1 // SPDX-License-Identifier: GPL-2.0
3 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
5 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
7 * Interactivity improvements by Mike Galbraith
8 * (C) 2007 Mike Galbraith <efault@gmx.de>
10 * Various enhancements by Dmitry Adamushko.
11 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
13 * Group scheduling enhancements by Srivatsa Vaddagiri
14 * Copyright IBM Corporation, 2007
15 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
17 * Scaled math optimizations by Thomas Gleixner
18 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
20 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
21 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
25 #include <trace/events/sched.h>
28 * Targeted preemption latency for CPU-bound tasks:
30 * NOTE: this latency value is not the same as the concept of
31 * 'timeslice length' - timeslices in CFS are of variable length
32 * and have no persistent notion like in traditional, time-slice
33 * based scheduling concepts.
35 * (to see the precise effective timeslice length of your workload,
36 * run vmstat and monitor the context-switches (cs) field)
38 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
40 unsigned int sysctl_sched_latency = 6000000ULL;
41 static unsigned int normalized_sysctl_sched_latency = 6000000ULL;
44 * The initial- and re-scaling of tunables is configurable
48 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
49 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
50 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
52 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
54 enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
57 * Minimal preemption granularity for CPU-bound tasks:
59 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
61 unsigned int sysctl_sched_min_granularity = 750000ULL;
62 static unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
65 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
67 static unsigned int sched_nr_latency = 8;
70 * After fork, child runs first. If set to 0 (default) then
71 * parent will (try to) run first.
73 unsigned int sysctl_sched_child_runs_first __read_mostly;
76 * SCHED_OTHER wake-up granularity.
78 * This option delays the preemption effects of decoupled workloads
79 * and reduces their over-scheduling. Synchronous workloads will still
80 * have immediate wakeup/sleep latencies.
82 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
84 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
85 static unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
87 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
91 * For asym packing, by default the lower numbered CPU has higher priority.
93 int __weak arch_asym_cpu_priority(int cpu)
99 * The margin used when comparing utilization with CPU capacity:
100 * util * margin < capacity * 1024
104 static unsigned int capacity_margin = 1280;
107 #ifdef CONFIG_CFS_BANDWIDTH
109 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
110 * each time a cfs_rq requests quota.
112 * Note: in the case that the slice exceeds the runtime remaining (either due
113 * to consumption or the quota being specified to be smaller than the slice)
114 * we will always only issue the remaining available time.
116 * (default: 5 msec, units: microseconds)
118 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
121 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
127 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
133 static inline void update_load_set(struct load_weight *lw, unsigned long w)
140 * Increase the granularity value when there are more CPUs,
141 * because with more CPUs the 'effective latency' as visible
142 * to users decreases. But the relationship is not linear,
143 * so pick a second-best guess by going with the log2 of the
146 * This idea comes from the SD scheduler of Con Kolivas:
148 static unsigned int get_update_sysctl_factor(void)
150 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
153 switch (sysctl_sched_tunable_scaling) {
154 case SCHED_TUNABLESCALING_NONE:
157 case SCHED_TUNABLESCALING_LINEAR:
160 case SCHED_TUNABLESCALING_LOG:
162 factor = 1 + ilog2(cpus);
169 static void update_sysctl(void)
171 unsigned int factor = get_update_sysctl_factor();
173 #define SET_SYSCTL(name) \
174 (sysctl_##name = (factor) * normalized_sysctl_##name)
175 SET_SYSCTL(sched_min_granularity);
176 SET_SYSCTL(sched_latency);
177 SET_SYSCTL(sched_wakeup_granularity);
181 void sched_init_granularity(void)
186 #define WMULT_CONST (~0U)
187 #define WMULT_SHIFT 32
189 static void __update_inv_weight(struct load_weight *lw)
193 if (likely(lw->inv_weight))
196 w = scale_load_down(lw->weight);
198 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
200 else if (unlikely(!w))
201 lw->inv_weight = WMULT_CONST;
203 lw->inv_weight = WMULT_CONST / w;
207 * delta_exec * weight / lw.weight
209 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
211 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
212 * we're guaranteed shift stays positive because inv_weight is guaranteed to
213 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
215 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
216 * weight/lw.weight <= 1, and therefore our shift will also be positive.
218 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
220 u64 fact = scale_load_down(weight);
221 int shift = WMULT_SHIFT;
223 __update_inv_weight(lw);
225 if (unlikely(fact >> 32)) {
232 /* hint to use a 32x32->64 mul */
233 fact = (u64)(u32)fact * lw->inv_weight;
240 return mul_u64_u32_shr(delta_exec, fact, shift);
244 const struct sched_class fair_sched_class;
246 /**************************************************************
247 * CFS operations on generic schedulable entities:
250 #ifdef CONFIG_FAIR_GROUP_SCHED
251 static inline struct task_struct *task_of(struct sched_entity *se)
253 SCHED_WARN_ON(!entity_is_task(se));
254 return container_of(se, struct task_struct, se);
257 /* Walk up scheduling entities hierarchy */
258 #define for_each_sched_entity(se) \
259 for (; se; se = se->parent)
261 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
266 /* runqueue on which this entity is (to be) queued */
267 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
272 /* runqueue "owned" by this group */
273 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
278 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
283 if (cfs_rq && task_group_is_autogroup(cfs_rq->tg))
284 autogroup_path(cfs_rq->tg, path, len);
285 else if (cfs_rq && cfs_rq->tg->css.cgroup)
286 cgroup_path(cfs_rq->tg->css.cgroup, path, len);
288 strlcpy(path, "(null)", len);
291 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
293 struct rq *rq = rq_of(cfs_rq);
294 int cpu = cpu_of(rq);
297 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
302 * Ensure we either appear before our parent (if already
303 * enqueued) or force our parent to appear after us when it is
304 * enqueued. The fact that we always enqueue bottom-up
305 * reduces this to two cases and a special case for the root
306 * cfs_rq. Furthermore, it also means that we will always reset
307 * tmp_alone_branch either when the branch is connected
308 * to a tree or when we reach the top of the tree
310 if (cfs_rq->tg->parent &&
311 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
313 * If parent is already on the list, we add the child
314 * just before. Thanks to circular linked property of
315 * the list, this means to put the child at the tail
316 * of the list that starts by parent.
318 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
319 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
321 * The branch is now connected to its tree so we can
322 * reset tmp_alone_branch to the beginning of the
325 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
329 if (!cfs_rq->tg->parent) {
331 * cfs rq without parent should be put
332 * at the tail of the list.
334 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
335 &rq->leaf_cfs_rq_list);
337 * We have reach the top of a tree so we can reset
338 * tmp_alone_branch to the beginning of the list.
340 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
345 * The parent has not already been added so we want to
346 * make sure that it will be put after us.
347 * tmp_alone_branch points to the begin of the branch
348 * where we will add parent.
350 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
352 * update tmp_alone_branch to points to the new begin
355 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
359 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
361 if (cfs_rq->on_list) {
362 struct rq *rq = rq_of(cfs_rq);
365 * With cfs_rq being unthrottled/throttled during an enqueue,
366 * it can happen the tmp_alone_branch points the a leaf that
367 * we finally want to del. In this case, tmp_alone_branch moves
368 * to the prev element but it will point to rq->leaf_cfs_rq_list
369 * at the end of the enqueue.
371 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
372 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
374 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
379 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
381 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
384 /* Iterate thr' all leaf cfs_rq's on a runqueue */
385 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
386 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
389 /* Do the two (enqueued) entities belong to the same group ? */
390 static inline struct cfs_rq *
391 is_same_group(struct sched_entity *se, struct sched_entity *pse)
393 if (se->cfs_rq == pse->cfs_rq)
399 static inline struct sched_entity *parent_entity(struct sched_entity *se)
405 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
407 int se_depth, pse_depth;
410 * preemption test can be made between sibling entities who are in the
411 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
412 * both tasks until we find their ancestors who are siblings of common
416 /* First walk up until both entities are at same depth */
417 se_depth = (*se)->depth;
418 pse_depth = (*pse)->depth;
420 while (se_depth > pse_depth) {
422 *se = parent_entity(*se);
425 while (pse_depth > se_depth) {
427 *pse = parent_entity(*pse);
430 while (!is_same_group(*se, *pse)) {
431 *se = parent_entity(*se);
432 *pse = parent_entity(*pse);
436 #else /* !CONFIG_FAIR_GROUP_SCHED */
438 static inline struct task_struct *task_of(struct sched_entity *se)
440 return container_of(se, struct task_struct, se);
443 #define for_each_sched_entity(se) \
444 for (; se; se = NULL)
446 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
448 return &task_rq(p)->cfs;
451 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
453 struct task_struct *p = task_of(se);
454 struct rq *rq = task_rq(p);
459 /* runqueue "owned" by this group */
460 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
465 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
468 strlcpy(path, "(null)", len);
471 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
476 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
480 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
484 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
485 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
487 static inline struct sched_entity *parent_entity(struct sched_entity *se)
493 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
497 #endif /* CONFIG_FAIR_GROUP_SCHED */
499 static __always_inline
500 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
502 /**************************************************************
503 * Scheduling class tree data structure manipulation methods:
506 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
508 s64 delta = (s64)(vruntime - max_vruntime);
510 max_vruntime = vruntime;
515 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
517 s64 delta = (s64)(vruntime - min_vruntime);
519 min_vruntime = vruntime;
524 static inline int entity_before(struct sched_entity *a,
525 struct sched_entity *b)
527 return (s64)(a->vruntime - b->vruntime) < 0;
530 static void update_min_vruntime(struct cfs_rq *cfs_rq)
532 struct sched_entity *curr = cfs_rq->curr;
533 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
535 u64 vruntime = cfs_rq->min_vruntime;
539 vruntime = curr->vruntime;
544 if (leftmost) { /* non-empty tree */
545 struct sched_entity *se;
546 se = rb_entry(leftmost, struct sched_entity, run_node);
549 vruntime = se->vruntime;
551 vruntime = min_vruntime(vruntime, se->vruntime);
554 /* ensure we never gain time by being placed backwards. */
555 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
558 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
563 * Enqueue an entity into the rb-tree:
565 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
567 struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node;
568 struct rb_node *parent = NULL;
569 struct sched_entity *entry;
570 bool leftmost = true;
573 * Find the right place in the rbtree:
577 entry = rb_entry(parent, struct sched_entity, run_node);
579 * We dont care about collisions. Nodes with
580 * the same key stay together.
582 if (entity_before(se, entry)) {
583 link = &parent->rb_left;
585 link = &parent->rb_right;
590 rb_link_node(&se->run_node, parent, link);
591 rb_insert_color_cached(&se->run_node,
592 &cfs_rq->tasks_timeline, leftmost);
595 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
597 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
600 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
602 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
607 return rb_entry(left, struct sched_entity, run_node);
610 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
612 struct rb_node *next = rb_next(&se->run_node);
617 return rb_entry(next, struct sched_entity, run_node);
620 #ifdef CONFIG_SCHED_DEBUG
621 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
623 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
628 return rb_entry(last, struct sched_entity, run_node);
631 /**************************************************************
632 * Scheduling class statistics methods:
635 int sched_proc_update_handler(struct ctl_table *table, int write,
636 void __user *buffer, size_t *lenp,
639 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
640 unsigned int factor = get_update_sysctl_factor();
645 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
646 sysctl_sched_min_granularity);
648 #define WRT_SYSCTL(name) \
649 (normalized_sysctl_##name = sysctl_##name / (factor))
650 WRT_SYSCTL(sched_min_granularity);
651 WRT_SYSCTL(sched_latency);
652 WRT_SYSCTL(sched_wakeup_granularity);
662 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
664 if (unlikely(se->load.weight != NICE_0_LOAD))
665 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
671 * The idea is to set a period in which each task runs once.
673 * When there are too many tasks (sched_nr_latency) we have to stretch
674 * this period because otherwise the slices get too small.
676 * p = (nr <= nl) ? l : l*nr/nl
678 static u64 __sched_period(unsigned long nr_running)
680 if (unlikely(nr_running > sched_nr_latency))
681 return nr_running * sysctl_sched_min_granularity;
683 return sysctl_sched_latency;
687 * We calculate the wall-time slice from the period by taking a part
688 * proportional to the weight.
692 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
694 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
696 for_each_sched_entity(se) {
697 struct load_weight *load;
698 struct load_weight lw;
700 cfs_rq = cfs_rq_of(se);
701 load = &cfs_rq->load;
703 if (unlikely(!se->on_rq)) {
706 update_load_add(&lw, se->load.weight);
709 slice = __calc_delta(slice, se->load.weight, load);
715 * We calculate the vruntime slice of a to-be-inserted task.
719 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
721 return calc_delta_fair(sched_slice(cfs_rq, se), se);
727 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
728 static unsigned long task_h_load(struct task_struct *p);
729 static unsigned long capacity_of(int cpu);
731 /* Give new sched_entity start runnable values to heavy its load in infant time */
732 void init_entity_runnable_average(struct sched_entity *se)
734 struct sched_avg *sa = &se->avg;
736 memset(sa, 0, sizeof(*sa));
739 * Tasks are initialized with full load to be seen as heavy tasks until
740 * they get a chance to stabilize to their real load level.
741 * Group entities are initialized with zero load to reflect the fact that
742 * nothing has been attached to the task group yet.
744 if (entity_is_task(se))
745 sa->runnable_load_avg = sa->load_avg = scale_load_down(se->load.weight);
747 se->runnable_weight = se->load.weight;
749 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
752 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
753 static void attach_entity_cfs_rq(struct sched_entity *se);
756 * With new tasks being created, their initial util_avgs are extrapolated
757 * based on the cfs_rq's current util_avg:
759 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
761 * However, in many cases, the above util_avg does not give a desired
762 * value. Moreover, the sum of the util_avgs may be divergent, such
763 * as when the series is a harmonic series.
765 * To solve this problem, we also cap the util_avg of successive tasks to
766 * only 1/2 of the left utilization budget:
768 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
770 * where n denotes the nth task and cpu_scale the CPU capacity.
772 * For example, for a CPU with 1024 of capacity, a simplest series from
773 * the beginning would be like:
775 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
776 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
778 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
779 * if util_avg > util_avg_cap.
781 void post_init_entity_util_avg(struct task_struct *p)
783 struct sched_entity *se = &p->se;
784 struct cfs_rq *cfs_rq = cfs_rq_of(se);
785 struct sched_avg *sa = &se->avg;
786 long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
787 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
790 if (cfs_rq->avg.util_avg != 0) {
791 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
792 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
794 if (sa->util_avg > cap)
801 if (p->sched_class != &fair_sched_class) {
803 * For !fair tasks do:
805 update_cfs_rq_load_avg(now, cfs_rq);
806 attach_entity_load_avg(cfs_rq, se, 0);
807 switched_from_fair(rq, p);
809 * such that the next switched_to_fair() has the
812 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
816 attach_entity_cfs_rq(se);
819 #else /* !CONFIG_SMP */
820 void init_entity_runnable_average(struct sched_entity *se)
823 void post_init_entity_util_avg(struct task_struct *p)
826 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
829 #endif /* CONFIG_SMP */
832 * Update the current task's runtime statistics.
834 static void update_curr(struct cfs_rq *cfs_rq)
836 struct sched_entity *curr = cfs_rq->curr;
837 u64 now = rq_clock_task(rq_of(cfs_rq));
843 delta_exec = now - curr->exec_start;
844 if (unlikely((s64)delta_exec <= 0))
847 curr->exec_start = now;
849 schedstat_set(curr->statistics.exec_max,
850 max(delta_exec, curr->statistics.exec_max));
852 curr->sum_exec_runtime += delta_exec;
853 schedstat_add(cfs_rq->exec_clock, delta_exec);
855 curr->vruntime += calc_delta_fair(delta_exec, curr);
856 update_min_vruntime(cfs_rq);
858 if (entity_is_task(curr)) {
859 struct task_struct *curtask = task_of(curr);
861 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
862 cgroup_account_cputime(curtask, delta_exec);
863 account_group_exec_runtime(curtask, delta_exec);
866 account_cfs_rq_runtime(cfs_rq, delta_exec);
869 static void update_curr_fair(struct rq *rq)
871 update_curr(cfs_rq_of(&rq->curr->se));
875 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
877 u64 wait_start, prev_wait_start;
879 if (!schedstat_enabled())
882 wait_start = rq_clock(rq_of(cfs_rq));
883 prev_wait_start = schedstat_val(se->statistics.wait_start);
885 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
886 likely(wait_start > prev_wait_start))
887 wait_start -= prev_wait_start;
889 __schedstat_set(se->statistics.wait_start, wait_start);
893 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
895 struct task_struct *p;
898 if (!schedstat_enabled())
901 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
903 if (entity_is_task(se)) {
905 if (task_on_rq_migrating(p)) {
907 * Preserve migrating task's wait time so wait_start
908 * time stamp can be adjusted to accumulate wait time
909 * prior to migration.
