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
252 /* cpu runqueue to which this cfs_rq is attached */
253 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
258 static inline struct task_struct *task_of(struct sched_entity *se)
260 SCHED_WARN_ON(!entity_is_task(se));
261 return container_of(se, struct task_struct, se);
264 /* Walk up scheduling entities hierarchy */
265 #define for_each_sched_entity(se) \
266 for (; se; se = se->parent)
268 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
273 /* runqueue on which this entity is (to be) queued */
274 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
279 /* runqueue "owned" by this group */
280 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
285 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
287 if (!cfs_rq->on_list) {
288 struct rq *rq = rq_of(cfs_rq);
289 int cpu = cpu_of(rq);
291 * Ensure we either appear before our parent (if already
292 * enqueued) or force our parent to appear after us when it is
293 * enqueued. The fact that we always enqueue bottom-up
294 * reduces this to two cases and a special case for the root
295 * cfs_rq. Furthermore, it also means that we will always reset
296 * tmp_alone_branch either when the branch is connected
297 * to a tree or when we reach the beg of the tree
299 if (cfs_rq->tg->parent &&
300 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
302 * If parent is already on the list, we add the child
303 * just before. Thanks to circular linked property of
304 * the list, this means to put the child at the tail
305 * of the list that starts by parent.
307 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
308 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
310 * The branch is now connected to its tree so we can
311 * reset tmp_alone_branch to the beginning of the
314 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
315 } else if (!cfs_rq->tg->parent) {
317 * cfs rq without parent should be put
318 * at the tail of the list.
320 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
321 &rq->leaf_cfs_rq_list);
323 * We have reach the beg of a tree so we can reset
324 * tmp_alone_branch to the beginning of the list.
326 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
329 * The parent has not already been added so we want to
330 * make sure that it will be put after us.
331 * tmp_alone_branch points to the beg of the branch
332 * where we will add parent.
334 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
335 rq->tmp_alone_branch);
337 * update tmp_alone_branch to points to the new beg
340 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
347 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
349 if (cfs_rq->on_list) {
350 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
355 /* Iterate thr' all leaf cfs_rq's on a runqueue */
356 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
357 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
360 /* Do the two (enqueued) entities belong to the same group ? */
361 static inline struct cfs_rq *
362 is_same_group(struct sched_entity *se, struct sched_entity *pse)
364 if (se->cfs_rq == pse->cfs_rq)
370 static inline struct sched_entity *parent_entity(struct sched_entity *se)
376 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
378 int se_depth, pse_depth;
381 * preemption test can be made between sibling entities who are in the
382 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
383 * both tasks until we find their ancestors who are siblings of common
387 /* First walk up until both entities are at same depth */
388 se_depth = (*se)->depth;
389 pse_depth = (*pse)->depth;
391 while (se_depth > pse_depth) {
393 *se = parent_entity(*se);
396 while (pse_depth > se_depth) {
398 *pse = parent_entity(*pse);
401 while (!is_same_group(*se, *pse)) {
402 *se = parent_entity(*se);
403 *pse = parent_entity(*pse);
407 #else /* !CONFIG_FAIR_GROUP_SCHED */
409 static inline struct task_struct *task_of(struct sched_entity *se)
411 return container_of(se, struct task_struct, se);
414 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
416 return container_of(cfs_rq, struct rq, cfs);
420 #define for_each_sched_entity(se) \
421 for (; se; se = NULL)
423 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
425 return &task_rq(p)->cfs;
428 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
430 struct task_struct *p = task_of(se);
431 struct rq *rq = task_rq(p);
436 /* runqueue "owned" by this group */
437 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
442 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
446 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
450 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
451 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
453 static inline struct sched_entity *parent_entity(struct sched_entity *se)
459 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
463 #endif /* CONFIG_FAIR_GROUP_SCHED */
465 static __always_inline
466 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
468 /**************************************************************
469 * Scheduling class tree data structure manipulation methods:
472 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
474 s64 delta = (s64)(vruntime - max_vruntime);
476 max_vruntime = vruntime;
481 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
483 s64 delta = (s64)(vruntime - min_vruntime);
485 min_vruntime = vruntime;
490 static inline int entity_before(struct sched_entity *a,
491 struct sched_entity *b)
493 return (s64)(a->vruntime - b->vruntime) < 0;
496 static void update_min_vruntime(struct cfs_rq *cfs_rq)
498 struct sched_entity *curr = cfs_rq->curr;
499 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
501 u64 vruntime = cfs_rq->min_vruntime;
505 vruntime = curr->vruntime;
510 if (leftmost) { /* non-empty tree */
511 struct sched_entity *se;
512 se = rb_entry(leftmost, struct sched_entity, run_node);
515 vruntime = se->vruntime;
517 vruntime = min_vruntime(vruntime, se->vruntime);
520 /* ensure we never gain time by being placed backwards. */
521 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
524 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
529 * Enqueue an entity into the rb-tree:
531 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
533 struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node;
534 struct rb_node *parent = NULL;
535 struct sched_entity *entry;
536 bool leftmost = true;
539 * Find the right place in the rbtree:
543 entry = rb_entry(parent, struct sched_entity, run_node);
545 * We dont care about collisions. Nodes with
546 * the same key stay together.
548 if (entity_before(se, entry)) {
549 link = &parent->rb_left;
551 link = &parent->rb_right;
556 rb_link_node(&se->run_node, parent, link);
557 rb_insert_color_cached(&se->run_node,
558 &cfs_rq->tasks_timeline, leftmost);
561 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
563 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
566 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
568 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
573 return rb_entry(left, struct sched_entity, run_node);
576 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
578 struct rb_node *next = rb_next(&se->run_node);
583 return rb_entry(next, struct sched_entity, run_node);
586 #ifdef CONFIG_SCHED_DEBUG
587 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
589 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
594 return rb_entry(last, struct sched_entity, run_node);
597 /**************************************************************
598 * Scheduling class statistics methods:
601 int sched_proc_update_handler(struct ctl_table *table, int write,
602 void __user *buffer, size_t *lenp,
605 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
606 unsigned int factor = get_update_sysctl_factor();
611 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
612 sysctl_sched_min_granularity);
614 #define WRT_SYSCTL(name) \
615 (normalized_sysctl_##name = sysctl_##name / (factor))
616 WRT_SYSCTL(sched_min_granularity);
617 WRT_SYSCTL(sched_latency);
618 WRT_SYSCTL(sched_wakeup_granularity);
628 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
630 if (unlikely(se->load.weight != NICE_0_LOAD))
631 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
637 * The idea is to set a period in which each task runs once.
639 * When there are too many tasks (sched_nr_latency) we have to stretch
640 * this period because otherwise the slices get too small.
642 * p = (nr <= nl) ? l : l*nr/nl
644 static u64 __sched_period(unsigned long nr_running)
646 if (unlikely(nr_running > sched_nr_latency))
647 return nr_running * sysctl_sched_min_granularity;
649 return sysctl_sched_latency;
653 * We calculate the wall-time slice from the period by taking a part
654 * proportional to the weight.
658 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
660 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
662 for_each_sched_entity(se) {
663 struct load_weight *load;
664 struct load_weight lw;
666 cfs_rq = cfs_rq_of(se);
667 load = &cfs_rq->load;
669 if (unlikely(!se->on_rq)) {
672 update_load_add(&lw, se->load.weight);
675 slice = __calc_delta(slice, se->load.weight, load);
681 * We calculate the vruntime slice of a to-be-inserted task.
685 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
687 return calc_delta_fair(sched_slice(cfs_rq, se), se);
692 #include "sched-pelt.h"
694 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
695 static unsigned long task_h_load(struct task_struct *p);
696 static unsigned long capacity_of(int cpu);
698 /* Give new sched_entity start runnable values to heavy its load in infant time */
699 void init_entity_runnable_average(struct sched_entity *se)
701 struct sched_avg *sa = &se->avg;
703 memset(sa, 0, sizeof(*sa));
706 * Tasks are initialized with full load to be seen as heavy tasks until
707 * they get a chance to stabilize to their real load level.
708 * Group entities are initialized with zero load to reflect the fact that
709 * nothing has been attached to the task group yet.
711 if (entity_is_task(se))
712 sa->runnable_load_avg = sa->load_avg = scale_load_down(se->load.weight);
714 se->runnable_weight = se->load.weight;
716 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
719 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
720 static void attach_entity_cfs_rq(struct sched_entity *se);
723 * With new tasks being created, their initial util_avgs are extrapolated
724 * based on the cfs_rq's current util_avg:
726 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
728 * However, in many cases, the above util_avg does not give a desired
729 * value. Moreover, the sum of the util_avgs may be divergent, such
730 * as when the series is a harmonic series.
732 * To solve this problem, we also cap the util_avg of successive tasks to
733 * only 1/2 of the left utilization budget:
735 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
737 * where n denotes the nth task and cpu_scale the CPU capacity.
739 * For example, for a CPU with 1024 of capacity, a simplest series from
740 * the beginning would be like:
742 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
743 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
745 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
746 * if util_avg > util_avg_cap.
748 void post_init_entity_util_avg(struct sched_entity *se)
750 struct cfs_rq *cfs_rq = cfs_rq_of(se);
751 struct sched_avg *sa = &se->avg;
752 long cpu_scale = arch_scale_cpu_capacity(NULL, cpu_of(rq_of(cfs_rq)));
753 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
756 if (cfs_rq->avg.util_avg != 0) {
757 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
758 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
760 if (sa->util_avg > cap)
767 if (entity_is_task(se)) {
768 struct task_struct *p = task_of(se);
769 if (p->sched_class != &fair_sched_class) {
771 * For !fair tasks do:
773 update_cfs_rq_load_avg(now, cfs_rq);
774 attach_entity_load_avg(cfs_rq, se, 0);
775 switched_from_fair(rq, p);
777 * such that the next switched_to_fair() has the
780 se->avg.last_update_time = cfs_rq_clock_task(cfs_rq);
785 attach_entity_cfs_rq(se);
788 #else /* !CONFIG_SMP */
789 void init_entity_runnable_average(struct sched_entity *se)
792 void post_init_entity_util_avg(struct sched_entity *se)
795 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
798 #endif /* CONFIG_SMP */
801 * Update the current task's runtime statistics.
803 static void update_curr(struct cfs_rq *cfs_rq)
805 struct sched_entity *curr = cfs_rq->curr;
806 u64 now = rq_clock_task(rq_of(cfs_rq));
812 delta_exec = now - curr->exec_start;
813 if (unlikely((s64)delta_exec <= 0))
816 curr->exec_start = now;
818 schedstat_set(curr->statistics.exec_max,
819 max(delta_exec, curr->statistics.exec_max));
821 curr->sum_exec_runtime += delta_exec;
822 schedstat_add(cfs_rq->exec_clock, delta_exec);
824 curr->vruntime += calc_delta_fair(delta_exec, curr);
825 update_min_vruntime(cfs_rq);
827 if (entity_is_task(curr)) {
828 struct task_struct *curtask = task_of(curr);
830 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
831 cgroup_account_cputime(curtask, delta_exec);
832 account_group_exec_runtime(curtask, delta_exec);
835 account_cfs_rq_runtime(cfs_rq, delta_exec);
838 static void update_curr_fair(struct rq *rq)
840 update_curr(cfs_rq_of(&rq->curr->se));
844 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
846 u64 wait_start, prev_wait_start;
848 if (!schedstat_enabled())
851 wait_start = rq_clock(rq_of(cfs_rq));
852 prev_wait_start = schedstat_val(se->statistics.wait_start);
854 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
855 likely(wait_start > prev_wait_start))
856 wait_start -= prev_wait_start;
858 __schedstat_set(se->statistics.wait_start, wait_start);
862 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
864 struct task_struct *p;
867 if (!schedstat_enabled())
870 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
872 if (entity_is_task(se)) {
874 if (task_on_rq_migrating(p)) {
876 * Preserve migrating task's wait time so wait_start
877 * time stamp can be adjusted to accumulate wait time
878 * prior to migration.
880 __schedstat_set(se->statistics.wait_start, delta);
883 trace_sched_stat_wait(p, delta);
886 __schedstat_set(se->statistics.wait_max,
887 max(schedstat_val(se->statistics.wait_max), delta));
888 __schedstat_inc(se->statistics.wait_count);
889 __schedstat_add(se->statistics.wait_sum, delta);
890 __schedstat_set(se->statistics.wait_start, 0);
894 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
896 struct task_struct *tsk = NULL;
897 u64 sleep_start, block_start;
899 if (!schedstat_enabled())
902 sleep_start = schedstat_val(se->statistics.sleep_start);
903 block_start = schedstat_val(se->statistics.block_start);
905 if (entity_is_task(se))
909 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
914 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
915 __schedstat_set(se->statistics.sleep_max, delta);
917 __schedstat_set(se->statistics.sleep_start, 0);
918 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
921 account_scheduler_latency(tsk, delta >> 10, 1);
922 trace_sched_stat_sleep(tsk, delta);
926 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
931 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
932 __schedstat_set(se->statistics.block_max, delta);
934 __schedstat_set(se->statistics.block_start, 0);
935 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
938 if (tsk->in_iowait) {
939 __schedstat_add(se->statistics.iowait_sum, delta);
940 __schedstat_inc(se->statistics.iowait_count);
941 trace_sched_stat_iowait(tsk, delta);
944 trace_sched_stat_blocked(tsk, delta);
947 * Blocking time is in units of nanosecs, so shift by
948 * 20 to get a milliseconds-range estimation of the
949 * amount of time that the task spent sleeping:
951 if (unlikely(prof_on == SLEEP_PROFILING)) {
952 profile_hits(SLEEP_PROFILING,
953 (void *)get_wchan(tsk),
956 account_scheduler_latency(tsk, delta >> 10, 0);
962 * Task is being enqueued - update stats:
965 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
967 if (!schedstat_enabled())
971 * Are we enqueueing a waiting task? (for current tasks
972 * a dequeue/enqueue event is a NOP)
974 if (se != cfs_rq->curr)
975 update_stats_wait_start(cfs_rq, se);
977 if (flags & ENQUEUE_WAKEUP)
978 update_stats_enqueue_sleeper(cfs_rq, se);
982 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
985 if (!schedstat_enabled())
989 * Mark the end of the wait period if dequeueing a
992 if (se != cfs_rq->curr)
993 update_stats_wait_end(cfs_rq, se);
995 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
996 struct task_struct *tsk = task_of(se);
998 if (tsk->state & TASK_INTERRUPTIBLE)
999 __schedstat_set(se->statistics.sleep_start,
1000 rq_clock(rq_of(cfs_rq)));
1001 if (tsk->state & TASK_UNINTERRUPTIBLE)
1002 __schedstat_set(se->statistics.block_start,
1003 rq_clock(rq_of(cfs_rq)));
1008 * We are picking a new current task - update its stats:
1011 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1014 * We are starting a new run period:
1016 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1019 /**************************************************
1020 * Scheduling class queueing methods:
1023 #ifdef CONFIG_NUMA_BALANCING
1025 * Approximate time to scan a full NUMA task in ms. The task scan period is
1026 * calculated based on the tasks virtual memory size and
1027 * numa_balancing_scan_size.
1029 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1030 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1032 /* Portion of address space to scan in MB */
1033 unsigned int sysctl_numa_balancing_scan_size = 256;
1035 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1036 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1041 spinlock_t lock; /* nr_tasks, tasks */
1046 struct rcu_head rcu;
1047 unsigned long total_faults;
1048 unsigned long max_faults_cpu;
1050 * Faults_cpu is used to decide whether memory should move
1051 * towards the CPU. As a consequence, these stats are weighted
1052 * more by CPU use than by memory faults.
1054 unsigned long *faults_cpu;
1055 unsigned long faults[0];
1058 static inline unsigned long group_faults_priv(struct numa_group *ng);
1059 static inline unsigned long group_faults_shared(struct numa_group *ng);
1061 static unsigned int task_nr_scan_windows(struct task_struct *p)
1063 unsigned long rss = 0;
1064 unsigned long nr_scan_pages;
1067 * Calculations based on RSS as non-present and empty pages are skipped
1068 * by the PTE scanner and NUMA hinting faults should be trapped based
1071 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1072 rss = get_mm_rss(p->mm);
1074 rss = nr_scan_pages;
1076 rss = round_up(rss, nr_scan_pages);
1077 return rss / nr_scan_pages;
1080 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1081 #define MAX_SCAN_WINDOW 2560
1083 static unsigned int task_scan_min(struct task_struct *p)
1085 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1086 unsigned int scan, floor;
1087 unsigned int windows = 1;
1089 if (scan_size < MAX_SCAN_WINDOW)
1090 windows = MAX_SCAN_WINDOW / scan_size;
1091 floor = 1000 / windows;
1093 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1094 return max_t(unsigned int, floor, scan);
1097 static unsigned int task_scan_start(struct task_struct *p)
1099 unsigned long smin = task_scan_min(p);
1100 unsigned long period = smin;
1102 /* Scale the maximum scan period with the amount of shared memory. */
1103 if (p->numa_group) {
1104 struct numa_group *ng = p->numa_group;
1105 unsigned long shared = group_faults_shared(ng);
1106 unsigned long private = group_faults_priv(ng);
1108 period *= atomic_read(&ng->refcount);
1109 period *= shared + 1;
1110 period /= private + shared + 1;
1113 return max(smin, period);
1116 static unsigned int task_scan_max(struct task_struct *p)
1118 unsigned long smin = task_scan_min(p);
1121 /* Watch for min being lower than max due to floor calculations */
1122 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1124 /* Scale the maximum scan period with the amount of shared memory. */
1125 if (p->numa_group) {
1126 struct numa_group *ng = p->numa_group;
1127 unsigned long shared = group_faults_shared(ng);
1128 unsigned long private = group_faults_priv(ng);
1129 unsigned long period = smax;
1131 period *= atomic_read(&ng->refcount);
1132 period *= shared + 1;
1133 period /= private + shared + 1;
1135 smax = max(smax, period);
1138 return max(smin, smax);
1141 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
1144 struct mm_struct *mm = p->mm;
1147 mm_users = atomic_read(&mm->mm_users);
1148 if (mm_users == 1) {
1149 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1150 mm->numa_scan_seq = 0;
1154 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
1155 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1156 p->numa_work.next = &p->numa_work;
1157 p->numa_faults = NULL;
1158 p->numa_group = NULL;
1159 p->last_task_numa_placement = 0;
1160 p->last_sum_exec_runtime = 0;
1162 /* New address space, reset the preferred nid */
1163 if (!(clone_flags & CLONE_VM)) {
1164 p->numa_preferred_nid = -1;
1169 * New thread, keep existing numa_preferred_nid which should be copied
1170 * already by arch_dup_task_struct but stagger when scans start.
1175 delay = min_t(unsigned int, task_scan_max(current),
1176 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
1177 delay += 2 * TICK_NSEC;
1178 p->node_stamp = delay;
1182 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1184 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1185 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1188 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1190 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1191 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1194 /* Shared or private faults. */
1195 #define NR_NUMA_HINT_FAULT_TYPES 2
1197 /* Memory and CPU locality */
1198 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1200 /* Averaged statistics, and temporary buffers. */
1201 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1203 pid_t task_numa_group_id(struct task_struct *p)
1205 return p->numa_group ? p->numa_group->gid : 0;
1209 * The averaged statistics, shared & private, memory & CPU,
1210 * occupy the first half of the array. The second half of the
1211 * array is for current counters, which are averaged into the
1212 * first set by task_numa_placement.
1214 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1216 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1219 static inline unsigned long task_faults(struct task_struct *p, int nid)
1221 if (!p->numa_faults)
1224 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1225 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1228 static inline unsigned long group_faults(struct task_struct *p, int nid)
1233 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1234 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1237 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1239 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1240 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1243 static inline unsigned long group_faults_priv(struct numa_group *ng)
1245 unsigned long faults = 0;
1248 for_each_online_node(node) {
1249 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1255 static inline unsigned long group_faults_shared(struct numa_group *ng)
1257 unsigned long faults = 0;
1260 for_each_online_node(node) {
1261 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1268 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1269 * considered part of a numa group's pseudo-interleaving set. Migrations
1270 * between these nodes are slowed down, to allow things to settle down.
1272 #define ACTIVE_NODE_FRACTION 3
1274 static bool numa_is_active_node(int nid, struct numa_group *ng)
1276 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1279 /* Handle placement on systems where not all nodes are directly connected. */
1280 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1281 int maxdist, bool task)
1283 unsigned long score = 0;
1287 * All nodes are directly connected, and the same distance
1288 * from each other. No need for fancy placement algorithms.
1290 if (sched_numa_topology_type == NUMA_DIRECT)
1294 * This code is called for each node, introducing N^2 complexity,
1295 * which should be ok given the number of nodes rarely exceeds 8.
1297 for_each_online_node(node) {
1298 unsigned long faults;
1299 int dist = node_distance(nid, node);
1302 * The furthest away nodes in the system are not interesting
1303 * for placement; nid was already counted.
1305 if (dist == sched_max_numa_distance || node == nid)
1309 * On systems with a backplane NUMA topology, compare groups
1310 * of nodes, and move tasks towards the group with the most
1311 * memory accesses. When comparing two nodes at distance
1312 * "hoplimit", only nodes closer by than "hoplimit" are part
1313 * of each group. Skip other nodes.
1315 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1319 /* Add up the faults from nearby nodes. */
1321 faults = task_faults(p, node);
1323 faults = group_faults(p, node);
1326 * On systems with a glueless mesh NUMA topology, there are
1327 * no fixed "groups of nodes". Instead, nodes that are not
1328 * directly connected bounce traffic through intermediate
1329 * nodes; a numa_group can occupy any set of nodes.
1330 * The further away a node is, the less the faults count.
1331 * This seems to result in good task placement.
1333 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1334 faults *= (sched_max_numa_distance - dist);
1335 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1345 * These return the fraction of accesses done by a particular task, or
1346 * task group, on a particular numa node. The group weight is given a
1347 * larger multiplier, in order to group tasks together that are almost
1348 * evenly spread out between numa nodes.
1350 static inline unsigned long task_weight(struct task_struct *p, int nid,
1353 unsigned long faults, total_faults;
1355 if (!p->numa_faults)
1358 total_faults = p->total_numa_faults;
1363 faults = task_faults(p, nid);
1364 faults += score_nearby_nodes(p, nid, dist, true);
1366 return 1000 * faults / total_faults;
1369 static inline unsigned long group_weight(struct task_struct *p, int nid,
1372 unsigned long faults, total_faults;
1377 total_faults = p->numa_group->total_faults;
1382 faults = group_faults(p, nid);
1383 faults += score_nearby_nodes(p, nid, dist, false);
1385 return 1000 * faults / total_faults;
1388 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1389 int src_nid, int dst_cpu)
1391 struct numa_group *ng = p->numa_group;
1392 int dst_nid = cpu_to_node(dst_cpu);
1393 int last_cpupid, this_cpupid;
1395 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1396 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1399 * Allow first faults or private faults to migrate immediately early in
1400 * the lifetime of a task. The magic number 4 is based on waiting for
1401 * two full passes of the "multi-stage node selection" test that is
1404 if ((p->numa_preferred_nid == -1 || p->numa_scan_seq <= 4) &&
1405 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1409 * Multi-stage node selection is used in conjunction with a periodic
1410 * migration fault to build a temporal task<->page relation. By using
1411 * a two-stage filter we remove short/unlikely relations.
1413 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1414 * a task's usage of a particular page (n_p) per total usage of this
1415 * page (n_t) (in a given time-span) to a probability.
1417 * Our periodic faults will sample this probability and getting the
1418 * same result twice in a row, given these samples are fully
1419 * independent, is then given by P(n)^2, provided our sample period
1420 * is sufficiently short compared to the usage pattern.
1422 * This quadric squishes small probabilities, making it less likely we
1423 * act on an unlikely task<->page relation.
1425 if (!cpupid_pid_unset(last_cpupid) &&
1426 cpupid_to_nid(last_cpupid) != dst_nid)
1429 /* Always allow migrate on private faults */
1430 if (cpupid_match_pid(p, last_cpupid))
1433 /* A shared fault, but p->numa_group has not been set up yet. */
1438 * Destination node is much more heavily used than the source
1439 * node? Allow migration.
1441 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1442 ACTIVE_NODE_FRACTION)
1446 * Distribute memory according to CPU & memory use on each node,
1447 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1449 * faults_cpu(dst) 3 faults_cpu(src)
1450 * --------------- * - > ---------------
1451 * faults_mem(dst) 4 faults_mem(src)
1453 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1454 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1457 static unsigned long weighted_cpuload(struct rq *rq);
1458 static unsigned long source_load(int cpu, int type);
1459 static unsigned long target_load(int cpu, int type);
1461 /* Cached statistics for all CPUs within a node */
1465 /* Total compute capacity of CPUs on a node */
1466 unsigned long compute_capacity;
1470 * XXX borrowed from update_sg_lb_stats
1472 static void update_numa_stats(struct numa_stats *ns, int nid)
1476 memset(ns, 0, sizeof(*ns));
1477 for_each_cpu(cpu, cpumask_of_node(nid)) {
1478 struct rq *rq = cpu_rq(cpu);
1480 ns->load += weighted_cpuload(rq);
1481 ns->compute_capacity += capacity_of(cpu);
1486 struct task_numa_env {
1487 struct task_struct *p;
1489 int src_cpu, src_nid;
1490 int dst_cpu, dst_nid;
1492 struct numa_stats src_stats, dst_stats;
1497 struct task_struct *best_task;
1502 static void task_numa_assign(struct task_numa_env *env,
1503 struct task_struct *p, long imp)
1505 struct rq *rq = cpu_rq(env->dst_cpu);
1507 /* Bail out if run-queue part of active NUMA balance. */
1508 if (xchg(&rq->numa_migrate_on, 1))
1512 * Clear previous best_cpu/rq numa-migrate flag, since task now
1513 * found a better CPU to move/swap.
1515 if (env->best_cpu != -1) {
1516 rq = cpu_rq(env->best_cpu);
1517 WRITE_ONCE(rq->numa_migrate_on, 0);
1521 put_task_struct(env->best_task);
1526 env->best_imp = imp;
1527 env->best_cpu = env->dst_cpu;
1530 static bool load_too_imbalanced(long src_load, long dst_load,
1531 struct task_numa_env *env)
1534 long orig_src_load, orig_dst_load;
1535 long src_capacity, dst_capacity;
1538 * The load is corrected for the CPU capacity available on each node.
1541 * ------------ vs ---------
1542 * src_capacity dst_capacity
1544 src_capacity = env->src_stats.compute_capacity;
1545 dst_capacity = env->dst_stats.compute_capacity;
1547 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1549 orig_src_load = env->src_stats.load;
1550 orig_dst_load = env->dst_stats.load;
1552 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1554 /* Would this change make things worse? */
1555 return (imb > old_imb);
1559 * Maximum NUMA importance can be 1998 (2*999);
1560 * SMALLIMP @ 30 would be close to 1998/64.
1561 * Used to deter task migration.
1566 * This checks if the overall compute and NUMA accesses of the system would
1567 * be improved if the source tasks was migrated to the target dst_cpu taking
1568 * into account that it might be best if task running on the dst_cpu should
1569 * be exchanged with the source task
1571 static void task_numa_compare(struct task_numa_env *env,
1572 long taskimp, long groupimp, bool maymove)
1574 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1575 struct task_struct *cur;
1576 long src_load, dst_load;
1578 long imp = env->p->numa_group ? groupimp : taskimp;
1580 int dist = env->dist;
1582 if (READ_ONCE(dst_rq->numa_migrate_on))
1586 cur = task_rcu_dereference(&dst_rq->curr);
1587 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1591 * Because we have preemption enabled we can get migrated around and
1592 * end try selecting ourselves (current == env->p) as a swap candidate.
1598 if (maymove && moveimp >= env->best_imp)
1605 * "imp" is the fault differential for the source task between the
1606 * source and destination node. Calculate the total differential for
1607 * the source task and potential destination task. The more negative
1608 * the value is, the more remote accesses that would be expected to
1609 * be incurred if the tasks were swapped.
1611 /* Skip this swap candidate if cannot move to the source cpu */
1612 if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed))
1616 * If dst and source tasks are in the same NUMA group, or not
1617 * in any group then look only at task weights.
1619 if (cur->numa_group == env->p->numa_group) {
1620 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1621 task_weight(cur, env->dst_nid, dist);
1623 * Add some hysteresis to prevent swapping the
1624 * tasks within a group over tiny differences.
1626 if (cur->numa_group)
1630 * Compare the group weights. If a task is all by itself
1631 * (not part of a group), use the task weight instead.
1633 if (cur->numa_group && env->p->numa_group)
1634 imp += group_weight(cur, env->src_nid, dist) -
1635 group_weight(cur, env->dst_nid, dist);
1637 imp += task_weight(cur, env->src_nid, dist) -
1638 task_weight(cur, env->dst_nid, dist);
1641 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1648 * If the NUMA importance is less than SMALLIMP,
1649 * task migration might only result in ping pong
1650 * of tasks and also hurt performance due to cache
1653 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1657 * In the overloaded case, try and keep the load balanced.
1659 load = task_h_load(env->p) - task_h_load(cur);
1663 dst_load = env->dst_stats.load + load;
1664 src_load = env->src_stats.load - load;
1666 if (load_too_imbalanced(src_load, dst_load, env))
1671 * One idle CPU per node is evaluated for a task numa move.
1672 * Call select_idle_sibling to maybe find a better one.
1676 * select_idle_siblings() uses an per-CPU cpumask that
1677 * can be used from IRQ context.
