weight */
10493 update_load_set(&se->load, NICE_0_LOAD);
10494 se->parent = parent;
10497 static DEFINE_MUTEX(shares_mutex);
10499 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10504 * We can't change the weight of the root cgroup.
10509 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10511 mutex_lock(&shares_mutex);
10512 if (tg->shares == shares)
10515 tg->shares = shares;
10516 for_each_possible_cpu(i) {
10517 struct rq *rq = cpu_rq(i);
10518 struct sched_entity *se = tg->se[i];
10519 struct rq_flags rf;
10521 /* Propagate contribution to hierarchy */
10522 rq_lock_irqsave(rq, &rf);
10523 update_rq_clock(rq);
10524 for_each_sched_entity(se) {
10525 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
10526 update_cfs_group(se);
10528 rq_unlock_irqrestore(rq, &rf);
10532 mutex_unlock(&shares_mutex);
10535 #else /* CONFIG_FAIR_GROUP_SCHED */
10537 void free_fair_sched_group(struct task_group *tg) { }
10539 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10544 void online_fair_sched_group(struct task_group *tg) { }
10546 void unregister_fair_sched_group(struct task_group *tg) { }
10548 #endif /* CONFIG_FAIR_GROUP_SCHED */
10551 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10553 struct sched_entity *se = &task->se;
10554 unsigned int rr_interval = 0;
10557 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10560 if (rq->cfs.load.weight)
10561 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10563 return rr_interval;
10567 * All the scheduling class methods:
10569 const struct sched_class fair_sched_class = {
10570 .next = &idle_sched_class,
10571 .enqueue_task = enqueue_task_fair,
10572 .dequeue_task = dequeue_task_fair,
10573 .yield_task = yield_task_fair,
10574 .yield_to_task = yield_to_task_fair,
10576 .check_preempt_curr = check_preempt_wakeup,
10578 .pick_next_task = pick_next_task_fair,
10579 .put_prev_task = put_prev_task_fair,
10582 .select_task_rq = select_task_rq_fair,
10583 .migrate_task_rq = migrate_task_rq_fair,
10585 .rq_online = rq_online_fair,
10586 .rq_offline = rq_offline_fair,
10588 .task_dead = task_dead_fair,
10589 .set_cpus_allowed = set_cpus_allowed_common,
10592 .set_curr_task = set_curr_task_fair,
10593 .task_tick = task_tick_fair,
10594 .task_fork = task_fork_fair,
10596 .prio_changed = prio_changed_fair,
10597 .switched_from = switched_from_fair,
10598 .switched_to = switched_to_fair,
10600 .get_rr_interval = get_rr_interval_fair,
10602 .update_curr = update_curr_fair,
10604 #ifdef CONFIG_FAIR_GROUP_SCHED
10605 .task_change_group = task_change_group_fair,
10609 #ifdef CONFIG_SCHED_DEBUG
10610 void print_cfs_stats(struct seq_file *m, int cpu)
10612 struct cfs_rq *cfs_rq;
10615 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
10616 print_cfs_rq(m, cpu, cfs_rq);
10620 #ifdef CONFIG_NUMA_BALANCING
10621 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10624 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10626 for_each_online_node(node) {
10627 if (p->numa_faults) {
10628 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10629 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10631 if (p->numa_group) {
10632 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
10633 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
10635 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10638 #endif /* CONFIG_NUMA_BALANCING */
10639 #endif /* CONFIG_SCHED_DEBUG */
10641 __init void init_sched_fair_class(void)
10644 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10646 #ifdef CONFIG_NO_HZ_COMMON
10647 nohz.next_balance = jiffies;
10648 nohz.next_blocked = jiffies;
10649 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);