911 __schedstat_set(se->statistics.wait_start, delta);
914 trace_sched_stat_wait(p, delta);
917 __schedstat_set(se->statistics.wait_max,
918 max(schedstat_val(se->statistics.wait_max), delta));
919 __schedstat_inc(se->statistics.wait_count);
920 __schedstat_add(se->statistics.wait_sum, delta);
921 __schedstat_set(se->statistics.wait_start, 0);
925 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
927 struct task_struct *tsk = NULL;
928 u64 sleep_start, block_start;
930 if (!schedstat_enabled())
933 sleep_start = schedstat_val(se->statistics.sleep_start);
934 block_start = schedstat_val(se->statistics.block_start);
936 if (entity_is_task(se))
940 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
945 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
946 __schedstat_set(se->statistics.sleep_max, delta);
948 __schedstat_set(se->statistics.sleep_start, 0);
949 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
952 account_scheduler_latency(tsk, delta >> 10, 1);
953 trace_sched_stat_sleep(tsk, delta);
957 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
962 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
963 __schedstat_set(se->statistics.block_max, delta);
965 __schedstat_set(se->statistics.block_start, 0);
966 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
969 if (tsk->in_iowait) {
970 __schedstat_add(se->statistics.iowait_sum, delta);
971 __schedstat_inc(se->statistics.iowait_count);
972 trace_sched_stat_iowait(tsk, delta);
975 trace_sched_stat_blocked(tsk, delta);
978 * Blocking time is in units of nanosecs, so shift by
979 * 20 to get a milliseconds-range estimation of the
980 * amount of time that the task spent sleeping:
982 if (unlikely(prof_on == SLEEP_PROFILING)) {
983 profile_hits(SLEEP_PROFILING,
984 (void *)get_wchan(tsk),
987 account_scheduler_latency(tsk, delta >> 10, 0);
993 * Task is being enqueued - update stats:
996 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
998 if (!schedstat_enabled())
1002 * Are we enqueueing a waiting task? (for current tasks
1003 * a dequeue/enqueue event is a NOP)
1005 if (se != cfs_rq->curr)
1006 update_stats_wait_start(cfs_rq, se);
1008 if (flags & ENQUEUE_WAKEUP)
1009 update_stats_enqueue_sleeper(cfs_rq, se);
1013 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1016 if (!schedstat_enabled())
1020 * Mark the end of the wait period if dequeueing a
1023 if (se != cfs_rq->curr)
1024 update_stats_wait_end(cfs_rq, se);
1026 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1027 struct task_struct *tsk = task_of(se);
1029 if (tsk->state & TASK_INTERRUPTIBLE)
1030 __schedstat_set(se->statistics.sleep_start,
1031 rq_clock(rq_of(cfs_rq)));
1032 if (tsk->state & TASK_UNINTERRUPTIBLE)
1033 __schedstat_set(se->statistics.block_start,
1034 rq_clock(rq_of(cfs_rq)));
1039 * We are picking a new current task - update its stats:
1042 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1045 * We are starting a new run period:
1047 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1050 /**************************************************
1051 * Scheduling class queueing methods:
1054 #ifdef CONFIG_NUMA_BALANCING
1056 * Approximate time to scan a full NUMA task in ms. The task scan period is
1057 * calculated based on the tasks virtual memory size and
1058 * numa_balancing_scan_size.
1060 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1061 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1063 /* Portion of address space to scan in MB */
1064 unsigned int sysctl_numa_balancing_scan_size = 256;
1066 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1067 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1070 refcount_t refcount;
1072 spinlock_t lock; /* nr_tasks, tasks */
1077 struct rcu_head rcu;
1078 unsigned long total_faults;
1079 unsigned long max_faults_cpu;
1081 * Faults_cpu is used to decide whether memory should move
1082 * towards the CPU. As a consequence, these stats are weighted
1083 * more by CPU use than by memory faults.
1085 unsigned long *faults_cpu;
1086 unsigned long faults[0];
1090 * For functions that can be called in multiple contexts that permit reading
1091 * ->numa_group (see struct task_struct for locking rules).
1093 static struct numa_group *deref_task_numa_group(struct task_struct *p)
1095 return rcu_dereference_check(p->numa_group, p == current ||
1096 (lockdep_is_held(&task_rq(p)->lock) && !READ_ONCE(p->on_cpu)));
1099 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1101 return rcu_dereference_protected(p->numa_group, p == current);
1104 static inline unsigned long group_faults_priv(struct numa_group *ng);
1105 static inline unsigned long group_faults_shared(struct numa_group *ng);
1107 static unsigned int task_nr_scan_windows(struct task_struct *p)
1109 unsigned long rss = 0;
1110 unsigned long nr_scan_pages;
1113 * Calculations based on RSS as non-present and empty pages are skipped
1114 * by the PTE scanner and NUMA hinting faults should be trapped based
1117 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1118 rss = get_mm_rss(p->mm);
1120 rss = nr_scan_pages;
1122 rss = round_up(rss, nr_scan_pages);
1123 return rss / nr_scan_pages;
1126 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1127 #define MAX_SCAN_WINDOW 2560
1129 static unsigned int task_scan_min(struct task_struct *p)
1131 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1132 unsigned int scan, floor;
1133 unsigned int windows = 1;
1135 if (scan_size < MAX_SCAN_WINDOW)
1136 windows = MAX_SCAN_WINDOW / scan_size;
1137 floor = 1000 / windows;
1139 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1140 return max_t(unsigned int, floor, scan);
1143 static unsigned int task_scan_start(struct task_struct *p)
1145 unsigned long smin = task_scan_min(p);
1146 unsigned long period = smin;
1147 struct numa_group *ng;
1149 /* Scale the maximum scan period with the amount of shared memory. */
1151 ng = rcu_dereference(p->numa_group);
1153 unsigned long shared = group_faults_shared(ng);
1154 unsigned long private = group_faults_priv(ng);
1156 period *= refcount_read(&ng->refcount);
1157 period *= shared + 1;
1158 period /= private + shared + 1;
1162 return max(smin, period);
1165 static unsigned int task_scan_max(struct task_struct *p)
1167 unsigned long smin = task_scan_min(p);
1169 struct numa_group *ng;
1171 /* Watch for min being lower than max due to floor calculations */
1172 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1174 /* Scale the maximum scan period with the amount of shared memory. */
1175 ng = deref_curr_numa_group(p);
1177 unsigned long shared = group_faults_shared(ng);
1178 unsigned long private = group_faults_priv(ng);
1179 unsigned long period = smax;
1181 period *= refcount_read(&ng->refcount);
1182 period *= shared + 1;
1183 period /= private + shared + 1;
1185 smax = max(smax, period);
1188 return max(smin, smax);
1191 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
1194 struct mm_struct *mm = p->mm;
1197 mm_users = atomic_read(&mm->mm_users);
1198 if (mm_users == 1) {
1199 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1200 mm->numa_scan_seq = 0;
1204 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
1205 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1206 p->numa_work.next = &p->numa_work;
1207 p->numa_faults = NULL;
1208 RCU_INIT_POINTER(p->numa_group, NULL);
1209 p->last_task_numa_placement = 0;
1210 p->last_sum_exec_runtime = 0;
1212 /* New address space, reset the preferred nid */
1213 if (!(clone_flags & CLONE_VM)) {
1214 p->numa_preferred_nid = NUMA_NO_NODE;
1219 * New thread, keep existing numa_preferred_nid which should be copied
1220 * already by arch_dup_task_struct but stagger when scans start.
1225 delay = min_t(unsigned int, task_scan_max(current),
1226 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
1227 delay += 2 * TICK_NSEC;
1228 p->node_stamp = delay;
1232 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1234 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1235 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1238 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1240 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1241 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1244 /* Shared or private faults. */
1245 #define NR_NUMA_HINT_FAULT_TYPES 2
1247 /* Memory and CPU locality */
1248 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1250 /* Averaged statistics, and temporary buffers. */
1251 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1253 pid_t task_numa_group_id(struct task_struct *p)
1255 struct numa_group *ng;
1259 ng = rcu_dereference(p->numa_group);
1268 * The averaged statistics, shared & private, memory & CPU,
1269 * occupy the first half of the array. The second half of the
1270 * array is for current counters, which are averaged into the
1271 * first set by task_numa_placement.
1273 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1275 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1278 static inline unsigned long task_faults(struct task_struct *p, int nid)
1280 if (!p->numa_faults)
1283 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1284 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1287 static inline unsigned long group_faults(struct task_struct *p, int nid)
1289 struct numa_group *ng = deref_task_numa_group(p);
1294 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1295 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1298 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1300 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1301 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1304 static inline unsigned long group_faults_priv(struct numa_group *ng)
1306 unsigned long faults = 0;
1309 for_each_online_node(node) {
1310 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1316 static inline unsigned long group_faults_shared(struct numa_group *ng)
1318 unsigned long faults = 0;
1321 for_each_online_node(node) {
1322 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1329 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1330 * considered part of a numa group's pseudo-interleaving set. Migrations
1331 * between these nodes are slowed down, to allow things to settle down.
1333 #define ACTIVE_NODE_FRACTION 3
1335 static bool numa_is_active_node(int nid, struct numa_group *ng)
1337 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1340 /* Handle placement on systems where not all nodes are directly connected. */
1341 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1342 int maxdist, bool task)
1344 unsigned long score = 0;
1348 * All nodes are directly connected, and the same distance
1349 * from each other. No need for fancy placement algorithms.
1351 if (sched_numa_topology_type == NUMA_DIRECT)
1355 * This code is called for each node, introducing N^2 complexity,
1356 * which should be ok given the number of nodes rarely exceeds 8.
1358 for_each_online_node(node) {
1359 unsigned long faults;
1360 int dist = node_distance(nid, node);
1363 * The furthest away nodes in the system are not interesting
1364 * for placement; nid was already counted.
1366 if (dist == sched_max_numa_distance || node == nid)
1370 * On systems with a backplane NUMA topology, compare groups
1371 * of nodes, and move tasks towards the group with the most
1372 * memory accesses. When comparing two nodes at distance
1373 * "hoplimit", only nodes closer by than "hoplimit" are part
1374 * of each group. Skip other nodes.
1376 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1380 /* Add up the faults from nearby nodes. */
1382 faults = task_faults(p, node);
1384 faults = group_faults(p, node);
1387 * On systems with a glueless mesh NUMA topology, there are
1388 * no fixed "groups of nodes". Instead, nodes that are not
1389 * directly connected bounce traffic through intermediate
1390 * nodes; a numa_group can occupy any set of nodes.
1391 * The further away a node is, the less the faults count.
1392 * This seems to result in good task placement.
1394 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1395 faults *= (sched_max_numa_distance - dist);
1396 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1406 * These return the fraction of accesses done by a particular task, or
1407 * task group, on a particular numa node. The group weight is given a
1408 * larger multiplier, in order to group tasks together that are almost
1409 * evenly spread out between numa nodes.
1411 static inline unsigned long task_weight(struct task_struct *p, int nid,
1414 unsigned long faults, total_faults;
1416 if (!p->numa_faults)
1419 total_faults = p->total_numa_faults;
1424 faults = task_faults(p, nid);
1425 faults += score_nearby_nodes(p, nid, dist, true);
1427 return 1000 * faults / total_faults;
1430 static inline unsigned long group_weight(struct task_struct *p, int nid,
1433 struct numa_group *ng = deref_task_numa_group(p);
1434 unsigned long faults, total_faults;
1439 total_faults = ng->total_faults;
1444 faults = group_faults(p, nid);
1445 faults += score_nearby_nodes(p, nid, dist, false);
1447 return 1000 * faults / total_faults;
1450 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1451 int src_nid, int dst_cpu)
1453 struct numa_group *ng = deref_curr_numa_group(p);
1454 int dst_nid = cpu_to_node(dst_cpu);
1455 int last_cpupid, this_cpupid;
1457 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1458 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1461 * Allow first faults or private faults to migrate immediately early in
1462 * the lifetime of a task. The magic number 4 is based on waiting for
1463 * two full passes of the "multi-stage node selection" test that is
1466 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1467 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1471 * Multi-stage node selection is used in conjunction with a periodic
1472 * migration fault to build a temporal task<->page relation. By using
1473 * a two-stage filter we remove short/unlikely relations.
1475 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1476 * a task's usage of a particular page (n_p) per total usage of this
1477 * page (n_t) (in a given time-span) to a probability.
1479 * Our periodic faults will sample this probability and getting the
1480 * same result twice in a row, given these samples are fully
1481 * independent, is then given by P(n)^2, provided our sample period
1482 * is sufficiently short compared to the usage pattern.
1484 * This quadric squishes small probabilities, making it less likely we
1485 * act on an unlikely task<->page relation.
1487 if (!cpupid_pid_unset(last_cpupid) &&
1488 cpupid_to_nid(last_cpupid) != dst_nid)
1491 /* Always allow migrate on private faults */
1492 if (cpupid_match_pid(p, last_cpupid))
1495 /* A shared fault, but p->numa_group has not been set up yet. */
1500 * Destination node is much more heavily used than the source
1501 * node? Allow migration.
1503 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1504 ACTIVE_NODE_FRACTION)
1508 * Distribute memory according to CPU & memory use on each node,
1509 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1511 * faults_cpu(dst) 3 faults_cpu(src)
1512 * --------------- * - > ---------------
1513 * faults_mem(dst) 4 faults_mem(src)
1515 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1516 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1519 static unsigned long cpu_runnable_load(struct rq *rq);
1521 /* Cached statistics for all CPUs within a node */
1525 /* Total compute capacity of CPUs on a node */
1526 unsigned long compute_capacity;
1530 * XXX borrowed from update_sg_lb_stats
1532 static void update_numa_stats(struct numa_stats *ns, int nid)
1536 memset(ns, 0, sizeof(*ns));
1537 for_each_cpu(cpu, cpumask_of_node(nid)) {
1538 struct rq *rq = cpu_rq(cpu);
1540 ns->load += cpu_runnable_load(rq);
1541 ns->compute_capacity += capacity_of(cpu);
1546 struct task_numa_env {
1547 struct task_struct *p;
1549 int src_cpu, src_nid;
1550 int dst_cpu, dst_nid;
1552 struct numa_stats src_stats, dst_stats;
1557 struct task_struct *best_task;
1562 static void task_numa_assign(struct task_numa_env *env,
1563 struct task_struct *p, long imp)
1565 struct rq *rq = cpu_rq(env->dst_cpu);
1567 /* Bail out if run-queue part of active NUMA balance. */
1568 if (xchg(&rq->numa_migrate_on, 1))
1572 * Clear previous best_cpu/rq numa-migrate flag, since task now
1573 * found a better CPU to move/swap.
1575 if (env->best_cpu != -1) {
1576 rq = cpu_rq(env->best_cpu);
1577 WRITE_ONCE(rq->numa_migrate_on, 0);
1581 put_task_struct(env->best_task);
1586 env->best_imp = imp;
1587 env->best_cpu = env->dst_cpu;
1590 static bool load_too_imbalanced(long src_load, long dst_load,
1591 struct task_numa_env *env)
1594 long orig_src_load, orig_dst_load;
1595 long src_capacity, dst_capacity;
1598 * The load is corrected for the CPU capacity available on each node.
1601 * ------------ vs ---------
1602 * src_capacity dst_capacity
1604 src_capacity = env->src_stats.compute_capacity;
1605 dst_capacity = env->dst_stats.compute_capacity;
1607 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1609 orig_src_load = env->src_stats.load;
1610 orig_dst_load = env->dst_stats.load;
1612 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1614 /* Would this change make things worse? */
1615 return (imb > old_imb);
1619 * Maximum NUMA importance can be 1998 (2*999);
1620 * SMALLIMP @ 30 would be close to 1998/64.
1621 * Used to deter task migration.
1626 * This checks if the overall compute and NUMA accesses of the system would
1627 * be improved if the source tasks was migrated to the target dst_cpu taking
1628 * into account that it might be best if task running on the dst_cpu should
1629 * be exchanged with the source task
1631 static void task_numa_compare(struct task_numa_env *env,
1632 long taskimp, long groupimp, bool maymove)
1634 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
1635 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1636 long imp = p_ng ? groupimp : taskimp;
1637 struct task_struct *cur;
1638 long src_load, dst_load;
1639 int dist = env->dist;
1643 if (READ_ONCE(dst_rq->numa_migrate_on))
1647 cur = task_rcu_dereference(&dst_rq->curr);
1648 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1652 * Because we have preemption enabled we can get migrated around and
1653 * end try selecting ourselves (current == env->p) as a swap candidate.
1659 if (maymove && moveimp >= env->best_imp)
1666 * "imp" is the fault differential for the source task between the
1667 * source and destination node. Calculate the total differential for
1668 * the source task and potential destination task. The more negative
1669 * the value is, the more remote accesses that would be expected to
1670 * be incurred if the tasks were swapped.
1672 /* Skip this swap candidate if cannot move to the source cpu */
1673 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
1677 * If dst and source tasks are in the same NUMA group, or not
1678 * in any group then look only at task weights.
1680 cur_ng = rcu_dereference(cur->numa_group);
1681 if (cur_ng == p_ng) {
1682 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1683 task_weight(cur, env->dst_nid, dist);
1685 * Add some hysteresis to prevent swapping the
1686 * tasks within a group over tiny differences.
1692 * Compare the group weights. If a task is all by itself
1693 * (not part of a group), use the task weight instead.
1696 imp += group_weight(cur, env->src_nid, dist) -
1697 group_weight(cur, env->dst_nid, dist);
1699 imp += task_weight(cur, env->src_nid, dist) -
1700 task_weight(cur, env->dst_nid, dist);
1703 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1710 * If the NUMA importance is less than SMALLIMP,
1711 * task migration might only result in ping pong
1712 * of tasks and also hurt performance due to cache
1715 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1719 * In the overloaded case, try and keep the load balanced.
1721 load = task_h_load(env->p) - task_h_load(cur);
1725 dst_load = env->dst_stats.load + load;
1726 src_load = env->src_stats.load - load;
1728 if (load_too_imbalanced(src_load, dst_load, env))
1733 * One idle CPU per node is evaluated for a task numa move.
1734 * Call select_idle_sibling to maybe find a better one.
1738 * select_idle_siblings() uses an per-CPU cpumask that
1739 * can be used from IRQ context.