1679 local_irq_disable();
1680 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1685 task_numa_assign(env, cur, imp);
1690 static void task_numa_find_cpu(struct task_numa_env *env,
1691 long taskimp, long groupimp)
1693 long src_load, dst_load, load;
1694 bool maymove = false;
1697 load = task_h_load(env->p);
1698 dst_load = env->dst_stats.load + load;
1699 src_load = env->src_stats.load - load;
1702 * If the improvement from just moving env->p direction is better
1703 * than swapping tasks around, check if a move is possible.
1705 maymove = !load_too_imbalanced(src_load, dst_load, env);
1707 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1708 /* Skip this CPU if the source task cannot migrate */
1709 if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
1713 task_numa_compare(env, taskimp, groupimp, maymove);
1717 static int task_numa_migrate(struct task_struct *p)
1719 struct task_numa_env env = {
1722 .src_cpu = task_cpu(p),
1723 .src_nid = task_node(p),
1725 .imbalance_pct = 112,
1731 struct sched_domain *sd;
1733 unsigned long taskweight, groupweight;
1735 long taskimp, groupimp;
1738 * Pick the lowest SD_NUMA domain, as that would have the smallest
1739 * imbalance and would be the first to start moving tasks about.
1741 * And we want to avoid any moving of tasks about, as that would create
1742 * random movement of tasks -- counter the numa conditions we're trying
1746 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1748 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1752 * Cpusets can break the scheduler domain tree into smaller
1753 * balance domains, some of which do not cross NUMA boundaries.
1754 * Tasks that are "trapped" in such domains cannot be migrated
1755 * elsewhere, so there is no point in (re)trying.
1757 if (unlikely(!sd)) {
1758 sched_setnuma(p, task_node(p));
1762 env.dst_nid = p->numa_preferred_nid;
1763 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1764 taskweight = task_weight(p, env.src_nid, dist);
1765 groupweight = group_weight(p, env.src_nid, dist);
1766 update_numa_stats(&env.src_stats, env.src_nid);
1767 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1768 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1769 update_numa_stats(&env.dst_stats, env.dst_nid);
1771 /* Try to find a spot on the preferred nid. */
1772 task_numa_find_cpu(&env, taskimp, groupimp);
1775 * Look at other nodes in these cases:
1776 * - there is no space available on the preferred_nid
1777 * - the task is part of a numa_group that is interleaved across
1778 * multiple NUMA nodes; in order to better consolidate the group,
1779 * we need to check other locations.
1781 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1782 for_each_online_node(nid) {
1783 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1786 dist = node_distance(env.src_nid, env.dst_nid);
1787 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1789 taskweight = task_weight(p, env.src_nid, dist);
1790 groupweight = group_weight(p, env.src_nid, dist);
1793 /* Only consider nodes where both task and groups benefit */
1794 taskimp = task_weight(p, nid, dist) - taskweight;
1795 groupimp = group_weight(p, nid, dist) - groupweight;
1796 if (taskimp < 0 && groupimp < 0)
1801 update_numa_stats(&env.dst_stats, env.dst_nid);
1802 task_numa_find_cpu(&env, taskimp, groupimp);
1807 * If the task is part of a workload that spans multiple NUMA nodes,
1808 * and is migrating into one of the workload's active nodes, remember
1809 * this node as the task's preferred numa node, so the workload can
1811 * A task that migrated to a second choice node will be better off
1812 * trying for a better one later. Do not set the preferred node here.
1814 if (p->numa_group) {
1815 if (env.best_cpu == -1)
1818 nid = cpu_to_node(env.best_cpu);
1820 if (nid != p->numa_preferred_nid)
1821 sched_setnuma(p, nid);
1824 /* No better CPU than the current one was found. */
1825 if (env.best_cpu == -1)
1828 best_rq = cpu_rq(env.best_cpu);
1829 if (env.best_task == NULL) {
1830 ret = migrate_task_to(p, env.best_cpu);
1831 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1833 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1837 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
1838 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1841 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1842 put_task_struct(env.best_task);
1846 /* Attempt to migrate a task to a CPU on the preferred node. */
1847 static void numa_migrate_preferred(struct task_struct *p)
1849 unsigned long interval = HZ;
1851 /* This task has no NUMA fault statistics yet */
1852 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1855 /* Periodically retry migrating the task to the preferred node */
1856 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1857 p->numa_migrate_retry = jiffies + interval;
1859 /* Success if task is already running on preferred CPU */
1860 if (task_node(p) == p->numa_preferred_nid)
1863 /* Otherwise, try migrate to a CPU on the preferred node */
1864 task_numa_migrate(p);
1868 * Find out how many nodes on the workload is actively running on. Do this by
1869 * tracking the nodes from which NUMA hinting faults are triggered. This can
1870 * be different from the set of nodes where the workload's memory is currently
1873 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1875 unsigned long faults, max_faults = 0;
1876 int nid, active_nodes = 0;
1878 for_each_online_node(nid) {
1879 faults = group_faults_cpu(numa_group, nid);
1880 if (faults > max_faults)
1881 max_faults = faults;
1884 for_each_online_node(nid) {
1885 faults = group_faults_cpu(numa_group, nid);
1886 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1890 numa_group->max_faults_cpu = max_faults;
1891 numa_group->active_nodes = active_nodes;
1895 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1896 * increments. The more local the fault statistics are, the higher the scan
1897 * period will be for the next scan window. If local/(local+remote) ratio is
1898 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1899 * the scan period will decrease. Aim for 70% local accesses.
1901 #define NUMA_PERIOD_SLOTS 10
1902 #define NUMA_PERIOD_THRESHOLD 7
1905 * Increase the scan period (slow down scanning) if the majority of
1906 * our memory is already on our local node, or if the majority of
1907 * the page accesses are shared with other processes.
1908 * Otherwise, decrease the scan period.
1910 static void update_task_scan_period(struct task_struct *p,
1911 unsigned long shared, unsigned long private)
1913 unsigned int period_slot;
1914 int lr_ratio, ps_ratio;
1917 unsigned long remote = p->numa_faults_locality[0];
1918 unsigned long local = p->numa_faults_locality[1];
1921 * If there were no record hinting faults then either the task is
1922 * completely idle or all activity is areas that are not of interest
1923 * to automatic numa balancing. Related to that, if there were failed
1924 * migration then it implies we are migrating too quickly or the local
1925 * node is overloaded. In either case, scan slower
1927 if (local + shared == 0 || p->numa_faults_locality[2]) {
1928 p->numa_scan_period = min(p->numa_scan_period_max,
1929 p->numa_scan_period << 1);
1931 p->mm->numa_next_scan = jiffies +
1932 msecs_to_jiffies(p->numa_scan_period);
1938 * Prepare to scale scan period relative to the current period.
1939 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1940 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1941 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1943 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1944 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1945 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1947 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1949 * Most memory accesses are local. There is no need to
1950 * do fast NUMA scanning, since memory is already local.
1952 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
1955 diff = slot * period_slot;
1956 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
1958 * Most memory accesses are shared with other tasks.
1959 * There is no point in continuing fast NUMA scanning,
1960 * since other tasks may just move the memory elsewhere.
1962 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
1965 diff = slot * period_slot;
1968 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
1969 * yet they are not on the local NUMA node. Speed up
1970 * NUMA scanning to get the memory moved over.
1972 int ratio = max(lr_ratio, ps_ratio);
1973 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1976 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1977 task_scan_min(p), task_scan_max(p));
1978 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1982 * Get the fraction of time the task has been running since the last
1983 * NUMA placement cycle. The scheduler keeps similar statistics, but
1984 * decays those on a 32ms period, which is orders of magnitude off
1985 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1986 * stats only if the task is so new there are no NUMA statistics yet.
1988 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1990 u64 runtime, delta, now;
1991 /* Use the start of this time slice to avoid calculations. */
1992 now = p->se.exec_start;
1993 runtime = p->se.sum_exec_runtime;
1995 if (p->last_task_numa_placement) {
1996 delta = runtime - p->last_sum_exec_runtime;
1997 *period = now - p->last_task_numa_placement;
1999 delta = p->se.avg.load_sum;
2000 *period = LOAD_AVG_MAX;
2003 p->last_sum_exec_runtime = runtime;
2004 p->last_task_numa_placement = now;
2010 * Determine the preferred nid for a task in a numa_group. This needs to
2011 * be done in a way that produces consistent results with group_weight,
2012 * otherwise workloads might not converge.
2014 static int preferred_group_nid(struct task_struct *p, int nid)
2019 /* Direct connections between all NUMA nodes. */
2020 if (sched_numa_topology_type == NUMA_DIRECT)
2024 * On a system with glueless mesh NUMA topology, group_weight
2025 * scores nodes according to the number of NUMA hinting faults on
2026 * both the node itself, and on nearby nodes.
2028 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2029 unsigned long score, max_score = 0;
2030 int node, max_node = nid;
2032 dist = sched_max_numa_distance;
2034 for_each_online_node(node) {
2035 score = group_weight(p, node, dist);
2036 if (score > max_score) {
2045 * Finding the preferred nid in a system with NUMA backplane
2046 * interconnect topology is more involved. The goal is to locate
2047 * tasks from numa_groups near each other in the system, and
2048 * untangle workloads from different sides of the system. This requires
2049 * searching down the hierarchy of node groups, recursively searching
2050 * inside the highest scoring group of nodes. The nodemask tricks
2051 * keep the complexity of the search down.
2053 nodes = node_online_map;
2054 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2055 unsigned long max_faults = 0;
2056 nodemask_t max_group = NODE_MASK_NONE;
2059 /* Are there nodes at this distance from each other? */
2060 if (!find_numa_distance(dist))
2063 for_each_node_mask(a, nodes) {
2064 unsigned long faults = 0;
2065 nodemask_t this_group;
2066 nodes_clear(this_group);
2068 /* Sum group's NUMA faults; includes a==b case. */
2069 for_each_node_mask(b, nodes) {
2070 if (node_distance(a, b) < dist) {
2071 faults += group_faults(p, b);
2072 node_set(b, this_group);
2073 node_clear(b, nodes);
2077 /* Remember the top group. */
2078 if (faults > max_faults) {
2079 max_faults = faults;
2080 max_group = this_group;
2082 * subtle: at the smallest distance there is
2083 * just one node left in each "group", the
2084 * winner is the preferred nid.
2089 /* Next round, evaluate the nodes within max_group. */
2097 static void task_numa_placement(struct task_struct *p)
2099 int seq, nid, max_nid = -1;
2100 unsigned long max_faults = 0;
2101 unsigned long fault_types[2] = { 0, 0 };
2102 unsigned long total_faults;
2103 u64 runtime, period;
2104 spinlock_t *group_lock = NULL;
2107 * The p->mm->numa_scan_seq field gets updated without
2108 * exclusive access. Use READ_ONCE() here to ensure
2109 * that the field is read in a single access:
2111 seq = READ_ONCE(p->mm->numa_scan_seq);
2112 if (p->numa_scan_seq == seq)
2114 p->numa_scan_seq = seq;
2115 p->numa_scan_period_max = task_scan_max(p);
2117 total_faults = p->numa_faults_locality[0] +
2118 p->numa_faults_locality[1];
2119 runtime = numa_get_avg_runtime(p, &period);
2121 /* If the task is part of a group prevent parallel updates to group stats */
2122 if (p->numa_group) {
2123 group_lock = &p->numa_group->lock;
2124 spin_lock_irq(group_lock);
2127 /* Find the node with the highest number of faults */
2128 for_each_online_node(nid) {
2129 /* Keep track of the offsets in numa_faults array */
2130 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2131 unsigned long faults = 0, group_faults = 0;
2134 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2135 long diff, f_diff, f_weight;
2137 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2138 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2139 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2140 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2142 /* Decay existing window, copy faults since last scan */
2143 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2144 fault_types[priv] += p->numa_faults[membuf_idx];
2145 p->numa_faults[membuf_idx] = 0;
2148 * Normalize the faults_from, so all tasks in a group
2149 * count according to CPU use, instead of by the raw
2150 * number of faults. Tasks with little runtime have
2151 * little over-all impact on throughput, and thus their
2152 * faults are less important.
2154 f_weight = div64_u64(runtime << 16, period + 1);
2155 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2157 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2158 p->numa_faults[cpubuf_idx] = 0;
2160 p->numa_faults[mem_idx] += diff;
2161 p->numa_faults[cpu_idx] += f_diff;
2162 faults += p->numa_faults[mem_idx];
2163 p->total_numa_faults += diff;
2164 if (p->numa_group) {
2166 * safe because we can only change our own group
2168 * mem_idx represents the offset for a given
2169 * nid and priv in a specific region because it
2170 * is at the beginning of the numa_faults array.
2172 p->numa_group->faults[mem_idx] += diff;
2173 p->numa_group->faults_cpu[mem_idx] += f_diff;
2174 p->numa_group->total_faults += diff;
2175 group_faults += p->numa_group->faults[mem_idx];
2179 if (!p->numa_group) {
2180 if (faults > max_faults) {
2181 max_faults = faults;
2184 } else if (group_faults > max_faults) {
2185 max_faults = group_faults;
2190 if (p->numa_group) {
2191 numa_group_count_active_nodes(p->numa_group);
2192 spin_unlock_irq(group_lock);
2193 max_nid = preferred_group_nid(p, max_nid);
2197 /* Set the new preferred node */
2198 if (max_nid != p->numa_preferred_nid)
2199 sched_setnuma(p, max_nid);
2202 update_task_scan_period(p, fault_types[0], fault_types[1]);
2205 static inline int get_numa_group(struct numa_group *grp)
2207 return atomic_inc_not_zero(&grp->refcount);
2210 static inline void put_numa_group(struct numa_group *grp)
2212 if (atomic_dec_and_test(&grp->refcount))
2213 kfree_rcu(grp, rcu);
2216 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2219 struct numa_group *grp, *my_grp;
2220 struct task_struct *tsk;
2222 int cpu = cpupid_to_cpu(cpupid);
2225 if (unlikely(!p->numa_group)) {
2226 unsigned int size = sizeof(struct numa_group) +
2227 4*nr_node_ids*sizeof(unsigned long);
2229 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2233 atomic_set(&grp->refcount, 1);
2234 grp->active_nodes = 1;
2235 grp->max_faults_cpu = 0;
2236 spin_lock_init(&grp->lock);
2238 /* Second half of the array tracks nids where faults happen */
2239 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2242 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2243 grp->faults[i] = p->numa_faults[i];
2245 grp->total_faults = p->total_numa_faults;
2248 rcu_assign_pointer(p->numa_group, grp);
2252 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2254 if (!cpupid_match_pid(tsk, cpupid))
2257 grp = rcu_dereference(tsk->numa_group);
2261 my_grp = p->numa_group;
2266 * Only join the other group if its bigger; if we're the bigger group,
2267 * the other task will join us.
2269 if (my_grp->nr_tasks > grp->nr_tasks)
2273 * Tie-break on the grp address.
2275 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2278 /* Always join threads in the same process. */
2279 if (tsk->mm == current->mm)
2282 /* Simple filter to avoid false positives due to PID collisions */
2283 if (flags & TNF_SHARED)
2286 /* Update priv based on whether false sharing was detected */
2289 if (join && !get_numa_group(grp))
2297 BUG_ON(irqs_disabled());
2298 double_lock_irq(&my_grp->lock, &grp->lock);
2300 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2301 my_grp->faults[i] -= p->numa_faults[i];
2302 grp->faults[i] += p->numa_faults[i];
2304 my_grp->total_faults -= p->total_numa_faults;
2305 grp->total_faults += p->total_numa_faults;
2310 spin_unlock(&my_grp->lock);
2311 spin_unlock_irq(&grp->lock);
2313 rcu_assign_pointer(p->numa_group, grp);
2315 put_numa_group(my_grp);
2323 void task_numa_free(struct task_struct *p)
2325 struct numa_group *grp = p->numa_group;
2326 void *numa_faults = p->numa_faults;
2327 unsigned long flags;
2331 spin_lock_irqsave(&grp->lock, flags);
2332 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2333 grp->faults[i] -= p->numa_faults[i];
2334 grp->total_faults -= p->total_numa_faults;
2337 spin_unlock_irqrestore(&grp->lock, flags);
2338 RCU_INIT_POINTER(p->numa_group, NULL);
2339 put_numa_group(grp);
2342 p->numa_faults = NULL;
2347 * Got a PROT_NONE fault for a page on @node.
2349 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2351 struct task_struct *p = current;
2352 bool migrated = flags & TNF_MIGRATED;
2353 int cpu_node = task_node(current);
2354 int local = !!(flags & TNF_FAULT_LOCAL);
2355 struct numa_group *ng;
2358 if (!static_branch_likely(&sched_numa_balancing))
2361 /* for example, ksmd faulting in a user's mm */
2365 /* Allocate buffer to track faults on a per-node basis */
2366 if (unlikely(!p->numa_faults)) {
2367 int size = sizeof(*p->numa_faults) *
2368 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2370 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2371 if (!p->numa_faults)
2374 p->total_numa_faults = 0;
2375 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2379 * First accesses are treated as private, otherwise consider accesses
2380 * to be private if the accessing pid has not changed
2382 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2385 priv = cpupid_match_pid(p, last_cpupid);
2386 if (!priv && !(flags & TNF_NO_GROUP))
2387 task_numa_group(p, last_cpupid, flags, &priv);
2391 * If a workload spans multiple NUMA nodes, a shared fault that
2392 * occurs wholly within the set of nodes that the workload is
2393 * actively using should be counted as local. This allows the
2394 * scan rate to slow down when a workload has settled down.
2397 if (!priv && !local && ng && ng->active_nodes > 1 &&
2398 numa_is_active_node(cpu_node, ng) &&
2399 numa_is_active_node(mem_node, ng))
2403 * Retry to migrate task to preferred node periodically, in case it
2404 * previously failed, or the scheduler moved us.
2406 if (time_after(jiffies, p->numa_migrate_retry)) {
2407 task_numa_placement(p);
2408 numa_migrate_preferred(p);
2412 p->numa_pages_migrated += pages;
2413 if (flags & TNF_MIGRATE_FAIL)
2414 p->numa_faults_locality[2] += pages;
2416 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2417 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2418 p->numa_faults_locality[local] += pages;
2421 static void reset_ptenuma_scan(struct task_struct *p)
2424 * We only did a read acquisition of the mmap sem, so
2425 * p->mm->numa_scan_seq is written to without exclusive access
2426 * and the update is not guaranteed to be atomic. That's not
2427 * much of an issue though, since this is just used for
2428 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2429 * expensive, to avoid any form of compiler optimizations:
2431 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2432 p->mm->numa_scan_offset = 0;
2436 * The expensive part of numa migration is done from task_work context.
2437 * Triggered from task_tick_numa().
2439 void task_numa_work(struct callback_head *work)
2441 unsigned long migrate, next_scan, now = jiffies;
2442 struct task_struct *p = current;
2443 struct mm_struct *mm = p->mm;
2444 u64 runtime = p->se.sum_exec_runtime;
2445 struct vm_area_struct *vma;
2446 unsigned long start, end;
2447 unsigned long nr_pte_updates = 0;
2448 long pages, virtpages;
2450 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2452 work->next = work; /* protect against double add */
2454 * Who cares about NUMA placement when they're dying.
2456 * NOTE: make sure not to dereference p->mm before this check,
2457 * exit_task_work() happens _after_ exit_mm() so we could be called
2458 * without p->mm even though we still had it when we enqueued this
2461 if (p->flags & PF_EXITING)
2464 if (!mm->numa_next_scan) {
2465 mm->numa_next_scan = now +
2466 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2470 * Enforce maximal scan/migration frequency..
2472 migrate = mm->numa_next_scan;
2473 if (time_before(now, migrate))
2476 if (p->numa_scan_period == 0) {
2477 p->numa_scan_period_max = task_scan_max(p);
2478 p->numa_scan_period = task_scan_start(p);
2481 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2482 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2486 * Delay this task enough that another task of this mm will likely win
2487 * the next time around.
2489 p->node_stamp += 2 * TICK_NSEC;
2491 start = mm->numa_scan_offset;
2492 pages = sysctl_numa_balancing_scan_size;
2493 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2494 virtpages = pages * 8; /* Scan up to this much virtual space */
2499 if (!down_read_trylock(&mm->mmap_sem))
2501 vma = find_vma(mm, start);
2503 reset_ptenuma_scan(p);
2507 for (; vma; vma = vma->vm_next) {
2508 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2509 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2514 * Shared library pages mapped by multiple processes are not
2515 * migrated as it is expected they are cache replicated. Avoid
2516 * hinting faults in read-only file-backed mappings or the vdso
2517 * as migrating the pages will be of marginal benefit.
2520 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2524 * Skip inaccessible VMAs to avoid any confusion between
2525 * PROT_NONE and NUMA hinting ptes
2527 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2531 start = max(start, vma->vm_start);
2532 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2533 end = min(end, vma->vm_end);
2534 nr_pte_updates = change_prot_numa(vma, start, end);
2537 * Try to scan sysctl_numa_balancing_size worth of
2538 * hpages that have at least one present PTE that
2539 * is not already pte-numa. If the VMA contains
2540 * areas that are unused or already full of prot_numa
2541 * PTEs, scan up to virtpages, to skip through those
2545 pages -= (end - start) >> PAGE_SHIFT;
2546 virtpages -= (end - start) >> PAGE_SHIFT;
2549 if (pages <= 0 || virtpages <= 0)
2553 } while (end != vma->vm_end);
2558 * It is possible to reach the end of the VMA list but the last few
2559 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2560 * would find the !migratable VMA on the next scan but not reset the
2561 * scanner to the start so check it now.
2564 mm->numa_scan_offset = start;
2566 reset_ptenuma_scan(p);
2567 up_read(&mm->mmap_sem);
2570 * Make sure tasks use at least 32x as much time to run other code
2571 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2572 * Usually update_task_scan_period slows down scanning enough; on an
2573 * overloaded system we need to limit overhead on a per task basis.
2575 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2576 u64 diff = p->se.sum_exec_runtime - runtime;
2577 p->node_stamp += 32 * diff;
2582 * Drive the periodic memory faults..
2584 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2586 struct callback_head *work = &curr->numa_work;
2590 * We don't care about NUMA placement if we don't have memory.
2592 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2596 * Using runtime rather than walltime has the dual advantage that
2597 * we (mostly) drive the selection from busy threads and that the
2598 * task needs to have done some actual work before we bother with
2601 now = curr->se.sum_exec_runtime;
2602 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2604 if (now > curr->node_stamp + period) {
2605 if (!curr->node_stamp)
2606 curr->numa_scan_period = task_scan_start(curr);
2607 curr->node_stamp += period;
2609 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2610 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2611 task_work_add(curr, work, true);
2616 static void update_scan_period(struct task_struct *p, int new_cpu)
2618 int src_nid = cpu_to_node(task_cpu(p));
2619 int dst_nid = cpu_to_node(new_cpu);
2621 if (!static_branch_likely(&sched_numa_balancing))
2624 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2627 if (src_nid == dst_nid)
2631 * Allow resets if faults have been trapped before one scan
2632 * has completed. This is most likely due to a new task that
2633 * is pulled cross-node due to wakeups or load balancing.
2635 if (p->numa_scan_seq) {
2637 * Avoid scan adjustments if moving to the preferred
2638 * node or if the task was not previously running on
2639 * the preferred node.
2641 if (dst_nid == p->numa_preferred_nid ||
2642 (p->numa_preferred_nid != -1 && src_nid != p->numa_preferred_nid))
2646 p->numa_scan_period = task_scan_start(p);
2650 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2654 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2658 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2662 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2666 #endif /* CONFIG_NUMA_BALANCING */
2669 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2671 update_load_add(&cfs_rq->load, se->load.weight);
2672 if (!parent_entity(se))
2673 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2675 if (entity_is_task(se)) {
2676 struct rq *rq = rq_of(cfs_rq);
2678 account_numa_enqueue(rq, task_of(se));
2679 list_add(&se->group_node, &rq->cfs_tasks);
2682 cfs_rq->nr_running++;
2686 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2688 update_load_sub(&cfs_rq->load, se->load.weight);
2689 if (!parent_entity(se))
2690 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2692 if (entity_is_task(se)) {
2693 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2694 list_del_init(&se->group_node);
2697 cfs_rq->nr_running--;
2701 * Signed add and clamp on underflow.
2703 * Explicitly do a load-store to ensure the intermediate value never hits
2704 * memory. This allows lockless observations without ever seeing the negative
2707 #define add_positive(_ptr, _val) do { \
2708 typeof(_ptr) ptr = (_ptr); \
2709 typeof(_val) val = (_val); \
2710 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2714 if (val < 0 && res > var) \
2717 WRITE_ONCE(*ptr, res); \
2721 * Unsigned subtract and clamp on underflow.
2723 * Explicitly do a load-store to ensure the intermediate value never hits
2724 * memory. This allows lockless observations without ever seeing the negative
2727 #define sub_positive(_ptr, _val) do { \
2728 typeof(_ptr) ptr = (_ptr); \
2729 typeof(*ptr) val = (_val); \
2730 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2734 WRITE_ONCE(*ptr, res); \
2738 * Remove and clamp on negative, from a local variable.
2740 * A variant of sub_positive(), which does not use explicit load-store
2741 * and is thus optimized for local variable updates.
2743 #define lsub_positive(_ptr, _val) do { \
2744 typeof(_ptr) ptr = (_ptr); \
2745 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
2750 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2752 cfs_rq->runnable_weight += se->runnable_weight;
2754 cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2755 cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2759 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2761 cfs_rq->runnable_weight -= se->runnable_weight;
2763 sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2764 sub_positive(&cfs_rq->avg.runnable_load_sum,
2765 se_runnable(se) * se->avg.runnable_load_sum);
2769 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2771 cfs_rq->avg.load_avg += se->avg.load_avg;
2772 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2776 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2778 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2779 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2783 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2785 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2787 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2789 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2792 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2793 unsigned long weight, unsigned long runnable)
2796 /* commit outstanding execution time */
2797 if (cfs_rq->curr == se)
2798 update_curr(cfs_rq);
2799 account_entity_dequeue(cfs_rq, se);
2800 dequeue_runnable_load_avg(cfs_rq, se);
2802 dequeue_load_avg(cfs_rq, se);
2804 se->runnable_weight = runnable;
2805 update_load_set(&se->load, weight);
2809 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2811 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2812 se->avg.runnable_load_avg =
2813 div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2817 enqueue_load_avg(cfs_rq, se);
2819 account_entity_enqueue(cfs_rq, se);
2820 enqueue_runnable_load_avg(cfs_rq, se);
2824 void reweight_task(struct task_struct *p, int prio)
2826 struct sched_entity *se = &p->se;
2827 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2828 struct load_weight *load = &se->load;
2829 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2831 reweight_entity(cfs_rq, se, weight, weight);
2832 load->inv_weight = sched_prio_to_wmult[prio];
2835 #ifdef CONFIG_FAIR_GROUP_SCHED
2838 * All this does is approximate the hierarchical proportion which includes that
2839 * global sum we all love to hate.
2841 * That is, the weight of a group entity, is the proportional share of the
2842 * group weight based on the group runqueue weights. That is:
2844 * tg->weight * grq->load.weight
2845 * ge->load.weight = ----------------------------- (1)
2846 * \Sum grq->load.weight
2848 * Now, because computing that sum is prohibitively expensive to compute (been
2849 * there, done that) we approximate it with this average stuff. The average
2850 * moves slower and therefore the approximation is cheaper and more stable.
2852 * So instead of the above, we substitute:
2854 * grq->load.weight -> grq->avg.load_avg (2)
2856 * which yields the following:
2858 * tg->weight * grq->avg.load_avg
2859 * ge->load.weight = ------------------------------ (3)
2862 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2864 * That is shares_avg, and it is right (given the approximation (2)).
2866 * The problem with it is that because the average is slow -- it was designed
2867 * to be exactly that of course -- this leads to transients in boundary
2868 * conditions. In specific, the case where the group was idle and we start the
2869 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2870 * yielding bad latency etc..
2872 * Now, in that special case (1) reduces to:
2874 * tg->weight * grq->load.weight
2875 * ge->load.weight = ----------------------------- = tg->weight (4)
2878 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2880 * So what we do is modify our approximation (3) to approach (4) in the (near)
2885 * tg->weight * grq->load.weight
2886 * --------------------------------------------------- (5)
2887 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2889 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2890 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2893 * tg->weight * grq->load.weight
2894 * ge->load.weight = ----------------------------- (6)
2899 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2900 * max(grq->load.weight, grq->avg.load_avg)
2902 * And that is shares_weight and is icky. In the (near) UP case it approaches
2903 * (4) while in the normal case it approaches (3). It consistently
2904 * overestimates the ge->load.weight and therefore:
2906 * \Sum ge->load.weight >= tg->weight
2910 static long calc_group_shares(struct cfs_rq *cfs_rq)
2912 long tg_weight, tg_shares, load, shares;
2913 struct task_group *tg = cfs_rq->tg;
2915 tg_shares = READ_ONCE(tg->shares);
2917 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
2919 tg_weight = atomic_long_read(&tg->load_avg);
2921 /* Ensure tg_weight >= load */
2922 tg_weight -= cfs_rq->tg_load_avg_contrib;
2925 shares = (tg_shares * load);
2927 shares /= tg_weight;
2930 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2931 * of a group with small tg->shares value. It is a floor value which is
2932 * assigned as a minimum load.weight to the sched_entity representing
2933 * the group on a CPU.
2935 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2936 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2937 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2938 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2941 return clamp_t(long, shares, MIN_SHARES, tg_shares);
2945 * This calculates the effective runnable weight for a group entity based on
2946 * the group entity weight calculated above.