1741 local_irq_disable();
1742 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1747 task_numa_assign(env, cur, imp);
1752 static void task_numa_find_cpu(struct task_numa_env *env,
1753 long taskimp, long groupimp)
1755 long src_load, dst_load, load;
1756 bool maymove = false;
1759 load = task_h_load(env->p);
1760 dst_load = env->dst_stats.load + load;
1761 src_load = env->src_stats.load - load;
1764 * If the improvement from just moving env->p direction is better
1765 * than swapping tasks around, check if a move is possible.
1767 maymove = !load_too_imbalanced(src_load, dst_load, env);
1769 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1770 /* Skip this CPU if the source task cannot migrate */
1771 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
1775 task_numa_compare(env, taskimp, groupimp, maymove);
1779 static int task_numa_migrate(struct task_struct *p)
1781 struct task_numa_env env = {
1784 .src_cpu = task_cpu(p),
1785 .src_nid = task_node(p),
1787 .imbalance_pct = 112,
1793 unsigned long taskweight, groupweight;
1794 struct sched_domain *sd;
1795 long taskimp, groupimp;
1796 struct numa_group *ng;
1801 * Pick the lowest SD_NUMA domain, as that would have the smallest
1802 * imbalance and would be the first to start moving tasks about.
1804 * And we want to avoid any moving of tasks about, as that would create
1805 * random movement of tasks -- counter the numa conditions we're trying
1809 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1811 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1815 * Cpusets can break the scheduler domain tree into smaller
1816 * balance domains, some of which do not cross NUMA boundaries.
1817 * Tasks that are "trapped" in such domains cannot be migrated
1818 * elsewhere, so there is no point in (re)trying.
1820 if (unlikely(!sd)) {
1821 sched_setnuma(p, task_node(p));
1825 env.dst_nid = p->numa_preferred_nid;
1826 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1827 taskweight = task_weight(p, env.src_nid, dist);
1828 groupweight = group_weight(p, env.src_nid, dist);
1829 update_numa_stats(&env.src_stats, env.src_nid);
1830 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1831 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1832 update_numa_stats(&env.dst_stats, env.dst_nid);
1834 /* Try to find a spot on the preferred nid. */
1835 task_numa_find_cpu(&env, taskimp, groupimp);
1838 * Look at other nodes in these cases:
1839 * - there is no space available on the preferred_nid
1840 * - the task is part of a numa_group that is interleaved across
1841 * multiple NUMA nodes; in order to better consolidate the group,
1842 * we need to check other locations.
1844 ng = deref_curr_numa_group(p);
1845 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
1846 for_each_online_node(nid) {
1847 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1850 dist = node_distance(env.src_nid, env.dst_nid);
1851 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1853 taskweight = task_weight(p, env.src_nid, dist);
1854 groupweight = group_weight(p, env.src_nid, dist);
1857 /* Only consider nodes where both task and groups benefit */
1858 taskimp = task_weight(p, nid, dist) - taskweight;
1859 groupimp = group_weight(p, nid, dist) - groupweight;
1860 if (taskimp < 0 && groupimp < 0)
1865 update_numa_stats(&env.dst_stats, env.dst_nid);
1866 task_numa_find_cpu(&env, taskimp, groupimp);
1871 * If the task is part of a workload that spans multiple NUMA nodes,
1872 * and is migrating into one of the workload's active nodes, remember
1873 * this node as the task's preferred numa node, so the workload can
1875 * A task that migrated to a second choice node will be better off
1876 * trying for a better one later. Do not set the preferred node here.
1879 if (env.best_cpu == -1)
1882 nid = cpu_to_node(env.best_cpu);
1884 if (nid != p->numa_preferred_nid)
1885 sched_setnuma(p, nid);
1888 /* No better CPU than the current one was found. */
1889 if (env.best_cpu == -1)
1892 best_rq = cpu_rq(env.best_cpu);
1893 if (env.best_task == NULL) {
1894 ret = migrate_task_to(p, env.best_cpu);
1895 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1897 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1901 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
1902 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1905 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1906 put_task_struct(env.best_task);
1910 /* Attempt to migrate a task to a CPU on the preferred node. */
1911 static void numa_migrate_preferred(struct task_struct *p)
1913 unsigned long interval = HZ;
1915 /* This task has no NUMA fault statistics yet */
1916 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
1919 /* Periodically retry migrating the task to the preferred node */
1920 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1921 p->numa_migrate_retry = jiffies + interval;
1923 /* Success if task is already running on preferred CPU */
1924 if (task_node(p) == p->numa_preferred_nid)
1927 /* Otherwise, try migrate to a CPU on the preferred node */
1928 task_numa_migrate(p);
1932 * Find out how many nodes on the workload is actively running on. Do this by
1933 * tracking the nodes from which NUMA hinting faults are triggered. This can
1934 * be different from the set of nodes where the workload's memory is currently
1937 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1939 unsigned long faults, max_faults = 0;
1940 int nid, active_nodes = 0;
1942 for_each_online_node(nid) {
1943 faults = group_faults_cpu(numa_group, nid);
1944 if (faults > max_faults)
1945 max_faults = faults;
1948 for_each_online_node(nid) {
1949 faults = group_faults_cpu(numa_group, nid);
1950 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1954 numa_group->max_faults_cpu = max_faults;
1955 numa_group->active_nodes = active_nodes;
1959 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1960 * increments. The more local the fault statistics are, the higher the scan
1961 * period will be for the next scan window. If local/(local+remote) ratio is
1962 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1963 * the scan period will decrease. Aim for 70% local accesses.
1965 #define NUMA_PERIOD_SLOTS 10
1966 #define NUMA_PERIOD_THRESHOLD 7
1969 * Increase the scan period (slow down scanning) if the majority of
1970 * our memory is already on our local node, or if the majority of
1971 * the page accesses are shared with other processes.
1972 * Otherwise, decrease the scan period.
1974 static void update_task_scan_period(struct task_struct *p,
1975 unsigned long shared, unsigned long private)
1977 unsigned int period_slot;
1978 int lr_ratio, ps_ratio;
1981 unsigned long remote = p->numa_faults_locality[0];
1982 unsigned long local = p->numa_faults_locality[1];
1985 * If there were no record hinting faults then either the task is
1986 * completely idle or all activity is areas that are not of interest
1987 * to automatic numa balancing. Related to that, if there were failed
1988 * migration then it implies we are migrating too quickly or the local
1989 * node is overloaded. In either case, scan slower
1991 if (local + shared == 0 || p->numa_faults_locality[2]) {
1992 p->numa_scan_period = min(p->numa_scan_period_max,
1993 p->numa_scan_period << 1);
1995 p->mm->numa_next_scan = jiffies +
1996 msecs_to_jiffies(p->numa_scan_period);
2002 * Prepare to scale scan period relative to the current period.
2003 * == NUMA_PERIOD_THRESHOLD scan period stays the same
2004 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2005 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2007 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
2008 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
2009 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
2011 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
2013 * Most memory accesses are local. There is no need to
2014 * do fast NUMA scanning, since memory is already local.
2016 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2019 diff = slot * period_slot;
2020 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2022 * Most memory accesses are shared with other tasks.
2023 * There is no point in continuing fast NUMA scanning,
2024 * since other tasks may just move the memory elsewhere.
2026 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2029 diff = slot * period_slot;
2032 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2033 * yet they are not on the local NUMA node. Speed up
2034 * NUMA scanning to get the memory moved over.
2036 int ratio = max(lr_ratio, ps_ratio);
2037 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2040 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2041 task_scan_min(p), task_scan_max(p));
2042 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2046 * Get the fraction of time the task has been running since the last
2047 * NUMA placement cycle. The scheduler keeps similar statistics, but
2048 * decays those on a 32ms period, which is orders of magnitude off
2049 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2050 * stats only if the task is so new there are no NUMA statistics yet.
2052 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2054 u64 runtime, delta, now;
2055 /* Use the start of this time slice to avoid calculations. */
2056 now = p->se.exec_start;
2057 runtime = p->se.sum_exec_runtime;
2059 if (p->last_task_numa_placement) {
2060 delta = runtime - p->last_sum_exec_runtime;
2061 *period = now - p->last_task_numa_placement;
2063 /* Avoid time going backwards, prevent potential divide error: */
2064 if (unlikely((s64)*period < 0))
2067 delta = p->se.avg.load_sum;
2068 *period = LOAD_AVG_MAX;
2071 p->last_sum_exec_runtime = runtime;
2072 p->last_task_numa_placement = now;
2078 * Determine the preferred nid for a task in a numa_group. This needs to
2079 * be done in a way that produces consistent results with group_weight,
2080 * otherwise workloads might not converge.
2082 static int preferred_group_nid(struct task_struct *p, int nid)
2087 /* Direct connections between all NUMA nodes. */
2088 if (sched_numa_topology_type == NUMA_DIRECT)
2092 * On a system with glueless mesh NUMA topology, group_weight
2093 * scores nodes according to the number of NUMA hinting faults on
2094 * both the node itself, and on nearby nodes.
2096 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2097 unsigned long score, max_score = 0;
2098 int node, max_node = nid;
2100 dist = sched_max_numa_distance;
2102 for_each_online_node(node) {
2103 score = group_weight(p, node, dist);
2104 if (score > max_score) {
2113 * Finding the preferred nid in a system with NUMA backplane
2114 * interconnect topology is more involved. The goal is to locate
2115 * tasks from numa_groups near each other in the system, and
2116 * untangle workloads from different sides of the system. This requires
2117 * searching down the hierarchy of node groups, recursively searching
2118 * inside the highest scoring group of nodes. The nodemask tricks
2119 * keep the complexity of the search down.
2121 nodes = node_online_map;
2122 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2123 unsigned long max_faults = 0;
2124 nodemask_t max_group = NODE_MASK_NONE;
2127 /* Are there nodes at this distance from each other? */
2128 if (!find_numa_distance(dist))
2131 for_each_node_mask(a, nodes) {
2132 unsigned long faults = 0;
2133 nodemask_t this_group;
2134 nodes_clear(this_group);
2136 /* Sum group's NUMA faults; includes a==b case. */
2137 for_each_node_mask(b, nodes) {
2138 if (node_distance(a, b) < dist) {
2139 faults += group_faults(p, b);
2140 node_set(b, this_group);
2141 node_clear(b, nodes);
2145 /* Remember the top group. */
2146 if (faults > max_faults) {
2147 max_faults = faults;
2148 max_group = this_group;
2150 * subtle: at the smallest distance there is
2151 * just one node left in each "group", the
2152 * winner is the preferred nid.
2157 /* Next round, evaluate the nodes within max_group. */
2165 static void task_numa_placement(struct task_struct *p)
2167 int seq, nid, max_nid = NUMA_NO_NODE;
2168 unsigned long max_faults = 0;
2169 unsigned long fault_types[2] = { 0, 0 };
2170 unsigned long total_faults;
2171 u64 runtime, period;
2172 spinlock_t *group_lock = NULL;
2173 struct numa_group *ng;
2176 * The p->mm->numa_scan_seq field gets updated without
2177 * exclusive access. Use READ_ONCE() here to ensure
2178 * that the field is read in a single access:
2180 seq = READ_ONCE(p->mm->numa_scan_seq);
2181 if (p->numa_scan_seq == seq)
2183 p->numa_scan_seq = seq;
2184 p->numa_scan_period_max = task_scan_max(p);
2186 total_faults = p->numa_faults_locality[0] +
2187 p->numa_faults_locality[1];
2188 runtime = numa_get_avg_runtime(p, &period);
2190 /* If the task is part of a group prevent parallel updates to group stats */
2191 ng = deref_curr_numa_group(p);
2193 group_lock = &ng->lock;
2194 spin_lock_irq(group_lock);
2197 /* Find the node with the highest number of faults */
2198 for_each_online_node(nid) {
2199 /* Keep track of the offsets in numa_faults array */
2200 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2201 unsigned long faults = 0, group_faults = 0;
2204 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2205 long diff, f_diff, f_weight;
2207 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2208 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2209 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2210 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2212 /* Decay existing window, copy faults since last scan */
2213 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2214 fault_types[priv] += p->numa_faults[membuf_idx];
2215 p->numa_faults[membuf_idx] = 0;
2218 * Normalize the faults_from, so all tasks in a group
2219 * count according to CPU use, instead of by the raw
2220 * number of faults. Tasks with little runtime have
2221 * little over-all impact on throughput, and thus their
2222 * faults are less important.
2224 f_weight = div64_u64(runtime << 16, period + 1);
2225 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2227 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2228 p->numa_faults[cpubuf_idx] = 0;
2230 p->numa_faults[mem_idx] += diff;
2231 p->numa_faults[cpu_idx] += f_diff;
2232 faults += p->numa_faults[mem_idx];
2233 p->total_numa_faults += diff;
2236 * safe because we can only change our own group
2238 * mem_idx represents the offset for a given
2239 * nid and priv in a specific region because it
2240 * is at the beginning of the numa_faults array.
2242 ng->faults[mem_idx] += diff;
2243 ng->faults_cpu[mem_idx] += f_diff;
2244 ng->total_faults += diff;
2245 group_faults += ng->faults[mem_idx];
2250 if (faults > max_faults) {
2251 max_faults = faults;
2254 } else if (group_faults > max_faults) {
2255 max_faults = group_faults;
2261 numa_group_count_active_nodes(ng);
2262 spin_unlock_irq(group_lock);
2263 max_nid = preferred_group_nid(p, max_nid);
2267 /* Set the new preferred node */
2268 if (max_nid != p->numa_preferred_nid)
2269 sched_setnuma(p, max_nid);
2272 update_task_scan_period(p, fault_types[0], fault_types[1]);
2275 static inline int get_numa_group(struct numa_group *grp)
2277 return refcount_inc_not_zero(&grp->refcount);
2280 static inline void put_numa_group(struct numa_group *grp)
2282 if (refcount_dec_and_test(&grp->refcount))
2283 kfree_rcu(grp, rcu);
2286 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2289 struct numa_group *grp, *my_grp;
2290 struct task_struct *tsk;
2292 int cpu = cpupid_to_cpu(cpupid);
2295 if (unlikely(!deref_curr_numa_group(p))) {
2296 unsigned int size = sizeof(struct numa_group) +
2297 4*nr_node_ids*sizeof(unsigned long);
2299 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2303 refcount_set(&grp->refcount, 1);
2304 grp->active_nodes = 1;
2305 grp->max_faults_cpu = 0;
2306 spin_lock_init(&grp->lock);
2308 /* Second half of the array tracks nids where faults happen */
2309 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2312 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2313 grp->faults[i] = p->numa_faults[i];
2315 grp->total_faults = p->total_numa_faults;
2318 rcu_assign_pointer(p->numa_group, grp);
2322 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2324 if (!cpupid_match_pid(tsk, cpupid))
2327 grp = rcu_dereference(tsk->numa_group);
2331 my_grp = deref_curr_numa_group(p);
2336 * Only join the other group if its bigger; if we're the bigger group,
2337 * the other task will join us.
2339 if (my_grp->nr_tasks > grp->nr_tasks)
2343 * Tie-break on the grp address.
2345 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2348 /* Always join threads in the same process. */
2349 if (tsk->mm == current->mm)
2352 /* Simple filter to avoid false positives due to PID collisions */
2353 if (flags & TNF_SHARED)
2356 /* Update priv based on whether false sharing was detected */
2359 if (join && !get_numa_group(grp))
2367 BUG_ON(irqs_disabled());
2368 double_lock_irq(&my_grp->lock, &grp->lock);
2370 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2371 my_grp->faults[i] -= p->numa_faults[i];
2372 grp->faults[i] += p->numa_faults[i];
2374 my_grp->total_faults -= p->total_numa_faults;
2375 grp->total_faults += p->total_numa_faults;
2380 spin_unlock(&my_grp->lock);
2381 spin_unlock_irq(&grp->lock);
2383 rcu_assign_pointer(p->numa_group, grp);
2385 put_numa_group(my_grp);
2394 * Get rid of NUMA staticstics associated with a task (either current or dead).
2395 * If @final is set, the task is dead and has reached refcount zero, so we can
2396 * safely free all relevant data structures. Otherwise, there might be
2397 * concurrent reads from places like load balancing and procfs, and we should
2398 * reset the data back to default state without freeing ->numa_faults.
2400 void task_numa_free(struct task_struct *p, bool final)
2402 /* safe: p either is current or is being freed by current */
2403 struct numa_group *grp = rcu_dereference_raw(p->numa_group);
2404 unsigned long *numa_faults = p->numa_faults;
2405 unsigned long flags;
2412 spin_lock_irqsave(&grp->lock, flags);
2413 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2414 grp->faults[i] -= p->numa_faults[i];
2415 grp->total_faults -= p->total_numa_faults;
2418 spin_unlock_irqrestore(&grp->lock, flags);
2419 RCU_INIT_POINTER(p->numa_group, NULL);
2420 put_numa_group(grp);
2424 p->numa_faults = NULL;
2427 p->total_numa_faults = 0;
2428 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2434 * Got a PROT_NONE fault for a page on @node.
2436 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2438 struct task_struct *p = current;
2439 bool migrated = flags & TNF_MIGRATED;
2440 int cpu_node = task_node(current);
2441 int local = !!(flags & TNF_FAULT_LOCAL);
2442 struct numa_group *ng;
2445 if (!static_branch_likely(&sched_numa_balancing))
2448 /* for example, ksmd faulting in a user's mm */
2452 /* Allocate buffer to track faults on a per-node basis */
2453 if (unlikely(!p->numa_faults)) {
2454 int size = sizeof(*p->numa_faults) *
2455 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2457 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2458 if (!p->numa_faults)
2461 p->total_numa_faults = 0;
2462 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2466 * First accesses are treated as private, otherwise consider accesses
2467 * to be private if the accessing pid has not changed
2469 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2472 priv = cpupid_match_pid(p, last_cpupid);
2473 if (!priv && !(flags & TNF_NO_GROUP))
2474 task_numa_group(p, last_cpupid, flags, &priv);
2478 * If a workload spans multiple NUMA nodes, a shared fault that
2479 * occurs wholly within the set of nodes that the workload is
2480 * actively using should be counted as local. This allows the
2481 * scan rate to slow down when a workload has settled down.
2483 ng = deref_curr_numa_group(p);
2484 if (!priv && !local && ng && ng->active_nodes > 1 &&
2485 numa_is_active_node(cpu_node, ng) &&
2486 numa_is_active_node(mem_node, ng))
2490 * Retry to migrate task to preferred node periodically, in case it
2491 * previously failed, or the scheduler moved us.
2493 if (time_after(jiffies, p->numa_migrate_retry)) {
2494 task_numa_placement(p);
2495 numa_migrate_preferred(p);
2499 p->numa_pages_migrated += pages;
2500 if (flags & TNF_MIGRATE_FAIL)
2501 p->numa_faults_locality[2] += pages;
2503 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2504 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2505 p->numa_faults_locality[local] += pages;
2508 static void reset_ptenuma_scan(struct task_struct *p)
2511 * We only did a read acquisition of the mmap sem, so
2512 * p->mm->numa_scan_seq is written to without exclusive access
2513 * and the update is not guaranteed to be atomic. That's not
2514 * much of an issue though, since this is just used for
2515 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2516 * expensive, to avoid any form of compiler optimizations:
2518 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2519 p->mm->numa_scan_offset = 0;
2523 * The expensive part of numa migration is done from task_work context.
2524 * Triggered from task_tick_numa().
2526 void task_numa_work(struct callback_head *work)
2528 unsigned long migrate, next_scan, now = jiffies;
2529 struct task_struct *p = current;
2530 struct mm_struct *mm = p->mm;
2531 u64 runtime = p->se.sum_exec_runtime;
2532 struct vm_area_struct *vma;
2533 unsigned long start, end;
2534 unsigned long nr_pte_updates = 0;
2535 long pages, virtpages;
2537 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2539 work->next = work; /* protect against double add */
2541 * Who cares about NUMA placement when they're dying.
2543 * NOTE: make sure not to dereference p->mm before this check,
2544 * exit_task_work() happens _after_ exit_mm() so we could be called
2545 * without p->mm even though we still had it when we enqueued this
2548 if (p->flags & PF_EXITING)
2551 if (!mm->numa_next_scan) {
2552 mm->numa_next_scan = now +
2553 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2557 * Enforce maximal scan/migration frequency..
2559 migrate = mm->numa_next_scan;
2560 if (time_before(now, migrate))
2563 if (p->numa_scan_period == 0) {
2564 p->numa_scan_period_max = task_scan_max(p);
2565 p->numa_scan_period = task_scan_start(p);
2568 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2569 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2573 * Delay this task enough that another task of this mm will likely win
2574 * the next time around.
2576 p->node_stamp += 2 * TICK_NSEC;
2578 start = mm->numa_scan_offset;
2579 pages = sysctl_numa_balancing_scan_size;
2580 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2581 virtpages = pages * 8; /* Scan up to this much virtual space */
2586 if (!down_read_trylock(&mm->mmap_sem))
2588 vma = find_vma(mm, start);
2590 reset_ptenuma_scan(p);
2594 for (; vma; vma = vma->vm_next) {
2595 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2596 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2601 * Shared library pages mapped by multiple processes are not
2602 * migrated as it is expected they are cache replicated. Avoid
2603 * hinting faults in read-only file-backed mappings or the vdso
2604 * as migrating the pages will be of marginal benefit.
2607 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2611 * Skip inaccessible VMAs to avoid any confusion between
2612 * PROT_NONE and NUMA hinting ptes
2614 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2618 start = max(start, vma->vm_start);
2619 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2620 end = min(end, vma->vm_end);
2621 nr_pte_updates = change_prot_numa(vma, start, end);
2624 * Try to scan sysctl_numa_balancing_size worth of
2625 * hpages that have at least one present PTE that
2626 * is not already pte-numa. If the VMA contains
2627 * areas that are unused or already full of prot_numa
2628 * PTEs, scan up to virtpages, to skip through those
2632 pages -= (end - start) >> PAGE_SHIFT;
2633 virtpages -= (end - start) >> PAGE_SHIFT;
2636 if (pages <= 0 || virtpages <= 0)
2640 } while (end != vma->vm_end);
2645 * It is possible to reach the end of the VMA list but the last few
2646 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2647 * would find the !migratable VMA on the next scan but not reset the
2648 * scanner to the start so check it now.
2651 mm->numa_scan_offset = start;
2653 reset_ptenuma_scan(p);
2654 up_read(&mm->mmap_sem);
2657 * Make sure tasks use at least 32x as much time to run other code
2658 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2659 * Usually update_task_scan_period slows down scanning enough; on an
2660 * overloaded system we need to limit overhead on a per task basis.
2662 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2663 u64 diff = p->se.sum_exec_runtime - runtime;
2664 p->node_stamp += 32 * diff;
2669 * Drive the periodic memory faults..
2671 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2673 struct callback_head *work = &curr->numa_work;
2677 * We don't care about NUMA placement if we don't have memory.
2679 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2683 * Using runtime rather than walltime has the dual advantage that
2684 * we (mostly) drive the selection from busy threads and that the
2685 * task needs to have done some actual work before we bother with
2688 now = curr->se.sum_exec_runtime;
2689 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2691 if (now > curr->node_stamp + period) {
2692 if (!curr->node_stamp)
2693 curr->numa_scan_period = task_scan_start(curr);
2694 curr->node_stamp += period;
2696 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2697 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2698 task_work_add(curr, work, true);
2703 static void update_scan_period(struct task_struct *p, int new_cpu)
2705 int src_nid = cpu_to_node(task_cpu(p));
2706 int dst_nid = cpu_to_node(new_cpu);
2708 if (!static_branch_likely(&sched_numa_balancing))
2711 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2714 if (src_nid == dst_nid)
2718 * Allow resets if faults have been trapped before one scan
2719 * has completed. This is most likely due to a new task that
2720 * is pulled cross-node due to wakeups or load balancing.
2722 if (p->numa_scan_seq) {
2724 * Avoid scan adjustments if moving to the preferred
2725 * node or if the task was not previously running on
2726 * the preferred node.
2728 if (dst_nid == p->numa_preferred_nid ||
2729 (p->numa_preferred_nid != NUMA_NO_NODE &&
2730 src_nid != p->numa_preferred_nid))
2734 p->numa_scan_period = task_scan_start(p);
2738 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2742 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2746 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2750 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2754 #endif /* CONFIG_NUMA_BALANCING */
2757 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2759 update_load_add(&cfs_rq->load, se->load.weight);
2761 if (entity_is_task(se)) {
2762 struct rq *rq = rq_of(cfs_rq);
2764 account_numa_enqueue(rq, task_of(se));
2765 list_add(&se->group_node, &rq->cfs_tasks);
2768 cfs_rq->nr_running++;
2772 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2774 update_load_sub(&cfs_rq->load, se->load.weight);
2776 if (entity_is_task(se)) {
2777 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2778 list_del_init(&se->group_node);
2781 cfs_rq->nr_running--;
2785 * Signed add and clamp on underflow.
2787 * Explicitly do a load-store to ensure the intermediate value never hits
2788 * memory. This allows lockless observations without ever seeing the negative
2791 #define add_positive(_ptr, _val) do { \
2792 typeof(_ptr) ptr = (_ptr); \
2793 typeof(_val) val = (_val); \
2794 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2798 if (val < 0 && res > var) \
2801 WRITE_ONCE(*ptr, res); \
2805 * Unsigned subtract and clamp on underflow.
2807 * Explicitly do a load-store to ensure the intermediate value never hits
2808 * memory. This allows lockless observations without ever seeing the negative
2811 #define sub_positive(_ptr, _val) do { \
2812 typeof(_ptr) ptr = (_ptr); \
2813 typeof(*ptr) val = (_val); \
2814 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2818 WRITE_ONCE(*ptr, res); \
2822 * Remove and clamp on negative, from a local variable.
2824 * A variant of sub_positive(), which does not use explicit load-store
2825 * and is thus optimized for local variable updates.
2827 #define lsub_positive(_ptr, _val) do { \
2828 typeof(_ptr) ptr = (_ptr); \
2829 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
2834 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2836 cfs_rq->runnable_weight += se->runnable_weight;
2838 cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2839 cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2843 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2845 cfs_rq->runnable_weight -= se->runnable_weight;
2847 sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2848 sub_positive(&cfs_rq->avg.runnable_load_sum,
2849 se_runnable(se) * se->avg.runnable_load_sum);
2853 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2855 cfs_rq->avg.load_avg += se->avg.load_avg;
2856 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2860 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2862 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2863 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2867 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2869 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2871 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2873 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2876 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2877 unsigned long weight, unsigned long runnable)
2880 /* commit outstanding execution time */
2881 if (cfs_rq->curr == se)
2882 update_curr(cfs_rq);
2883 account_entity_dequeue(cfs_rq, se);
2884 dequeue_runnable_load_avg(cfs_rq, se);
2886 dequeue_load_avg(cfs_rq, se);
2888 se->runnable_weight = runnable;
2889 update_load_set(&se->load, weight);
2893 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2895 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2896 se->avg.runnable_load_avg =
2897 div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2901 enqueue_load_avg(cfs_rq, se);
2903 account_entity_enqueue(cfs_rq, se);
2904 enqueue_runnable_load_avg(cfs_rq, se);
2908 void reweight_task(struct task_struct *p, int prio)
2910 struct sched_entity *se = &p->se;
2911 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2912 struct load_weight *load = &se->load;
2913 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2915 reweight_entity(cfs_rq, se, weight, weight);
2916 load->inv_weight = sched_prio_to_wmult[prio];
2919 #ifdef CONFIG_FAIR_GROUP_SCHED
2922 * All this does is approximate the hierarchical proportion which includes that
2923 * global sum we all love to hate.
2925 * That is, the weight of a group entity, is the proportional share of the
2926 * group weight based on the group runqueue weights. That is:
2928 * tg->weight * grq->load.weight
2929 * ge->load.weight = ----------------------------- (1)
2930 * \Sum grq->load.weight
2932 * Now, because computing that sum is prohibitively expensive to compute (been
2933 * there, done that) we approximate it with this average stuff. The average
2934 * moves slower and therefore the approximation is cheaper and more stable.
2936 * So instead of the above, we substitute:
2938 * grq->load.weight -> grq->avg.load_avg (2)
2940 * which yields the following:
2942 * tg->weight * grq->avg.load_avg
2943 * ge->load.weight = ------------------------------ (3)
2946 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2948 * That is shares_avg, and it is right (given the approximation (2)).
2950 * The problem with it is that because the average is slow -- it was designed
2951 * to be exactly that of course -- this leads to transients in boundary
2952 * conditions. In specific, the case where the group was idle and we start the
2953 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2954 * yielding bad latency etc..
2956 * Now, in that special case (1) reduces to:
2958 * tg->weight * grq->load.weight
2959 * ge->load.weight = ----------------------------- = tg->weight (4)
2962 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2964 * So what we do is modify our approximation (3) to approach (4) in the (near)
2969 * tg->weight * grq->load.weight
2970 * --------------------------------------------------- (5)
2971 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2973 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2974 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2977 * tg->weight * grq->load.weight
2978 * ge->load.weight = ----------------------------- (6)
2983 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2984 * max(grq->load.weight, grq->avg.load_avg)
2986 * And that is shares_weight and is icky. In the (near) UP case it approaches
2987 * (4) while in the normal case it approaches (3). It consistently
2988 * overestimates the ge->load.weight and therefore:
2990 * \Sum ge->load.weight >= tg->weight
2994 static long calc_group_shares(struct cfs_rq *cfs_rq)
2996 long tg_weight, tg_shares, load, shares;
2997 struct task_group *tg = cfs_rq->tg;
2999 tg_shares = READ_ONCE(tg->shares);
3001 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3003 tg_weight = atomic_long_read(&tg->load_avg);
3005 /* Ensure tg_weight >= load */
3006 tg_weight -= cfs_rq->tg_load_avg_contrib;
3009 shares = (tg_shares * load);
3011 shares /= tg_weight;
3014 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3015 * of a group with small tg->shares value. It is a floor value which is
3016 * assigned as a minimum load.weight to the sched_entity representing
3017 * the group on a CPU.
3019 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3020 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3021 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3022 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3025 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3029 * This calculates the effective runnable weight for a group entity based on
3030 * the group entity weight calculated above.
3032 * Because of the above approximation (2), our group entity weight is
3033 * an load_avg based ratio (3). This means that it includes blocked load and
3034 * does not represent the runnable weight.
3036 * Approximate the group entity's runnable weight per ratio from the group
3039 * grq->avg.runnable_load_avg
3040 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
3043 * However, analogous to above, since the avg numbers are slow, this leads to
3044 * transients in the from-idle case. Instead we use:
3046 * ge->runnable_weight = ge->load.weight *
3048 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
3049 * ----------------------------------------------------- (8)
3050 * max(grq->avg.load_avg, grq->load.weight)
3052 * Where these max() serve both to use the 'instant' values to fix the slow
3053 * from-idle and avoid the /0 on to-idle, similar to (6).
3055 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
3057 long runnable, load_avg;
3059 load_avg = max(cfs_rq->avg.load_avg,
3060 scale_load_down(cfs_rq->load.weight));
3062 runnable = max(cfs_rq->avg.runnable_load_avg,
3063 scale_load_down(cfs_rq->runnable_weight));
3067 runnable /= load_avg;
3069 return clamp_t(long, runnable, MIN_SHARES, shares);
3071 #endif /* CONFIG_SMP */
3073 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3076 * Recomputes the group entity based on the current state of its group
3079 static void update_cfs_group(struct sched_entity *se)
3081 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3082 long shares, runnable;
3087 if (throttled_hierarchy(gcfs_rq))
3091 runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
3093 if (likely(se->load.weight == shares))
3096 shares = calc_group_shares(gcfs_rq);
3097 runnable = calc_group_runnable(gcfs_rq, shares);
3100 reweight_entity(cfs_rq_of(se), se, shares, runnable);
3103 #else /* CONFIG_FAIR_GROUP_SCHED */
3104 static inline void update_cfs_group(struct sched_entity *se)
3107 #endif /* CONFIG_FAIR_GROUP_SCHED */
3109 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3111 struct rq *rq = rq_of(cfs_rq);
3113 if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) {
3115 * There are a few boundary cases this might miss but it should
3116 * get called often enough that that should (hopefully) not be
3119 * It will not get called when we go idle, because the idle
3120 * thread is a different class (!fair), nor will the utilization
3121 * number include things like RT tasks.
3123 * As is, the util number is not freq-invariant (we'd have to
3124 * implement arch_scale_freq_capacity() for that).
3128 cpufreq_update_util(rq, flags);
3133 #ifdef CONFIG_FAIR_GROUP_SCHED
3135 * update_tg_load_avg - update the tg's load avg
3136 * @cfs_rq: the cfs_rq whose avg changed
3137 * @force: update regardless of how small the difference
3139 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3140 * However, because tg->load_avg is a global value there are performance
3143 * In order to avoid having to look at the other cfs_rq's, we use a
3144 * differential update where we store the last value we propagated. This in
3145 * turn allows skipping updates if the differential is 'small'.
3147 * Updating tg's load_avg is necessary before update_cfs_share().
3149 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3151 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3154 * No need to update load_avg for root_task_group as it is not used.
3156 if (cfs_rq->tg == &root_task_group)
3159 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3160 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3161 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3166 * Called within set_task_rq() right before setting a task's CPU. The
3167 * caller only guarantees p->pi_lock is held; no other assumptions,
3168 * including the state of rq->lock, should be made.
3170 void set_task_rq_fair(struct sched_entity *se,
3171 struct cfs_rq *prev, struct cfs_rq *next)
3173 u64 p_last_update_time;
3174 u64 n_last_update_time;
3176 if (!sched_feat(ATTACH_AGE_LOAD))
3180 * We are supposed to update the task to "current" time, then its up to
3181 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3182 * getting what current time is, so simply throw away the out-of-date
3183 * time. This will result in the wakee task is less decayed, but giving
3184 * the wakee more load sounds not bad.
3186 if (!(se->avg.last_update_time && prev))
3189 #ifndef CONFIG_64BIT
3191 u64 p_last_update_time_copy;
3192 u64 n_last_update_time_copy;
3195 p_last_update_time_copy = prev->load_last_update_time_copy;
3196 n_last_update_time_copy = next->load_last_update_time_copy;
3200 p_last_update_time = prev->avg.last_update_time;
3201 n_last_update_time = next->avg.last_update_time;
3203 } while (p_last_update_time != p_last_update_time_copy ||
3204 n_last_update_time != n_last_update_time_copy);
3207 p_last_update_time = prev->avg.last_update_time;
3208 n_last_update_time = next->avg.last_update_time;
3210 __update_load_avg_blocked_se(p_last_update_time, se);
3211 se->avg.last_update_time = n_last_update_time;
3216 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3217 * propagate its contribution. The key to this propagation is the invariant
3218 * that for each group:
3220 * ge->avg == grq->avg (1)
3222 * _IFF_ we look at the pure running and runnable sums. Because they
3223 * represent the very same entity, just at different points in the hierarchy.