2948 * Because of the above approximation (2), our group entity weight is
2949 * an load_avg based ratio (3). This means that it includes blocked load and
2950 * does not represent the runnable weight.
2952 * Approximate the group entity's runnable weight per ratio from the group
2955 * grq->avg.runnable_load_avg
2956 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
2959 * However, analogous to above, since the avg numbers are slow, this leads to
2960 * transients in the from-idle case. Instead we use:
2962 * ge->runnable_weight = ge->load.weight *
2964 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
2965 * ----------------------------------------------------- (8)
2966 * max(grq->avg.load_avg, grq->load.weight)
2968 * Where these max() serve both to use the 'instant' values to fix the slow
2969 * from-idle and avoid the /0 on to-idle, similar to (6).
2971 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
2973 long runnable, load_avg;
2975 load_avg = max(cfs_rq->avg.load_avg,
2976 scale_load_down(cfs_rq->load.weight));
2978 runnable = max(cfs_rq->avg.runnable_load_avg,
2979 scale_load_down(cfs_rq->runnable_weight));
2983 runnable /= load_avg;
2985 return clamp_t(long, runnable, MIN_SHARES, shares);
2987 #endif /* CONFIG_SMP */
2989 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2992 * Recomputes the group entity based on the current state of its group
2995 static void update_cfs_group(struct sched_entity *se)
2997 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
2998 long shares, runnable;
3003 if (throttled_hierarchy(gcfs_rq))
3007 runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
3009 if (likely(se->load.weight == shares))
3012 shares = calc_group_shares(gcfs_rq);
3013 runnable = calc_group_runnable(gcfs_rq, shares);
3016 reweight_entity(cfs_rq_of(se), se, shares, runnable);
3019 #else /* CONFIG_FAIR_GROUP_SCHED */
3020 static inline void update_cfs_group(struct sched_entity *se)
3023 #endif /* CONFIG_FAIR_GROUP_SCHED */
3025 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3027 struct rq *rq = rq_of(cfs_rq);
3029 if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) {
3031 * There are a few boundary cases this might miss but it should
3032 * get called often enough that that should (hopefully) not be
3035 * It will not get called when we go idle, because the idle
3036 * thread is a different class (!fair), nor will the utilization
3037 * number include things like RT tasks.
3039 * As is, the util number is not freq-invariant (we'd have to
3040 * implement arch_scale_freq_capacity() for that).
3044 cpufreq_update_util(rq, flags);
3049 #ifdef CONFIG_FAIR_GROUP_SCHED
3051 * update_tg_load_avg - update the tg's load avg
3052 * @cfs_rq: the cfs_rq whose avg changed
3053 * @force: update regardless of how small the difference
3055 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3056 * However, because tg->load_avg is a global value there are performance
3059 * In order to avoid having to look at the other cfs_rq's, we use a
3060 * differential update where we store the last value we propagated. This in
3061 * turn allows skipping updates if the differential is 'small'.
3063 * Updating tg's load_avg is necessary before update_cfs_share().
3065 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3067 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3070 * No need to update load_avg for root_task_group as it is not used.
3072 if (cfs_rq->tg == &root_task_group)
3075 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3076 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3077 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3082 * Called within set_task_rq() right before setting a task's CPU. The
3083 * caller only guarantees p->pi_lock is held; no other assumptions,
3084 * including the state of rq->lock, should be made.
3086 void set_task_rq_fair(struct sched_entity *se,
3087 struct cfs_rq *prev, struct cfs_rq *next)
3089 u64 p_last_update_time;
3090 u64 n_last_update_time;
3092 if (!sched_feat(ATTACH_AGE_LOAD))
3096 * We are supposed to update the task to "current" time, then its up to
3097 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3098 * getting what current time is, so simply throw away the out-of-date
3099 * time. This will result in the wakee task is less decayed, but giving
3100 * the wakee more load sounds not bad.
3102 if (!(se->avg.last_update_time && prev))
3105 #ifndef CONFIG_64BIT
3107 u64 p_last_update_time_copy;
3108 u64 n_last_update_time_copy;
3111 p_last_update_time_copy = prev->load_last_update_time_copy;
3112 n_last_update_time_copy = next->load_last_update_time_copy;
3116 p_last_update_time = prev->avg.last_update_time;
3117 n_last_update_time = next->avg.last_update_time;
3119 } while (p_last_update_time != p_last_update_time_copy ||
3120 n_last_update_time != n_last_update_time_copy);
3123 p_last_update_time = prev->avg.last_update_time;
3124 n_last_update_time = next->avg.last_update_time;
3126 __update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se);
3127 se->avg.last_update_time = n_last_update_time;
3132 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3133 * propagate its contribution. The key to this propagation is the invariant
3134 * that for each group:
3136 * ge->avg == grq->avg (1)
3138 * _IFF_ we look at the pure running and runnable sums. Because they
3139 * represent the very same entity, just at different points in the hierarchy.
3141 * Per the above update_tg_cfs_util() is trivial and simply copies the running
3142 * sum over (but still wrong, because the group entity and group rq do not have
3143 * their PELT windows aligned).
3145 * However, update_tg_cfs_runnable() is more complex. So we have:
3147 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3149 * And since, like util, the runnable part should be directly transferable,
3150 * the following would _appear_ to be the straight forward approach:
3152 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3154 * And per (1) we have:
3156 * ge->avg.runnable_avg == grq->avg.runnable_avg
3160 * ge->load.weight * grq->avg.load_avg
3161 * ge->avg.load_avg = ----------------------------------- (4)
3164 * Except that is wrong!
3166 * Because while for entities historical weight is not important and we
3167 * really only care about our future and therefore can consider a pure
3168 * runnable sum, runqueues can NOT do this.
3170 * We specifically want runqueues to have a load_avg that includes
3171 * historical weights. Those represent the blocked load, the load we expect
3172 * to (shortly) return to us. This only works by keeping the weights as
3173 * integral part of the sum. We therefore cannot decompose as per (3).
3175 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3176 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3177 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3178 * runnable section of these tasks overlap (or not). If they were to perfectly
3179 * align the rq as a whole would be runnable 2/3 of the time. If however we
3180 * always have at least 1 runnable task, the rq as a whole is always runnable.
3182 * So we'll have to approximate.. :/
3184 * Given the constraint:
3186 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3188 * We can construct a rule that adds runnable to a rq by assuming minimal
3191 * On removal, we'll assume each task is equally runnable; which yields:
3193 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3195 * XXX: only do this for the part of runnable > running ?
3200 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3202 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3204 /* Nothing to update */
3209 * The relation between sum and avg is:
3211 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3213 * however, the PELT windows are not aligned between grq and gse.
3216 /* Set new sched_entity's utilization */
3217 se->avg.util_avg = gcfs_rq->avg.util_avg;
3218 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3220 /* Update parent cfs_rq utilization */
3221 add_positive(&cfs_rq->avg.util_avg, delta);
3222 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3226 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3228 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3229 unsigned long runnable_load_avg, load_avg;
3230 u64 runnable_load_sum, load_sum = 0;
3236 gcfs_rq->prop_runnable_sum = 0;
3238 if (runnable_sum >= 0) {
3240 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3241 * the CPU is saturated running == runnable.
3243 runnable_sum += se->avg.load_sum;
3244 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3247 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3248 * assuming all tasks are equally runnable.
3250 if (scale_load_down(gcfs_rq->load.weight)) {
3251 load_sum = div_s64(gcfs_rq->avg.load_sum,
3252 scale_load_down(gcfs_rq->load.weight));
3255 /* But make sure to not inflate se's runnable */
3256 runnable_sum = min(se->avg.load_sum, load_sum);
3260 * runnable_sum can't be lower than running_sum
3261 * As running sum is scale with CPU capacity wehreas the runnable sum
3262 * is not we rescale running_sum 1st
3264 running_sum = se->avg.util_sum /
3265 arch_scale_cpu_capacity(NULL, cpu_of(rq_of(cfs_rq)));
3266 runnable_sum = max(runnable_sum, running_sum);
3268 load_sum = (s64)se_weight(se) * runnable_sum;
3269 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3271 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3272 delta_avg = load_avg - se->avg.load_avg;
3274 se->avg.load_sum = runnable_sum;
3275 se->avg.load_avg = load_avg;
3276 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3277 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3279 runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3280 runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3281 delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3282 delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3284 se->avg.runnable_load_sum = runnable_sum;
3285 se->avg.runnable_load_avg = runnable_load_avg;
3288 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3289 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3293 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3295 cfs_rq->propagate = 1;
3296 cfs_rq->prop_runnable_sum += runnable_sum;
3299 /* Update task and its cfs_rq load average */
3300 static inline int propagate_entity_load_avg(struct sched_entity *se)
3302 struct cfs_rq *cfs_rq, *gcfs_rq;
3304 if (entity_is_task(se))
3307 gcfs_rq = group_cfs_rq(se);
3308 if (!gcfs_rq->propagate)
3311 gcfs_rq->propagate = 0;
3313 cfs_rq = cfs_rq_of(se);
3315 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3317 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3318 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3324 * Check if we need to update the load and the utilization of a blocked
3327 static inline bool skip_blocked_update(struct sched_entity *se)
3329 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3332 * If sched_entity still have not zero load or utilization, we have to
3335 if (se->avg.load_avg || se->avg.util_avg)
3339 * If there is a pending propagation, we have to update the load and
3340 * the utilization of the sched_entity:
3342 if (gcfs_rq->propagate)
3346 * Otherwise, the load and the utilization of the sched_entity is
3347 * already zero and there is no pending propagation, so it will be a
3348 * waste of time to try to decay it:
3353 #else /* CONFIG_FAIR_GROUP_SCHED */
3355 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3357 static inline int propagate_entity_load_avg(struct sched_entity *se)
3362 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3364 #endif /* CONFIG_FAIR_GROUP_SCHED */
3367 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3368 * @now: current time, as per cfs_rq_clock_task()
3369 * @cfs_rq: cfs_rq to update
3371 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3372 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3373 * post_init_entity_util_avg().
3375 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3377 * Returns true if the load decayed or we removed load.
3379 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3380 * call update_tg_load_avg() when this function returns true.
3383 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3385 unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3386 struct sched_avg *sa = &cfs_rq->avg;
3389 if (cfs_rq->removed.nr) {
3391 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3393 raw_spin_lock(&cfs_rq->removed.lock);
3394 swap(cfs_rq->removed.util_avg, removed_util);
3395 swap(cfs_rq->removed.load_avg, removed_load);
3396 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3397 cfs_rq->removed.nr = 0;
3398 raw_spin_unlock(&cfs_rq->removed.lock);
3401 sub_positive(&sa->load_avg, r);
3402 sub_positive(&sa->load_sum, r * divider);
3405 sub_positive(&sa->util_avg, r);
3406 sub_positive(&sa->util_sum, r * divider);
3408 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3413 decayed |= __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq);
3415 #ifndef CONFIG_64BIT
3417 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3421 cfs_rq_util_change(cfs_rq, 0);
3427 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3428 * @cfs_rq: cfs_rq to attach to
3429 * @se: sched_entity to attach
3430 * @flags: migration hints
3432 * Must call update_cfs_rq_load_avg() before this, since we rely on
3433 * cfs_rq->avg.last_update_time being current.
3435 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3437 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3440 * When we attach the @se to the @cfs_rq, we must align the decay
3441 * window because without that, really weird and wonderful things can
3446 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3447 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3450 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3451 * period_contrib. This isn't strictly correct, but since we're
3452 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3455 se->avg.util_sum = se->avg.util_avg * divider;
3457 se->avg.load_sum = divider;
3458 if (se_weight(se)) {
3460 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3463 se->avg.runnable_load_sum = se->avg.load_sum;
3465 enqueue_load_avg(cfs_rq, se);
3466 cfs_rq->avg.util_avg += se->avg.util_avg;
3467 cfs_rq->avg.util_sum += se->avg.util_sum;
3469 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3471 cfs_rq_util_change(cfs_rq, flags);
3475 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3476 * @cfs_rq: cfs_rq to detach from
3477 * @se: sched_entity to detach
3479 * Must call update_cfs_rq_load_avg() before this, since we rely on
3480 * cfs_rq->avg.last_update_time being current.
3482 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3484 dequeue_load_avg(cfs_rq, se);
3485 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3486 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3488 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3490 cfs_rq_util_change(cfs_rq, 0);
3494 * Optional action to be done while updating the load average
3496 #define UPDATE_TG 0x1
3497 #define SKIP_AGE_LOAD 0x2
3498 #define DO_ATTACH 0x4
3500 /* Update task and its cfs_rq load average */
3501 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3503 u64 now = cfs_rq_clock_task(cfs_rq);
3504 struct rq *rq = rq_of(cfs_rq);
3505 int cpu = cpu_of(rq);
3509 * Track task load average for carrying it to new CPU after migrated, and
3510 * track group sched_entity load average for task_h_load calc in migration
3512 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3513 __update_load_avg_se(now, cpu, cfs_rq, se);
3515 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3516 decayed |= propagate_entity_load_avg(se);
3518 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3521 * DO_ATTACH means we're here from enqueue_entity().
3522 * !last_update_time means we've passed through
3523 * migrate_task_rq_fair() indicating we migrated.
3525 * IOW we're enqueueing a task on a new CPU.
3527 attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
3528 update_tg_load_avg(cfs_rq, 0);
3530 } else if (decayed && (flags & UPDATE_TG))
3531 update_tg_load_avg(cfs_rq, 0);
3534 #ifndef CONFIG_64BIT
3535 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3537 u64 last_update_time_copy;
3538 u64 last_update_time;
3541 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3543 last_update_time = cfs_rq->avg.last_update_time;
3544 } while (last_update_time != last_update_time_copy);
3546 return last_update_time;
3549 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3551 return cfs_rq->avg.last_update_time;
3556 * Synchronize entity load avg of dequeued entity without locking
3559 void sync_entity_load_avg(struct sched_entity *se)
3561 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3562 u64 last_update_time;
3564 last_update_time = cfs_rq_last_update_time(cfs_rq);
3565 __update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se);
3569 * Task first catches up with cfs_rq, and then subtract
3570 * itself from the cfs_rq (task must be off the queue now).
3572 void remove_entity_load_avg(struct sched_entity *se)
3574 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3575 unsigned long flags;
3578 * tasks cannot exit without having gone through wake_up_new_task() ->
3579 * post_init_entity_util_avg() which will have added things to the
3580 * cfs_rq, so we can remove unconditionally.
3582 * Similarly for groups, they will have passed through
3583 * post_init_entity_util_avg() before unregister_sched_fair_group()
3587 sync_entity_load_avg(se);
3589 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3590 ++cfs_rq->removed.nr;
3591 cfs_rq->removed.util_avg += se->avg.util_avg;
3592 cfs_rq->removed.load_avg += se->avg.load_avg;
3593 cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */
3594 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3597 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3599 return cfs_rq->avg.runnable_load_avg;
3602 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3604 return cfs_rq->avg.load_avg;
3607 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3609 static inline unsigned long task_util(struct task_struct *p)
3611 return READ_ONCE(p->se.avg.util_avg);
3614 static inline unsigned long _task_util_est(struct task_struct *p)
3616 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3618 return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED);
3621 static inline unsigned long task_util_est(struct task_struct *p)
3623 return max(task_util(p), _task_util_est(p));
3626 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3627 struct task_struct *p)
3629 unsigned int enqueued;
3631 if (!sched_feat(UTIL_EST))
3634 /* Update root cfs_rq's estimated utilization */
3635 enqueued = cfs_rq->avg.util_est.enqueued;
3636 enqueued += _task_util_est(p);
3637 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3641 * Check if a (signed) value is within a specified (unsigned) margin,
3642 * based on the observation that:
3644 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3646 * NOTE: this only works when value + maring < INT_MAX.
3648 static inline bool within_margin(int value, int margin)
3650 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3654 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3656 long last_ewma_diff;
3659 if (!sched_feat(UTIL_EST))
3662 /* Update root cfs_rq's estimated utilization */
3663 ue.enqueued = cfs_rq->avg.util_est.enqueued;
3664 ue.enqueued -= min_t(unsigned int, ue.enqueued, _task_util_est(p));
3665 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3668 * Skip update of task's estimated utilization when the task has not
3669 * yet completed an activation, e.g. being migrated.
3675 * If the PELT values haven't changed since enqueue time,
3676 * skip the util_est update.
3678 ue = p->se.avg.util_est;
3679 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3683 * Skip update of task's estimated utilization when its EWMA is
3684 * already ~1% close to its last activation value.
3686 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3687 last_ewma_diff = ue.enqueued - ue.ewma;
3688 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3692 * Update Task's estimated utilization
3694 * When *p completes an activation we can consolidate another sample
3695 * of the task size. This is done by storing the current PELT value
3696 * as ue.enqueued and by using this value to update the Exponential
3697 * Weighted Moving Average (EWMA):
3699 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
3700 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
3701 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
3702 * = w * ( last_ewma_diff ) + ewma(t-1)
3703 * = w * (last_ewma_diff + ewma(t-1) / w)
3705 * Where 'w' is the weight of new samples, which is configured to be
3706 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3708 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
3709 ue.ewma += last_ewma_diff;
3710 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
3711 WRITE_ONCE(p->se.avg.util_est, ue);
3714 static inline int task_fits_capacity(struct task_struct *p, long capacity)
3716 return capacity * 1024 > task_util_est(p) * capacity_margin;
3719 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
3721 if (!static_branch_unlikely(&sched_asym_cpucapacity))
3725 rq->misfit_task_load = 0;
3729 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
3730 rq->misfit_task_load = 0;
3734 rq->misfit_task_load = task_h_load(p);
3737 #else /* CONFIG_SMP */
3739 #define UPDATE_TG 0x0
3740 #define SKIP_AGE_LOAD 0x0
3741 #define DO_ATTACH 0x0
3743 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3745 cfs_rq_util_change(cfs_rq, 0);
3748 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3751 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
3753 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3755 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3761 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
3764 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
3766 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
3768 #endif /* CONFIG_SMP */
3770 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3772 #ifdef CONFIG_SCHED_DEBUG
3773 s64 d = se->vruntime - cfs_rq->min_vruntime;
3778 if (d > 3*sysctl_sched_latency)
3779 schedstat_inc(cfs_rq->nr_spread_over);
3784 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3786 u64 vruntime = cfs_rq->min_vruntime;
3789 * The 'current' period is already promised to the current tasks,
3790 * however the extra weight of the new task will slow them down a
3791 * little, place the new task so that it fits in the slot that
3792 * stays open at the end.
3794 if (initial && sched_feat(START_DEBIT))
3795 vruntime += sched_vslice(cfs_rq, se);
3797 /* sleeps up to a single latency don't count. */
3799 unsigned long thresh = sysctl_sched_latency;
3802 * Halve their sleep time's effect, to allow
3803 * for a gentler effect of sleepers:
3805 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3811 /* ensure we never gain time by being placed backwards. */
3812 se->vruntime = max_vruntime(se->vruntime, vruntime);
3815 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3817 static inline void check_schedstat_required(void)
3819 #ifdef CONFIG_SCHEDSTATS
3820 if (schedstat_enabled())
3823 /* Force schedstat enabled if a dependent tracepoint is active */
3824 if (trace_sched_stat_wait_enabled() ||
3825 trace_sched_stat_sleep_enabled() ||
3826 trace_sched_stat_iowait_enabled() ||
3827 trace_sched_stat_blocked_enabled() ||
3828 trace_sched_stat_runtime_enabled()) {
3829 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3830 "stat_blocked and stat_runtime require the "
3831 "kernel parameter schedstats=enable or "
3832 "kernel.sched_schedstats=1\n");
3843 * update_min_vruntime()
3844 * vruntime -= min_vruntime
3848 * update_min_vruntime()
3849 * vruntime += min_vruntime
3851 * this way the vruntime transition between RQs is done when both
3852 * min_vruntime are up-to-date.
3856 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3857 * vruntime -= min_vruntime
3861 * update_min_vruntime()
3862 * vruntime += min_vruntime
3864 * this way we don't have the most up-to-date min_vruntime on the originating
3865 * CPU and an up-to-date min_vruntime on the destination CPU.
3869 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3871 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3872 bool curr = cfs_rq->curr == se;
3875 * If we're the current task, we must renormalise before calling
3879 se->vruntime += cfs_rq->min_vruntime;
3881 update_curr(cfs_rq);
3884 * Otherwise, renormalise after, such that we're placed at the current
3885 * moment in time, instead of some random moment in the past. Being
3886 * placed in the past could significantly boost this task to the
3887 * fairness detriment of existing tasks.
3889 if (renorm && !curr)
3890 se->vruntime += cfs_rq->min_vruntime;
3893 * When enqueuing a sched_entity, we must:
3894 * - Update loads to have both entity and cfs_rq synced with now.
3895 * - Add its load to cfs_rq->runnable_avg
3896 * - For group_entity, update its weight to reflect the new share of
3898 * - Add its new weight to cfs_rq->load.weight
3900 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
3901 update_cfs_group(se);
3902 enqueue_runnable_load_avg(cfs_rq, se);
3903 account_entity_enqueue(cfs_rq, se);
3905 if (flags & ENQUEUE_WAKEUP)
3906 place_entity(cfs_rq, se, 0);
3908 check_schedstat_required();
3909 update_stats_enqueue(cfs_rq, se, flags);
3910 check_spread(cfs_rq, se);
3912 __enqueue_entity(cfs_rq, se);
3915 if (cfs_rq->nr_running == 1) {
3916 list_add_leaf_cfs_rq(cfs_rq);
3917 check_enqueue_throttle(cfs_rq);
3921 static void __clear_buddies_last(struct sched_entity *se)
3923 for_each_sched_entity(se) {
3924 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3925 if (cfs_rq->last != se)
3928 cfs_rq->last = NULL;
3932 static void __clear_buddies_next(struct sched_entity *se)
3934 for_each_sched_entity(se) {
3935 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3936 if (cfs_rq->next != se)
3939 cfs_rq->next = NULL;
3943 static void __clear_buddies_skip(struct sched_entity *se)
3945 for_each_sched_entity(se) {
3946 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3947 if (cfs_rq->skip != se)
3950 cfs_rq->skip = NULL;
3954 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3956 if (cfs_rq->last == se)
3957 __clear_buddies_last(se);
3959 if (cfs_rq->next == se)
3960 __clear_buddies_next(se);
3962 if (cfs_rq->skip == se)
3963 __clear_buddies_skip(se);
3966 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3969 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3972 * Update run-time statistics of the 'current'.
3974 update_curr(cfs_rq);
3977 * When dequeuing a sched_entity, we must:
3978 * - Update loads to have both entity and cfs_rq synced with now.
3979 * - Subtract its load from the cfs_rq->runnable_avg.
3980 * - Subtract its previous weight from cfs_rq->load.weight.
3981 * - For group entity, update its weight to reflect the new share
3982 * of its group cfs_rq.
3984 update_load_avg(cfs_rq, se, UPDATE_TG);
3985 dequeue_runnable_load_avg(cfs_rq, se);
3987 update_stats_dequeue(cfs_rq, se, flags);
3989 clear_buddies(cfs_rq, se);
3991 if (se != cfs_rq->curr)
3992 __dequeue_entity(cfs_rq, se);
3994 account_entity_dequeue(cfs_rq, se);
3997 * Normalize after update_curr(); which will also have moved
3998 * min_vruntime if @se is the one holding it back. But before doing
3999 * update_min_vruntime() again, which will discount @se's position and
4000 * can move min_vruntime forward still more.
4002 if (!(flags & DEQUEUE_SLEEP))
4003 se->vruntime -= cfs_rq->min_vruntime;
4005 /* return excess runtime on last dequeue */
4006 return_cfs_rq_runtime(cfs_rq);
4008 update_cfs_group(se);
4011 * Now advance min_vruntime if @se was the entity holding it back,
4012 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4013 * put back on, and if we advance min_vruntime, we'll be placed back
4014 * further than we started -- ie. we'll be penalized.
4016 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4017 update_min_vruntime(cfs_rq);
4021 * Preempt the current task with a newly woken task if needed:
4024 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4026 unsigned long ideal_runtime, delta_exec;
4027 struct sched_entity *se;
4030 ideal_runtime = sched_slice(cfs_rq, curr);
4031 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4032 if (delta_exec > ideal_runtime) {
4033 resched_curr(rq_of(cfs_rq));
4035 * The current task ran long enough, ensure it doesn't get
4036 * re-elected due to buddy favours.
4038 clear_buddies(cfs_rq, curr);
4043 * Ensure that a task that missed wakeup preemption by a
4044 * narrow margin doesn't have to wait for a full slice.
4045 * This also mitigates buddy induced latencies under load.
4047 if (delta_exec < sysctl_sched_min_granularity)
4050 se = __pick_first_entity(cfs_rq);
4051 delta = curr->vruntime - se->vruntime;
4056 if (delta > ideal_runtime)
4057 resched_curr(rq_of(cfs_rq));
4061 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4063 /* 'current' is not kept within the tree. */
4066 * Any task has to be enqueued before it get to execute on
4067 * a CPU. So account for the time it spent waiting on the
4070 update_stats_wait_end(cfs_rq, se);
4071 __dequeue_entity(cfs_rq, se);
4072 update_load_avg(cfs_rq, se, UPDATE_TG);
4075 update_stats_curr_start(cfs_rq, se);
4079 * Track our maximum slice length, if the CPU's load is at
4080 * least twice that of our own weight (i.e. dont track it
4081 * when there are only lesser-weight tasks around):
4083 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
4084 schedstat_set(se->statistics.slice_max,
4085 max((u64)schedstat_val(se->statistics.slice_max),
4086 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4089 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4093 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4096 * Pick the next process, keeping these things in mind, in this order:
4097 * 1) keep things fair between processes/task groups
4098 * 2) pick the "next" process, since someone really wants that to run
4099 * 3) pick the "last" process, for cache locality
4100 * 4) do not run the "skip" process, if something else is available
4102 static struct sched_entity *
4103 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4105 struct sched_entity *left = __pick_first_entity(cfs_rq);
4106 struct sched_entity *se;
4109 * If curr is set we have to see if its left of the leftmost entity
4110 * still in the tree, provided there was anything in the tree at all.
4112 if (!left || (curr && entity_before(curr, left)))
4115 se = left; /* ideally we run the leftmost entity */
4118 * Avoid running the skip buddy, if running something else can
4119 * be done without getting too unfair.
4121 if (cfs_rq->skip == se) {
4122 struct sched_entity *second;
4125 second = __pick_first_entity(cfs_rq);
4127 second = __pick_next_entity(se);
4128 if (!second || (curr && entity_before(curr, second)))
4132 if (second && wakeup_preempt_entity(second, left) < 1)
4137 * Prefer last buddy, try to return the CPU to a preempted task.
4139 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4143 * Someone really wants this to run. If it's not unfair, run it.
4145 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4148 clear_buddies(cfs_rq, se);
4153 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4155 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4158 * If still on the runqueue then deactivate_task()
4159 * was not called and update_curr() has to be done:
4162 update_curr(cfs_rq);
4164 /* throttle cfs_rqs exceeding runtime */
4165 check_cfs_rq_runtime(cfs_rq);
4167 check_spread(cfs_rq, prev);
4170 update_stats_wait_start(cfs_rq, prev);
4171 /* Put 'current' back into the tree. */
4172 __enqueue_entity(cfs_rq, prev);
4173 /* in !on_rq case, update occurred at dequeue */
4174 update_load_avg(cfs_rq, prev, 0);
4176 cfs_rq->curr = NULL;
4180 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4183 * Update run-time statistics of the 'current'.
4185 update_curr(cfs_rq);
4188 * Ensure that runnable average is periodically updated.
4190 update_load_avg(cfs_rq, curr, UPDATE_TG);
4191 update_cfs_group(curr);
4193 #ifdef CONFIG_SCHED_HRTICK
4195 * queued ticks are scheduled to match the slice, so don't bother
4196 * validating it and just reschedule.
4199 resched_curr(rq_of(cfs_rq));
4203 * don't let the period tick interfere with the hrtick preemption
4205 if (!sched_feat(DOUBLE_TICK) &&
4206 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4210 if (cfs_rq->nr_running > 1)
4211 check_preempt_tick(cfs_rq, curr);
4215 /**************************************************
4216 * CFS bandwidth control machinery
4219 #ifdef CONFIG_CFS_BANDWIDTH
4221 #ifdef HAVE_JUMP_LABEL
4222 static struct static_key __cfs_bandwidth_used;
4224 static inline bool cfs_bandwidth_used(void)
4226 return static_key_false(&__cfs_bandwidth_used);
4229 void cfs_bandwidth_usage_inc(void)
4231 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4234 void cfs_bandwidth_usage_dec(void)
4236 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4238 #else /* HAVE_JUMP_LABEL */
4239 static bool cfs_bandwidth_used(void)
4244 void cfs_bandwidth_usage_inc(void) {}
4245 void cfs_bandwidth_usage_dec(void) {}
4246 #endif /* HAVE_JUMP_LABEL */
4249 * default period for cfs group bandwidth.
4250 * default: 0.1s, units: nanoseconds
4252 static inline u64 default_cfs_period(void)
4254 return 100000000ULL;
4257 static inline u64 sched_cfs_bandwidth_slice(void)
4259 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4263 * Replenish runtime according to assigned quota and update expiration time.
4264 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4265 * additional synchronization around rq->lock.
4267 * requires cfs_b->lock
4269 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4273 if (cfs_b->quota == RUNTIME_INF)
4276 now = sched_clock_cpu(smp_processor_id());
4277 cfs_b->runtime = cfs_b->quota;
4278 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4279 cfs_b->expires_seq++;
4282 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4284 return &tg->cfs_bandwidth;
4287 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4288 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4290 if (unlikely(cfs_rq->throttle_count))
4291 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4293 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4296 /* returns 0 on failure to allocate runtime */
4297 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4299 struct task_group *tg = cfs_rq->tg;
4300 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4301 u64 amount = 0, min_amount, expires;
4304 /* note: this is a positive sum as runtime_remaining <= 0 */
4305 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4307 raw_spin_lock(&cfs_b->lock);
4308 if (cfs_b->quota == RUNTIME_INF)
4309 amount = min_amount;
4311 start_cfs_bandwidth(cfs_b);
4313 if (cfs_b->runtime > 0) {
4314 amount = min(cfs_b->runtime, min_amount);
4315 cfs_b->runtime -= amount;
4319 expires_seq = cfs_b->expires_seq;
4320 expires = cfs_b->runtime_expires;
4321 raw_spin_unlock(&cfs_b->lock);
4323 cfs_rq->runtime_remaining += amount;
4325 * we may have advanced our local expiration to account for allowed
4326 * spread between our sched_clock and the one on which runtime was
4329 if (cfs_rq->expires_seq != expires_seq) {
4330 cfs_rq->expires_seq = expires_seq;
4331 cfs_rq->runtime_expires = expires;
4334 return cfs_rq->runtime_remaining > 0;
4338 * Note: This depends on the synchronization provided by sched_clock and the
4339 * fact that rq->clock snapshots this value.
4341 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4343 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4345 /* if the deadline is ahead of our clock, nothing to do */
4346 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4349 if (cfs_rq->runtime_remaining < 0)
4353 * If the local deadline has passed we have to consider the
4354 * possibility that our sched_clock is 'fast' and the global deadline
4355 * has not truly expired.
4357 * Fortunately we can check determine whether this the case by checking
4358 * whether the global deadline(cfs_b->expires_seq) has advanced.
4360 if (cfs_rq->expires_seq == cfs_b->expires_seq) {
4361 /* extend local deadline, drift is bounded above by 2 ticks */
4362 cfs_rq->runtime_expires += TICK_NSEC;
4364 /* global deadline is ahead, expiration has passed */
4365 cfs_rq->runtime_remaining = 0;
4369 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4371 /* dock delta_exec before expiring quota (as it could span periods) */
4372 cfs_rq->runtime_remaining -= delta_exec;
4373 expire_cfs_rq_runtime(cfs_rq);
4375 if (likely(cfs_rq->runtime_remaining > 0))
4379 * if we're unable to extend our runtime we resched so that the active
4380 * hierarchy can be throttled
4382 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4383 resched_curr(rq_of(cfs_rq));
4386 static __always_inline
4387 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4389 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4392 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4395 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4397 return cfs_bandwidth_used() && cfs_rq->throttled;
4400 /* check whether cfs_rq, or any parent, is throttled */
4401 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4403 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4407 * Ensure that neither of the group entities corresponding to src_cpu or
4408 * dest_cpu are members of a throttled hierarchy when performing group
4409 * load-balance operations.
4411 static inline int throttled_lb_pair(struct task_group *tg,
4412 int src_cpu, int dest_cpu)
4414 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4416 src_cfs_rq = tg->cfs_rq[src_cpu];
4417 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4419 return throttled_hierarchy(src_cfs_rq) ||
4420 throttled_hierarchy(dest_cfs_rq);
4423 static int tg_unthrottle_up(struct task_group *tg, void *data)
4425 struct rq *rq = data;
4426 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4428 cfs_rq->throttle_count--;
4429 if (!cfs_rq->throttle_count) {
4430 /* adjust cfs_rq_clock_task() */
4431 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4432 cfs_rq->throttled_clock_task;
4438 static int tg_throttle_down(struct task_group *tg, void *data)
4440 struct rq *rq = data;
4441 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4443 /* group is entering throttled state, stop time */
4444 if (!cfs_rq->throttle_count)
4445 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4446 cfs_rq->throttle_count++;
4451 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4453 struct rq *rq = rq_of(cfs_rq);
4454 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4455 struct sched_entity *se;
4456 long task_delta, dequeue = 1;
4459 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4461 /* freeze hierarchy runnable averages while throttled */
4463 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4466 task_delta = cfs_rq->h_nr_running;
4467 for_each_sched_entity(se) {
4468 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4469 /* throttled entity or throttle-on-deactivate */
4474 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4475 qcfs_rq->h_nr_running -= task_delta;
4477 if (qcfs_rq->load.weight)
4482 sub_nr_running(rq, task_delta);
4484 cfs_rq->throttled = 1;
4485 cfs_rq->throttled_clock = rq_clock(rq);
4486 raw_spin_lock(&cfs_b->lock);
4487 empty = list_empty(&cfs_b->throttled_cfs_rq);
4490 * Add to the _head_ of the list, so that an already-started
4491 * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
4492 * not running add to the tail so that later runqueues don't get starved.
4494 if (cfs_b->distribute_running)
4495 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4497 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4500 * If we're the first throttled task, make sure the bandwidth
4504 start_cfs_bandwidth(cfs_b);
4506 raw_spin_unlock(&cfs_b->lock);
4509 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4511 struct rq *rq = rq_of(cfs_rq);
4512 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4513 struct sched_entity *se;
4517 se = cfs_rq->tg->se[cpu_of(rq)];
4519 cfs_rq->throttled = 0;
4521 update_rq_clock(rq);
4523 raw_spin_lock(&cfs_b->lock);
4524 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4525 list_del_rcu(&cfs_rq->throttled_list);
4526 raw_spin_unlock(&cfs_b->lock);
4528 /* update hierarchical throttle state */
4529 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4531 if (!cfs_rq->load.weight)
4534 task_delta = cfs_rq->h_nr_running;
4535 for_each_sched_entity(se) {
4539 cfs_rq = cfs_rq_of(se);
4541 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4542 cfs_rq->h_nr_running += task_delta;
4544 if (cfs_rq_throttled(cfs_rq))
4549 add_nr_running(rq, task_delta);
4551 /* Determine whether we need to wake up potentially idle CPU: */
4552 if (rq->curr == rq->idle && rq->cfs.nr_running)
4556 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4557 u64 remaining, u64 expires)
4559 struct cfs_rq *cfs_rq;
4561 u64 starting_runtime = remaining;
4564 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4566 struct rq *rq = rq_of(cfs_rq);
4570 if (!cfs_rq_throttled(cfs_rq))
4573 runtime = -cfs_rq->runtime_remaining + 1;
4574 if (runtime > remaining)
4575 runtime = remaining;
4576 remaining -= runtime;
4578 cfs_rq->runtime_remaining += runtime;
4579 cfs_rq->runtime_expires = expires;
4581 /* we check whether we're throttled above */
4582 if (cfs_rq->runtime_remaining > 0)
4583 unthrottle_cfs_rq(cfs_rq);
4593 return starting_runtime - remaining;
4597 * Responsible for refilling a task_group's bandwidth and unthrottling its
4598 * cfs_rqs as appropriate. If there has been no activity within the last
4599 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4600 * used to track this state.
4602 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4604 u64 runtime, runtime_expires;
4607 /* no need to continue the timer with no bandwidth constraint */
4608 if (cfs_b->quota == RUNTIME_INF)
4609 goto out_deactivate;
4611 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4612 cfs_b->nr_periods += overrun;
4615 * idle depends on !throttled (for the case of a large deficit), and if
4616 * we're going inactive then everything else can be deferred
4618 if (cfs_b->idle && !throttled)
4619 goto out_deactivate;
4621 __refill_cfs_bandwidth_runtime(cfs_b);
4624 /* mark as potentially idle for the upcoming period */
4629 /* account preceding periods in which throttling occurred */
4630 cfs_b->nr_throttled += overrun;
4632 runtime_expires = cfs_b->runtime_expires;
4635 * This check is repeated as we are holding onto the new bandwidth while
4636 * we unthrottle. This can potentially race with an unthrottled group
4637 * trying to acquire new bandwidth from the global pool. This can result
4638 * in us over-using our runtime if it is all used during this loop, but
4639 * only by limited amounts in that extreme case.
4641 while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
4642 runtime = cfs_b->runtime;
4643 cfs_b->distribute_running = 1;
4644 raw_spin_unlock(&cfs_b->lock);
4645 /* we can't nest cfs_b->lock while distributing bandwidth */
4646 runtime = distribute_cfs_runtime(cfs_b, runtime,
4648 raw_spin_lock(&cfs_b->lock);
4650 cfs_b->distribute_running = 0;
4651 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4653 lsub_positive(&cfs_b->runtime, runtime);
4657 * While we are ensured activity in the period following an
4658 * unthrottle, this also covers the case in which the new bandwidth is
4659 * insufficient to cover the existing bandwidth deficit. (Forcing the
4660 * timer to remain active while there are any throttled entities.)
4670 /* a cfs_rq won't donate quota below this amount */
4671 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4672 /* minimum remaining period time to redistribute slack quota */
4673 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4674 /* how long we wait to gather additional slack before distributing */
4675 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4678 * Are we near the end of the current quota period?
4680 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4681 * hrtimer base being cleared by hrtimer_start. In the case of
4682 * migrate_hrtimers, base is never cleared, so we are fine.
4684 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4686 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4689 /* if the call-back is running a quota refresh is already occurring */
4690 if (hrtimer_callback_running(refresh_timer))
4693 /* is a quota refresh about to occur? */
4694 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4695 if (remaining < min_expire)
4701 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4703 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4705 /* if there's a quota refresh soon don't bother with slack */
4706 if (runtime_refresh_within(cfs_b, min_left))
4709 hrtimer_start(&cfs_b->slack_timer,
4710 ns_to_ktime(cfs_bandwidth_slack_period),
4714 /* we know any runtime found here is valid as update_curr() precedes return */
4715 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4717 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4718 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4720 if (slack_runtime <= 0)
4723 raw_spin_lock(&cfs_b->lock);
4724 if (cfs_b->quota != RUNTIME_INF &&
4725 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4726 cfs_b->runtime += slack_runtime;
4728 /* we are under rq->lock, defer unthrottling using a timer */
4729 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4730 !list_empty(&cfs_b->throttled_cfs_rq))
4731 start_cfs_slack_bandwidth(cfs_b);
4733 raw_spin_unlock(&cfs_b->lock);
4735 /* even if it's not valid for return we don't want to try again */
4736 cfs_rq->runtime_remaining -= slack_runtime;
4739 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4741 if (!cfs_bandwidth_used())
4744 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4747 __return_cfs_rq_runtime(cfs_rq);
4751 * This is done with a timer (instead of inline with bandwidth return) since
4752 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4754 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4756 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4759 /* confirm we're still not at a refresh boundary */
4760 raw_spin_lock(&cfs_b->lock);
4761 if (cfs_b->distribute_running) {
4762 raw_spin_unlock(&cfs_b->lock);
4766 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4767 raw_spin_unlock(&cfs_b->lock);
4771 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4772 runtime = cfs_b->runtime;
4774 expires = cfs_b->runtime_expires;
4776 cfs_b->distribute_running = 1;
4778 raw_spin_unlock(&cfs_b->lock);
4783 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4785 raw_spin_lock(&cfs_b->lock);
4786 if (expires == cfs_b->runtime_expires)
4787 lsub_positive(&cfs_b->runtime, runtime);
4788 cfs_b->distribute_running = 0;
4789 raw_spin_unlock(&cfs_b->lock);
4793 * When a group wakes up we want to make sure that its quota is not already
4794 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4795 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4797 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4799 if (!cfs_bandwidth_used())
4802 /* an active group must be handled by the update_curr()->put() path */
4803 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4806 /* ensure the group is not already throttled */
4807 if (cfs_rq_throttled(cfs_rq))
4810 /* update runtime allocation */
4811 account_cfs_rq_runtime(cfs_rq, 0);
4812 if (cfs_rq->runtime_remaining <= 0)
4813 throttle_cfs_rq(cfs_rq);
4816 static void sync_throttle(struct task_group *tg, int cpu)
4818 struct cfs_rq *pcfs_rq, *cfs_rq;
4820 if (!cfs_bandwidth_used())
4826 cfs_rq = tg->cfs_rq[cpu];
4827 pcfs_rq = tg->parent->cfs_rq[cpu];
4829 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4830 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4833 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4834 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4836 if (!cfs_bandwidth_used())
4839 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4843 * it's possible for a throttled entity to be forced into a running
4844 * state (e.g. set_curr_task), in this case we're finished.
4846 if (cfs_rq_throttled(cfs_rq))
4849 throttle_cfs_rq(cfs_rq);
4853 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4855 struct cfs_bandwidth *cfs_b =
4856 container_of(timer, struct cfs_bandwidth, slack_timer);
4858 do_sched_cfs_slack_timer(cfs_b);
4860 return HRTIMER_NORESTART;
4863 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4865 struct cfs_bandwidth *cfs_b =
4866 container_of(timer, struct cfs_bandwidth, period_timer);
4870 raw_spin_lock(&cfs_b->lock);
4872 overrun = hrtimer_forward_now(timer, cfs_b->period);
4876 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4879 cfs_b->period_active = 0;
4880 raw_spin_unlock(&cfs_b->lock);
4882 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4885 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4887 raw_spin_lock_init(&cfs_b->lock);
4889 cfs_b->quota = RUNTIME_INF;
4890 cfs_b->period = ns_to_ktime(default_cfs_period());
4892 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4893 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4894 cfs_b->period_timer.function = sched_cfs_period_timer;
4895 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4896 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4897 cfs_b->distribute_running = 0;
4900 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4902 cfs_rq->runtime_enabled = 0;
4903 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4906 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4910 lockdep_assert_held(&cfs_b->lock);
4912 if (cfs_b->period_active)
4915 cfs_b->period_active = 1;
4916 overrun = hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4917 cfs_b->runtime_expires += (overrun + 1) * ktime_to_ns(cfs_b->period);
4918 cfs_b->expires_seq++;
4919 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4922 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4924 /* init_cfs_bandwidth() was not called */
4925 if (!cfs_b->throttled_cfs_rq.next)
4928 hrtimer_cancel(&cfs_b->period_timer);
4929 hrtimer_cancel(&cfs_b->slack_timer);
4933 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
4935 * The race is harmless, since modifying bandwidth settings of unhooked group
4936 * bits doesn't do much.
4939 /* cpu online calback */
4940 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4942 struct task_group *tg;
4944 lockdep_assert_held(&rq->lock);
4947 list_for_each_entry_rcu(tg, &task_groups, list) {
4948 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
4949 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4951 raw_spin_lock(&cfs_b->lock);
4952 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4953 raw_spin_unlock(&cfs_b->lock);
4958 /* cpu offline callback */
4959 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4961 struct task_group *tg;
4963 lockdep_assert_held(&rq->lock);
4966 list_for_each_entry_rcu(tg, &task_groups, list) {
4967 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4969 if (!cfs_rq->runtime_enabled)
4973 * clock_task is not advancing so we just need to make sure
4974 * there's some valid quota amount
4976 cfs_rq->runtime_remaining = 1;
4978 * Offline rq is schedulable till CPU is completely disabled
4979 * in take_cpu_down(), so we prevent new cfs throttling here.
4981 cfs_rq->runtime_enabled = 0;
4983 if (cfs_rq_throttled(cfs_rq))
4984 unthrottle_cfs_rq(cfs_rq);
4989 #else /* CONFIG_CFS_BANDWIDTH */
4990 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4992 return rq_clock_task(rq_of(cfs_rq));
4995 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4996 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4997 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4998 static inline void sync_throttle(struct task_group *tg, int cpu) {}
4999 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5001 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5006 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5011 static inline int throttled_lb_pair(struct task_group *tg,
5012 int src_cpu, int dest_cpu)
5017 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5019 #ifdef CONFIG_FAIR_GROUP_SCHED
5020 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5023 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5027 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5028 static inline void update_runtime_enabled(struct rq *rq) {}
5029 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5031 #endif /* CONFIG_CFS_BANDWIDTH */
5033 /**************************************************
5034 * CFS operations on tasks:
5037 #ifdef CONFIG_SCHED_HRTICK
5038 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5040 struct sched_entity *se = &p->se;
5041 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5043 SCHED_WARN_ON(task_rq(p) != rq);
5045 if (rq->cfs.h_nr_running > 1) {
5046 u64 slice = sched_slice(cfs_rq, se);
5047 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5048 s64 delta = slice - ran;
5055 hrtick_start(rq, delta);
5060 * called from enqueue/dequeue and updates the hrtick when the
5061 * current task is from our class and nr_running is low enough
5064 static void hrtick_update(struct rq *rq)
5066 struct task_struct *curr = rq->curr;
5068 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5071 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5072 hrtick_start_fair(rq, curr);
5074 #else /* !CONFIG_SCHED_HRTICK */
5076 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5080 static inline void hrtick_update(struct rq *rq)
5086 static inline unsigned long cpu_util(int cpu);
5087 static unsigned long capacity_of(int cpu);
5089 static inline bool cpu_overutilized(int cpu)
5091 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5094 static inline void update_overutilized_status(struct rq *rq)
5096 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu))
5097 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5100 static inline void update_overutilized_status(struct rq *rq) { }
5104 * The enqueue_task method is called before nr_running is
5105 * increased. Here we update the fair scheduling stats and
5106 * then put the task into the rbtree:
5109 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5111 struct cfs_rq *cfs_rq;
5112 struct sched_entity *se = &p->se;
5115 * The code below (indirectly) updates schedutil which looks at
5116 * the cfs_rq utilization to select a frequency.
5117 * Let's add the task's estimated utilization to the cfs_rq's
5118 * estimated utilization, before we update schedutil.
5120 util_est_enqueue(&rq->cfs, p);
5123 * If in_iowait is set, the code below may not trigger any cpufreq
5124 * utilization updates, so do it here explicitly with the IOWAIT flag
5128 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5130 for_each_sched_entity(se) {
5133 cfs_rq = cfs_rq_of(se);
5134 enqueue_entity(cfs_rq, se, flags);
5137 * end evaluation on encountering a throttled cfs_rq
5139 * note: in the case of encountering a throttled cfs_rq we will
5140 * post the final h_nr_running increment below.
5142 if (cfs_rq_throttled(cfs_rq))
5144 cfs_rq->h_nr_running++;
5146 flags = ENQUEUE_WAKEUP;
5149 for_each_sched_entity(se) {
5150 cfs_rq = cfs_rq_of(se);
5151 cfs_rq->h_nr_running++;
5153 if (cfs_rq_throttled(cfs_rq))
5156 update_load_avg(cfs_rq, se, UPDATE_TG);
5157 update_cfs_group(se);
5161 add_nr_running(rq, 1);
5163 * Since new tasks are assigned an initial util_avg equal to
5164 * half of the spare capacity of their CPU, tiny tasks have the
5165 * ability to cross the overutilized threshold, which will
5166 * result in the load balancer ruining all the task placement
5167 * done by EAS. As a way to mitigate that effect, do not account
5168 * for the first enqueue operation of new tasks during the
5169 * overutilized flag detection.
5171 * A better way of solving this problem would be to wait for
5172 * the PELT signals of tasks to converge before taking them
5173 * into account, but that is not straightforward to implement,
5174 * and the following generally works well enough in practice.
5176 if (flags & ENQUEUE_WAKEUP)
5177 update_overutilized_status(rq);
5184 static void set_next_buddy(struct sched_entity *se);
5187 * The dequeue_task method is called before nr_running is
5188 * decreased. We remove the task from the rbtree and
5189 * update the fair scheduling stats:
5191 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5193 struct cfs_rq *cfs_rq;
5194 struct sched_entity *se = &p->se;
5195 int task_sleep = flags & DEQUEUE_SLEEP;
5197 for_each_sched_entity(se) {
5198 cfs_rq = cfs_rq_of(se);
5199 dequeue_entity(cfs_rq, se, flags);
5202 * end evaluation on encountering a throttled cfs_rq
5204 * note: in the case of encountering a throttled cfs_rq we will
5205 * post the final h_nr_running decrement below.
5207 if (cfs_rq_throttled(cfs_rq))
5209 cfs_rq->h_nr_running--;
5211 /* Don't dequeue parent if it has other entities besides us */
5212 if (cfs_rq->load.weight) {
5213 /* Avoid re-evaluating load for this entity: */
5214 se = parent_entity(se);
5216 * Bias pick_next to pick a task from this cfs_rq, as
5217 * p is sleeping when it is within its sched_slice.
5219 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5223 flags |= DEQUEUE_SLEEP;
5226 for_each_sched_entity(se) {
5227 cfs_rq = cfs_rq_of(se);
5228 cfs_rq->h_nr_running--;
5230 if (cfs_rq_throttled(cfs_rq))
5233 update_load_avg(cfs_rq, se, UPDATE_TG);
5234 update_cfs_group(se);
5238 sub_nr_running(rq, 1);
5240 util_est_dequeue(&rq->cfs, p, task_sleep);
5246 /* Working cpumask for: load_balance, load_balance_newidle. */
5247 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5248 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5250 #ifdef CONFIG_NO_HZ_COMMON
5252 * per rq 'load' arrray crap; XXX kill this.
5256 * The exact cpuload calculated at every tick would be:
5258 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5260 * If a CPU misses updates for n ticks (as it was idle) and update gets
5261 * called on the n+1-th tick when CPU may be busy, then we have:
5263 * load_n = (1 - 1/2^i)^n * load_0
5264 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
5266 * decay_load_missed() below does efficient calculation of
5268 * load' = (1 - 1/2^i)^n * load
5270 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5271 * This allows us to precompute the above in said factors, thereby allowing the
5272 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5273 * fixed_power_int())
5275 * The calculation is approximated on a 128 point scale.
5277 #define DEGRADE_SHIFT 7
5279 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5280 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5281 { 0, 0, 0, 0, 0, 0, 0, 0 },
5282 { 64, 32, 8, 0, 0, 0, 0, 0 },
5283 { 96, 72, 40, 12, 1, 0, 0, 0 },
5284 { 112, 98, 75, 43, 15, 1, 0, 0 },
5285 { 120, 112, 98, 76, 45, 16, 2, 0 }
5289 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5290 * would be when CPU is idle and so we just decay the old load without
5291 * adding any new load.
5293 static unsigned long
5294 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5298 if (!missed_updates)
5301 if (missed_updates >= degrade_zero_ticks[idx])
5305 return load >> missed_updates;
5307 while (missed_updates) {
5308 if (missed_updates % 2)
5309 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5311 missed_updates >>= 1;
5318 cpumask_var_t idle_cpus_mask;
5320 int has_blocked; /* Idle CPUS has blocked load */
5321 unsigned long next_balance; /* in jiffy units */
5322 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5323 } nohz ____cacheline_aligned;
5325 #endif /* CONFIG_NO_HZ_COMMON */
5328 * __cpu_load_update - update the rq->cpu_load[] statistics
5329 * @this_rq: The rq to update statistics for
5330 * @this_load: The current load
5331 * @pending_updates: The number of missed updates
5333 * Update rq->cpu_load[] statistics. This function is usually called every
5334 * scheduler tick (TICK_NSEC).
5336 * This function computes a decaying average:
5338 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5340 * Because of NOHZ it might not get called on every tick which gives need for
5341 * the @pending_updates argument.
5343 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5344 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5345 * = A * (A * load[i]_n-2 + B) + B
5346 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5347 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5348 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5349 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5350 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5352 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5353 * any change in load would have resulted in the tick being turned back on.
5355 * For regular NOHZ, this reduces to:
5357 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5359 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5362 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5363 unsigned long pending_updates)
5365 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5368 this_rq->nr_load_updates++;
5370 /* Update our load: */
5371 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5372 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5373 unsigned long old_load, new_load;
5375 /* scale is effectively 1 << i now, and >> i divides by scale */
5377 old_load = this_rq->cpu_load[i];
5378 #ifdef CONFIG_NO_HZ_COMMON
5379 old_load = decay_load_missed(old_load, pending_updates - 1, i);
5380 if (tickless_load) {
5381 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5383 * old_load can never be a negative value because a
5384 * decayed tickless_load cannot be greater than the
5385 * original tickless_load.
5387 old_load += tickless_load;
5390 new_load = this_load;
5392 * Round up the averaging division if load is increasing. This
5393 * prevents us from getting stuck on 9 if the load is 10, for
5396 if (new_load > old_load)
5397 new_load += scale - 1;
5399 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5403 /* Used instead of source_load when we know the type == 0 */
5404 static unsigned long weighted_cpuload(struct rq *rq)
5406 return cfs_rq_runnable_load_avg(&rq->cfs);
5409 #ifdef CONFIG_NO_HZ_COMMON
5411 * There is no sane way to deal with nohz on smp when using jiffies because the
5412 * CPU doing the jiffies update might drift wrt the CPU doing the jiffy reading
5413 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5415 * Therefore we need to avoid the delta approach from the regular tick when
5416 * possible since that would seriously skew the load calculation. This is why we
5417 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5418 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5419 * loop exit, nohz_idle_balance, nohz full exit...)
5421 * This means we might still be one tick off for nohz periods.
5424 static void cpu_load_update_nohz(struct rq *this_rq,
5425 unsigned long curr_jiffies,
5428 unsigned long pending_updates;
5430 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5431 if (pending_updates) {
5432 this_rq->last_load_update_tick = curr_jiffies;
5434 * In the regular NOHZ case, we were idle, this means load 0.
5435 * In the NOHZ_FULL case, we were non-idle, we should consider
5436 * its weighted load.
5438 cpu_load_update(this_rq, load, pending_updates);
5443 * Called from nohz_idle_balance() to update the load ratings before doing the
5446 static void cpu_load_update_idle(struct rq *this_rq)
5449 * bail if there's load or we're actually up-to-date.
5451 if (weighted_cpuload(this_rq))
5454 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5458 * Record CPU load on nohz entry so we know the tickless load to account
5459 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5460 * than other cpu_load[idx] but it should be fine as cpu_load readers
5461 * shouldn't rely into synchronized cpu_load[*] updates.
5463 void cpu_load_update_nohz_start(void)
5465 struct rq *this_rq = this_rq();
5468 * This is all lockless but should be fine. If weighted_cpuload changes
5469 * concurrently we'll exit nohz. And cpu_load write can race with
5470 * cpu_load_update_idle() but both updater would be writing the same.
5472 this_rq->cpu_load[0] = weighted_cpuload(this_rq);
5476 * Account the tickless load in the end of a nohz frame.
5478 void cpu_load_update_nohz_stop(void)
5480 unsigned long curr_jiffies = READ_ONCE(jiffies);
5481 struct rq *this_rq = this_rq();
5485 if (curr_jiffies == this_rq->last_load_update_tick)
5488 load = weighted_cpuload(this_rq);
5489 rq_lock(this_rq, &rf);
5490 update_rq_clock(this_rq);
5491 cpu_load_update_nohz(this_rq, curr_jiffies, load);
5492 rq_unlock(this_rq, &rf);
5494 #else /* !CONFIG_NO_HZ_COMMON */
5495 static inline void cpu_load_update_nohz(struct rq *this_rq,
5496 unsigned long curr_jiffies,
5497 unsigned long load) { }
5498 #endif /* CONFIG_NO_HZ_COMMON */
5500 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5502 #ifdef CONFIG_NO_HZ_COMMON
5503 /* See the mess around cpu_load_update_nohz(). */
5504 this_rq->last_load_update_tick = READ_ONCE(jiffies);
5506 cpu_load_update(this_rq, load, 1);
5510 * Called from scheduler_tick()
5512 void cpu_load_update_active(struct rq *this_rq)
5514 unsigned long load = weighted_cpuload(this_rq);
5516 if (tick_nohz_tick_stopped())
5517 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5519 cpu_load_update_periodic(this_rq, load);
5523 * Return a low guess at the load of a migration-source CPU weighted
5524 * according to the scheduling class and "nice" value.
5526 * We want to under-estimate the load of migration sources, to
5527 * balance conservatively.
5529 static unsigned long source_load(int cpu, int type)
5531 struct rq *rq = cpu_rq(cpu);
5532 unsigned long total = weighted_cpuload(rq);
5534 if (type == 0 || !sched_feat(LB_BIAS))
5537 return min(rq->cpu_load[type-1], total);
5541 * Return a high guess at the load of a migration-target CPU weighted
5542 * according to the scheduling class and "nice" value.
5544 static unsigned long target_load(int cpu, int type)
5546 struct rq *rq = cpu_rq(cpu);
5547 unsigned long total = weighted_cpuload(rq);
5549 if (type == 0 || !sched_feat(LB_BIAS))
5552 return max(rq->cpu_load[type-1], total);
5555 static unsigned long capacity_of(int cpu)
5557 return cpu_rq(cpu)->cpu_capacity;
5560 static unsigned long capacity_orig_of(int cpu)
5562 return cpu_rq(cpu)->cpu_capacity_orig;
5565 static unsigned long cpu_avg_load_per_task(int cpu)
5567 struct rq *rq = cpu_rq(cpu);
5568 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5569 unsigned long load_avg = weighted_cpuload(rq);
5572 return load_avg / nr_running;
5577 static void record_wakee(struct task_struct *p)
5580 * Only decay a single time; tasks that have less then 1 wakeup per
5581 * jiffy will not have built up many flips.
5583 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5584 current->wakee_flips >>= 1;
5585 current->wakee_flip_decay_ts = jiffies;
5588 if (current->last_wakee != p) {
5589 current->last_wakee = p;
5590 current->wakee_flips++;
5595 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5597 * A waker of many should wake a different task than the one last awakened
5598 * at a frequency roughly N times higher than one of its wakees.
5600 * In order to determine whether we should let the load spread vs consolidating
5601 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5602 * partner, and a factor of lls_size higher frequency in the other.