3225 * Per the above update_tg_cfs_util() is trivial and simply copies the running
3226 * sum over (but still wrong, because the group entity and group rq do not have
3227 * their PELT windows aligned).
3229 * However, update_tg_cfs_runnable() is more complex. So we have:
3231 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3233 * And since, like util, the runnable part should be directly transferable,
3234 * the following would _appear_ to be the straight forward approach:
3236 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3238 * And per (1) we have:
3240 * ge->avg.runnable_avg == grq->avg.runnable_avg
3244 * ge->load.weight * grq->avg.load_avg
3245 * ge->avg.load_avg = ----------------------------------- (4)
3248 * Except that is wrong!
3250 * Because while for entities historical weight is not important and we
3251 * really only care about our future and therefore can consider a pure
3252 * runnable sum, runqueues can NOT do this.
3254 * We specifically want runqueues to have a load_avg that includes
3255 * historical weights. Those represent the blocked load, the load we expect
3256 * to (shortly) return to us. This only works by keeping the weights as
3257 * integral part of the sum. We therefore cannot decompose as per (3).
3259 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3260 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3261 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3262 * runnable section of these tasks overlap (or not). If they were to perfectly
3263 * align the rq as a whole would be runnable 2/3 of the time. If however we
3264 * always have at least 1 runnable task, the rq as a whole is always runnable.
3266 * So we'll have to approximate.. :/
3268 * Given the constraint:
3270 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3272 * We can construct a rule that adds runnable to a rq by assuming minimal
3275 * On removal, we'll assume each task is equally runnable; which yields:
3277 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3279 * XXX: only do this for the part of runnable > running ?
3284 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3286 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3288 /* Nothing to update */
3293 * The relation between sum and avg is:
3295 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3297 * however, the PELT windows are not aligned between grq and gse.
3300 /* Set new sched_entity's utilization */
3301 se->avg.util_avg = gcfs_rq->avg.util_avg;
3302 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3304 /* Update parent cfs_rq utilization */
3305 add_positive(&cfs_rq->avg.util_avg, delta);
3306 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3310 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3312 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3313 unsigned long runnable_load_avg, load_avg;
3314 u64 runnable_load_sum, load_sum = 0;
3320 gcfs_rq->prop_runnable_sum = 0;
3322 if (runnable_sum >= 0) {
3324 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3325 * the CPU is saturated running == runnable.
3327 runnable_sum += se->avg.load_sum;
3328 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3331 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3332 * assuming all tasks are equally runnable.
3334 if (scale_load_down(gcfs_rq->load.weight)) {
3335 load_sum = div_s64(gcfs_rq->avg.load_sum,
3336 scale_load_down(gcfs_rq->load.weight));
3339 /* But make sure to not inflate se's runnable */
3340 runnable_sum = min(se->avg.load_sum, load_sum);
3344 * runnable_sum can't be lower than running_sum
3345 * Rescale running sum to be in the same range as runnable sum
3346 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3347 * runnable_sum is in [0 : LOAD_AVG_MAX]
3349 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3350 runnable_sum = max(runnable_sum, running_sum);
3352 load_sum = (s64)se_weight(se) * runnable_sum;
3353 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3355 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3356 delta_avg = load_avg - se->avg.load_avg;
3358 se->avg.load_sum = runnable_sum;
3359 se->avg.load_avg = load_avg;
3360 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3361 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3363 runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3364 runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3365 delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3366 delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3368 se->avg.runnable_load_sum = runnable_sum;
3369 se->avg.runnable_load_avg = runnable_load_avg;
3372 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3373 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3377 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3379 cfs_rq->propagate = 1;
3380 cfs_rq->prop_runnable_sum += runnable_sum;
3383 /* Update task and its cfs_rq load average */
3384 static inline int propagate_entity_load_avg(struct sched_entity *se)
3386 struct cfs_rq *cfs_rq, *gcfs_rq;
3388 if (entity_is_task(se))
3391 gcfs_rq = group_cfs_rq(se);
3392 if (!gcfs_rq->propagate)
3395 gcfs_rq->propagate = 0;
3397 cfs_rq = cfs_rq_of(se);
3399 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3401 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3402 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3404 trace_pelt_cfs_tp(cfs_rq);
3405 trace_pelt_se_tp(se);
3411 * Check if we need to update the load and the utilization of a blocked
3414 static inline bool skip_blocked_update(struct sched_entity *se)
3416 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3419 * If sched_entity still have not zero load or utilization, we have to
3422 if (se->avg.load_avg || se->avg.util_avg)
3426 * If there is a pending propagation, we have to update the load and
3427 * the utilization of the sched_entity:
3429 if (gcfs_rq->propagate)
3433 * Otherwise, the load and the utilization of the sched_entity is
3434 * already zero and there is no pending propagation, so it will be a
3435 * waste of time to try to decay it:
3440 #else /* CONFIG_FAIR_GROUP_SCHED */
3442 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3444 static inline int propagate_entity_load_avg(struct sched_entity *se)
3449 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3451 #endif /* CONFIG_FAIR_GROUP_SCHED */
3454 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3455 * @now: current time, as per cfs_rq_clock_pelt()
3456 * @cfs_rq: cfs_rq to update
3458 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3459 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3460 * post_init_entity_util_avg().
3462 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3464 * Returns true if the load decayed or we removed load.
3466 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3467 * call update_tg_load_avg() when this function returns true.
3470 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3472 unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3473 struct sched_avg *sa = &cfs_rq->avg;
3476 if (cfs_rq->removed.nr) {
3478 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3480 raw_spin_lock(&cfs_rq->removed.lock);
3481 swap(cfs_rq->removed.util_avg, removed_util);
3482 swap(cfs_rq->removed.load_avg, removed_load);
3483 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3484 cfs_rq->removed.nr = 0;
3485 raw_spin_unlock(&cfs_rq->removed.lock);
3488 sub_positive(&sa->load_avg, r);
3489 sub_positive(&sa->load_sum, r * divider);
3492 sub_positive(&sa->util_avg, r);
3493 sub_positive(&sa->util_sum, r * divider);
3495 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3500 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
3502 #ifndef CONFIG_64BIT
3504 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3508 cfs_rq_util_change(cfs_rq, 0);
3514 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3515 * @cfs_rq: cfs_rq to attach to
3516 * @se: sched_entity to attach
3517 * @flags: migration hints
3519 * Must call update_cfs_rq_load_avg() before this, since we rely on
3520 * cfs_rq->avg.last_update_time being current.
3522 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3524 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3527 * When we attach the @se to the @cfs_rq, we must align the decay
3528 * window because without that, really weird and wonderful things can
3533 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3534 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3537 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3538 * period_contrib. This isn't strictly correct, but since we're
3539 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3542 se->avg.util_sum = se->avg.util_avg * divider;
3544 se->avg.load_sum = divider;
3545 if (se_weight(se)) {
3547 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3550 se->avg.runnable_load_sum = se->avg.load_sum;
3552 enqueue_load_avg(cfs_rq, se);
3553 cfs_rq->avg.util_avg += se->avg.util_avg;
3554 cfs_rq->avg.util_sum += se->avg.util_sum;
3556 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3558 cfs_rq_util_change(cfs_rq, flags);
3560 trace_pelt_cfs_tp(cfs_rq);
3564 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3565 * @cfs_rq: cfs_rq to detach from
3566 * @se: sched_entity to detach
3568 * Must call update_cfs_rq_load_avg() before this, since we rely on
3569 * cfs_rq->avg.last_update_time being current.
3571 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3573 dequeue_load_avg(cfs_rq, se);
3574 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3575 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3577 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3579 cfs_rq_util_change(cfs_rq, 0);
3581 trace_pelt_cfs_tp(cfs_rq);
3585 * Optional action to be done while updating the load average
3587 #define UPDATE_TG 0x1
3588 #define SKIP_AGE_LOAD 0x2
3589 #define DO_ATTACH 0x4
3591 /* Update task and its cfs_rq load average */
3592 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3594 u64 now = cfs_rq_clock_pelt(cfs_rq);
3598 * Track task load average for carrying it to new CPU after migrated, and
3599 * track group sched_entity load average for task_h_load calc in migration
3601 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3602 __update_load_avg_se(now, cfs_rq, se);
3604 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3605 decayed |= propagate_entity_load_avg(se);
3607 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3610 * DO_ATTACH means we're here from enqueue_entity().
3611 * !last_update_time means we've passed through
3612 * migrate_task_rq_fair() indicating we migrated.
3614 * IOW we're enqueueing a task on a new CPU.
3616 attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
3617 update_tg_load_avg(cfs_rq, 0);
3619 } else if (decayed && (flags & UPDATE_TG))
3620 update_tg_load_avg(cfs_rq, 0);
3623 #ifndef CONFIG_64BIT
3624 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3626 u64 last_update_time_copy;
3627 u64 last_update_time;
3630 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3632 last_update_time = cfs_rq->avg.last_update_time;
3633 } while (last_update_time != last_update_time_copy);
3635 return last_update_time;
3638 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3640 return cfs_rq->avg.last_update_time;
3645 * Synchronize entity load avg of dequeued entity without locking
3648 static void sync_entity_load_avg(struct sched_entity *se)
3650 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3651 u64 last_update_time;
3653 last_update_time = cfs_rq_last_update_time(cfs_rq);
3654 __update_load_avg_blocked_se(last_update_time, se);
3658 * Task first catches up with cfs_rq, and then subtract
3659 * itself from the cfs_rq (task must be off the queue now).
3661 static void remove_entity_load_avg(struct sched_entity *se)
3663 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3664 unsigned long flags;
3667 * tasks cannot exit without having gone through wake_up_new_task() ->
3668 * post_init_entity_util_avg() which will have added things to the
3669 * cfs_rq, so we can remove unconditionally.
3672 sync_entity_load_avg(se);
3674 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3675 ++cfs_rq->removed.nr;
3676 cfs_rq->removed.util_avg += se->avg.util_avg;
3677 cfs_rq->removed.load_avg += se->avg.load_avg;
3678 cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */
3679 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3682 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3684 return cfs_rq->avg.runnable_load_avg;
3687 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3689 return cfs_rq->avg.load_avg;
3692 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3694 static inline unsigned long task_util(struct task_struct *p)
3696 return READ_ONCE(p->se.avg.util_avg);
3699 static inline unsigned long _task_util_est(struct task_struct *p)
3701 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3703 return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED);
3706 static inline unsigned long task_util_est(struct task_struct *p)
3708 return max(task_util(p), _task_util_est(p));
3711 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3712 struct task_struct *p)
3714 unsigned int enqueued;
3716 if (!sched_feat(UTIL_EST))
3719 /* Update root cfs_rq's estimated utilization */
3720 enqueued = cfs_rq->avg.util_est.enqueued;
3721 enqueued += _task_util_est(p);
3722 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3726 * Check if a (signed) value is within a specified (unsigned) margin,
3727 * based on the observation that:
3729 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3731 * NOTE: this only works when value + maring < INT_MAX.
3733 static inline bool within_margin(int value, int margin)
3735 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3739 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3741 long last_ewma_diff;
3745 if (!sched_feat(UTIL_EST))
3748 /* Update root cfs_rq's estimated utilization */
3749 ue.enqueued = cfs_rq->avg.util_est.enqueued;
3750 ue.enqueued -= min_t(unsigned int, ue.enqueued, _task_util_est(p));
3751 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3754 * Skip update of task's estimated utilization when the task has not
3755 * yet completed an activation, e.g. being migrated.
3761 * If the PELT values haven't changed since enqueue time,
3762 * skip the util_est update.
3764 ue = p->se.avg.util_est;
3765 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3769 * Skip update of task's estimated utilization when its EWMA is
3770 * already ~1% close to its last activation value.
3772 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3773 last_ewma_diff = ue.enqueued - ue.ewma;
3774 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3778 * To avoid overestimation of actual task utilization, skip updates if
3779 * we cannot grant there is idle time in this CPU.
3781 cpu = cpu_of(rq_of(cfs_rq));
3782 if (task_util(p) > capacity_orig_of(cpu))
3786 * Update Task's estimated utilization
3788 * When *p completes an activation we can consolidate another sample
3789 * of the task size. This is done by storing the current PELT value
3790 * as ue.enqueued and by using this value to update the Exponential
3791 * Weighted Moving Average (EWMA):
3793 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
3794 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
3795 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
3796 * = w * ( last_ewma_diff ) + ewma(t-1)
3797 * = w * (last_ewma_diff + ewma(t-1) / w)
3799 * Where 'w' is the weight of new samples, which is configured to be
3800 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3802 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
3803 ue.ewma += last_ewma_diff;
3804 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
3805 WRITE_ONCE(p->se.avg.util_est, ue);
3808 static inline int task_fits_capacity(struct task_struct *p, long capacity)
3810 return capacity * 1024 > task_util_est(p) * capacity_margin;
3813 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
3815 if (!static_branch_unlikely(&sched_asym_cpucapacity))
3819 rq->misfit_task_load = 0;
3823 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
3824 rq->misfit_task_load = 0;
3828 rq->misfit_task_load = task_h_load(p);
3831 #else /* CONFIG_SMP */
3833 #define UPDATE_TG 0x0
3834 #define SKIP_AGE_LOAD 0x0
3835 #define DO_ATTACH 0x0
3837 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3839 cfs_rq_util_change(cfs_rq, 0);
3842 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3845 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
3847 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3849 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3855 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
3858 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
3860 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
3862 #endif /* CONFIG_SMP */
3864 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3866 #ifdef CONFIG_SCHED_DEBUG
3867 s64 d = se->vruntime - cfs_rq->min_vruntime;
3872 if (d > 3*sysctl_sched_latency)
3873 schedstat_inc(cfs_rq->nr_spread_over);
3878 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3880 u64 vruntime = cfs_rq->min_vruntime;
3883 * The 'current' period is already promised to the current tasks,
3884 * however the extra weight of the new task will slow them down a
3885 * little, place the new task so that it fits in the slot that
3886 * stays open at the end.
3888 if (initial && sched_feat(START_DEBIT))
3889 vruntime += sched_vslice(cfs_rq, se);
3891 /* sleeps up to a single latency don't count. */
3893 unsigned long thresh = sysctl_sched_latency;
3896 * Halve their sleep time's effect, to allow
3897 * for a gentler effect of sleepers:
3899 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3905 /* ensure we never gain time by being placed backwards. */
3906 se->vruntime = max_vruntime(se->vruntime, vruntime);
3909 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3911 static inline void check_schedstat_required(void)
3913 #ifdef CONFIG_SCHEDSTATS
3914 if (schedstat_enabled())
3917 /* Force schedstat enabled if a dependent tracepoint is active */
3918 if (trace_sched_stat_wait_enabled() ||
3919 trace_sched_stat_sleep_enabled() ||
3920 trace_sched_stat_iowait_enabled() ||
3921 trace_sched_stat_blocked_enabled() ||
3922 trace_sched_stat_runtime_enabled()) {
3923 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3924 "stat_blocked and stat_runtime require the "
3925 "kernel parameter schedstats=enable or "
3926 "kernel.sched_schedstats=1\n");
3937 * update_min_vruntime()
3938 * vruntime -= min_vruntime
3942 * update_min_vruntime()
3943 * vruntime += min_vruntime
3945 * this way the vruntime transition between RQs is done when both
3946 * min_vruntime are up-to-date.
3950 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3951 * vruntime -= min_vruntime
3955 * update_min_vruntime()
3956 * vruntime += min_vruntime
3958 * this way we don't have the most up-to-date min_vruntime on the originating
3959 * CPU and an up-to-date min_vruntime on the destination CPU.
3963 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3965 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3966 bool curr = cfs_rq->curr == se;
3969 * If we're the current task, we must renormalise before calling
3973 se->vruntime += cfs_rq->min_vruntime;
3975 update_curr(cfs_rq);
3978 * Otherwise, renormalise after, such that we're placed at the current
3979 * moment in time, instead of some random moment in the past. Being
3980 * placed in the past could significantly boost this task to the
3981 * fairness detriment of existing tasks.
3983 if (renorm && !curr)
3984 se->vruntime += cfs_rq->min_vruntime;
3987 * When enqueuing a sched_entity, we must:
3988 * - Update loads to have both entity and cfs_rq synced with now.
3989 * - Add its load to cfs_rq->runnable_avg
3990 * - For group_entity, update its weight to reflect the new share of
3992 * - Add its new weight to cfs_rq->load.weight
3994 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
3995 update_cfs_group(se);
3996 enqueue_runnable_load_avg(cfs_rq, se);
3997 account_entity_enqueue(cfs_rq, se);
3999 if (flags & ENQUEUE_WAKEUP)
4000 place_entity(cfs_rq, se, 0);
4002 check_schedstat_required();
4003 update_stats_enqueue(cfs_rq, se, flags);
4004 check_spread(cfs_rq, se);
4006 __enqueue_entity(cfs_rq, se);
4009 if (cfs_rq->nr_running == 1) {
4010 list_add_leaf_cfs_rq(cfs_rq);
4011 check_enqueue_throttle(cfs_rq);
4015 static void __clear_buddies_last(struct sched_entity *se)
4017 for_each_sched_entity(se) {
4018 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4019 if (cfs_rq->last != se)
4022 cfs_rq->last = NULL;
4026 static void __clear_buddies_next(struct sched_entity *se)
4028 for_each_sched_entity(se) {
4029 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4030 if (cfs_rq->next != se)
4033 cfs_rq->next = NULL;
4037 static void __clear_buddies_skip(struct sched_entity *se)
4039 for_each_sched_entity(se) {
4040 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4041 if (cfs_rq->skip != se)
4044 cfs_rq->skip = NULL;
4048 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4050 if (cfs_rq->last == se)
4051 __clear_buddies_last(se);
4053 if (cfs_rq->next == se)
4054 __clear_buddies_next(se);
4056 if (cfs_rq->skip == se)
4057 __clear_buddies_skip(se);
4060 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4063 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4066 * Update run-time statistics of the 'current'.