5604 * With both conditions met, we can be relatively sure that the relationship is
5605 * non-monogamous, with partner count exceeding socket size.
5607 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5608 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5611 static int wake_wide(struct task_struct *p)
5613 unsigned int master = current->wakee_flips;
5614 unsigned int slave = p->wakee_flips;
5615 int factor = this_cpu_read(sd_llc_size);
5618 swap(master, slave);
5619 if (slave < factor || master < slave * factor)
5625 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5626 * soonest. For the purpose of speed we only consider the waking and previous
5629 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5630 * cache-affine and is (or will be) idle.
5632 * wake_affine_weight() - considers the weight to reflect the average
5633 * scheduling latency of the CPUs. This seems to work
5634 * for the overloaded case.
5637 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5640 * If this_cpu is idle, it implies the wakeup is from interrupt
5641 * context. Only allow the move if cache is shared. Otherwise an
5642 * interrupt intensive workload could force all tasks onto one
5643 * node depending on the IO topology or IRQ affinity settings.
5645 * If the prev_cpu is idle and cache affine then avoid a migration.
5646 * There is no guarantee that the cache hot data from an interrupt
5647 * is more important than cache hot data on the prev_cpu and from
5648 * a cpufreq perspective, it's better to have higher utilisation
5651 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5652 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5654 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5657 return nr_cpumask_bits;
5661 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5662 int this_cpu, int prev_cpu, int sync)
5664 s64 this_eff_load, prev_eff_load;
5665 unsigned long task_load;
5667 this_eff_load = target_load(this_cpu, sd->wake_idx);
5670 unsigned long current_load = task_h_load(current);
5672 if (current_load > this_eff_load)
5675 this_eff_load -= current_load;
5678 task_load = task_h_load(p);
5680 this_eff_load += task_load;
5681 if (sched_feat(WA_BIAS))
5682 this_eff_load *= 100;
5683 this_eff_load *= capacity_of(prev_cpu);
5685 prev_eff_load = source_load(prev_cpu, sd->wake_idx);
5686 prev_eff_load -= task_load;
5687 if (sched_feat(WA_BIAS))
5688 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5689 prev_eff_load *= capacity_of(this_cpu);
5692 * If sync, adjust the weight of prev_eff_load such that if
5693 * prev_eff == this_eff that select_idle_sibling() will consider
5694 * stacking the wakee on top of the waker if no other CPU is
5700 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5703 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5704 int this_cpu, int prev_cpu, int sync)
5706 int target = nr_cpumask_bits;
5708 if (sched_feat(WA_IDLE))
5709 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5711 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5712 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5714 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5715 if (target == nr_cpumask_bits)
5718 schedstat_inc(sd->ttwu_move_affine);
5719 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5723 static unsigned long cpu_util_without(int cpu, struct task_struct *p);
5725 static unsigned long capacity_spare_without(int cpu, struct task_struct *p)
5727 return max_t(long, capacity_of(cpu) - cpu_util_without(cpu, p), 0);
5731 * find_idlest_group finds and returns the least busy CPU group within the
5734 * Assumes p is allowed on at least one CPU in sd.
5736 static struct sched_group *
5737 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5738 int this_cpu, int sd_flag)
5740 struct sched_group *idlest = NULL, *group = sd->groups;
5741 struct sched_group *most_spare_sg = NULL;
5742 unsigned long min_runnable_load = ULONG_MAX;
5743 unsigned long this_runnable_load = ULONG_MAX;
5744 unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5745 unsigned long most_spare = 0, this_spare = 0;
5746 int load_idx = sd->forkexec_idx;
5747 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5748 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5749 (sd->imbalance_pct-100) / 100;
5751 if (sd_flag & SD_BALANCE_WAKE)
5752 load_idx = sd->wake_idx;
5755 unsigned long load, avg_load, runnable_load;
5756 unsigned long spare_cap, max_spare_cap;
5760 /* Skip over this group if it has no CPUs allowed */
5761 if (!cpumask_intersects(sched_group_span(group),
5765 local_group = cpumask_test_cpu(this_cpu,
5766 sched_group_span(group));
5769 * Tally up the load of all CPUs in the group and find
5770 * the group containing the CPU with most spare capacity.
5776 for_each_cpu(i, sched_group_span(group)) {
5777 /* Bias balancing toward CPUs of our domain */
5779 load = source_load(i, load_idx);
5781 load = target_load(i, load_idx);
5783 runnable_load += load;
5785 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5787 spare_cap = capacity_spare_without(i, p);
5789 if (spare_cap > max_spare_cap)
5790 max_spare_cap = spare_cap;
5793 /* Adjust by relative CPU capacity of the group */
5794 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5795 group->sgc->capacity;
5796 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5797 group->sgc->capacity;
5800 this_runnable_load = runnable_load;
5801 this_avg_load = avg_load;
5802 this_spare = max_spare_cap;
5804 if (min_runnable_load > (runnable_load + imbalance)) {
5806 * The runnable load is significantly smaller
5807 * so we can pick this new CPU:
5809 min_runnable_load = runnable_load;
5810 min_avg_load = avg_load;
5812 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5813 (100*min_avg_load > imbalance_scale*avg_load)) {
5815 * The runnable loads are close so take the
5816 * blocked load into account through avg_load:
5818 min_avg_load = avg_load;
5822 if (most_spare < max_spare_cap) {
5823 most_spare = max_spare_cap;
5824 most_spare_sg = group;
5827 } while (group = group->next, group != sd->groups);
5830 * The cross-over point between using spare capacity or least load
5831 * is too conservative for high utilization tasks on partially
5832 * utilized systems if we require spare_capacity > task_util(p),
5833 * so we allow for some task stuffing by using
5834 * spare_capacity > task_util(p)/2.
5836 * Spare capacity can't be used for fork because the utilization has
5837 * not been set yet, we must first select a rq to compute the initial
5840 if (sd_flag & SD_BALANCE_FORK)
5843 if (this_spare > task_util(p) / 2 &&
5844 imbalance_scale*this_spare > 100*most_spare)
5847 if (most_spare > task_util(p) / 2)
5848 return most_spare_sg;
5855 * When comparing groups across NUMA domains, it's possible for the
5856 * local domain to be very lightly loaded relative to the remote
5857 * domains but "imbalance" skews the comparison making remote CPUs
5858 * look much more favourable. When considering cross-domain, add
5859 * imbalance to the runnable load on the remote node and consider
5862 if ((sd->flags & SD_NUMA) &&
5863 min_runnable_load + imbalance >= this_runnable_load)
5866 if (min_runnable_load > (this_runnable_load + imbalance))
5869 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5870 (100*this_avg_load < imbalance_scale*min_avg_load))
5877 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5880 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5882 unsigned long load, min_load = ULONG_MAX;
5883 unsigned int min_exit_latency = UINT_MAX;
5884 u64 latest_idle_timestamp = 0;
5885 int least_loaded_cpu = this_cpu;
5886 int shallowest_idle_cpu = -1;
5889 /* Check if we have any choice: */
5890 if (group->group_weight == 1)
5891 return cpumask_first(sched_group_span(group));
5893 /* Traverse only the allowed CPUs */
5894 for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
5895 if (available_idle_cpu(i)) {
5896 struct rq *rq = cpu_rq(i);
5897 struct cpuidle_state *idle = idle_get_state(rq);
5898 if (idle && idle->exit_latency < min_exit_latency) {
5900 * We give priority to a CPU whose idle state
5901 * has the smallest exit latency irrespective
5902 * of any idle timestamp.
5904 min_exit_latency = idle->exit_latency;
5905 latest_idle_timestamp = rq->idle_stamp;
5906 shallowest_idle_cpu = i;
5907 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5908 rq->idle_stamp > latest_idle_timestamp) {
5910 * If equal or no active idle state, then
5911 * the most recently idled CPU might have
5914 latest_idle_timestamp = rq->idle_stamp;
5915 shallowest_idle_cpu = i;
5917 } else if (shallowest_idle_cpu == -1) {
5918 load = weighted_cpuload(cpu_rq(i));
5919 if (load < min_load) {
5921 least_loaded_cpu = i;
5926 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5929 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5930 int cpu, int prev_cpu, int sd_flag)
5934 if (!cpumask_intersects(sched_domain_span(sd), &p->cpus_allowed))
5938 * We need task's util for capacity_spare_without, sync it up to
5939 * prev_cpu's last_update_time.
5941 if (!(sd_flag & SD_BALANCE_FORK))
5942 sync_entity_load_avg(&p->se);
5945 struct sched_group *group;
5946 struct sched_domain *tmp;
5949 if (!(sd->flags & sd_flag)) {
5954 group = find_idlest_group(sd, p, cpu, sd_flag);
5960 new_cpu = find_idlest_group_cpu(group, p, cpu);
5961 if (new_cpu == cpu) {
5962 /* Now try balancing at a lower domain level of 'cpu': */
5967 /* Now try balancing at a lower domain level of 'new_cpu': */
5969 weight = sd->span_weight;
5971 for_each_domain(cpu, tmp) {
5972 if (weight <= tmp->span_weight)
5974 if (tmp->flags & sd_flag)
5982 #ifdef CONFIG_SCHED_SMT
5983 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5985 static inline void set_idle_cores(int cpu, int val)
5987 struct sched_domain_shared *sds;
5989 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5991 WRITE_ONCE(sds->has_idle_cores, val);
5994 static inline bool test_idle_cores(int cpu, bool def)
5996 struct sched_domain_shared *sds;
5998 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6000 return READ_ONCE(sds->has_idle_cores);
6006 * Scans the local SMT mask to see if the entire core is idle, and records this
6007 * information in sd_llc_shared->has_idle_cores.
6009 * Since SMT siblings share all cache levels, inspecting this limited remote
6010 * state should be fairly cheap.
6012 void __update_idle_core(struct rq *rq)
6014 int core = cpu_of(rq);
6018 if (test_idle_cores(core, true))
6021 for_each_cpu(cpu, cpu_smt_mask(core)) {
6025 if (!available_idle_cpu(cpu))
6029 set_idle_cores(core, 1);
6035 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6036 * there are no idle cores left in the system; tracked through
6037 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6039 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6041 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6044 if (!static_branch_likely(&sched_smt_present))
6047 if (!test_idle_cores(target, false))
6050 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
6052 for_each_cpu_wrap(core, cpus, target) {
6055 for_each_cpu(cpu, cpu_smt_mask(core)) {
6056 cpumask_clear_cpu(cpu, cpus);
6057 if (!available_idle_cpu(cpu))
6066 * Failed to find an idle core; stop looking for one.
6068 set_idle_cores(target, 0);
6074 * Scan the local SMT mask for idle CPUs.
6076 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6080 if (!static_branch_likely(&sched_smt_present))
6083 for_each_cpu(cpu, cpu_smt_mask(target)) {
6084 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6086 if (available_idle_cpu(cpu))
6093 #else /* CONFIG_SCHED_SMT */
6095 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6100 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6105 #endif /* CONFIG_SCHED_SMT */
6108 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6109 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6110 * average idle time for this rq (as found in rq->avg_idle).
6112 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
6114 struct sched_domain *this_sd;
6115 u64 avg_cost, avg_idle;
6118 int cpu, nr = INT_MAX;
6120 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6125 * Due to large variance we need a large fuzz factor; hackbench in
6126 * particularly is sensitive here.
6128 avg_idle = this_rq()->avg_idle / 512;
6129 avg_cost = this_sd->avg_scan_cost + 1;
6131 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
6134 if (sched_feat(SIS_PROP)) {
6135 u64 span_avg = sd->span_weight * avg_idle;
6136 if (span_avg > 4*avg_cost)
6137 nr = div_u64(span_avg, avg_cost);
6142 time = local_clock();
6144 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
6147 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6149 if (available_idle_cpu(cpu))
6153 time = local_clock() - time;
6154 cost = this_sd->avg_scan_cost;
6155 delta = (s64)(time - cost) / 8;
6156 this_sd->avg_scan_cost += delta;
6162 * Try and locate an idle core/thread in the LLC cache domain.
6164 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6166 struct sched_domain *sd;
6167 int i, recent_used_cpu;
6169 if (available_idle_cpu(target))
6173 * If the previous CPU is cache affine and idle, don't be stupid:
6175 if (prev != target && cpus_share_cache(prev, target) && available_idle_cpu(prev))
6178 /* Check a recently used CPU as a potential idle candidate: */
6179 recent_used_cpu = p->recent_used_cpu;
6180 if (recent_used_cpu != prev &&
6181 recent_used_cpu != target &&
6182 cpus_share_cache(recent_used_cpu, target) &&
6183 available_idle_cpu(recent_used_cpu) &&
6184 cpumask_test_cpu(p->recent_used_cpu, &p->cpus_allowed)) {
6186 * Replace recent_used_cpu with prev as it is a potential
6187 * candidate for the next wake:
6189 p->recent_used_cpu = prev;
6190 return recent_used_cpu;
6193 sd = rcu_dereference(per_cpu(sd_llc, target));
6197 i = select_idle_core(p, sd, target);
6198 if ((unsigned)i < nr_cpumask_bits)
6201 i = select_idle_cpu(p, sd, target);
6202 if ((unsigned)i < nr_cpumask_bits)
6205 i = select_idle_smt(p, sd, target);
6206 if ((unsigned)i < nr_cpumask_bits)
6213 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6214 * @cpu: the CPU to get the utilization of
6216 * The unit of the return value must be the one of capacity so we can compare
6217 * the utilization with the capacity of the CPU that is available for CFS task
6218 * (ie cpu_capacity).
6220 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6221 * recent utilization of currently non-runnable tasks on a CPU. It represents
6222 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6223 * capacity_orig is the cpu_capacity available at the highest frequency
6224 * (arch_scale_freq_capacity()).
6225 * The utilization of a CPU converges towards a sum equal to or less than the
6226 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6227 * the running time on this CPU scaled by capacity_curr.
6229 * The estimated utilization of a CPU is defined to be the maximum between its
6230 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6231 * currently RUNNABLE on that CPU.
6232 * This allows to properly represent the expected utilization of a CPU which
6233 * has just got a big task running since a long sleep period. At the same time
6234 * however it preserves the benefits of the "blocked utilization" in
6235 * describing the potential for other tasks waking up on the same CPU.
6237 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6238 * higher than capacity_orig because of unfortunate rounding in
6239 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6240 * the average stabilizes with the new running time. We need to check that the
6241 * utilization stays within the range of [0..capacity_orig] and cap it if
6242 * necessary. Without utilization capping, a group could be seen as overloaded
6243 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6244 * available capacity. We allow utilization to overshoot capacity_curr (but not
6245 * capacity_orig) as it useful for predicting the capacity required after task
6246 * migrations (scheduler-driven DVFS).
6248 * Return: the (estimated) utilization for the specified CPU
6250 static inline unsigned long cpu_util(int cpu)
6252 struct cfs_rq *cfs_rq;
6255 cfs_rq = &cpu_rq(cpu)->cfs;
6256 util = READ_ONCE(cfs_rq->avg.util_avg);
6258 if (sched_feat(UTIL_EST))
6259 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6261 return min_t(unsigned long, util, capacity_orig_of(cpu));
6265 * cpu_util_without: compute cpu utilization without any contributions from *p
6266 * @cpu: the CPU which utilization is requested
6267 * @p: the task which utilization should be discounted
6269 * The utilization of a CPU is defined by the utilization of tasks currently
6270 * enqueued on that CPU as well as tasks which are currently sleeping after an
6271 * execution on that CPU.
6273 * This method returns the utilization of the specified CPU by discounting the
6274 * utilization of the specified task, whenever the task is currently
6275 * contributing to the CPU utilization.
6277 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6279 struct cfs_rq *cfs_rq;
6282 /* Task has no contribution or is new */
6283 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6284 return cpu_util(cpu);
6286 cfs_rq = &cpu_rq(cpu)->cfs;
6287 util = READ_ONCE(cfs_rq->avg.util_avg);
6289 /* Discount task's util from CPU's util */
6290 lsub_positive(&util, task_util(p));
6295 * a) if *p is the only task sleeping on this CPU, then:
6296 * cpu_util (== task_util) > util_est (== 0)
6297 * and thus we return:
6298 * cpu_util_without = (cpu_util - task_util) = 0
6300 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6302 * cpu_util >= task_util
6303 * cpu_util > util_est (== 0)
6304 * and thus we discount *p's blocked utilization to return:
6305 * cpu_util_without = (cpu_util - task_util) >= 0
6307 * c) if other tasks are RUNNABLE on that CPU and
6308 * util_est > cpu_util
6309 * then we use util_est since it returns a more restrictive
6310 * estimation of the spare capacity on that CPU, by just
6311 * considering the expected utilization of tasks already
6312 * runnable on that CPU.
6314 * Cases a) and b) are covered by the above code, while case c) is
6315 * covered by the following code when estimated utilization is
6318 if (sched_feat(UTIL_EST)) {
6319 unsigned int estimated =
6320 READ_ONCE(cfs_rq->avg.util_est.enqueued);
6323 * Despite the following checks we still have a small window
6324 * for a possible race, when an execl's select_task_rq_fair()
6325 * races with LB's detach_task():
6328 * p->on_rq = TASK_ON_RQ_MIGRATING;
6329 * ---------------------------------- A
6330 * deactivate_task() \
6331 * dequeue_task() + RaceTime
6332 * util_est_dequeue() /
6333 * ---------------------------------- B
6335 * The additional check on "current == p" it's required to
6336 * properly fix the execl regression and it helps in further
6337 * reducing the chances for the above race.
6339 if (unlikely(task_on_rq_queued(p) || current == p))
6340 lsub_positive(&estimated, _task_util_est(p));
6342 util = max(util, estimated);
6346 * Utilization (estimated) can exceed the CPU capacity, thus let's
6347 * clamp to the maximum CPU capacity to ensure consistency with
6348 * the cpu_util call.
6350 return min_t(unsigned long, util, capacity_orig_of(cpu));
6354 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6355 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6357 * In that case WAKE_AFFINE doesn't make sense and we'll let
6358 * BALANCE_WAKE sort things out.
6360 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6362 long min_cap, max_cap;
6364 if (!static_branch_unlikely(&sched_asym_cpucapacity))
6367 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6368 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6370 /* Minimum capacity is close to max, no need to abort wake_affine */
6371 if (max_cap - min_cap < max_cap >> 3)
6374 /* Bring task utilization in sync with prev_cpu */
6375 sync_entity_load_avg(&p->se);
6377 return !task_fits_capacity(p, min_cap);
6381 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6384 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6386 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6387 unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
6390 * If @p migrates from @cpu to another, remove its contribution. Or,
6391 * if @p migrates from another CPU to @cpu, add its contribution. In
6392 * the other cases, @cpu is not impacted by the migration, so the
6393 * util_avg should already be correct.
6395 if (task_cpu(p) == cpu && dst_cpu != cpu)
6396 sub_positive(&util, task_util(p));
6397 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6398 util += task_util(p);
6400 if (sched_feat(UTIL_EST)) {
6401 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6404 * During wake-up, the task isn't enqueued yet and doesn't
6405 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6406 * so just add it (if needed) to "simulate" what will be
6407 * cpu_util() after the task has been enqueued.
6410 util_est += _task_util_est(p);
6412 util = max(util, util_est);
6415 return min(util, capacity_orig_of(cpu));
6419 * compute_energy(): Estimates the energy that would be consumed if @p was
6420 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6421 * landscape of the * CPUs after the task migration, and uses the Energy Model
6422 * to compute what would be the energy if we decided to actually migrate that
6426 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
6428 long util, max_util, sum_util, energy = 0;
6431 for (; pd; pd = pd->next) {
6432 max_util = sum_util = 0;
6434 * The capacity state of CPUs of the current rd can be driven by
6435 * CPUs of another rd if they belong to the same performance
6436 * domain. So, account for the utilization of these CPUs too
6437 * by masking pd with cpu_online_mask instead of the rd span.
6439 * If an entire performance domain is outside of the current rd,
6440 * it will not appear in its pd list and will not be accounted
6441 * by compute_energy().
6443 for_each_cpu_and(cpu, perf_domain_span(pd), cpu_online_mask) {
6444 util = cpu_util_next(cpu, p, dst_cpu);
6445 util = schedutil_energy_util(cpu, util);
6446 max_util = max(util, max_util);
6450 energy += em_pd_energy(pd->em_pd, max_util, sum_util);
6457 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6458 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6459 * spare capacity in each performance domain and uses it as a potential
6460 * candidate to execute the task. Then, it uses the Energy Model to figure
6461 * out which of the CPU candidates is the most energy-efficient.
6463 * The rationale for this heuristic is as follows. In a performance domain,
6464 * all the most energy efficient CPU candidates (according to the Energy
6465 * Model) are those for which we'll request a low frequency. When there are
6466 * several CPUs for which the frequency request will be the same, we don't
6467 * have enough data to break the tie between them, because the Energy Model
6468 * only includes active power costs. With this model, if we assume that
6469 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6470 * the maximum spare capacity in a performance domain is guaranteed to be among
6471 * the best candidates of the performance domain.
6473 * In practice, it could be preferable from an energy standpoint to pack
6474 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6475 * but that could also hurt our chances to go cluster idle, and we have no
6476 * ways to tell with the current Energy Model if this is actually a good
6477 * idea or not. So, find_energy_efficient_cpu() basically favors
6478 * cluster-packing, and spreading inside a cluster. That should at least be
6479 * a good thing for latency, and this is consistent with the idea that most
6480 * of the energy savings of EAS come from the asymmetry of the system, and
6481 * not so much from breaking the tie between identical CPUs. That's also the
6482 * reason why EAS is enabled in the topology code only for systems where
6483 * SD_ASYM_CPUCAPACITY is set.
6485 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6486 * they don't have any useful utilization data yet and it's not possible to
6487 * forecast their impact on energy consumption. Consequently, they will be
6488 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6489 * to be energy-inefficient in some use-cases. The alternative would be to
6490 * bias new tasks towards specific types of CPUs first, or to try to infer
6491 * their util_avg from the parent task, but those heuristics could hurt
6492 * other use-cases too. So, until someone finds a better way to solve this,
6493 * let's keep things simple by re-using the existing slow path.
6496 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6498 unsigned long prev_energy = ULONG_MAX, best_energy = ULONG_MAX;
6499 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6500 int cpu, best_energy_cpu = prev_cpu;
6501 struct perf_domain *head, *pd;
6502 unsigned long cpu_cap, util;
6503 struct sched_domain *sd;
6506 pd = rcu_dereference(rd->pd);
6507 if (!pd || READ_ONCE(rd->overutilized))
6512 * Energy-aware wake-up happens on the lowest sched_domain starting
6513 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6515 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6516 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6521 sync_entity_load_avg(&p->se);
6522 if (!task_util_est(p))
6525 for (; pd; pd = pd->next) {
6526 unsigned long cur_energy, spare_cap, max_spare_cap = 0;
6527 int max_spare_cap_cpu = -1;
6529 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
6530 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6533 /* Skip CPUs that will be overutilized. */
6534 util = cpu_util_next(cpu, p, cpu);
6535 cpu_cap = capacity_of(cpu);
6536 if (cpu_cap * 1024 < util * capacity_margin)
6539 /* Always use prev_cpu as a candidate. */
6540 if (cpu == prev_cpu) {
6541 prev_energy = compute_energy(p, prev_cpu, head);
6542 best_energy = min(best_energy, prev_energy);
6547 * Find the CPU with the maximum spare capacity in
6548 * the performance domain
6550 spare_cap = cpu_cap - util;
6551 if (spare_cap > max_spare_cap) {
6552 max_spare_cap = spare_cap;
6553 max_spare_cap_cpu = cpu;
6557 /* Evaluate the energy impact of using this CPU. */
6558 if (max_spare_cap_cpu >= 0) {
6559 cur_energy = compute_energy(p, max_spare_cap_cpu, head);
6560 if (cur_energy < best_energy) {
6561 best_energy = cur_energy;
6562 best_energy_cpu = max_spare_cap_cpu;
6570 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6571 * least 6% of the energy used by prev_cpu.
6573 if (prev_energy == ULONG_MAX)
6574 return best_energy_cpu;
6576 if ((prev_energy - best_energy) > (prev_energy >> 4))
6577 return best_energy_cpu;
6588 * select_task_rq_fair: Select target runqueue for the waking task in domains
6589 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6590 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6592 * Balances load by selecting the idlest CPU in the idlest group, or under
6593 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6595 * Returns the target CPU number.
6597 * preempt must be disabled.
6600 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6602 struct sched_domain *tmp, *sd = NULL;
6603 int cpu = smp_processor_id();
6604 int new_cpu = prev_cpu;
6605 int want_affine = 0;
6606 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6608 if (sd_flag & SD_BALANCE_WAKE) {
6611 if (static_branch_unlikely(&sched_energy_present)) {
6612 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
6618 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu) &&
6619 cpumask_test_cpu(cpu, &p->cpus_allowed);
6623 for_each_domain(cpu, tmp) {
6624 if (!(tmp->flags & SD_LOAD_BALANCE))
6628 * If both 'cpu' and 'prev_cpu' are part of this domain,
6629 * cpu is a valid SD_WAKE_AFFINE target.
6631 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6632 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6633 if (cpu != prev_cpu)
6634 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6636 sd = NULL; /* Prefer wake_affine over balance flags */
6640 if (tmp->flags & sd_flag)
6642 else if (!want_affine)
6648 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6649 } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6652 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6655 current->recent_used_cpu = cpu;
6662 static void detach_entity_cfs_rq(struct sched_entity *se);
6665 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6666 * cfs_rq_of(p) references at time of call are still valid and identify the
6667 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6669 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6672 * As blocked tasks retain absolute vruntime the migration needs to
6673 * deal with this by subtracting the old and adding the new
6674 * min_vruntime -- the latter is done by enqueue_entity() when placing
6675 * the task on the new runqueue.
6677 if (p->state == TASK_WAKING) {
6678 struct sched_entity *se = &p->se;
6679 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6682 #ifndef CONFIG_64BIT
6683 u64 min_vruntime_copy;
6686 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6688 min_vruntime = cfs_rq->min_vruntime;
6689 } while (min_vruntime != min_vruntime_copy);
6691 min_vruntime = cfs_rq->min_vruntime;
6694 se->vruntime -= min_vruntime;
6697 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6699 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6700 * rq->lock and can modify state directly.
6702 lockdep_assert_held(&task_rq(p)->lock);
6703 detach_entity_cfs_rq(&p->se);
6707 * We are supposed to update the task to "current" time, then
6708 * its up to date and ready to go to new CPU/cfs_rq. But we
6709 * have difficulty in getting what current time is, so simply
6710 * throw away the out-of-date time. This will result in the
6711 * wakee task is less decayed, but giving the wakee more load
6714 remove_entity_load_avg(&p->se);
6717 /* Tell new CPU we are migrated */
6718 p->se.avg.last_update_time = 0;
6720 /* We have migrated, no longer consider this task hot */
6721 p->se.exec_start = 0;
6723 update_scan_period(p, new_cpu);
6726 static void task_dead_fair(struct task_struct *p)
6728 remove_entity_load_avg(&p->se);
6730 #endif /* CONFIG_SMP */
6732 static unsigned long wakeup_gran(struct sched_entity *se)
6734 unsigned long gran = sysctl_sched_wakeup_granularity;
6737 * Since its curr running now, convert the gran from real-time
6738 * to virtual-time in his units.
6740 * By using 'se' instead of 'curr' we penalize light tasks, so
6741 * they get preempted easier. That is, if 'se' < 'curr' then
6742 * the resulting gran will be larger, therefore penalizing the
6743 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6744 * be smaller, again penalizing the lighter task.
6746 * This is especially important for buddies when the leftmost
6747 * task is higher priority than the buddy.
6749 return calc_delta_fair(gran, se);
6753 * Should 'se' preempt 'curr'.
6767 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6769 s64 gran, vdiff = curr->vruntime - se->vruntime;
6774 gran = wakeup_gran(se);
6781 static void set_last_buddy(struct sched_entity *se)
6783 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6786 for_each_sched_entity(se) {
6787 if (SCHED_WARN_ON(!se->on_rq))
6789 cfs_rq_of(se)->last = se;
6793 static void set_next_buddy(struct sched_entity *se)
6795 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6798 for_each_sched_entity(se) {
6799 if (SCHED_WARN_ON(!se->on_rq))
6801 cfs_rq_of(se)->next = se;
6805 static void set_skip_buddy(struct sched_entity *se)
6807 for_each_sched_entity(se)
6808 cfs_rq_of(se)->skip = se;
6812 * Preempt the current task with a newly woken task if needed:
6814 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6816 struct task_struct *curr = rq->curr;
6817 struct sched_entity *se = &curr->se, *pse = &p->se;
6818 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6819 int scale = cfs_rq->nr_running >= sched_nr_latency;
6820 int next_buddy_marked = 0;
6822 if (unlikely(se == pse))
6826 * This is possible from callers such as attach_tasks(), in which we
6827 * unconditionally check_prempt_curr() after an enqueue (which may have
6828 * lead to a throttle). This both saves work and prevents false
6829 * next-buddy nomination below.
6831 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6834 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6835 set_next_buddy(pse);
6836 next_buddy_marked = 1;
6840 * We can come here with TIF_NEED_RESCHED already set from new task
6843 * Note: this also catches the edge-case of curr being in a throttled
6844 * group (e.g. via set_curr_task), since update_curr() (in the
6845 * enqueue of curr) will have resulted in resched being set. This
6846 * prevents us from potentially nominating it as a false LAST_BUDDY
6849 if (test_tsk_need_resched(curr))
6852 /* Idle tasks are by definition preempted by non-idle tasks. */
6853 if (unlikely(task_has_idle_policy(curr)) &&
6854 likely(!task_has_idle_policy(p)))
6858 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6859 * is driven by the tick):
6861 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6864 find_matching_se(&se, &pse);
6865 update_curr(cfs_rq_of(se));
6867 if (wakeup_preempt_entity(se, pse) == 1) {
6869 * Bias pick_next to pick the sched entity that is
6870 * triggering this preemption.