4068 update_curr(cfs_rq);
4071 * When dequeuing a sched_entity, we must:
4072 * - Update loads to have both entity and cfs_rq synced with now.
4073 * - Subtract its load from the cfs_rq->runnable_avg.
4074 * - Subtract its previous weight from cfs_rq->load.weight.
4075 * - For group entity, update its weight to reflect the new share
4076 * of its group cfs_rq.
4078 update_load_avg(cfs_rq, se, UPDATE_TG);
4079 dequeue_runnable_load_avg(cfs_rq, se);
4081 update_stats_dequeue(cfs_rq, se, flags);
4083 clear_buddies(cfs_rq, se);
4085 if (se != cfs_rq->curr)
4086 __dequeue_entity(cfs_rq, se);
4088 account_entity_dequeue(cfs_rq, se);
4091 * Normalize after update_curr(); which will also have moved
4092 * min_vruntime if @se is the one holding it back. But before doing
4093 * update_min_vruntime() again, which will discount @se's position and
4094 * can move min_vruntime forward still more.
4096 if (!(flags & DEQUEUE_SLEEP))
4097 se->vruntime -= cfs_rq->min_vruntime;
4099 /* return excess runtime on last dequeue */
4100 return_cfs_rq_runtime(cfs_rq);
4102 update_cfs_group(se);
4105 * Now advance min_vruntime if @se was the entity holding it back,
4106 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4107 * put back on, and if we advance min_vruntime, we'll be placed back
4108 * further than we started -- ie. we'll be penalized.
4110 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4111 update_min_vruntime(cfs_rq);
4115 * Preempt the current task with a newly woken task if needed:
4118 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4120 unsigned long ideal_runtime, delta_exec;
4121 struct sched_entity *se;
4124 ideal_runtime = sched_slice(cfs_rq, curr);
4125 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4126 if (delta_exec > ideal_runtime) {
4127 resched_curr(rq_of(cfs_rq));
4129 * The current task ran long enough, ensure it doesn't get
4130 * re-elected due to buddy favours.
4132 clear_buddies(cfs_rq, curr);
4137 * Ensure that a task that missed wakeup preemption by a
4138 * narrow margin doesn't have to wait for a full slice.
4139 * This also mitigates buddy induced latencies under load.
4141 if (delta_exec < sysctl_sched_min_granularity)
4144 se = __pick_first_entity(cfs_rq);
4145 delta = curr->vruntime - se->vruntime;
4150 if (delta > ideal_runtime)
4151 resched_curr(rq_of(cfs_rq));
4155 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4157 /* 'current' is not kept within the tree. */
4160 * Any task has to be enqueued before it get to execute on
4161 * a CPU. So account for the time it spent waiting on the
4164 update_stats_wait_end(cfs_rq, se);
4165 __dequeue_entity(cfs_rq, se);
4166 update_load_avg(cfs_rq, se, UPDATE_TG);
4169 update_stats_curr_start(cfs_rq, se);
4173 * Track our maximum slice length, if the CPU's load is at
4174 * least twice that of our own weight (i.e. dont track it
4175 * when there are only lesser-weight tasks around):
4177 if (schedstat_enabled() &&
4178 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
4179 schedstat_set(se->statistics.slice_max,
4180 max((u64)schedstat_val(se->statistics.slice_max),
4181 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4184 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4188 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4191 * Pick the next process, keeping these things in mind, in this order:
4192 * 1) keep things fair between processes/task groups
4193 * 2) pick the "next" process, since someone really wants that to run
4194 * 3) pick the "last" process, for cache locality
4195 * 4) do not run the "skip" process, if something else is available
4197 static struct sched_entity *
4198 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4200 struct sched_entity *left = __pick_first_entity(cfs_rq);
4201 struct sched_entity *se;
4204 * If curr is set we have to see if its left of the leftmost entity
4205 * still in the tree, provided there was anything in the tree at all.
4207 if (!left || (curr && entity_before(curr, left)))
4210 se = left; /* ideally we run the leftmost entity */
4213 * Avoid running the skip buddy, if running something else can
4214 * be done without getting too unfair.
4216 if (cfs_rq->skip == se) {
4217 struct sched_entity *second;
4220 second = __pick_first_entity(cfs_rq);
4222 second = __pick_next_entity(se);
4223 if (!second || (curr && entity_before(curr, second)))
4227 if (second && wakeup_preempt_entity(second, left) < 1)
4232 * Prefer last buddy, try to return the CPU to a preempted task.
4234 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4238 * Someone really wants this to run. If it's not unfair, run it.
4240 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4243 clear_buddies(cfs_rq, se);
4248 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4250 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4253 * If still on the runqueue then deactivate_task()
4254 * was not called and update_curr() has to be done:
4257 update_curr(cfs_rq);
4259 /* throttle cfs_rqs exceeding runtime */
4260 check_cfs_rq_runtime(cfs_rq);
4262 check_spread(cfs_rq, prev);
4265 update_stats_wait_start(cfs_rq, prev);
4266 /* Put 'current' back into the tree. */
4267 __enqueue_entity(cfs_rq, prev);
4268 /* in !on_rq case, update occurred at dequeue */
4269 update_load_avg(cfs_rq, prev, 0);
4271 cfs_rq->curr = NULL;
4275 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4278 * Update run-time statistics of the 'current'.
4280 update_curr(cfs_rq);
4283 * Ensure that runnable average is periodically updated.
4285 update_load_avg(cfs_rq, curr, UPDATE_TG);
4286 update_cfs_group(curr);
4288 #ifdef CONFIG_SCHED_HRTICK
4290 * queued ticks are scheduled to match the slice, so don't bother
4291 * validating it and just reschedule.
4294 resched_curr(rq_of(cfs_rq));
4298 * don't let the period tick interfere with the hrtick preemption
4300 if (!sched_feat(DOUBLE_TICK) &&
4301 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4305 if (cfs_rq->nr_running > 1)
4306 check_preempt_tick(cfs_rq, curr);
4310 /**************************************************
4311 * CFS bandwidth control machinery
4314 #ifdef CONFIG_CFS_BANDWIDTH
4316 #ifdef CONFIG_JUMP_LABEL
4317 static struct static_key __cfs_bandwidth_used;
4319 static inline bool cfs_bandwidth_used(void)
4321 return static_key_false(&__cfs_bandwidth_used);
4324 void cfs_bandwidth_usage_inc(void)
4326 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4329 void cfs_bandwidth_usage_dec(void)
4331 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4333 #else /* CONFIG_JUMP_LABEL */
4334 static bool cfs_bandwidth_used(void)
4339 void cfs_bandwidth_usage_inc(void) {}
4340 void cfs_bandwidth_usage_dec(void) {}
4341 #endif /* CONFIG_JUMP_LABEL */
4344 * default period for cfs group bandwidth.
4345 * default: 0.1s, units: nanoseconds
4347 static inline u64 default_cfs_period(void)
4349 return 100000000ULL;
4352 static inline u64 sched_cfs_bandwidth_slice(void)
4354 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4358 * Replenish runtime according to assigned quota and update expiration time.
4359 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4360 * additional synchronization around rq->lock.
4362 * requires cfs_b->lock
4364 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4368 if (cfs_b->quota == RUNTIME_INF)
4371 now = sched_clock_cpu(smp_processor_id());
4372 cfs_b->runtime = cfs_b->quota;
4373 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4374 cfs_b->expires_seq++;
4377 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4379 return &tg->cfs_bandwidth;
4382 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4383 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4385 if (unlikely(cfs_rq->throttle_count))
4386 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4388 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4391 /* returns 0 on failure to allocate runtime */
4392 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4394 struct task_group *tg = cfs_rq->tg;
4395 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4396 u64 amount = 0, min_amount, expires;
4399 /* note: this is a positive sum as runtime_remaining <= 0 */
4400 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4402 raw_spin_lock(&cfs_b->lock);
4403 if (cfs_b->quota == RUNTIME_INF)
4404 amount = min_amount;
4406 start_cfs_bandwidth(cfs_b);
4408 if (cfs_b->runtime > 0) {
4409 amount = min(cfs_b->runtime, min_amount);
4410 cfs_b->runtime -= amount;
4414 expires_seq = cfs_b->expires_seq;
4415 expires = cfs_b->runtime_expires;
4416 raw_spin_unlock(&cfs_b->lock);
4418 cfs_rq->runtime_remaining += amount;
4420 * we may have advanced our local expiration to account for allowed
4421 * spread between our sched_clock and the one on which runtime was
4424 if (cfs_rq->expires_seq != expires_seq) {
4425 cfs_rq->expires_seq = expires_seq;
4426 cfs_rq->runtime_expires = expires;
4429 return cfs_rq->runtime_remaining > 0;
4433 * Note: This depends on the synchronization provided by sched_clock and the
4434 * fact that rq->clock snapshots this value.
4436 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4438 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4440 /* if the deadline is ahead of our clock, nothing to do */
4441 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4444 if (cfs_rq->runtime_remaining < 0)
4448 * If the local deadline has passed we have to consider the
4449 * possibility that our sched_clock is 'fast' and the global deadline
4450 * has not truly expired.
4452 * Fortunately we can check determine whether this the case by checking
4453 * whether the global deadline(cfs_b->expires_seq) has advanced.
4455 if (cfs_rq->expires_seq == cfs_b->expires_seq) {
4456 /* extend local deadline, drift is bounded above by 2 ticks */
4457 cfs_rq->runtime_expires += TICK_NSEC;
4459 /* global deadline is ahead, expiration has passed */
4460 cfs_rq->runtime_remaining = 0;
4464 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4466 /* dock delta_exec before expiring quota (as it could span periods) */
4467 cfs_rq->runtime_remaining -= delta_exec;
4468 expire_cfs_rq_runtime(cfs_rq);
4470 if (likely(cfs_rq->runtime_remaining > 0))
4473 if (cfs_rq->throttled)
4476 * if we're unable to extend our runtime we resched so that the active
4477 * hierarchy can be throttled
4479 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4480 resched_curr(rq_of(cfs_rq));
4483 static __always_inline
4484 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4486 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4489 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4492 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4494 return cfs_bandwidth_used() && cfs_rq->throttled;
4497 /* check whether cfs_rq, or any parent, is throttled */
4498 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4500 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4504 * Ensure that neither of the group entities corresponding to src_cpu or
4505 * dest_cpu are members of a throttled hierarchy when performing group
4506 * load-balance operations.
4508 static inline int throttled_lb_pair(struct task_group *tg,
4509 int src_cpu, int dest_cpu)
4511 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4513 src_cfs_rq = tg->cfs_rq[src_cpu];
4514 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4516 return throttled_hierarchy(src_cfs_rq) ||
4517 throttled_hierarchy(dest_cfs_rq);
4520 static int tg_unthrottle_up(struct task_group *tg, void *data)
4522 struct rq *rq = data;
4523 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4525 cfs_rq->throttle_count--;
4526 if (!cfs_rq->throttle_count) {
4527 /* adjust cfs_rq_clock_task() */
4528 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4529 cfs_rq->throttled_clock_task;
4531 /* Add cfs_rq with already running entity in the list */
4532 if (cfs_rq->nr_running >= 1)
4533 list_add_leaf_cfs_rq(cfs_rq);
4539 static int tg_throttle_down(struct task_group *tg, void *data)
4541 struct rq *rq = data;
4542 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4544 /* group is entering throttled state, stop time */
4545 if (!cfs_rq->throttle_count) {
4546 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4547 list_del_leaf_cfs_rq(cfs_rq);
4549 cfs_rq->throttle_count++;
4554 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4556 struct rq *rq = rq_of(cfs_rq);
4557 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4558 struct sched_entity *se;
4559 long task_delta, dequeue = 1;
4562 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4564 /* freeze hierarchy runnable averages while throttled */
4566 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4569 task_delta = cfs_rq->h_nr_running;
4570 for_each_sched_entity(se) {
4571 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4572 /* throttled entity or throttle-on-deactivate */
4577 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4578 qcfs_rq->h_nr_running -= task_delta;
4580 if (qcfs_rq->load.weight)
4585 sub_nr_running(rq, task_delta);
4587 cfs_rq->throttled = 1;
4588 cfs_rq->throttled_clock = rq_clock(rq);
4589 raw_spin_lock(&cfs_b->lock);
4590 empty = list_empty(&cfs_b->throttled_cfs_rq);
4593 * Add to the _head_ of the list, so that an already-started
4594 * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
4595 * not running add to the tail so that later runqueues don't get starved.
4597 if (cfs_b->distribute_running)
4598 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4600 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4603 * If we're the first throttled task, make sure the bandwidth
4607 start_cfs_bandwidth(cfs_b);
4609 raw_spin_unlock(&cfs_b->lock);
4612 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4614 struct rq *rq = rq_of(cfs_rq);
4615 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4616 struct sched_entity *se;
4620 se = cfs_rq->tg->se[cpu_of(rq)];
4622 cfs_rq->throttled = 0;
4624 update_rq_clock(rq);
4626 raw_spin_lock(&cfs_b->lock);
4627 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4628 list_del_rcu(&cfs_rq->throttled_list);
4629 raw_spin_unlock(&cfs_b->lock);
4631 /* update hierarchical throttle state */
4632 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4634 if (!cfs_rq->load.weight)
4637 task_delta = cfs_rq->h_nr_running;
4638 for_each_sched_entity(se) {
4642 cfs_rq = cfs_rq_of(se);
4644 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4645 cfs_rq->h_nr_running += task_delta;
4647 if (cfs_rq_throttled(cfs_rq))
4651 assert_list_leaf_cfs_rq(rq);
4654 add_nr_running(rq, task_delta);
4656 /* Determine whether we need to wake up potentially idle CPU: */
4657 if (rq->curr == rq->idle && rq->cfs.nr_running)
4661 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4662 u64 remaining, u64 expires)
4664 struct cfs_rq *cfs_rq;
4666 u64 starting_runtime = remaining;
4669 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4671 struct rq *rq = rq_of(cfs_rq);
4674 rq_lock_irqsave(rq, &rf);
4675 if (!cfs_rq_throttled(cfs_rq))
4678 /* By the above check, this should never be true */
4679 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
4681 runtime = -cfs_rq->runtime_remaining + 1;
4682 if (runtime > remaining)
4683 runtime = remaining;
4684 remaining -= runtime;
4686 cfs_rq->runtime_remaining += runtime;
4687 cfs_rq->runtime_expires = expires;
4689 /* we check whether we're throttled above */
4690 if (cfs_rq->runtime_remaining > 0)
4691 unthrottle_cfs_rq(cfs_rq);
4694 rq_unlock_irqrestore(rq, &rf);
4701 return starting_runtime - remaining;
4705 * Responsible for refilling a task_group's bandwidth and unthrottling its
4706 * cfs_rqs as appropriate. If there has been no activity within the last
4707 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4708 * used to track this state.
4710 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
4712 u64 runtime, runtime_expires;
4715 /* no need to continue the timer with no bandwidth constraint */
4716 if (cfs_b->quota == RUNTIME_INF)
4717 goto out_deactivate;
4719 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4720 cfs_b->nr_periods += overrun;
4723 * idle depends on !throttled (for the case of a large deficit), and if
4724 * we're going inactive then everything else can be deferred
4726 if (cfs_b->idle && !throttled)
4727 goto out_deactivate;
4729 __refill_cfs_bandwidth_runtime(cfs_b);
4732 /* mark as potentially idle for the upcoming period */
4737 /* account preceding periods in which throttling occurred */
4738 cfs_b->nr_throttled += overrun;
4740 runtime_expires = cfs_b->runtime_expires;
4743 * This check is repeated as we are holding onto the new bandwidth while
4744 * we unthrottle. This can potentially race with an unthrottled group
4745 * trying to acquire new bandwidth from the global pool. This can result
4746 * in us over-using our runtime if it is all used during this loop, but
4747 * only by limited amounts in that extreme case.
4749 while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
4750 runtime = cfs_b->runtime;
4751 cfs_b->distribute_running = 1;
4752 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4753 /* we can't nest cfs_b->lock while distributing bandwidth */
4754 runtime = distribute_cfs_runtime(cfs_b, runtime,
4756 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4758 cfs_b->distribute_running = 0;
4759 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4761 lsub_positive(&cfs_b->runtime, runtime);
4765 * While we are ensured activity in the period following an
4766 * unthrottle, this also covers the case in which the new bandwidth is
4767 * insufficient to cover the existing bandwidth deficit. (Forcing the
4768 * timer to remain active while there are any throttled entities.)
4778 /* a cfs_rq won't donate quota below this amount */
4779 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4780 /* minimum remaining period time to redistribute slack quota */
4781 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4782 /* how long we wait to gather additional slack before distributing */
4783 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4786 * Are we near the end of the current quota period?
4788 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4789 * hrtimer base being cleared by hrtimer_start. In the case of
4790 * migrate_hrtimers, base is never cleared, so we are fine.