6872 if (!next_buddy_marked)
6873 set_next_buddy(pse);
6882 * Only set the backward buddy when the current task is still
6883 * on the rq. This can happen when a wakeup gets interleaved
6884 * with schedule on the ->pre_schedule() or idle_balance()
6885 * point, either of which can * drop the rq lock.
6887 * Also, during early boot the idle thread is in the fair class,
6888 * for obvious reasons its a bad idea to schedule back to it.
6890 if (unlikely(!se->on_rq || curr == rq->idle))
6893 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6897 static struct task_struct *
6898 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6900 struct cfs_rq *cfs_rq = &rq->cfs;
6901 struct sched_entity *se;
6902 struct task_struct *p;
6906 if (!cfs_rq->nr_running)
6909 #ifdef CONFIG_FAIR_GROUP_SCHED
6910 if (prev->sched_class != &fair_sched_class)
6914 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6915 * likely that a next task is from the same cgroup as the current.
6917 * Therefore attempt to avoid putting and setting the entire cgroup
6918 * hierarchy, only change the part that actually changes.
6922 struct sched_entity *curr = cfs_rq->curr;
6925 * Since we got here without doing put_prev_entity() we also
6926 * have to consider cfs_rq->curr. If it is still a runnable
6927 * entity, update_curr() will update its vruntime, otherwise
6928 * forget we've ever seen it.
6932 update_curr(cfs_rq);
6937 * This call to check_cfs_rq_runtime() will do the
6938 * throttle and dequeue its entity in the parent(s).
6939 * Therefore the nr_running test will indeed
6942 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6945 if (!cfs_rq->nr_running)
6952 se = pick_next_entity(cfs_rq, curr);
6953 cfs_rq = group_cfs_rq(se);
6959 * Since we haven't yet done put_prev_entity and if the selected task
6960 * is a different task than we started out with, try and touch the
6961 * least amount of cfs_rqs.
6964 struct sched_entity *pse = &prev->se;
6966 while (!(cfs_rq = is_same_group(se, pse))) {
6967 int se_depth = se->depth;
6968 int pse_depth = pse->depth;
6970 if (se_depth <= pse_depth) {
6971 put_prev_entity(cfs_rq_of(pse), pse);
6972 pse = parent_entity(pse);
6974 if (se_depth >= pse_depth) {
6975 set_next_entity(cfs_rq_of(se), se);
6976 se = parent_entity(se);
6980 put_prev_entity(cfs_rq, pse);
6981 set_next_entity(cfs_rq, se);
6988 put_prev_task(rq, prev);
6991 se = pick_next_entity(cfs_rq, NULL);
6992 set_next_entity(cfs_rq, se);
6993 cfs_rq = group_cfs_rq(se);
6998 done: __maybe_unused;
7001 * Move the next running task to the front of
7002 * the list, so our cfs_tasks list becomes MRU
7005 list_move(&p->se.group_node, &rq->cfs_tasks);
7008 if (hrtick_enabled(rq))
7009 hrtick_start_fair(rq, p);
7011 update_misfit_status(p, rq);
7016 update_misfit_status(NULL, rq);
7017 new_tasks = idle_balance(rq, rf);
7020 * Because idle_balance() releases (and re-acquires) rq->lock, it is
7021 * possible for any higher priority task to appear. In that case we
7022 * must re-start the pick_next_entity() loop.
7034 * Account for a descheduled task:
7036 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
7038 struct sched_entity *se = &prev->se;
7039 struct cfs_rq *cfs_rq;
7041 for_each_sched_entity(se) {
7042 cfs_rq = cfs_rq_of(se);
7043 put_prev_entity(cfs_rq, se);
7048 * sched_yield() is very simple
7050 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7052 static void yield_task_fair(struct rq *rq)
7054 struct task_struct *curr = rq->curr;
7055 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7056 struct sched_entity *se = &curr->se;
7059 * Are we the only task in the tree?
7061 if (unlikely(rq->nr_running == 1))
7064 clear_buddies(cfs_rq, se);
7066 if (curr->policy != SCHED_BATCH) {
7067 update_rq_clock(rq);
7069 * Update run-time statistics of the 'current'.
7071 update_curr(cfs_rq);
7073 * Tell update_rq_clock() that we've just updated,
7074 * so we don't do microscopic update in schedule()
7075 * and double the fastpath cost.
7077 rq_clock_skip_update(rq);
7083 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
7085 struct sched_entity *se = &p->se;
7087 /* throttled hierarchies are not runnable */
7088 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
7091 /* Tell the scheduler that we'd really like pse to run next. */
7094 yield_task_fair(rq);
7100 /**************************************************
7101 * Fair scheduling class load-balancing methods.
7105 * The purpose of load-balancing is to achieve the same basic fairness the
7106 * per-CPU scheduler provides, namely provide a proportional amount of compute
7107 * time to each task. This is expressed in the following equation:
7109 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7111 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7112 * W_i,0 is defined as:
7114 * W_i,0 = \Sum_j w_i,j (2)
7116 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7117 * is derived from the nice value as per sched_prio_to_weight[].
7119 * The weight average is an exponential decay average of the instantaneous
7122 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7124 * C_i is the compute capacity of CPU i, typically it is the
7125 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7126 * can also include other factors [XXX].
7128 * To achieve this balance we define a measure of imbalance which follows
7129 * directly from (1):
7131 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7133 * We them move tasks around to minimize the imbalance. In the continuous
7134 * function space it is obvious this converges, in the discrete case we get
7135 * a few fun cases generally called infeasible weight scenarios.
7138 * - infeasible weights;
7139 * - local vs global optima in the discrete case. ]
7144 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7145 * for all i,j solution, we create a tree of CPUs that follows the hardware
7146 * topology where each level pairs two lower groups (or better). This results
7147 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7148 * tree to only the first of the previous level and we decrease the frequency
7149 * of load-balance at each level inv. proportional to the number of CPUs in
7155 * \Sum { --- * --- * 2^i } = O(n) (5)
7157 * `- size of each group
7158 * | | `- number of CPUs doing load-balance
7160 * `- sum over all levels
7162 * Coupled with a limit on how many tasks we can migrate every balance pass,
7163 * this makes (5) the runtime complexity of the balancer.
7165 * An important property here is that each CPU is still (indirectly) connected
7166 * to every other CPU in at most O(log n) steps:
7168 * The adjacency matrix of the resulting graph is given by:
7171 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7174 * And you'll find that:
7176 * A^(log_2 n)_i,j != 0 for all i,j (7)
7178 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7179 * The task movement gives a factor of O(m), giving a convergence complexity
7182 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7187 * In order to avoid CPUs going idle while there's still work to do, new idle
7188 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7189 * tree itself instead of relying on other CPUs to bring it work.
7191 * This adds some complexity to both (5) and (8) but it reduces the total idle
7199 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7202 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7207 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7209 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7211 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7214 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7215 * rewrite all of this once again.]
7218 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7220 enum fbq_type { regular, remote, all };
7229 #define LBF_ALL_PINNED 0x01
7230 #define LBF_NEED_BREAK 0x02
7231 #define LBF_DST_PINNED 0x04
7232 #define LBF_SOME_PINNED 0x08
7233 #define LBF_NOHZ_STATS 0x10
7234 #define LBF_NOHZ_AGAIN 0x20
7237 struct sched_domain *sd;
7245 struct cpumask *dst_grpmask;
7247 enum cpu_idle_type idle;
7249 /* The set of CPUs under consideration for load-balancing */
7250 struct cpumask *cpus;
7255 unsigned int loop_break;
7256 unsigned int loop_max;
7258 enum fbq_type fbq_type;
7259 enum group_type src_grp_type;
7260 struct list_head tasks;
7264 * Is this task likely cache-hot:
7266 static int task_hot(struct task_struct *p, struct lb_env *env)
7270 lockdep_assert_held(&env->src_rq->lock);
7272 if (p->sched_class != &fair_sched_class)
7275 if (unlikely(task_has_idle_policy(p)))
7279 * Buddy candidates are cache hot:
7281 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7282 (&p->se == cfs_rq_of(&p->se)->next ||
7283 &p->se == cfs_rq_of(&p->se)->last))
7286 if (sysctl_sched_migration_cost == -1)
7288 if (sysctl_sched_migration_cost == 0)
7291 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7293 return delta < (s64)sysctl_sched_migration_cost;
7296 #ifdef CONFIG_NUMA_BALANCING
7298 * Returns 1, if task migration degrades locality
7299 * Returns 0, if task migration improves locality i.e migration preferred.
7300 * Returns -1, if task migration is not affected by locality.
7302 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7304 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7305 unsigned long src_weight, dst_weight;
7306 int src_nid, dst_nid, dist;
7308 if (!static_branch_likely(&sched_numa_balancing))
7311 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7314 src_nid = cpu_to_node(env->src_cpu);
7315 dst_nid = cpu_to_node(env->dst_cpu);
7317 if (src_nid == dst_nid)
7320 /* Migrating away from the preferred node is always bad. */
7321 if (src_nid == p->numa_preferred_nid) {
7322 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7328 /* Encourage migration to the preferred node. */
7329 if (dst_nid == p->numa_preferred_nid)
7332 /* Leaving a core idle is often worse than degrading locality. */
7333 if (env->idle == CPU_IDLE)
7336 dist = node_distance(src_nid, dst_nid);
7338 src_weight = group_weight(p, src_nid, dist);
7339 dst_weight = group_weight(p, dst_nid, dist);
7341 src_weight = task_weight(p, src_nid, dist);
7342 dst_weight = task_weight(p, dst_nid, dist);
7345 return dst_weight < src_weight;
7349 static inline int migrate_degrades_locality(struct task_struct *p,
7357 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7360 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7364 lockdep_assert_held(&env->src_rq->lock);
7367 * We do not migrate tasks that are:
7368 * 1) throttled_lb_pair, or
7369 * 2) cannot be migrated to this CPU due to cpus_allowed, or
7370 * 3) running (obviously), or
7371 * 4) are cache-hot on their current CPU.
7373 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7376 if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) {
7379 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7381 env->flags |= LBF_SOME_PINNED;
7384 * Remember if this task can be migrated to any other CPU in
7385 * our sched_group. We may want to revisit it if we couldn't
7386 * meet load balance goals by pulling other tasks on src_cpu.
7388 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7389 * already computed one in current iteration.
7391 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7394 /* Prevent to re-select dst_cpu via env's CPUs: */
7395 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7396 if (cpumask_test_cpu(cpu, &p->cpus_allowed)) {
7397 env->flags |= LBF_DST_PINNED;
7398 env->new_dst_cpu = cpu;
7406 /* Record that we found atleast one task that could run on dst_cpu */
7407 env->flags &= ~LBF_ALL_PINNED;
7409 if (task_running(env->src_rq, p)) {
7410 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7415 * Aggressive migration if:
7416 * 1) destination numa is preferred
7417 * 2) task is cache cold, or
7418 * 3) too many balance attempts have failed.
7420 tsk_cache_hot = migrate_degrades_locality(p, env);
7421 if (tsk_cache_hot == -1)
7422 tsk_cache_hot = task_hot(p, env);
7424 if (tsk_cache_hot <= 0 ||
7425 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7426 if (tsk_cache_hot == 1) {
7427 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7428 schedstat_inc(p->se.statistics.nr_forced_migrations);
7433 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7438 * detach_task() -- detach the task for the migration specified in env
7440 static void detach_task(struct task_struct *p, struct lb_env *env)
7442 lockdep_assert_held(&env->src_rq->lock);
7444 p->on_rq = TASK_ON_RQ_MIGRATING;
7445 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7446 set_task_cpu(p, env->dst_cpu);
7450 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7451 * part of active balancing operations within "domain".
7453 * Returns a task if successful and NULL otherwise.
7455 static struct task_struct *detach_one_task(struct lb_env *env)
7457 struct task_struct *p;
7459 lockdep_assert_held(&env->src_rq->lock);
7461 list_for_each_entry_reverse(p,
7462 &env->src_rq->cfs_tasks, se.group_node) {
7463 if (!can_migrate_task(p, env))
7466 detach_task(p, env);
7469 * Right now, this is only the second place where
7470 * lb_gained[env->idle] is updated (other is detach_tasks)
7471 * so we can safely collect stats here rather than
7472 * inside detach_tasks().
7474 schedstat_inc(env->sd->lb_gained[env->idle]);
7480 static const unsigned int sched_nr_migrate_break = 32;
7483 * detach_tasks() -- tries to detach up to imbalance weighted load from
7484 * busiest_rq, as part of a balancing operation within domain "sd".
7486 * Returns number of detached tasks if successful and 0 otherwise.
7488 static int detach_tasks(struct lb_env *env)
7490 struct list_head *tasks = &env->src_rq->cfs_tasks;
7491 struct task_struct *p;
7495 lockdep_assert_held(&env->src_rq->lock);
7497 if (env->imbalance <= 0)
7500 while (!list_empty(tasks)) {
7502 * We don't want to steal all, otherwise we may be treated likewise,
7503 * which could at worst lead to a livelock crash.
7505 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7508 p = list_last_entry(tasks, struct task_struct, se.group_node);
7511 /* We've more or less seen every task there is, call it quits */
7512 if (env->loop > env->loop_max)
7515 /* take a breather every nr_migrate tasks */
7516 if (env->loop > env->loop_break) {
7517 env->loop_break += sched_nr_migrate_break;
7518 env->flags |= LBF_NEED_BREAK;
7522 if (!can_migrate_task(p, env))
7525 load = task_h_load(p);
7527 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7530 if ((load / 2) > env->imbalance)
7533 detach_task(p, env);
7534 list_add(&p->se.group_node, &env->tasks);
7537 env->imbalance -= load;
7539 #ifdef CONFIG_PREEMPT
7541 * NEWIDLE balancing is a source of latency, so preemptible
7542 * kernels will stop after the first task is detached to minimize
7543 * the critical section.
7545 if (env->idle == CPU_NEWLY_IDLE)
7550 * We only want to steal up to the prescribed amount of
7553 if (env->imbalance <= 0)
7558 list_move(&p->se.group_node, tasks);
7562 * Right now, this is one of only two places we collect this stat
7563 * so we can safely collect detach_one_task() stats here rather
7564 * than inside detach_one_task().
7566 schedstat_add(env->sd->lb_gained[env->idle], detached);
7572 * attach_task() -- attach the task detached by detach_task() to its new rq.
7574 static void attach_task(struct rq *rq, struct task_struct *p)
7576 lockdep_assert_held(&rq->lock);
7578 BUG_ON(task_rq(p) != rq);
7579 activate_task(rq, p, ENQUEUE_NOCLOCK);
7580 p->on_rq = TASK_ON_RQ_QUEUED;
7581 check_preempt_curr(rq, p, 0);
7585 * attach_one_task() -- attaches the task returned from detach_one_task() to
7588 static void attach_one_task(struct rq *rq, struct task_struct *p)
7593 update_rq_clock(rq);
7599 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7602 static void attach_tasks(struct lb_env *env)
7604 struct list_head *tasks = &env->tasks;
7605 struct task_struct *p;
7608 rq_lock(env->dst_rq, &rf);
7609 update_rq_clock(env->dst_rq);
7611 while (!list_empty(tasks)) {
7612 p = list_first_entry(tasks, struct task_struct, se.group_node);
7613 list_del_init(&p->se.group_node);
7615 attach_task(env->dst_rq, p);
7618 rq_unlock(env->dst_rq, &rf);
7621 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7623 if (cfs_rq->avg.load_avg)
7626 if (cfs_rq->avg.util_avg)
7632 static inline bool others_have_blocked(struct rq *rq)
7634 if (READ_ONCE(rq->avg_rt.util_avg))
7637 if (READ_ONCE(rq->avg_dl.util_avg))
7640 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7641 if (READ_ONCE(rq->avg_irq.util_avg))
7648 #ifdef CONFIG_FAIR_GROUP_SCHED
7650 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7652 if (cfs_rq->load.weight)
7655 if (cfs_rq->avg.load_sum)
7658 if (cfs_rq->avg.util_sum)
7661 if (cfs_rq->avg.runnable_load_sum)
7667 static void update_blocked_averages(int cpu)
7669 struct rq *rq = cpu_rq(cpu);
7670 struct cfs_rq *cfs_rq, *pos;
7671 const struct sched_class *curr_class;
7675 rq_lock_irqsave(rq, &rf);
7676 update_rq_clock(rq);
7679 * Iterates the task_group tree in a bottom up fashion, see
7680 * list_add_leaf_cfs_rq() for details.
7682 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7683 struct sched_entity *se;
7685 /* throttled entities do not contribute to load */
7686 if (throttled_hierarchy(cfs_rq))
7689 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
7690 update_tg_load_avg(cfs_rq, 0);
7692 /* Propagate pending load changes to the parent, if any: */
7693 se = cfs_rq->tg->se[cpu];
7694 if (se && !skip_blocked_update(se))
7695 update_load_avg(cfs_rq_of(se), se, 0);
7698 * There can be a lot of idle CPU cgroups. Don't let fully
7699 * decayed cfs_rqs linger on the list.
7701 if (cfs_rq_is_decayed(cfs_rq))
7702 list_del_leaf_cfs_rq(cfs_rq);
7704 /* Don't need periodic decay once load/util_avg are null */
7705 if (cfs_rq_has_blocked(cfs_rq))
7709 curr_class = rq->curr->sched_class;
7710 update_rt_rq_load_avg(rq_clock_task(rq), rq, curr_class == &rt_sched_class);
7711 update_dl_rq_load_avg(rq_clock_task(rq), rq, curr_class == &dl_sched_class);
7712 update_irq_load_avg(rq, 0);
7713 /* Don't need periodic decay once load/util_avg are null */
7714 if (others_have_blocked(rq))
7717 #ifdef CONFIG_NO_HZ_COMMON
7718 rq->last_blocked_load_update_tick = jiffies;
7720 rq->has_blocked_load = 0;
7722 rq_unlock_irqrestore(rq, &rf);
7726 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7727 * This needs to be done in a top-down fashion because the load of a child
7728 * group is a fraction of its parents load.
7730 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7732 struct rq *rq = rq_of(cfs_rq);
7733 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7734 unsigned long now = jiffies;
7737 if (cfs_rq->last_h_load_update == now)
7740 cfs_rq->h_load_next = NULL;
7741 for_each_sched_entity(se) {
7742 cfs_rq = cfs_rq_of(se);
7743 cfs_rq->h_load_next = se;
7744 if (cfs_rq->last_h_load_update == now)
7749 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7750 cfs_rq->last_h_load_update = now;
7753 while ((se = cfs_rq->h_load_next) != NULL) {
7754 load = cfs_rq->h_load;
7755 load = div64_ul(load * se->avg.load_avg,
7756 cfs_rq_load_avg(cfs_rq) + 1);
7757 cfs_rq = group_cfs_rq(se);
7758 cfs_rq->h_load = load;
7759 cfs_rq->last_h_load_update = now;
7763 static unsigned long task_h_load(struct task_struct *p)
7765 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7767 update_cfs_rq_h_load(cfs_rq);
7768 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7769 cfs_rq_load_avg(cfs_rq) + 1);
7772 static inline void update_blocked_averages(int cpu)
7774 struct rq *rq = cpu_rq(cpu);
7775 struct cfs_rq *cfs_rq = &rq->cfs;
7776 const struct sched_class *curr_class;
7779 rq_lock_irqsave(rq, &rf);
7780 update_rq_clock(rq);
7781 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
7783 curr_class = rq->curr->sched_class;
7784 update_rt_rq_load_avg(rq_clock_task(rq), rq, curr_class == &rt_sched_class);
7785 update_dl_rq_load_avg(rq_clock_task(rq), rq, curr_class == &dl_sched_class);
7786 update_irq_load_avg(rq, 0);
7787 #ifdef CONFIG_NO_HZ_COMMON
7788 rq->last_blocked_load_update_tick = jiffies;
7789 if (!cfs_rq_has_blocked(cfs_rq) && !others_have_blocked(rq))
7790 rq->has_blocked_load = 0;
7792 rq_unlock_irqrestore(rq, &rf);
7795 static unsigned long task_h_load(struct task_struct *p)
7797 return p->se.avg.load_avg;
7801 /********** Helpers for find_busiest_group ************************/
7804 * sg_lb_stats - stats of a sched_group required for load_balancing
7806 struct sg_lb_stats {
7807 unsigned long avg_load; /*Avg load across the CPUs of the group */
7808 unsigned long group_load; /* Total load over the CPUs of the group */
7809 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7810 unsigned long load_per_task;
7811 unsigned long group_capacity;
7812 unsigned long group_util; /* Total utilization of the group */
7813 unsigned int sum_nr_running; /* Nr tasks running in the group */
7814 unsigned int idle_cpus;
7815 unsigned int group_weight;
7816 enum group_type group_type;
7817 int group_no_capacity;
7818 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
7819 #ifdef CONFIG_NUMA_BALANCING
7820 unsigned int nr_numa_running;
7821 unsigned int nr_preferred_running;
7826 * sd_lb_stats - Structure to store the statistics of a sched_domain
7827 * during load balancing.
7829 struct sd_lb_stats {
7830 struct sched_group *busiest; /* Busiest group in this sd */
7831 struct sched_group *local; /* Local group in this sd */
7832 unsigned long total_running;
7833 unsigned long total_load; /* Total load of all groups in sd */
7834 unsigned long total_capacity; /* Total capacity of all groups in sd */
7835 unsigned long avg_load; /* Average load across all groups in sd */
7837 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7838 struct sg_lb_stats local_stat; /* Statistics of the local group */
7841 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7844 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7845 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7846 * We must however clear busiest_stat::avg_load because
7847 * update_sd_pick_busiest() reads this before assignment.
7849 *sds = (struct sd_lb_stats){
7852 .total_running = 0UL,
7854 .total_capacity = 0UL,
7857 .sum_nr_running = 0,
7858 .group_type = group_other,
7864 * get_sd_load_idx - Obtain the load index for a given sched domain.
7865 * @sd: The sched_domain whose load_idx is to be obtained.
7866 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7868 * Return: The load index.
7870 static inline int get_sd_load_idx(struct sched_domain *sd,
7871 enum cpu_idle_type idle)
7877 load_idx = sd->busy_idx;
7880 case CPU_NEWLY_IDLE:
7881 load_idx = sd->newidle_idx;
7884 load_idx = sd->idle_idx;
7891 static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu)
7893 struct rq *rq = cpu_rq(cpu);
7894 unsigned long max = arch_scale_cpu_capacity(sd, cpu);
7895 unsigned long used, free;
7898 irq = cpu_util_irq(rq);
7900 if (unlikely(irq >= max))
7903 used = READ_ONCE(rq->avg_rt.util_avg);
7904 used += READ_ONCE(rq->avg_dl.util_avg);
7906 if (unlikely(used >= max))
7911 return scale_irq_capacity(free, irq, max);
7914 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7916 unsigned long capacity = scale_rt_capacity(sd, cpu);
7917 struct sched_group *sdg = sd->groups;
7919 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(sd, cpu);
7924 cpu_rq(cpu)->cpu_capacity = capacity;
7925 sdg->sgc->capacity = capacity;
7926 sdg->sgc->min_capacity = capacity;
7927 sdg->sgc->max_capacity = capacity;
7930 void update_group_capacity(struct sched_domain *sd, int cpu)
7932 struct sched_domain *child = sd->child;
7933 struct sched_group *group, *sdg = sd->groups;
7934 unsigned long capacity, min_capacity, max_capacity;
7935 unsigned long interval;
7937 interval = msecs_to_jiffies(sd->balance_interval);
7938 interval = clamp(interval, 1UL, max_load_balance_interval);
7939 sdg->sgc->next_update = jiffies + interval;
7942 update_cpu_capacity(sd, cpu);
7947 min_capacity = ULONG_MAX;
7950 if (child->flags & SD_OVERLAP) {
7952 * SD_OVERLAP domains cannot assume that child groups
7953 * span the current group.
7956 for_each_cpu(cpu, sched_group_span(sdg)) {
7957 struct sched_group_capacity *sgc;
7958 struct rq *rq = cpu_rq(cpu);
7961 * build_sched_domains() -> init_sched_groups_capacity()
7962 * gets here before we've attached the domains to the
7965 * Use capacity_of(), which is set irrespective of domains
7966 * in update_cpu_capacity().
7968 * This avoids capacity from being 0 and
7969 * causing divide-by-zero issues on boot.
7971 if (unlikely(!rq->sd)) {
7972 capacity += capacity_of(cpu);
7974 sgc = rq->sd->groups->sgc;
7975 capacity += sgc->capacity;
7978 min_capacity = min(capacity, min_capacity);
7979 max_capacity = max(capacity, max_capacity);
7983 * !SD_OVERLAP domains can assume that child groups
7984 * span the current group.
7987 group = child->groups;
7989 struct sched_group_capacity *sgc = group->sgc;
7991 capacity += sgc->capacity;
7992 min_capacity = min(sgc->min_capacity, min_capacity);
7993 max_capacity = max(sgc->max_capacity, max_capacity);
7994 group = group->next;
7995 } while (group != child->groups);
7998 sdg->sgc->capacity = capacity;
7999 sdg->sgc->min_capacity = min_capacity;
8000 sdg->sgc->max_capacity = max_capacity;
8004 * Check whether the capacity of the rq has been noticeably reduced by side
8005 * activity. The imbalance_pct is used for the threshold.
8006 * Return true is the capacity is reduced
8009 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
8011 return ((rq->cpu_capacity * sd->imbalance_pct) <
8012 (rq->cpu_capacity_orig * 100));
8016 * Group imbalance indicates (and tries to solve) the problem where balancing
8017 * groups is inadequate due to ->cpus_allowed constraints.
8019 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
8020 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
8023 * { 0 1 2 3 } { 4 5 6 7 }
8026 * If we were to balance group-wise we'd place two tasks in the first group and
8027 * two tasks in the second group. Clearly this is undesired as it will overload
8028 * cpu 3 and leave one of the CPUs in the second group unused.
8030 * The current solution to this issue is detecting the skew in the first group
8031 * by noticing the lower domain failed to reach balance and had difficulty
8032 * moving tasks due to affinity constraints.
8034 * When this is so detected; this group becomes a candidate for busiest; see
8035 * update_sd_pick_busiest(). And calculate_imbalance() and
8036 * find_busiest_group() avoid some of the usual balance conditions to allow it
8037 * to create an effective group imbalance.
8039 * This is a somewhat tricky proposition since the next run might not find the
8040 * group imbalance and decide the groups need to be balanced again. A most
8041 * subtle and fragile situation.
8044 static inline int sg_imbalanced(struct sched_group *group)
8046 return group->sgc->imbalance;
8050 * group_has_capacity returns true if the group has spare capacity that could
8051 * be used by some tasks.
8052 * We consider that a group has spare capacity if the * number of task is
8053 * smaller than the number of CPUs or if the utilization is lower than the
8054 * available capacity for CFS tasks.
8055 * For the latter, we use a threshold to stabilize the state, to take into
8056 * account the variance of the tasks' load and to return true if the available
8057 * capacity in meaningful for the load balancer.
8058 * As an example, an available capacity of 1% can appear but it doesn't make
8059 * any benefit for the load balance.
8062 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
8064 if (sgs->sum_nr_running < sgs->group_weight)
8067 if ((sgs->group_capacity * 100) >
8068 (sgs->group_util * env->sd->imbalance_pct))
8075 * group_is_overloaded returns true if the group has more tasks than it can
8077 * group_is_overloaded is not equals to !group_has_capacity because a group
8078 * with the exact right number of tasks, has no more spare capacity but is not
8079 * overloaded so both group_has_capacity and group_is_overloaded return
8083 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
8085 if (sgs->sum_nr_running <= sgs->group_weight)
8088 if ((sgs->group_capacity * 100) <
8089 (sgs->group_util * env->sd->imbalance_pct))
8096 * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller
8097 * per-CPU capacity than sched_group ref.
8100 group_smaller_min_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
8102 return sg->sgc->min_capacity * capacity_margin <
8103 ref->sgc->min_capacity * 1024;
8107 * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller
8108 * per-CPU capacity_orig than sched_group ref.
8111 group_smaller_max_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
8113 return sg->sgc->max_capacity * capacity_margin <
8114 ref->sgc->max_capacity * 1024;
8118 group_type group_classify(struct sched_group *group,
8119 struct sg_lb_stats *sgs)
8121 if (sgs->group_no_capacity)
8122 return group_overloaded;
8124 if (sg_imbalanced(group))
8125 return group_imbalanced;
8127 if (sgs->group_misfit_task_load)
8128 return group_misfit_task;
8133 static bool update_nohz_stats(struct rq *rq, bool force)
8135 #ifdef CONFIG_NO_HZ_COMMON
8136 unsigned int cpu = rq->cpu;
8138 if (!rq->has_blocked_load)
8141 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
8144 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
8147 update_blocked_averages(cpu);
8149 return rq->has_blocked_load;
8156 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8157 * @env: The load balancing environment.
8158 * @group: sched_group whose statistics are to be updated.
8159 * @sgs: variable to hold the statistics for this group.