4792 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4794 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4797 /* if the call-back is running a quota refresh is already occurring */
4798 if (hrtimer_callback_running(refresh_timer))
4801 /* is a quota refresh about to occur? */
4802 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4803 if (remaining < min_expire)
4809 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4811 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4813 /* if there's a quota refresh soon don't bother with slack */
4814 if (runtime_refresh_within(cfs_b, min_left))
4817 /* don't push forwards an existing deferred unthrottle */
4818 if (cfs_b->slack_started)
4820 cfs_b->slack_started = true;
4822 hrtimer_start(&cfs_b->slack_timer,
4823 ns_to_ktime(cfs_bandwidth_slack_period),
4827 /* we know any runtime found here is valid as update_curr() precedes return */
4828 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4830 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4831 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4833 if (slack_runtime <= 0)
4836 raw_spin_lock(&cfs_b->lock);
4837 if (cfs_b->quota != RUNTIME_INF &&
4838 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4839 cfs_b->runtime += slack_runtime;
4841 /* we are under rq->lock, defer unthrottling using a timer */
4842 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4843 !list_empty(&cfs_b->throttled_cfs_rq))
4844 start_cfs_slack_bandwidth(cfs_b);
4846 raw_spin_unlock(&cfs_b->lock);
4848 /* even if it's not valid for return we don't want to try again */
4849 cfs_rq->runtime_remaining -= slack_runtime;
4852 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4854 if (!cfs_bandwidth_used())
4857 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4860 __return_cfs_rq_runtime(cfs_rq);
4864 * This is done with a timer (instead of inline with bandwidth return) since
4865 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4867 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4869 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4870 unsigned long flags;
4873 /* confirm we're still not at a refresh boundary */
4874 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4875 cfs_b->slack_started = false;
4876 if (cfs_b->distribute_running) {
4877 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4881 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4882 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4886 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4887 runtime = cfs_b->runtime;
4889 expires = cfs_b->runtime_expires;
4891 cfs_b->distribute_running = 1;
4893 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4898 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4900 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4901 if (expires == cfs_b->runtime_expires)
4902 lsub_positive(&cfs_b->runtime, runtime);
4903 cfs_b->distribute_running = 0;
4904 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4908 * When a group wakes up we want to make sure that its quota is not already
4909 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4910 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4912 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4914 if (!cfs_bandwidth_used())
4917 /* an active group must be handled by the update_curr()->put() path */
4918 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4921 /* ensure the group is not already throttled */
4922 if (cfs_rq_throttled(cfs_rq))
4925 /* update runtime allocation */
4926 account_cfs_rq_runtime(cfs_rq, 0);
4927 if (cfs_rq->runtime_remaining <= 0)
4928 throttle_cfs_rq(cfs_rq);
4931 static void sync_throttle(struct task_group *tg, int cpu)
4933 struct cfs_rq *pcfs_rq, *cfs_rq;
4935 if (!cfs_bandwidth_used())
4941 cfs_rq = tg->cfs_rq[cpu];
4942 pcfs_rq = tg->parent->cfs_rq[cpu];
4944 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4945 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4948 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4949 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4951 if (!cfs_bandwidth_used())
4954 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4958 * it's possible for a throttled entity to be forced into a running
4959 * state (e.g. set_curr_task), in this case we're finished.
4961 if (cfs_rq_throttled(cfs_rq))
4964 throttle_cfs_rq(cfs_rq);
4968 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4970 struct cfs_bandwidth *cfs_b =
4971 container_of(timer, struct cfs_bandwidth, slack_timer);
4973 do_sched_cfs_slack_timer(cfs_b);
4975 return HRTIMER_NORESTART;
4978 extern const u64 max_cfs_quota_period;
4980 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4982 struct cfs_bandwidth *cfs_b =
4983 container_of(timer, struct cfs_bandwidth, period_timer);
4984 unsigned long flags;
4989 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4991 overrun = hrtimer_forward_now(timer, cfs_b->period);
4996 u64 new, old = ktime_to_ns(cfs_b->period);
4998 new = (old * 147) / 128; /* ~115% */
4999 new = min(new, max_cfs_quota_period);
5001 cfs_b->period = ns_to_ktime(new);
5003 /* since max is 1s, this is limited to 1e9^2, which fits in u64 */
5004 cfs_b->quota *= new;
5005 cfs_b->quota = div64_u64(cfs_b->quota, old);
5007 pr_warn_ratelimited(
5008 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us %lld, cfs_quota_us = %lld)\n",
5010 div_u64(new, NSEC_PER_USEC),
5011 div_u64(cfs_b->quota, NSEC_PER_USEC));
5013 /* reset count so we don't come right back in here */
5017 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
5020 cfs_b->period_active = 0;
5021 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5023 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
5026 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5028 raw_spin_lock_init(&cfs_b->lock);
5030 cfs_b->quota = RUNTIME_INF;
5031 cfs_b->period = ns_to_ktime(default_cfs_period());
5033 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
5034 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
5035 cfs_b->period_timer.function = sched_cfs_period_timer;
5036 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5037 cfs_b->slack_timer.function = sched_cfs_slack_timer;
5038 cfs_b->distribute_running = 0;
5039 cfs_b->slack_started = false;
5042 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5044 cfs_rq->runtime_enabled = 0;
5045 INIT_LIST_HEAD(&cfs_rq->throttled_list);
5048 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5052 lockdep_assert_held(&cfs_b->lock);
5054 if (cfs_b->period_active)
5057 cfs_b->period_active = 1;
5058 overrun = hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5059 cfs_b->runtime_expires += (overrun + 1) * ktime_to_ns(cfs_b->period);
5060 cfs_b->expires_seq++;
5061 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5064 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5066 /* init_cfs_bandwidth() was not called */
5067 if (!cfs_b->throttled_cfs_rq.next)
5070 hrtimer_cancel(&cfs_b->period_timer);
5071 hrtimer_cancel(&cfs_b->slack_timer);
5075 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5077 * The race is harmless, since modifying bandwidth settings of unhooked group
5078 * bits doesn't do much.
5081 /* cpu online calback */
5082 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5084 struct task_group *tg;
5086 lockdep_assert_held(&rq->lock);
5089 list_for_each_entry_rcu(tg, &task_groups, list) {
5090 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5091 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5093 raw_spin_lock(&cfs_b->lock);
5094 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5095 raw_spin_unlock(&cfs_b->lock);
5100 /* cpu offline callback */
5101 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5103 struct task_group *tg;
5105 lockdep_assert_held(&rq->lock);
5108 list_for_each_entry_rcu(tg, &task_groups, list) {
5109 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5111 if (!cfs_rq->runtime_enabled)
5115 * clock_task is not advancing so we just need to make sure
5116 * there's some valid quota amount
5118 cfs_rq->runtime_remaining = 1;
5120 * Offline rq is schedulable till CPU is completely disabled
5121 * in take_cpu_down(), so we prevent new cfs throttling here.
5123 cfs_rq->runtime_enabled = 0;
5125 if (cfs_rq_throttled(cfs_rq))
5126 unthrottle_cfs_rq(cfs_rq);
5131 #else /* CONFIG_CFS_BANDWIDTH */
5133 static inline bool cfs_bandwidth_used(void)
5138 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
5140 return rq_clock_task(rq_of(cfs_rq));
5143 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5144 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5145 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5146 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5147 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5149 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5154 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5159 static inline int throttled_lb_pair(struct task_group *tg,
5160 int src_cpu, int dest_cpu)
5165 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5167 #ifdef CONFIG_FAIR_GROUP_SCHED
5168 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5171 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5175 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5176 static inline void update_runtime_enabled(struct rq *rq) {}
5177 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5179 #endif /* CONFIG_CFS_BANDWIDTH */
5181 /**************************************************
5182 * CFS operations on tasks:
5185 #ifdef CONFIG_SCHED_HRTICK
5186 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5188 struct sched_entity *se = &p->se;
5189 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5191 SCHED_WARN_ON(task_rq(p) != rq);
5193 if (rq->cfs.h_nr_running > 1) {
5194 u64 slice = sched_slice(cfs_rq, se);
5195 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5196 s64 delta = slice - ran;
5203 hrtick_start(rq, delta);
5208 * called from enqueue/dequeue and updates the hrtick when the
5209 * current task is from our class and nr_running is low enough
5212 static void hrtick_update(struct rq *rq)
5214 struct task_struct *curr = rq->curr;
5216 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5219 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5220 hrtick_start_fair(rq, curr);
5222 #else /* !CONFIG_SCHED_HRTICK */
5224 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5228 static inline void hrtick_update(struct rq *rq)
5234 static inline unsigned long cpu_util(int cpu);
5236 static inline bool cpu_overutilized(int cpu)
5238 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5241 static inline void update_overutilized_status(struct rq *rq)
5243 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
5244 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5245 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
5249 static inline void update_overutilized_status(struct rq *rq) { }
5253 * The enqueue_task method is called before nr_running is
5254 * increased. Here we update the fair scheduling stats and
5255 * then put the task into the rbtree:
5258 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5260 struct cfs_rq *cfs_rq;
5261 struct sched_entity *se = &p->se;
5264 * The code below (indirectly) updates schedutil which looks at
5265 * the cfs_rq utilization to select a frequency.
5266 * Let's add the task's estimated utilization to the cfs_rq's
5267 * estimated utilization, before we update schedutil.
5269 util_est_enqueue(&rq->cfs, p);
5272 * If in_iowait is set, the code below may not trigger any cpufreq
5273 * utilization updates, so do it here explicitly with the IOWAIT flag
5277 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5279 for_each_sched_entity(se) {
5282 cfs_rq = cfs_rq_of(se);
5283 enqueue_entity(cfs_rq, se, flags);
5286 * end evaluation on encountering a throttled cfs_rq
5288 * note: in the case of encountering a throttled cfs_rq we will
5289 * post the final h_nr_running increment below.
5291 if (cfs_rq_throttled(cfs_rq))
5293 cfs_rq->h_nr_running++;
5295 flags = ENQUEUE_WAKEUP;
5298 for_each_sched_entity(se) {
5299 cfs_rq = cfs_rq_of(se);
5300 cfs_rq->h_nr_running++;
5302 if (cfs_rq_throttled(cfs_rq))
5305 update_load_avg(cfs_rq, se, UPDATE_TG);
5306 update_cfs_group(se);
5310 add_nr_running(rq, 1);
5312 * Since new tasks are assigned an initial util_avg equal to
5313 * half of the spare capacity of their CPU, tiny tasks have the
5314 * ability to cross the overutilized threshold, which will
5315 * result in the load balancer ruining all the task placement
5316 * done by EAS. As a way to mitigate that effect, do not account
5317 * for the first enqueue operation of new tasks during the
5318 * overutilized flag detection.
5320 * A better way of solving this problem would be to wait for
5321 * the PELT signals of tasks to converge before taking them
5322 * into account, but that is not straightforward to implement,
5323 * and the following generally works well enough in practice.
5325 if (flags & ENQUEUE_WAKEUP)
5326 update_overutilized_status(rq);
5330 if (cfs_bandwidth_used()) {
5332 * When bandwidth control is enabled; the cfs_rq_throttled()
5333 * breaks in the above iteration can result in incomplete
5334 * leaf list maintenance, resulting in triggering the assertion
5337 for_each_sched_entity(se) {
5338 cfs_rq = cfs_rq_of(se);
5340 if (list_add_leaf_cfs_rq(cfs_rq))
5345 assert_list_leaf_cfs_rq(rq);
5350 static void set_next_buddy(struct sched_entity *se);
5353 * The dequeue_task method is called before nr_running is
5354 * decreased. We remove the task from the rbtree and
5355 * update the fair scheduling stats:
5357 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5359 struct cfs_rq *cfs_rq;
5360 struct sched_entity *se = &p->se;
5361 int task_sleep = flags & DEQUEUE_SLEEP;
5363 for_each_sched_entity(se) {
5364 cfs_rq = cfs_rq_of(se);
5365 dequeue_entity(cfs_rq, se, flags);
5368 * end evaluation on encountering a throttled cfs_rq
5370 * note: in the case of encountering a throttled cfs_rq we will
5371 * post the final h_nr_running decrement below.
5373 if (cfs_rq_throttled(cfs_rq))
5375 cfs_rq->h_nr_running--;
5377 /* Don't dequeue parent if it has other entities besides us */
5378 if (cfs_rq->load.weight) {
5379 /* Avoid re-evaluating load for this entity: */
5380 se = parent_entity(se);
5382 * Bias pick_next to pick a task from this cfs_rq, as
5383 * p is sleeping when it is within its sched_slice.
5385 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5389 flags |= DEQUEUE_SLEEP;
5392 for_each_sched_entity(se) {
5393 cfs_rq = cfs_rq_of(se);
5394 cfs_rq->h_nr_running--;
5396 if (cfs_rq_throttled(cfs_rq))
5399 update_load_avg(cfs_rq, se, UPDATE_TG);
5400 update_cfs_group(se);
5404 sub_nr_running(rq, 1);
5406 util_est_dequeue(&rq->cfs, p, task_sleep);
5412 /* Working cpumask for: load_balance, load_balance_newidle. */
5413 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5414 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5416 #ifdef CONFIG_NO_HZ_COMMON
5419 cpumask_var_t idle_cpus_mask;
5421 int has_blocked; /* Idle CPUS has blocked load */
5422 unsigned long next_balance; /* in jiffy units */
5423 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5424 } nohz ____cacheline_aligned;
5426 #endif /* CONFIG_NO_HZ_COMMON */
5428 static unsigned long cpu_runnable_load(struct rq *rq)
5430 return cfs_rq_runnable_load_avg(&rq->cfs);
5433 static unsigned long capacity_of(int cpu)
5435 return cpu_rq(cpu)->cpu_capacity;
5438 static unsigned long cpu_avg_load_per_task(int cpu)
5440 struct rq *rq = cpu_rq(cpu);
5441 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5442 unsigned long load_avg = cpu_runnable_load(rq);
5445 return load_avg / nr_running;
5450 static void record_wakee(struct task_struct *p)
5453 * Only decay a single time; tasks that have less then 1 wakeup per
5454 * jiffy will not have built up many flips.
5456 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5457 current->wakee_flips >>= 1;
5458 current->wakee_flip_decay_ts = jiffies;
5461 if (current->last_wakee != p) {
5462 current->last_wakee = p;
5463 current->wakee_flips++;
5468 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5470 * A waker of many should wake a different task than the one last awakened
5471 * at a frequency roughly N times higher than one of its wakees.
5473 * In order to determine whether we should let the load spread vs consolidating
5474 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5475 * partner, and a factor of lls_size higher frequency in the other.
5477 * With both conditions met, we can be relatively sure that the relationship is
5478 * non-monogamous, with partner count exceeding socket size.
5480 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5481 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5484 static int wake_wide(struct task_struct *p)
5486 unsigned int master = current->wakee_flips;
5487 unsigned int slave = p->wakee_flips;
5488 int factor = this_cpu_read(sd_llc_size);
5491 swap(master, slave);
5492 if (slave < factor || master < slave * factor)
5498 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5499 * soonest. For the purpose of speed we only consider the waking and previous
5502 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5503 * cache-affine and is (or will be) idle.
5505 * wake_affine_weight() - considers the weight to reflect the average
5506 * scheduling latency of the CPUs. This seems to work
5507 * for the overloaded case.
5510 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5513 * If this_cpu is idle, it implies the wakeup is from interrupt
5514 * context. Only allow the move if cache is shared. Otherwise an
5515 * interrupt intensive workload could force all tasks onto one
5516 * node depending on the IO topology or IRQ affinity settings.
5518 * If the prev_cpu is idle and cache affine then avoid a migration.
5519 * There is no guarantee that the cache hot data from an interrupt
5520 * is more important than cache hot data on the prev_cpu and from
5521 * a cpufreq perspective, it's better to have higher utilisation
5524 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5525 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5527 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5530 return nr_cpumask_bits;
5534 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5535 int this_cpu, int prev_cpu, int sync)
5537 s64 this_eff_load, prev_eff_load;
5538 unsigned long task_load;
5540 this_eff_load = cpu_runnable_load(cpu_rq(this_cpu));
5543 unsigned long current_load = task_h_load(current);
5545 if (current_load > this_eff_load)
5548 this_eff_load -= current_load;
5551 task_load = task_h_load(p);
5553 this_eff_load += task_load;
5554 if (sched_feat(WA_BIAS))
5555 this_eff_load *= 100;
5556 this_eff_load *= capacity_of(prev_cpu);
5558 prev_eff_load = cpu_runnable_load(cpu_rq(prev_cpu));
5559 prev_eff_load -= task_load;
5560 if (sched_feat(WA_BIAS))
5561 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5562 prev_eff_load *= capacity_of(this_cpu);
5565 * If sync, adjust the weight of prev_eff_load such that if
5566 * prev_eff == this_eff that select_idle_sibling() will consider
5567 * stacking the wakee on top of the waker if no other CPU is
5573 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5576 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5577 int this_cpu, int prev_cpu, int sync)
5579 int target = nr_cpumask_bits;
5581 if (sched_feat(WA_IDLE))
5582 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5584 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5585 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5587 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5588 if (target == nr_cpumask_bits)
5591 schedstat_inc(sd->ttwu_move_affine);
5592 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5596 static unsigned long cpu_util_without(int cpu, struct task_struct *p);
5598 static unsigned long capacity_spare_without(int cpu, struct task_struct *p)
5600 return max_t(long, capacity_of(cpu) - cpu_util_without(cpu, p), 0);
5604 * find_idlest_group finds and returns the least busy CPU group within the
5607 * Assumes p is allowed on at least one CPU in sd.