8160 * @sg_status: Holds flag indicating the status of the sched_group
8162 static inline void update_sg_lb_stats(struct lb_env *env,
8163 struct sched_group *group,
8164 struct sg_lb_stats *sgs,
8167 int local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(group));
8168 int load_idx = get_sd_load_idx(env->sd, env->idle);
8172 memset(sgs, 0, sizeof(*sgs));
8174 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8175 struct rq *rq = cpu_rq(i);
8177 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
8178 env->flags |= LBF_NOHZ_AGAIN;
8180 /* Bias balancing toward CPUs of our domain: */
8182 load = target_load(i, load_idx);
8184 load = source_load(i, load_idx);
8186 sgs->group_load += load;
8187 sgs->group_util += cpu_util(i);
8188 sgs->sum_nr_running += rq->cfs.h_nr_running;
8190 nr_running = rq->nr_running;
8192 *sg_status |= SG_OVERLOAD;
8194 if (cpu_overutilized(i))
8195 *sg_status |= SG_OVERUTILIZED;
8197 #ifdef CONFIG_NUMA_BALANCING
8198 sgs->nr_numa_running += rq->nr_numa_running;
8199 sgs->nr_preferred_running += rq->nr_preferred_running;
8201 sgs->sum_weighted_load += weighted_cpuload(rq);
8203 * No need to call idle_cpu() if nr_running is not 0
8205 if (!nr_running && idle_cpu(i))
8208 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8209 sgs->group_misfit_task_load < rq->misfit_task_load) {
8210 sgs->group_misfit_task_load = rq->misfit_task_load;
8211 *sg_status |= SG_OVERLOAD;
8215 /* Adjust by relative CPU capacity of the group */
8216 sgs->group_capacity = group->sgc->capacity;
8217 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
8219 if (sgs->sum_nr_running)
8220 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
8222 sgs->group_weight = group->group_weight;
8224 sgs->group_no_capacity = group_is_overloaded(env, sgs);
8225 sgs->group_type = group_classify(group, sgs);
8229 * update_sd_pick_busiest - return 1 on busiest group
8230 * @env: The load balancing environment.
8231 * @sds: sched_domain statistics
8232 * @sg: sched_group candidate to be checked for being the busiest
8233 * @sgs: sched_group statistics
8235 * Determine if @sg is a busier group than the previously selected
8238 * Return: %true if @sg is a busier group than the previously selected
8239 * busiest group. %false otherwise.
8241 static bool update_sd_pick_busiest(struct lb_env *env,
8242 struct sd_lb_stats *sds,
8243 struct sched_group *sg,
8244 struct sg_lb_stats *sgs)
8246 struct sg_lb_stats *busiest = &sds->busiest_stat;
8249 * Don't try to pull misfit tasks we can't help.
8250 * We can use max_capacity here as reduction in capacity on some
8251 * CPUs in the group should either be possible to resolve
8252 * internally or be covered by avg_load imbalance (eventually).
8254 if (sgs->group_type == group_misfit_task &&
8255 (!group_smaller_max_cpu_capacity(sg, sds->local) ||
8256 !group_has_capacity(env, &sds->local_stat)))
8259 if (sgs->group_type > busiest->group_type)
8262 if (sgs->group_type < busiest->group_type)
8265 if (sgs->avg_load <= busiest->avg_load)
8268 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
8272 * Candidate sg has no more than one task per CPU and
8273 * has higher per-CPU capacity. Migrating tasks to less
8274 * capable CPUs may harm throughput. Maximize throughput,
8275 * power/energy consequences are not considered.
8277 if (sgs->sum_nr_running <= sgs->group_weight &&
8278 group_smaller_min_cpu_capacity(sds->local, sg))
8282 * If we have more than one misfit sg go with the biggest misfit.
8284 if (sgs->group_type == group_misfit_task &&
8285 sgs->group_misfit_task_load < busiest->group_misfit_task_load)
8289 /* This is the busiest node in its class. */
8290 if (!(env->sd->flags & SD_ASYM_PACKING))
8293 /* No ASYM_PACKING if target CPU is already busy */
8294 if (env->idle == CPU_NOT_IDLE)
8297 * ASYM_PACKING needs to move all the work to the highest
8298 * prority CPUs in the group, therefore mark all groups
8299 * of lower priority than ourself as busy.
8301 if (sgs->sum_nr_running &&
8302 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
8306 /* Prefer to move from lowest priority CPU's work */
8307 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
8308 sg->asym_prefer_cpu))
8315 #ifdef CONFIG_NUMA_BALANCING
8316 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8318 if (sgs->sum_nr_running > sgs->nr_numa_running)
8320 if (sgs->sum_nr_running > sgs->nr_preferred_running)
8325 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8327 if (rq->nr_running > rq->nr_numa_running)
8329 if (rq->nr_running > rq->nr_preferred_running)
8334 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8339 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8343 #endif /* CONFIG_NUMA_BALANCING */
8346 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8347 * @env: The load balancing environment.
8348 * @sds: variable to hold the statistics for this sched_domain.
8350 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8352 struct sched_domain *child = env->sd->child;
8353 struct sched_group *sg = env->sd->groups;
8354 struct sg_lb_stats *local = &sds->local_stat;
8355 struct sg_lb_stats tmp_sgs;
8356 bool prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
8359 #ifdef CONFIG_NO_HZ_COMMON
8360 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
8361 env->flags |= LBF_NOHZ_STATS;
8365 struct sg_lb_stats *sgs = &tmp_sgs;
8368 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
8373 if (env->idle != CPU_NEWLY_IDLE ||
8374 time_after_eq(jiffies, sg->sgc->next_update))
8375 update_group_capacity(env->sd, env->dst_cpu);
8378 update_sg_lb_stats(env, sg, sgs, &sg_status);
8384 * In case the child domain prefers tasks go to siblings
8385 * first, lower the sg capacity so that we'll try
8386 * and move all the excess tasks away. We lower the capacity
8387 * of a group only if the local group has the capacity to fit
8388 * these excess tasks. The extra check prevents the case where
8389 * you always pull from the heaviest group when it is already
8390 * under-utilized (possible with a large weight task outweighs
8391 * the tasks on the system).
8393 if (prefer_sibling && sds->local &&
8394 group_has_capacity(env, local) &&
8395 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
8396 sgs->group_no_capacity = 1;
8397 sgs->group_type = group_classify(sg, sgs);
8400 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8402 sds->busiest_stat = *sgs;
8406 /* Now, start updating sd_lb_stats */
8407 sds->total_running += sgs->sum_nr_running;
8408 sds->total_load += sgs->group_load;
8409 sds->total_capacity += sgs->group_capacity;
8412 } while (sg != env->sd->groups);
8414 #ifdef CONFIG_NO_HZ_COMMON
8415 if ((env->flags & LBF_NOHZ_AGAIN) &&
8416 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8418 WRITE_ONCE(nohz.next_blocked,
8419 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8423 if (env->sd->flags & SD_NUMA)
8424 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8426 if (!env->sd->parent) {
8427 struct root_domain *rd = env->dst_rq->rd;
8429 /* update overload indicator if we are at root domain */
8430 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
8432 /* Update over-utilization (tipping point, U >= 0) indicator */
8433 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
8434 } else if (sg_status & SG_OVERUTILIZED) {
8435 WRITE_ONCE(env->dst_rq->rd->overutilized, SG_OVERUTILIZED);
8440 * check_asym_packing - Check to see if the group is packed into the
8443 * This is primarily intended to used at the sibling level. Some
8444 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8445 * case of POWER7, it can move to lower SMT modes only when higher
8446 * threads are idle. When in lower SMT modes, the threads will
8447 * perform better since they share less core resources. Hence when we
8448 * have idle threads, we want them to be the higher ones.
8450 * This packing function is run on idle threads. It checks to see if
8451 * the busiest CPU in this domain (core in the P7 case) has a higher
8452 * CPU number than the packing function is being run on. Here we are
8453 * assuming lower CPU number will be equivalent to lower a SMT thread
8456 * Return: 1 when packing is required and a task should be moved to
8457 * this CPU. The amount of the imbalance is returned in env->imbalance.
8459 * @env: The load balancing environment.
8460 * @sds: Statistics of the sched_domain which is to be packed
8462 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8466 if (!(env->sd->flags & SD_ASYM_PACKING))
8469 if (env->idle == CPU_NOT_IDLE)
8475 busiest_cpu = sds->busiest->asym_prefer_cpu;
8476 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
8479 env->imbalance = DIV_ROUND_CLOSEST(
8480 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
8481 SCHED_CAPACITY_SCALE);
8487 * fix_small_imbalance - Calculate the minor imbalance that exists
8488 * amongst the groups of a sched_domain, during
8490 * @env: The load balancing environment.
8491 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8494 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8496 unsigned long tmp, capa_now = 0, capa_move = 0;
8497 unsigned int imbn = 2;
8498 unsigned long scaled_busy_load_per_task;
8499 struct sg_lb_stats *local, *busiest;
8501 local = &sds->local_stat;
8502 busiest = &sds->busiest_stat;
8504 if (!local->sum_nr_running)
8505 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8506 else if (busiest->load_per_task > local->load_per_task)
8509 scaled_busy_load_per_task =
8510 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8511 busiest->group_capacity;
8513 if (busiest->avg_load + scaled_busy_load_per_task >=
8514 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8515 env->imbalance = busiest->load_per_task;
8520 * OK, we don't have enough imbalance to justify moving tasks,
8521 * however we may be able to increase total CPU capacity used by
8525 capa_now += busiest->group_capacity *
8526 min(busiest->load_per_task, busiest->avg_load);
8527 capa_now += local->group_capacity *
8528 min(local->load_per_task, local->avg_load);
8529 capa_now /= SCHED_CAPACITY_SCALE;
8531 /* Amount of load we'd subtract */
8532 if (busiest->avg_load > scaled_busy_load_per_task) {
8533 capa_move += busiest->group_capacity *
8534 min(busiest->load_per_task,
8535 busiest->avg_load - scaled_busy_load_per_task);
8538 /* Amount of load we'd add */
8539 if (busiest->avg_load * busiest->group_capacity <
8540 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8541 tmp = (busiest->avg_load * busiest->group_capacity) /
8542 local->group_capacity;
8544 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8545 local->group_capacity;
8547 capa_move += local->group_capacity *
8548 min(local->load_per_task, local->avg_load + tmp);
8549 capa_move /= SCHED_CAPACITY_SCALE;
8551 /* Move if we gain throughput */
8552 if (capa_move > capa_now)
8553 env->imbalance = busiest->load_per_task;
8557 * calculate_imbalance - Calculate the amount of imbalance present within the
8558 * groups of a given sched_domain during load balance.
8559 * @env: load balance environment
8560 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8562 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8564 unsigned long max_pull, load_above_capacity = ~0UL;
8565 struct sg_lb_stats *local, *busiest;
8567 local = &sds->local_stat;
8568 busiest = &sds->busiest_stat;
8570 if (busiest->group_type == group_imbalanced) {
8572 * In the group_imb case we cannot rely on group-wide averages
8573 * to ensure CPU-load equilibrium, look at wider averages. XXX
8575 busiest->load_per_task =
8576 min(busiest->load_per_task, sds->avg_load);
8580 * Avg load of busiest sg can be less and avg load of local sg can
8581 * be greater than avg load across all sgs of sd because avg load
8582 * factors in sg capacity and sgs with smaller group_type are
8583 * skipped when updating the busiest sg:
8585 if (busiest->group_type != group_misfit_task &&
8586 (busiest->avg_load <= sds->avg_load ||
8587 local->avg_load >= sds->avg_load)) {
8589 return fix_small_imbalance(env, sds);
8593 * If there aren't any idle CPUs, avoid creating some.
8595 if (busiest->group_type == group_overloaded &&
8596 local->group_type == group_overloaded) {
8597 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
8598 if (load_above_capacity > busiest->group_capacity) {
8599 load_above_capacity -= busiest->group_capacity;
8600 load_above_capacity *= scale_load_down(NICE_0_LOAD);
8601 load_above_capacity /= busiest->group_capacity;
8603 load_above_capacity = ~0UL;
8607 * We're trying to get all the CPUs to the average_load, so we don't
8608 * want to push ourselves above the average load, nor do we wish to
8609 * reduce the max loaded CPU below the average load. At the same time,
8610 * we also don't want to reduce the group load below the group
8611 * capacity. Thus we look for the minimum possible imbalance.
8613 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8615 /* How much load to actually move to equalise the imbalance */
8616 env->imbalance = min(
8617 max_pull * busiest->group_capacity,
8618 (sds->avg_load - local->avg_load) * local->group_capacity
8619 ) / SCHED_CAPACITY_SCALE;
8621 /* Boost imbalance to allow misfit task to be balanced. */
8622 if (busiest->group_type == group_misfit_task) {
8623 env->imbalance = max_t(long, env->imbalance,
8624 busiest->group_misfit_task_load);
8628 * if *imbalance is less than the average load per runnable task
8629 * there is no guarantee that any tasks will be moved so we'll have
8630 * a think about bumping its value to force at least one task to be
8633 if (env->imbalance < busiest->load_per_task)
8634 return fix_small_imbalance(env, sds);
8637 /******* find_busiest_group() helpers end here *********************/
8640 * find_busiest_group - Returns the busiest group within the sched_domain
8641 * if there is an imbalance.
8643 * Also calculates the amount of weighted load which should be moved
8644 * to restore balance.
8646 * @env: The load balancing environment.
8648 * Return: - The busiest group if imbalance exists.
8650 static struct sched_group *find_busiest_group(struct lb_env *env)
8652 struct sg_lb_stats *local, *busiest;
8653 struct sd_lb_stats sds;
8655 init_sd_lb_stats(&sds);
8658 * Compute the various statistics relavent for load balancing at
8661 update_sd_lb_stats(env, &sds);
8663 if (static_branch_unlikely(&sched_energy_present)) {
8664 struct root_domain *rd = env->dst_rq->rd;
8666 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
8670 local = &sds.local_stat;
8671 busiest = &sds.busiest_stat;
8673 /* ASYM feature bypasses nice load balance check */
8674 if (check_asym_packing(env, &sds))
8677 /* There is no busy sibling group to pull tasks from */
8678 if (!sds.busiest || busiest->sum_nr_running == 0)
8681 /* XXX broken for overlapping NUMA groups */
8682 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8683 / sds.total_capacity;
8686 * If the busiest group is imbalanced the below checks don't
8687 * work because they assume all things are equal, which typically
8688 * isn't true due to cpus_allowed constraints and the like.
8690 if (busiest->group_type == group_imbalanced)
8694 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8695 * capacities from resulting in underutilization due to avg_load.
8697 if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8698 busiest->group_no_capacity)
8701 /* Misfit tasks should be dealt with regardless of the avg load */
8702 if (busiest->group_type == group_misfit_task)
8706 * If the local group is busier than the selected busiest group
8707 * don't try and pull any tasks.
8709 if (local->avg_load >= busiest->avg_load)
8713 * Don't pull any tasks if this group is already above the domain
8716 if (local->avg_load >= sds.avg_load)
8719 if (env->idle == CPU_IDLE) {
8721 * This CPU is idle. If the busiest group is not overloaded
8722 * and there is no imbalance between this and busiest group
8723 * wrt idle CPUs, it is balanced. The imbalance becomes
8724 * significant if the diff is greater than 1 otherwise we
8725 * might end up to just move the imbalance on another group
8727 if ((busiest->group_type != group_overloaded) &&
8728 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8732 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8733 * imbalance_pct to be conservative.
8735 if (100 * busiest->avg_load <=
8736 env->sd->imbalance_pct * local->avg_load)
8741 /* Looks like there is an imbalance. Compute it */
8742 env->src_grp_type = busiest->group_type;
8743 calculate_imbalance(env, &sds);
8744 return env->imbalance ? sds.busiest : NULL;
8752 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
8754 static struct rq *find_busiest_queue(struct lb_env *env,
8755 struct sched_group *group)
8757 struct rq *busiest = NULL, *rq;
8758 unsigned long busiest_load = 0, busiest_capacity = 1;
8761 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8762 unsigned long capacity, wl;
8766 rt = fbq_classify_rq(rq);
8769 * We classify groups/runqueues into three groups:
8770 * - regular: there are !numa tasks
8771 * - remote: there are numa tasks that run on the 'wrong' node
8772 * - all: there is no distinction
8774 * In order to avoid migrating ideally placed numa tasks,
8775 * ignore those when there's better options.
8777 * If we ignore the actual busiest queue to migrate another
8778 * task, the next balance pass can still reduce the busiest
8779 * queue by moving tasks around inside the node.
8781 * If we cannot move enough load due to this classification
8782 * the next pass will adjust the group classification and
8783 * allow migration of more tasks.
8785 * Both cases only affect the total convergence complexity.
8787 if (rt > env->fbq_type)
8791 * For ASYM_CPUCAPACITY domains with misfit tasks we simply
8792 * seek the "biggest" misfit task.
8794 if (env->src_grp_type == group_misfit_task) {
8795 if (rq->misfit_task_load > busiest_load) {
8796 busiest_load = rq->misfit_task_load;
8803 capacity = capacity_of(i);
8806 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
8807 * eventually lead to active_balancing high->low capacity.
8808 * Higher per-CPU capacity is considered better than balancing
8811 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8812 capacity_of(env->dst_cpu) < capacity &&
8813 rq->nr_running == 1)
8816 wl = weighted_cpuload(rq);
8819 * When comparing with imbalance, use weighted_cpuload()
8820 * which is not scaled with the CPU capacity.
8823 if (rq->nr_running == 1 && wl > env->imbalance &&
8824 !check_cpu_capacity(rq, env->sd))
8828 * For the load comparisons with the other CPU's, consider
8829 * the weighted_cpuload() scaled with the CPU capacity, so
8830 * that the load can be moved away from the CPU that is
8831 * potentially running at a lower capacity.
8833 * Thus we're looking for max(wl_i / capacity_i), crosswise
8834 * multiplication to rid ourselves of the division works out
8835 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8836 * our previous maximum.
8838 if (wl * busiest_capacity > busiest_load * capacity) {
8840 busiest_capacity = capacity;
8849 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8850 * so long as it is large enough.
8852 #define MAX_PINNED_INTERVAL 512
8854 static int need_active_balance(struct lb_env *env)
8856 struct sched_domain *sd = env->sd;
8858 if (env->idle == CPU_NEWLY_IDLE) {
8861 * ASYM_PACKING needs to force migrate tasks from busy but
8862 * lower priority CPUs in order to pack all tasks in the
8863 * highest priority CPUs.
8865 if ((sd->flags & SD_ASYM_PACKING) &&
8866 sched_asym_prefer(env->dst_cpu, env->src_cpu))
8871 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8872 * It's worth migrating the task if the src_cpu's capacity is reduced
8873 * because of other sched_class or IRQs if more capacity stays
8874 * available on dst_cpu.
8876 if ((env->idle != CPU_NOT_IDLE) &&
8877 (env->src_rq->cfs.h_nr_running == 1)) {
8878 if ((check_cpu_capacity(env->src_rq, sd)) &&
8879 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8883 if (env->src_grp_type == group_misfit_task)
8886 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8889 static int active_load_balance_cpu_stop(void *data);
8891 static int should_we_balance(struct lb_env *env)
8893 struct sched_group *sg = env->sd->groups;
8894 int cpu, balance_cpu = -1;
8897 * Ensure the balancing environment is consistent; can happen
8898 * when the softirq triggers 'during' hotplug.
8900 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8904 * In the newly idle case, we will allow all the CPUs
8905 * to do the newly idle load balance.
8907 if (env->idle == CPU_NEWLY_IDLE)
8910 /* Try to find first idle CPU */
8911 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8919 if (balance_cpu == -1)
8920 balance_cpu = group_balance_cpu(sg);
8923 * First idle CPU or the first CPU(busiest) in this sched group
8924 * is eligible for doing load balancing at this and above domains.
8926 return balance_cpu == env->dst_cpu;
8930 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8931 * tasks if there is an imbalance.
8933 static int load_balance(int this_cpu, struct rq *this_rq,
8934 struct sched_domain *sd, enum cpu_idle_type idle,
8935 int *continue_balancing)
8937 int ld_moved, cur_ld_moved, active_balance = 0;
8938 struct sched_domain *sd_parent = sd->parent;
8939 struct sched_group *group;
8942 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8944 struct lb_env env = {
8946 .dst_cpu = this_cpu,
8948 .dst_grpmask = sched_group_span(sd->groups),
8950 .loop_break = sched_nr_migrate_break,
8953 .tasks = LIST_HEAD_INIT(env.tasks),
8956 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8958 schedstat_inc(sd->lb_count[idle]);
8961 if (!should_we_balance(&env)) {
8962 *continue_balancing = 0;
8966 group = find_busiest_group(&env);
8968 schedstat_inc(sd->lb_nobusyg[idle]);
8972 busiest = find_busiest_queue(&env, group);
8974 schedstat_inc(sd->lb_nobusyq[idle]);
8978 BUG_ON(busiest == env.dst_rq);
8980 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8982 env.src_cpu = busiest->cpu;
8983 env.src_rq = busiest;
8986 if (busiest->nr_running > 1) {
8988 * Attempt to move tasks. If find_busiest_group has found
8989 * an imbalance but busiest->nr_running <= 1, the group is
8990 * still unbalanced. ld_moved simply stays zero, so it is
8991 * correctly treated as an imbalance.
8993 env.flags |= LBF_ALL_PINNED;
8994 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8997 rq_lock_irqsave(busiest, &rf);
8998 update_rq_clock(busiest);
9001 * cur_ld_moved - load moved in current iteration
9002 * ld_moved - cumulative load moved across iterations
9004 cur_ld_moved = detach_tasks(&env);
9007 * We've detached some tasks from busiest_rq. Every
9008 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
9009 * unlock busiest->lock, and we are able to be sure
9010 * that nobody can manipulate the tasks in parallel.
9011 * See task_rq_lock() family for the details.
9014 rq_unlock(busiest, &rf);
9018 ld_moved += cur_ld_moved;
9021 local_irq_restore(rf.flags);
9023 if (env.flags & LBF_NEED_BREAK) {
9024 env.flags &= ~LBF_NEED_BREAK;
9029 * Revisit (affine) tasks on src_cpu that couldn't be moved to
9030 * us and move them to an alternate dst_cpu in our sched_group
9031 * where they can run. The upper limit on how many times we
9032 * iterate on same src_cpu is dependent on number of CPUs in our
9035 * This changes load balance semantics a bit on who can move
9036 * load to a given_cpu. In addition to the given_cpu itself
9037 * (or a ilb_cpu acting on its behalf where given_cpu is
9038 * nohz-idle), we now have balance_cpu in a position to move
9039 * load to given_cpu. In rare situations, this may cause
9040 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
9041 * _independently_ and at _same_ time to move some load to
9042 * given_cpu) causing exceess load to be moved to given_cpu.
9043 * This however should not happen so much in practice and
9044 * moreover subsequent load balance cycles should correct the
9045 * excess load moved.
9047 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
9049 /* Prevent to re-select dst_cpu via env's CPUs */
9050 cpumask_clear_cpu(env.dst_cpu, env.cpus);
9052 env.dst_rq = cpu_rq(env.new_dst_cpu);
9053 env.dst_cpu = env.new_dst_cpu;
9054 env.flags &= ~LBF_DST_PINNED;
9056 env.loop_break = sched_nr_migrate_break;
9059 * Go back to "more_balance" rather than "redo" since we
9060 * need to continue with same src_cpu.
9066 * We failed to reach balance because of affinity.
9069 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9071 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
9072 *group_imbalance = 1;
9075 /* All tasks on this runqueue were pinned by CPU affinity */
9076 if (unlikely(env.flags & LBF_ALL_PINNED)) {
9077 cpumask_clear_cpu(cpu_of(busiest), cpus);
9079 * Attempting to continue load balancing at the current
9080 * sched_domain level only makes sense if there are
9081 * active CPUs remaining as possible busiest CPUs to
9082 * pull load from which are not contained within the
9083 * destination group that is receiving any migrated
9086 if (!cpumask_subset(cpus, env.dst_grpmask)) {
9088 env.loop_break = sched_nr_migrate_break;
9091 goto out_all_pinned;
9096 schedstat_inc(sd->lb_failed[idle]);
9098 * Increment the failure counter only on periodic balance.
9099 * We do not want newidle balance, which can be very
9100 * frequent, pollute the failure counter causing
9101 * excessive cache_hot migrations and active balances.
9103 if (idle != CPU_NEWLY_IDLE)
9104 sd->nr_balance_failed++;
9106 if (need_active_balance(&env)) {
9107 unsigned long flags;
9109 raw_spin_lock_irqsave(&busiest->lock, flags);
9112 * Don't kick the active_load_balance_cpu_stop,
9113 * if the curr task on busiest CPU can't be
9114 * moved to this_cpu:
9116 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
9117 raw_spin_unlock_irqrestore(&busiest->lock,
9119 env.flags |= LBF_ALL_PINNED;
9120 goto out_one_pinned;
9124 * ->active_balance synchronizes accesses to
9125 * ->active_balance_work. Once set, it's cleared
9126 * only after active load balance is finished.
9128 if (!busiest->active_balance) {
9129 busiest->active_balance = 1;
9130 busiest->push_cpu = this_cpu;
9133 raw_spin_unlock_irqrestore(&busiest->lock, flags);
9135 if (active_balance) {
9136 stop_one_cpu_nowait(cpu_of(busiest),
9137 active_load_balance_cpu_stop, busiest,
9138 &busiest->active_balance_work);
9141 /* We've kicked active balancing, force task migration. */
9142 sd->nr_balance_failed = sd->cache_nice_tries+1;
9145 sd->nr_balance_failed = 0;
9147 if (likely(!active_balance)) {
9148 /* We were unbalanced, so reset the balancing interval */
9149 sd->balance_interval = sd->min_interval;
9152 * If we've begun active balancing, start to back off. This
9153 * case may not be covered by the all_pinned logic if there
9154 * is only 1 task on the busy runqueue (because we don't call
9157 if (sd->balance_interval < sd->max_interval)
9158 sd->balance_interval *= 2;
9165 * We reach balance although we may have faced some affinity
9166 * constraints. Clear the imbalance flag if it was set.
9169 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9171 if (*group_imbalance)
9172 *group_imbalance = 0;
9177 * We reach balance because all tasks are pinned at this level so
9178 * we can't migrate them. Let the imbalance flag set so parent level
9179 * can try to migrate them.
9181 schedstat_inc(sd->lb_balanced[idle]);
9183 sd->nr_balance_failed = 0;
9189 * idle_balance() disregards balance intervals, so we could repeatedly
9190 * reach this code, which would lead to balance_interval skyrocketting
9191 * in a short amount of time. Skip the balance_interval increase logic
9194 if (env.idle == CPU_NEWLY_IDLE)
9197 /* tune up the balancing interval */
9198 if ((env.flags & LBF_ALL_PINNED &&
9199 sd->balance_interval < MAX_PINNED_INTERVAL) ||
9200 sd->balance_interval < sd->max_interval)
9201 sd->balance_interval *= 2;
9206 static inline unsigned long
9207 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
9209 unsigned long interval = sd->balance_interval;
9212 interval *= sd->busy_factor;
9214 /* scale ms to jiffies */
9215 interval = msecs_to_jiffies(interval);
9216 interval = clamp(interval, 1UL, max_load_balance_interval);
9222 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
9224 unsigned long interval, next;
9226 /* used by idle balance, so cpu_busy = 0 */
9227 interval = get_sd_balance_interval(sd, 0);
9228 next = sd->last_balance + interval;
9230 if (time_after(*next_balance, next))
9231 *next_balance = next;
9235 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9236 * running tasks off the busiest CPU onto idle CPUs. It requires at
9237 * least 1 task to be running on each physical CPU where possible, and
9238 * avoids physical / logical imbalances.
9240 static int active_load_balance_cpu_stop(void *data)
9242 struct rq *busiest_rq = data;
9243 int busiest_cpu = cpu_of(busiest_rq);
9244 int target_cpu = busiest_rq->push_cpu;
9245 struct rq *target_rq = cpu_rq(target_cpu);
9246 struct sched_domain *sd;
9247 struct task_struct *p = NULL;
9250 rq_lock_irq(busiest_rq, &rf);
9252 * Between queueing the stop-work and running it is a hole in which
9253 * CPUs can become inactive. We should not move tasks from or to
9256 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
9259 /* Make sure the requested CPU hasn't gone down in the meantime: */
9260 if (unlikely(busiest_cpu != smp_processor_id() ||
9261 !busiest_rq->active_balance))
9264 /* Is there any task to move? */
9265 if (busiest_rq->nr_running <= 1)
9269 * This condition is "impossible", if it occurs
9270 * we need to fix it. Originally reported by
9271 * Bjorn Helgaas on a 128-CPU setup.
9273 BUG_ON(busiest_rq == target_rq);
9275 /* Search for an sd spanning us and the target CPU. */
9277 for_each_domain(target_cpu, sd) {
9278 if ((sd->flags & SD_LOAD_BALANCE) &&
9279 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9284 struct lb_env env = {
9286 .dst_cpu = target_cpu,
9287 .dst_rq = target_rq,
9288 .src_cpu = busiest_rq->cpu,
9289 .src_rq = busiest_rq,
9292 * can_migrate_task() doesn't need to compute new_dst_cpu
9293 * for active balancing. Since we have CPU_IDLE, but no
9294 * @dst_grpmask we need to make that test go away with lying
9297 .flags = LBF_DST_PINNED,
9300 schedstat_inc(sd->alb_count);
9301 update_rq_clock(busiest_rq);
9303 p = detach_one_task(&env);
9305 schedstat_inc(sd->alb_pushed);
9306 /* Active balancing done, reset the failure counter. */
9307 sd->nr_balance_failed = 0;
9309 schedstat_inc(sd->alb_failed);
9314 busiest_rq->active_balance = 0;
9315 rq_unlock(busiest_rq, &rf);
9318 attach_one_task(target_rq, p);
9325 static DEFINE_SPINLOCK(balancing);
9328 * Scale the max load_balance interval with the number of CPUs in the system.
9329 * This trades load-balance latency on larger machines for less cross talk.
9331 void update_max_interval(void)
9333 max_load_balance_interval = HZ*num_online_cpus()/10;
9337 * It checks each scheduling domain to see if it is due to be balanced,
9338 * and initiates a balancing operation if so.
9340 * Balancing parameters are set up in init_sched_domains.
9342 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9344 int continue_balancing = 1;
9346 unsigned long interval;
9347 struct sched_domain *sd;
9348 /* Earliest time when we have to do rebalance again */
9349 unsigned long next_balance = jiffies + 60*HZ;
9350 int update_next_balance = 0;
9351 int need_serialize, need_decay = 0;
9355 for_each_domain(cpu, sd) {
9357 * Decay the newidle max times here because this is a regular
9358 * visit to all the domains. Decay ~1% per second.
9360 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9361 sd->max_newidle_lb_cost =
9362 (sd->max_newidle_lb_cost * 253) / 256;
9363 sd->next_decay_max_lb_cost = jiffies + HZ;
9366 max_cost += sd->max_newidle_lb_cost;
9368 if (!(sd->flags & SD_LOAD_BALANCE))
9372 * Stop the load balance at this level. There is another
9373 * CPU in our sched group which is doing load balancing more
9376 if (!continue_balancing) {
9382 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9384 need_serialize = sd->flags & SD_SERIALIZE;
9385 if (need_serialize) {
9386 if (!spin_trylock(&balancing))
9390 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9391 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9393 * The LBF_DST_PINNED logic could have changed
9394 * env->dst_cpu, so we can't know our idle
9395 * state even if we migrated tasks. Update it.
9397 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9399 sd->last_balance = jiffies;
9400 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9403 spin_unlock(&balancing);
9405 if (time_after(next_balance, sd->last_balance + interval)) {
9406 next_balance = sd->last_balance + interval;
9407 update_next_balance = 1;
9412 * Ensure the rq-wide value also decays but keep it at a
9413 * reasonable floor to avoid funnies with rq->avg_idle.
9415 rq->max_idle_balance_cost =
9416 max((u64)sysctl_sched_migration_cost, max_cost);
9421 * next_balance will be updated only when there is a need.
9422 * When the cpu is attached to null domain for ex, it will not be
9425 if (likely(update_next_balance)) {
9426 rq->next_balance = next_balance;
9428 #ifdef CONFIG_NO_HZ_COMMON
9430 * If this CPU has been elected to perform the nohz idle
9431 * balance. Other idle CPUs have already rebalanced with
9432 * nohz_idle_balance() and nohz.next_balance has been
9433 * updated accordingly. This CPU is now running the idle load
9434 * balance for itself and we need to update the
9435 * nohz.next_balance accordingly.
9437 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9438 nohz.next_balance = rq->next_balance;
9443 static inline int on_null_domain(struct rq *rq)
9445 return unlikely(!rcu_dereference_sched(rq->sd));
9448 #ifdef CONFIG_NO_HZ_COMMON
9450 * idle load balancing details
9451 * - When one of the busy CPUs notice that there may be an idle rebalancing
9452 * needed, they will kick the idle load balancer, which then does idle
9453 * load balancing for all the idle CPUs.
9456 static inline int find_new_ilb(void)
9458 int ilb = cpumask_first(nohz.idle_cpus_mask);
9460 if (ilb < nr_cpu_ids && idle_cpu(ilb))
9467 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
9468 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
9469 * CPU (if there is one).
9471 static void kick_ilb(unsigned int flags)
9475 nohz.next_balance++;
9477 ilb_cpu = find_new_ilb();
9479 if (ilb_cpu >= nr_cpu_ids)
9482 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
9483 if (flags & NOHZ_KICK_MASK)
9487 * Use smp_send_reschedule() instead of resched_cpu().
9488 * This way we generate a sched IPI on the target CPU which
9489 * is idle. And the softirq performing nohz idle load balance
9490 * will be run before returning from the IPI.
9492 smp_send_reschedule(ilb_cpu);
9496 * Current heuristic for kicking the idle load balancer in the presence
9497 * of an idle cpu in the system.
9498 * - This rq has more than one task.
9499 * - This rq has at least one CFS task and the capacity of the CPU is
9500 * significantly reduced because of RT tasks or IRQs.
9501 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9502 * multiple busy cpu.
9503 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9504 * domain span are idle.
9506 static void nohz_balancer_kick(struct rq *rq)
9508 unsigned long now = jiffies;
9509 struct sched_domain_shared *sds;
9510 struct sched_domain *sd;
9511 int nr_busy, i, cpu = rq->cpu;
9512 unsigned int flags = 0;
9514 if (unlikely(rq->idle_balance))
9518 * We may be recently in ticked or tickless idle mode. At the first
9519 * busy tick after returning from idle, we will update the busy stats.
9521 nohz_balance_exit_idle(rq);
9524 * None are in tickless mode and hence no need for NOHZ idle load
9527 if (likely(!atomic_read(&nohz.nr_cpus)))
9530 if (READ_ONCE(nohz.has_blocked) &&
9531 time_after(now, READ_ONCE(nohz.next_blocked)))
9532 flags = NOHZ_STATS_KICK;
9534 if (time_before(now, nohz.next_balance))
9537 if (rq->nr_running >= 2 || rq->misfit_task_load) {
9538 flags = NOHZ_KICK_MASK;
9543 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9546 * XXX: write a coherent comment on why we do this.
9547 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
9549 nr_busy = atomic_read(&sds->nr_busy_cpus);
9551 flags = NOHZ_KICK_MASK;
9557 sd = rcu_dereference(rq->sd);
9559 if ((rq->cfs.h_nr_running >= 1) &&
9560 check_cpu_capacity(rq, sd)) {
9561 flags = NOHZ_KICK_MASK;
9566 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
9568 for_each_cpu(i, sched_domain_span(sd)) {
9570 !cpumask_test_cpu(i, nohz.idle_cpus_mask))
9573 if (sched_asym_prefer(i, cpu)) {
9574 flags = NOHZ_KICK_MASK;
9586 static void set_cpu_sd_state_busy(int cpu)
9588 struct sched_domain *sd;
9591 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9593 if (!sd || !sd->nohz_idle)
9597 atomic_inc(&sd->shared->nr_busy_cpus);
9602 void nohz_balance_exit_idle(struct rq *rq)
9604 SCHED_WARN_ON(rq != this_rq());
9606 if (likely(!rq->nohz_tick_stopped))
9609 rq->nohz_tick_stopped = 0;
9610 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
9611 atomic_dec(&nohz.nr_cpus);
9613 set_cpu_sd_state_busy(rq->cpu);
9616 static void set_cpu_sd_state_idle(int cpu)
9618 struct sched_domain *sd;
9621 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9623 if (!sd || sd->nohz_idle)
9627 atomic_dec(&sd->shared->nr_busy_cpus);
9633 * This routine will record that the CPU is going idle with tick stopped.
9634 * This info will be used in performing idle load balancing in the future.
9636 void nohz_balance_enter_idle(int cpu)
9638 struct rq *rq = cpu_rq(cpu);
9640 SCHED_WARN_ON(cpu != smp_processor_id());
9642 /* If this CPU is going down, then nothing needs to be done: */
9643 if (!cpu_active(cpu))
9646 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9647 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9651 * Can be set safely without rq->lock held
9652 * If a clear happens, it will have evaluated last additions because
9653 * rq->lock is held during the check and the clear
9655 rq->has_blocked_load = 1;
9658 * The tick is still stopped but load could have been added in the
9659 * meantime. We set the nohz.has_blocked flag to trig a check of the
9660 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
9661 * of nohz.has_blocked can only happen after checking the new load
9663 if (rq->nohz_tick_stopped)
9666 /* If we're a completely isolated CPU, we don't play: */
9667 if (on_null_domain(rq))
9670 rq->nohz_tick_stopped = 1;
9672 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9673 atomic_inc(&nohz.nr_cpus);
9676 * Ensures that if nohz_idle_balance() fails to observe our
9677 * @idle_cpus_mask store, it must observe the @has_blocked
9680 smp_mb__after_atomic();
9682 set_cpu_sd_state_idle(cpu);
9686 * Each time a cpu enter idle, we assume that it has blocked load and
9687 * enable the periodic update of the load of idle cpus
9689 WRITE_ONCE(nohz.has_blocked, 1);
9693 * Internal function that runs load balance for all idle cpus. The load balance
9694 * can be a simple update of blocked load or a complete load balance with
9695 * tasks movement depending of flags.
9696 * The function returns false if the loop has stopped before running
9697 * through all idle CPUs.
9699 static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
9700 enum cpu_idle_type idle)
9702 /* Earliest time when we have to do rebalance again */
9703 unsigned long now = jiffies;
9704 unsigned long next_balance = now + 60*HZ;
9705 bool has_blocked_load = false;
9706 int update_next_balance = 0;
9707 int this_cpu = this_rq->cpu;
9712 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
9715 * We assume there will be no idle load after this update and clear
9716 * the has_blocked flag. If a cpu enters idle in the mean time, it will
9717 * set the has_blocked flag and trig another update of idle load.
9718 * Because a cpu that becomes idle, is added to idle_cpus_mask before
9719 * setting the flag, we are sure to not clear the state and not
9720 * check the load of an idle cpu.
9722 WRITE_ONCE(nohz.has_blocked, 0);
9725 * Ensures that if we miss the CPU, we must see the has_blocked
9726 * store from nohz_balance_enter_idle().
9730 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9731 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9735 * If this CPU gets work to do, stop the load balancing
9736 * work being done for other CPUs. Next load
9737 * balancing owner will pick it up.
9739 if (need_resched()) {
9740 has_blocked_load = true;
9744 rq = cpu_rq(balance_cpu);
9746 has_blocked_load |= update_nohz_stats(rq, true);
9749 * If time for next balance is due,
9752 if (time_after_eq(jiffies, rq->next_balance)) {
9755 rq_lock_irqsave(rq, &rf);
9756 update_rq_clock(rq);
9757 cpu_load_update_idle(rq);
9758 rq_unlock_irqrestore(rq, &rf);
9760 if (flags & NOHZ_BALANCE_KICK)
9761 rebalance_domains(rq, CPU_IDLE);
9764 if (time_after(next_balance, rq->next_balance)) {
9765 next_balance = rq->next_balance;
9766 update_next_balance = 1;
9770 /* Newly idle CPU doesn't need an update */
9771 if (idle != CPU_NEWLY_IDLE) {
9772 update_blocked_averages(this_cpu);
9773 has_blocked_load |= this_rq->has_blocked_load;
9776 if (flags & NOHZ_BALANCE_KICK)
9777 rebalance_domains(this_rq, CPU_IDLE);
9779 WRITE_ONCE(nohz.next_blocked,
9780 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
9782 /* The full idle balance loop has been done */
9786 /* There is still blocked load, enable periodic update */
9787 if (has_blocked_load)
9788 WRITE_ONCE(nohz.has_blocked, 1);
9791 * next_balance will be updated only when there is a need.
9792 * When the CPU is attached to null domain for ex, it will not be
9795 if (likely(update_next_balance))
9796 nohz.next_balance = next_balance;
9802 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9803 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9805 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9807 int this_cpu = this_rq->cpu;
9810 if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
9813 if (idle != CPU_IDLE) {
9814 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9818 /* could be _relaxed() */
9819 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9820 if (!(flags & NOHZ_KICK_MASK))
9823 _nohz_idle_balance(this_rq, flags, idle);
9828 static void nohz_newidle_balance(struct rq *this_rq)
9830 int this_cpu = this_rq->cpu;
9833 * This CPU doesn't want to be disturbed by scheduler
9836 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
9839 /* Will wake up very soon. No time for doing anything else*/
9840 if (this_rq->avg_idle < sysctl_sched_migration_cost)
9843 /* Don't need to update blocked load of idle CPUs*/
9844 if (!READ_ONCE(nohz.has_blocked) ||
9845 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
9848 raw_spin_unlock(&this_rq->lock);
9850 * This CPU is going to be idle and blocked load of idle CPUs
9851 * need to be updated. Run the ilb locally as it is a good
9852 * candidate for ilb instead of waking up another idle CPU.
9853 * Kick an normal ilb if we failed to do the update.
9855 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
9856 kick_ilb(NOHZ_STATS_KICK);
9857 raw_spin_lock(&this_rq->lock);
9860 #else /* !CONFIG_NO_HZ_COMMON */
9861 static inline void nohz_balancer_kick(struct rq *rq) { }
9863 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9868 static inline void nohz_newidle_balance(struct rq *this_rq) { }
9869 #endif /* CONFIG_NO_HZ_COMMON */
9872 * idle_balance is called by schedule() if this_cpu is about to become
9873 * idle. Attempts to pull tasks from other CPUs.
9875 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
9877 unsigned long next_balance = jiffies + HZ;
9878 int this_cpu = this_rq->cpu;
9879 struct sched_domain *sd;
9880 int pulled_task = 0;
9884 * We must set idle_stamp _before_ calling idle_balance(), such that we
9885 * measure the duration of idle_balance() as idle time.
9887 this_rq->idle_stamp = rq_clock(this_rq);
9890 * Do not pull tasks towards !active CPUs...
9892 if (!cpu_active(this_cpu))
9896 * This is OK, because current is on_cpu, which avoids it being picked
9897 * for load-balance and preemption/IRQs are still disabled avoiding
9898 * further scheduler activity on it and we're being very careful to
9899 * re-start the picking loop.
9901 rq_unpin_lock(this_rq, rf);
9903 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
9904 !READ_ONCE(this_rq->rd->overload)) {
9907 sd = rcu_dereference_check_sched_domain(this_rq->sd);
9909 update_next_balance(sd, &next_balance);
9912 nohz_newidle_balance(this_rq);
9917 raw_spin_unlock(&this_rq->lock);
9919 update_blocked_averages(this_cpu);
9921 for_each_domain(this_cpu, sd) {
9922 int continue_balancing = 1;
9923 u64 t0, domain_cost;
9925 if (!(sd->flags & SD_LOAD_BALANCE))
9928 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
9929 update_next_balance(sd, &next_balance);
9933 if (sd->flags & SD_BALANCE_NEWIDLE) {
9934 t0 = sched_clock_cpu(this_cpu);
9936 pulled_task = load_balance(this_cpu, this_rq,
9938 &continue_balancing);
9940 domain_cost = sched_clock_cpu(this_cpu) - t0;
9941 if (domain_cost > sd->max_newidle_lb_cost)
9942 sd->max_newidle_lb_cost = domain_cost;
9944 curr_cost += domain_cost;
9947 update_next_balance(sd, &next_balance);
9950 * Stop searching for tasks to pull if there are
9951 * now runnable tasks on this rq.
9953 if (pulled_task || this_rq->nr_running > 0)
9958 raw_spin_lock(&this_rq->lock);
9960 if (curr_cost > this_rq->max_idle_balance_cost)
9961 this_rq->max_idle_balance_cost = curr_cost;
9965 * While browsing the domains, we released the rq lock, a task could
9966 * have been enqueued in the meantime. Since we're not going idle,
9967 * pretend we pulled a task.
9969 if (this_rq->cfs.h_nr_running && !pulled_task)
9972 /* Move the next balance forward */
9973 if (time_after(this_rq->next_balance, next_balance))
9974 this_rq->next_balance = next_balance;
9976 /* Is there a task of a high priority class? */
9977 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
9981 this_rq->idle_stamp = 0;
9983 rq_repin_lock(this_rq, rf);
9989 * run_rebalance_domains is triggered when needed from the scheduler tick.
9990 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9992 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9994 struct rq *this_rq = this_rq();
9995 enum cpu_idle_type idle = this_rq->idle_balance ?
9996 CPU_IDLE : CPU_NOT_IDLE;
9999 * If this CPU has a pending nohz_balance_kick, then do the
10000 * balancing on behalf of the other idle CPUs whose ticks are
10001 * stopped. Do nohz_idle_balance *before* rebalance_domains to
10002 * give the idle CPUs a chance to load balance. Else we may
10003 * load balance only within the local sched_domain hierarchy
10004 * and abort nohz_idle_balance altogether if we pull some load.
10006 if (nohz_idle_balance(this_rq, idle))
10009 /* normal load balance */
10010 update_blocked_averages(this_rq->cpu);
10011 rebalance_domains(this_rq, idle);
10015 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
10017 void trigger_load_balance(struct rq *rq)
10019 /* Don't need to rebalance while attached to NULL domain */
10020 if (unlikely(on_null_domain(rq)))
10023 if (time_after_eq(jiffies, rq->next_balance))
10024 raise_softirq(SCHED_SOFTIRQ);
10026 nohz_balancer_kick(rq);
10029 static void rq_online_fair(struct rq *rq)
10033 update_runtime_enabled(rq);
10036 static void rq_offline_fair(struct rq *rq)
10040 /* Ensure any throttled groups are reachable by pick_next_task */
10041 unthrottle_offline_cfs_rqs(rq);
10044 #endif /* CONFIG_SMP */
10047 * scheduler tick hitting a task of our scheduling class.
10049 * NOTE: This function can be called remotely by the tick offload that
10050 * goes along full dynticks. Therefore no local assumption can be made
10051 * and everything must be accessed through the @rq and @curr passed in
10054 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
10056 struct cfs_rq *cfs_rq;
10057 struct sched_entity *se = &curr->se;
10059 for_each_sched_entity(se) {
10060 cfs_rq = cfs_rq_of(se);
10061 entity_tick(cfs_rq, se, queued);
10064 if (static_branch_unlikely(&sched_numa_balancing))
10065 task_tick_numa(rq, curr);
10067 update_misfit_status(curr, rq);
10068 update_overutilized_status(task_rq(curr));
10072 * called on fork with the child task as argument from the parent's context
10073 * - child not yet on the tasklist
10074 * - preemption disabled
10076 static void task_fork_fair(struct task_struct *p)
10078 struct cfs_rq *cfs_rq;
10079 struct sched_entity *se = &p->se, *curr;
10080 struct rq *rq = this_rq();
10081 struct rq_flags rf;
10084 update_rq_clock(rq);
10086 cfs_rq = task_cfs_rq(current);
10087 curr = cfs_rq->curr;
10089 update_curr(cfs_rq);
10090 se->vruntime = curr->vruntime;
10092 place_entity(cfs_rq, se, 1);
10094 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
10096 * Upon rescheduling, sched_class::put_prev_task() will place
10097 * 'current' within the tree based on its new key value.
10099 swap(curr->vruntime, se->vruntime);
10103 se->vruntime -= cfs_rq->min_vruntime;
10104 rq_unlock(rq, &rf);
10108 * Priority of the task has changed. Check to see if we preempt
10109 * the current task.
10112 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
10114 if (!task_on_rq_queued(p))
10118 * Reschedule if we are currently running on this runqueue and
10119 * our priority decreased, or if we are not currently running on
10120 * this runqueue and our priority is higher than the current's
10122 if (rq->curr == p) {
10123 if (p->prio > oldprio)
10126 check_preempt_curr(rq, p, 0);
10129 static inline bool vruntime_normalized(struct task_struct *p)
10131 struct sched_entity *se = &p->se;
10134 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
10135 * the dequeue_entity(.flags=0) will already have normalized the
10142 * When !on_rq, vruntime of the task has usually NOT been normalized.
10143 * But there are some cases where it has already been normalized:
10145 * - A forked child which is waiting for being woken up by
10146 * wake_up_new_task().
10147 * - A task which has been woken up by try_to_wake_up() and
10148 * waiting for actually being woken up by sched_ttwu_pending().
10150 if (!se->sum_exec_runtime ||
10151 (p->state == TASK_WAKING && p->sched_remote_wakeup))
10157 #ifdef CONFIG_FAIR_GROUP_SCHED
10159 * Propagate the changes of the sched_entity across the tg tree to make it
10160 * visible to the root
10162 static void propagate_entity_cfs_rq(struct sched_entity *se)
10164 struct cfs_rq *cfs_rq;
10166 /* Start to propagate at parent */
10169 for_each_sched_entity(se) {
10170 cfs_rq = cfs_rq_of(se);
10172 if (cfs_rq_throttled(cfs_rq))
10175 update_load_avg(cfs_rq, se, UPDATE_TG);
10179 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
10182 static void detach_entity_cfs_rq(struct sched_entity *se)
10184 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10186 /* Catch up with the cfs_rq and remove our load when we leave */
10187 update_load_avg(cfs_rq, se, 0);
10188 detach_entity_load_avg(cfs_rq, se);
10189 update_tg_load_avg(cfs_rq, false);
10190 propagate_entity_cfs_rq(se);
10193 static void attach_entity_cfs_rq(struct sched_entity *se)
10195 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10197 #ifdef CONFIG_FAIR_GROUP_SCHED
10199 * Since the real-depth could have been changed (only FAIR
10200 * class maintain depth value), reset depth properly.
10202 se->depth = se->parent ? se->parent->depth + 1 : 0;
10205 /* Synchronize entity with its cfs_rq */
10206 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
10207 attach_entity_load_avg(cfs_rq, se, 0);
10208 update_tg_load_avg(cfs_rq, false);
10209 propagate_entity_cfs_rq(se);
10212 static void detach_task_cfs_rq(struct task_struct *p)
10214 struct sched_entity *se = &p->se;
10215 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10217 if (!vruntime_normalized(p)) {
10219 * Fix up our vruntime so that the current sleep doesn't
10220 * cause 'unlimited' sleep bonus.
10222 place_entity(cfs_rq, se, 0);
10223 se->vruntime -= cfs_rq->min_vruntime;
10226 detach_entity_cfs_rq(se);
10229 static void attach_task_cfs_rq(struct task_struct *p)
10231 struct sched_entity *se = &p->se;
10232 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10234 attach_entity_cfs_rq(se);
10236 if (!vruntime_normalized(p))
10237 se->vruntime += cfs_rq->min_vruntime;
10240 static void switched_from_fair(struct rq *rq, struct task_struct *p)
10242 detach_task_cfs_rq(p);
10245 static void switched_to_fair(struct rq *rq, struct task_struct *p)
10247 attach_task_cfs_rq(p);
10249 if (task_on_rq_queued(p)) {
10251 * We were most likely switched from sched_rt, so
10252 * kick off the schedule if running, otherwise just see
10253 * if we can still preempt the current task.
10258 check_preempt_curr(rq, p, 0);
10262 /* Account for a task changing its policy or group.
10264 * This routine is mostly called to set cfs_rq->curr field when a task
10265 * migrates between groups/classes.
10267 static void set_curr_task_fair(struct rq *rq)
10269 struct sched_entity *se = &rq->curr->se;
10271 for_each_sched_entity(se) {
10272 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10274 set_next_entity(cfs_rq, se);
10275 /* ensure bandwidth has been allocated on our new cfs_rq */
10276 account_cfs_rq_runtime(cfs_rq, 0);
10280 void init_cfs_rq(struct cfs_rq *cfs_rq)
10282 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
10283 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
10284 #ifndef CONFIG_64BIT
10285 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
10288 raw_spin_lock_init(&cfs_rq->removed.lock);
10292 #ifdef CONFIG_FAIR_GROUP_SCHED
10293 static void task_set_group_fair(struct task_struct *p)
10295 struct sched_entity *se = &p->se;
10297 set_task_rq(p, task_cpu(p));
10298 se->depth = se->parent ? se->parent->depth + 1 : 0;
10301 static void task_move_group_fair(struct task_struct *p)
10303 detach_task_cfs_rq(p);
10304 set_task_rq(p, task_cpu(p));
10307 /* Tell se's cfs_rq has been changed -- migrated */
10308 p->se.avg.last_update_time = 0;
10310 attach_task_cfs_rq(p);
10313 static void task_change_group_fair(struct task_struct *p, int type)
10316 case TASK_SET_GROUP:
10317 task_set_group_fair(p);
10320 case TASK_MOVE_GROUP:
10321 task_move_group_fair(p);
10326 void free_fair_sched_group(struct task_group *tg)
10330 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
10332 for_each_possible_cpu(i) {
10334 kfree(tg->cfs_rq[i]);
10343 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10345 struct sched_entity *se;
10346 struct cfs_rq *cfs_rq;
10349 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
10352 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
10356 tg->shares = NICE_0_LOAD;
10358 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
10360 for_each_possible_cpu(i) {
10361 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
10362 GFP_KERNEL, cpu_to_node(i));
10366 se = kzalloc_node(sizeof(struct sched_entity),
10367 GFP_KERNEL, cpu_to_node(i));
10371 init_cfs_rq(cfs_rq);
10372 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
10373 init_entity_runnable_average(se);
10384 void online_fair_sched_group(struct task_group *tg)
10386 struct sched_entity *se;
10390 for_each_possible_cpu(i) {
10394 raw_spin_lock_irq(&rq->lock);
10395 update_rq_clock(rq);
10396 attach_entity_cfs_rq(se);
10397 sync_throttle(tg, i);
10398 raw_spin_unlock_irq(&rq->lock);
10402 void unregister_fair_sched_group(struct task_group *tg)
10404 unsigned long flags;
10408 for_each_possible_cpu(cpu) {
10410 remove_entity_load_avg(tg->se[cpu]);
10413 * Only empty task groups can be destroyed; so we can speculatively
10414 * check on_list without danger of it being re-added.
10416 if (!tg->cfs_rq[cpu]->on_list)
10421 raw_spin_lock_irqsave(&rq->lock, flags);
10422 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
10423 raw_spin_unlock_irqrestore(&rq->lock, flags);
10427 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
10428 struct sched_entity *se, int cpu,
10429 struct sched_entity *parent)
10431 struct rq *rq = cpu_rq(cpu);
10435 init_cfs_rq_runtime(cfs_rq);
10437 tg->cfs_rq[cpu] = cfs_rq;
10440 /* se could be NULL for root_task_group */
10445 se->cfs_rq = &rq->cfs;
10448 se->cfs_rq = parent->my_q;
10449 se->depth = parent->depth + 1;
10453 /* guarantee group entities always have weight */
10454 update_load_set(&se->load, NICE_0_LOAD);
10455 se->parent = parent;
10458 static DEFINE_MUTEX(shares_mutex);
10460 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10465 * We can't change the weight of the root cgroup.
10470 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10472 mutex_lock(&shares_mutex);
10473 if (tg->shares == shares)
10476 tg->shares = shares;
10477 for_each_possible_cpu(i) {
10478 struct rq *rq = cpu_rq(i);
10479 struct sched_entity *se = tg->se[i];
10480 struct rq_flags rf;
10482 /* Propagate contribution to hierarchy */
10483 rq_lock_irqsave(rq, &rf);
10484 update_rq_clock(rq);
10485 for_each_sched_entity(se) {
10486 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
10487 update_cfs_group(se);
10489 rq_unlock_irqrestore(rq, &rf);
10493 mutex_unlock(&shares_mutex);
10496 #else /* CONFIG_FAIR_GROUP_SCHED */
10498 void free_fair_sched_group(struct task_group *tg) { }
10500 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10505 void online_fair_sched_group(struct task_group *tg) { }
10507 void unregister_fair_sched_group(struct task_group *tg) { }
10509 #endif /* CONFIG_FAIR_GROUP_SCHED */
10512 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10514 struct sched_entity *se = &task->se;
10515 unsigned int rr_interval = 0;
10518 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10521 if (rq->cfs.load.weight)
10522 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10524 return rr_interval;
10528 * All the scheduling class methods:
10530 const struct sched_class fair_sched_class = {
10531 .next = &idle_sched_class,
10532 .enqueue_task = enqueue_task_fair,
10533 .dequeue_task = dequeue_task_fair,
10534 .yield_task = yield_task_fair,
10535 .yield_to_task = yield_to_task_fair,
10537 .check_preempt_curr = check_preempt_wakeup,
10539 .pick_next_task = pick_next_task_fair,
10540 .put_prev_task = put_prev_task_fair,
10543 .select_task_rq = select_task_rq_fair,
10544 .migrate_task_rq = migrate_task_rq_fair,
10546 .rq_online = rq_online_fair,
10547 .rq_offline = rq_offline_fair,
10549 .task_dead = task_dead_fair,
10550 .set_cpus_allowed = set_cpus_allowed_common,
10553 .set_curr_task = set_curr_task_fair,
10554 .task_tick = task_tick_fair,
10555 .task_fork = task_fork_fair,
10557 .prio_changed = prio_changed_fair,
10558 .switched_from = switched_from_fair,
10559 .switched_to = switched_to_fair,
10561 .get_rr_interval = get_rr_interval_fair,
10563 .update_curr = update_curr_fair,
10565 #ifdef CONFIG_FAIR_GROUP_SCHED
10566 .task_change_group = task_change_group_fair,
10570 #ifdef CONFIG_SCHED_DEBUG
10571 void print_cfs_stats(struct seq_file *m, int cpu)
10573 struct cfs_rq *cfs_rq, *pos;
10576 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
10577 print_cfs_rq(m, cpu, cfs_rq);
10581 #ifdef CONFIG_NUMA_BALANCING
10582 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10585 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10587 for_each_online_node(node) {
10588 if (p->numa_faults) {
10589 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10590 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10592 if (p->numa_group) {
10593 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
10594 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
10596 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10599 #endif /* CONFIG_NUMA_BALANCING */
10600 #endif /* CONFIG_SCHED_DEBUG */
10602 __init void init_sched_fair_class(void)
10605 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10607 #ifdef CONFIG_NO_HZ_COMMON
10608 nohz.next_balance = jiffies;
10609 nohz.next_blocked = jiffies;
10610 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);