5609 static struct sched_group *
5610 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5611 int this_cpu, int sd_flag)
5613 struct sched_group *idlest = NULL, *group = sd->groups;
5614 struct sched_group *most_spare_sg = NULL;
5615 unsigned long min_runnable_load = ULONG_MAX;
5616 unsigned long this_runnable_load = ULONG_MAX;
5617 unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5618 unsigned long most_spare = 0, this_spare = 0;
5619 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5620 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5621 (sd->imbalance_pct-100) / 100;
5624 unsigned long load, avg_load, runnable_load;
5625 unsigned long spare_cap, max_spare_cap;
5629 /* Skip over this group if it has no CPUs allowed */
5630 if (!cpumask_intersects(sched_group_span(group),
5634 local_group = cpumask_test_cpu(this_cpu,
5635 sched_group_span(group));
5638 * Tally up the load of all CPUs in the group and find
5639 * the group containing the CPU with most spare capacity.
5645 for_each_cpu(i, sched_group_span(group)) {
5646 load = cpu_runnable_load(cpu_rq(i));
5647 runnable_load += load;
5649 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5651 spare_cap = capacity_spare_without(i, p);
5653 if (spare_cap > max_spare_cap)
5654 max_spare_cap = spare_cap;
5657 /* Adjust by relative CPU capacity of the group */
5658 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5659 group->sgc->capacity;
5660 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5661 group->sgc->capacity;
5664 this_runnable_load = runnable_load;
5665 this_avg_load = avg_load;
5666 this_spare = max_spare_cap;
5668 if (min_runnable_load > (runnable_load + imbalance)) {
5670 * The runnable load is significantly smaller
5671 * so we can pick this new CPU:
5673 min_runnable_load = runnable_load;
5674 min_avg_load = avg_load;
5676 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5677 (100*min_avg_load > imbalance_scale*avg_load)) {
5679 * The runnable loads are close so take the
5680 * blocked load into account through avg_load:
5682 min_avg_load = avg_load;
5686 if (most_spare < max_spare_cap) {
5687 most_spare = max_spare_cap;
5688 most_spare_sg = group;
5691 } while (group = group->next, group != sd->groups);
5694 * The cross-over point between using spare capacity or least load
5695 * is too conservative for high utilization tasks on partially
5696 * utilized systems if we require spare_capacity > task_util(p),
5697 * so we allow for some task stuffing by using
5698 * spare_capacity > task_util(p)/2.
5700 * Spare capacity can't be used for fork because the utilization has
5701 * not been set yet, we must first select a rq to compute the initial
5704 if (sd_flag & SD_BALANCE_FORK)
5707 if (this_spare > task_util(p) / 2 &&
5708 imbalance_scale*this_spare > 100*most_spare)
5711 if (most_spare > task_util(p) / 2)
5712 return most_spare_sg;
5719 * When comparing groups across NUMA domains, it's possible for the
5720 * local domain to be very lightly loaded relative to the remote
5721 * domains but "imbalance" skews the comparison making remote CPUs
5722 * look much more favourable. When considering cross-domain, add
5723 * imbalance to the runnable load on the remote node and consider
5726 if ((sd->flags & SD_NUMA) &&
5727 min_runnable_load + imbalance >= this_runnable_load)
5730 if (min_runnable_load > (this_runnable_load + imbalance))
5733 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5734 (100*this_avg_load < imbalance_scale*min_avg_load))
5741 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5744 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5746 unsigned long load, min_load = ULONG_MAX;
5747 unsigned int min_exit_latency = UINT_MAX;
5748 u64 latest_idle_timestamp = 0;
5749 int least_loaded_cpu = this_cpu;
5750 int shallowest_idle_cpu = -1;
5753 /* Check if we have any choice: */
5754 if (group->group_weight == 1)
5755 return cpumask_first(sched_group_span(group));
5757 /* Traverse only the allowed CPUs */
5758 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
5759 if (available_idle_cpu(i)) {
5760 struct rq *rq = cpu_rq(i);
5761 struct cpuidle_state *idle = idle_get_state(rq);
5762 if (idle && idle->exit_latency < min_exit_latency) {
5764 * We give priority to a CPU whose idle state
5765 * has the smallest exit latency irrespective
5766 * of any idle timestamp.
5768 min_exit_latency = idle->exit_latency;
5769 latest_idle_timestamp = rq->idle_stamp;
5770 shallowest_idle_cpu = i;
5771 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5772 rq->idle_stamp > latest_idle_timestamp) {
5774 * If equal or no active idle state, then
5775 * the most recently idled CPU might have
5778 latest_idle_timestamp = rq->idle_stamp;
5779 shallowest_idle_cpu = i;
5781 } else if (shallowest_idle_cpu == -1) {
5782 load = cpu_runnable_load(cpu_rq(i));
5783 if (load < min_load) {
5785 least_loaded_cpu = i;
5790 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5793 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5794 int cpu, int prev_cpu, int sd_flag)
5798 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
5802 * We need task's util for capacity_spare_without, sync it up to
5803 * prev_cpu's last_update_time.
5805 if (!(sd_flag & SD_BALANCE_FORK))
5806 sync_entity_load_avg(&p->se);
5809 struct sched_group *group;
5810 struct sched_domain *tmp;
5813 if (!(sd->flags & sd_flag)) {
5818 group = find_idlest_group(sd, p, cpu, sd_flag);
5824 new_cpu = find_idlest_group_cpu(group, p, cpu);
5825 if (new_cpu == cpu) {
5826 /* Now try balancing at a lower domain level of 'cpu': */
5831 /* Now try balancing at a lower domain level of 'new_cpu': */
5833 weight = sd->span_weight;
5835 for_each_domain(cpu, tmp) {
5836 if (weight <= tmp->span_weight)
5838 if (tmp->flags & sd_flag)
5846 #ifdef CONFIG_SCHED_SMT
5847 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5848 EXPORT_SYMBOL_GPL(sched_smt_present);
5850 static inline void set_idle_cores(int cpu, int val)
5852 struct sched_domain_shared *sds;
5854 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5856 WRITE_ONCE(sds->has_idle_cores, val);
5859 static inline bool test_idle_cores(int cpu, bool def)
5861 struct sched_domain_shared *sds;
5863 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5865 return READ_ONCE(sds->has_idle_cores);
5871 * Scans the local SMT mask to see if the entire core is idle, and records this
5872 * information in sd_llc_shared->has_idle_cores.
5874 * Since SMT siblings share all cache levels, inspecting this limited remote
5875 * state should be fairly cheap.
5877 void __update_idle_core(struct rq *rq)
5879 int core = cpu_of(rq);
5883 if (test_idle_cores(core, true))
5886 for_each_cpu(cpu, cpu_smt_mask(core)) {
5890 if (!available_idle_cpu(cpu))
5894 set_idle_cores(core, 1);
5900 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5901 * there are no idle cores left in the system; tracked through
5902 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5904 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5906 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5909 if (!static_branch_likely(&sched_smt_present))
5912 if (!test_idle_cores(target, false))
5915 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
5917 for_each_cpu_wrap(core, cpus, target) {
5920 for_each_cpu(cpu, cpu_smt_mask(core)) {
5921 __cpumask_clear_cpu(cpu, cpus);
5922 if (!available_idle_cpu(cpu))
5931 * Failed to find an idle core; stop looking for one.
5933 set_idle_cores(target, 0);
5939 * Scan the local SMT mask for idle CPUs.
5941 static int select_idle_smt(struct task_struct *p, int target)
5945 if (!static_branch_likely(&sched_smt_present))
5948 for_each_cpu(cpu, cpu_smt_mask(target)) {
5949 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
5951 if (available_idle_cpu(cpu))
5958 #else /* CONFIG_SCHED_SMT */
5960 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5965 static inline int select_idle_smt(struct task_struct *p, int target)
5970 #endif /* CONFIG_SCHED_SMT */
5973 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5974 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5975 * average idle time for this rq (as found in rq->avg_idle).
5977 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
5979 struct sched_domain *this_sd;
5980 u64 avg_cost, avg_idle;
5983 int cpu, nr = INT_MAX;
5984 int this = smp_processor_id();
5986 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
5991 * Due to large variance we need a large fuzz factor; hackbench in
5992 * particularly is sensitive here.
5994 avg_idle = this_rq()->avg_idle / 512;
5995 avg_cost = this_sd->avg_scan_cost + 1;
5997 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
6000 if (sched_feat(SIS_PROP)) {
6001 u64 span_avg = sd->span_weight * avg_idle;
6002 if (span_avg > 4*avg_cost)
6003 nr = div_u64(span_avg, avg_cost);
6008 time = cpu_clock(this);
6010 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
6013 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6015 if (available_idle_cpu(cpu))
6019 time = cpu_clock(this) - time;
6020 cost = this_sd->avg_scan_cost;
6021 delta = (s64)(time - cost) / 8;
6022 this_sd->avg_scan_cost += delta;
6028 * Try and locate an idle core/thread in the LLC cache domain.
6030 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6032 struct sched_domain *sd;
6033 int i, recent_used_cpu;
6035 if (available_idle_cpu(target))
6039 * If the previous CPU is cache affine and idle, don't be stupid:
6041 if (prev != target && cpus_share_cache(prev, target) && available_idle_cpu(prev))
6044 /* Check a recently used CPU as a potential idle candidate: */
6045 recent_used_cpu = p->recent_used_cpu;
6046 if (recent_used_cpu != prev &&
6047 recent_used_cpu != target &&
6048 cpus_share_cache(recent_used_cpu, target) &&
6049 available_idle_cpu(recent_used_cpu) &&
6050 cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr)) {
6052 * Replace recent_used_cpu with prev as it is a potential
6053 * candidate for the next wake:
6055 p->recent_used_cpu = prev;
6056 return recent_used_cpu;
6059 sd = rcu_dereference(per_cpu(sd_llc, target));
6063 i = select_idle_core(p, sd, target);
6064 if ((unsigned)i < nr_cpumask_bits)
6067 i = select_idle_cpu(p, sd, target);
6068 if ((unsigned)i < nr_cpumask_bits)
6071 i = select_idle_smt(p, target);
6072 if ((unsigned)i < nr_cpumask_bits)
6079 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6080 * @cpu: the CPU to get the utilization of
6082 * The unit of the return value must be the one of capacity so we can compare
6083 * the utilization with the capacity of the CPU that is available for CFS task
6084 * (ie cpu_capacity).
6086 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6087 * recent utilization of currently non-runnable tasks on a CPU. It represents
6088 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6089 * capacity_orig is the cpu_capacity available at the highest frequency
6090 * (arch_scale_freq_capacity()).
6091 * The utilization of a CPU converges towards a sum equal to or less than the
6092 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6093 * the running time on this CPU scaled by capacity_curr.
6095 * The estimated utilization of a CPU is defined to be the maximum between its
6096 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6097 * currently RUNNABLE on that CPU.
6098 * This allows to properly represent the expected utilization of a CPU which
6099 * has just got a big task running since a long sleep period. At the same time
6100 * however it preserves the benefits of the "blocked utilization" in
6101 * describing the potential for other tasks waking up on the same CPU.
6103 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6104 * higher than capacity_orig because of unfortunate rounding in
6105 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6106 * the average stabilizes with the new running time. We need to check that the
6107 * utilization stays within the range of [0..capacity_orig] and cap it if
6108 * necessary. Without utilization capping, a group could be seen as overloaded
6109 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6110 * available capacity. We allow utilization to overshoot capacity_curr (but not
6111 * capacity_orig) as it useful for predicting the capacity required after task
6112 * migrations (scheduler-driven DVFS).
6114 * Return: the (estimated) utilization for the specified CPU
6116 static inline unsigned long cpu_util(int cpu)
6118 struct cfs_rq *cfs_rq;
6121 cfs_rq = &cpu_rq(cpu)->cfs;
6122 util = READ_ONCE(cfs_rq->avg.util_avg);
6124 if (sched_feat(UTIL_EST))
6125 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6127 return min_t(unsigned long, util, capacity_orig_of(cpu));
6131 * cpu_util_without: compute cpu utilization without any contributions from *p
6132 * @cpu: the CPU which utilization is requested
6133 * @p: the task which utilization should be discounted
6135 * The utilization of a CPU is defined by the utilization of tasks currently
6136 * enqueued on that CPU as well as tasks which are currently sleeping after an
6137 * execution on that CPU.
6139 * This method returns the utilization of the specified CPU by discounting the
6140 * utilization of the specified task, whenever the task is currently
6141 * contributing to the CPU utilization.
6143 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6145 struct cfs_rq *cfs_rq;
6148 /* Task has no contribution or is new */
6149 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6150 return cpu_util(cpu);
6152 cfs_rq = &cpu_rq(cpu)->cfs;
6153 util = READ_ONCE(cfs_rq->avg.util_avg);
6155 /* Discount task's util from CPU's util */
6156 lsub_positive(&util, task_util(p));
6161 * a) if *p is the only task sleeping on this CPU, then:
6162 * cpu_util (== task_util) > util_est (== 0)
6163 * and thus we return:
6164 * cpu_util_without = (cpu_util - task_util) = 0
6166 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6168 * cpu_util >= task_util
6169 * cpu_util > util_est (== 0)
6170 * and thus we discount *p's blocked utilization to return:
6171 * cpu_util_without = (cpu_util - task_util) >= 0
6173 * c) if other tasks are RUNNABLE on that CPU and
6174 * util_est > cpu_util
6175 * then we use util_est since it returns a more restrictive
6176 * estimation of the spare capacity on that CPU, by just
6177 * considering the expected utilization of tasks already
6178 * runnable on that CPU.
6180 * Cases a) and b) are covered by the above code, while case c) is
6181 * covered by the following code when estimated utilization is
6184 if (sched_feat(UTIL_EST)) {
6185 unsigned int estimated =
6186 READ_ONCE(cfs_rq->avg.util_est.enqueued);
6189 * Despite the following checks we still have a small window
6190 * for a possible race, when an execl's select_task_rq_fair()
6191 * races with LB's detach_task():
6194 * p->on_rq = TASK_ON_RQ_MIGRATING;
6195 * ---------------------------------- A
6196 * deactivate_task() \
6197 * dequeue_task() + RaceTime
6198 * util_est_dequeue() /
6199 * ---------------------------------- B
6201 * The additional check on "current == p" it's required to
6202 * properly fix the execl regression and it helps in further
6203 * reducing the chances for the above race.
6205 if (unlikely(task_on_rq_queued(p) || current == p))
6206 lsub_positive(&estimated, _task_util_est(p));
6208 util = max(util, estimated);
6212 * Utilization (estimated) can exceed the CPU capacity, thus let's
6213 * clamp to the maximum CPU capacity to ensure consistency with
6214 * the cpu_util call.
6216 return min_t(unsigned long, util, capacity_orig_of(cpu));
6220 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6221 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6223 * In that case WAKE_AFFINE doesn't make sense and we'll let
6224 * BALANCE_WAKE sort things out.
6226 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6228 long min_cap, max_cap;
6230 if (!static_branch_unlikely(&sched_asym_cpucapacity))
6233 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6234 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6236 /* Minimum capacity is close to max, no need to abort wake_affine */
6237 if (max_cap - min_cap < max_cap >> 3)
6240 /* Bring task utilization in sync with prev_cpu */
6241 sync_entity_load_avg(&p->se);
6243 return !task_fits_capacity(p, min_cap);
6247 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6250 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6252 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6253 unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
6256 * If @p migrates from @cpu to another, remove its contribution. Or,
6257 * if @p migrates from another CPU to @cpu, add its contribution. In
6258 * the other cases, @cpu is not impacted by the migration, so the
6259 * util_avg should already be correct.
6261 if (task_cpu(p) == cpu && dst_cpu != cpu)
6262 sub_positive(&util, task_util(p));
6263 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6264 util += task_util(p);
6266 if (sched_feat(UTIL_EST)) {
6267 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6270 * During wake-up, the task isn't enqueued yet and doesn't
6271 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6272 * so just add it (if needed) to "simulate" what will be
6273 * cpu_util() after the task has been enqueued.
6276 util_est += _task_util_est(p);
6278 util = max(util, util_est);
6281 return min(util, capacity_orig_of(cpu));
6285 * compute_energy(): Estimates the energy that would be consumed if @p was
6286 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6287 * landscape of the * CPUs after the task migration, and uses the Energy Model
6288 * to compute what would be the energy if we decided to actually migrate that
6292 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
6294 unsigned int max_util, util_cfs, cpu_util, cpu_cap;
6295 unsigned long sum_util, energy = 0;
6296 struct task_struct *tsk;
6299 for (; pd; pd = pd->next) {
6300 struct cpumask *pd_mask = perf_domain_span(pd);
6303 * The energy model mandates all the CPUs of a performance
6304 * domain have the same capacity.
6306 cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask));
6307 max_util = sum_util = 0;
6310 * The capacity state of CPUs of the current rd can be driven by
6311 * CPUs of another rd if they belong to the same performance
6312 * domain. So, account for the utilization of these CPUs too
6313 * by masking pd with cpu_online_mask instead of the rd span.
6315 * If an entire performance domain is outside of the current rd,
6316 * it will not appear in its pd list and will not be accounted
6317 * by compute_energy().
6319 for_each_cpu_and(cpu, pd_mask, cpu_online_mask) {
6320 util_cfs = cpu_util_next(cpu, p, dst_cpu);
6323 * Busy time computation: utilization clamping is not
6324 * required since the ratio (sum_util / cpu_capacity)
6325 * is already enough to scale the EM reported power
6326 * consumption at the (eventually clamped) cpu_capacity.
6328 sum_util += schedutil_cpu_util(cpu, util_cfs, cpu_cap,
6332 * Performance domain frequency: utilization clamping
6333 * must be considered since it affects the selection
6334 * of the performance domain frequency.
6335 * NOTE: in case RT tasks are running, by default the
6336 * FREQUENCY_UTIL's utilization can be max OPP.
6338 tsk = cpu == dst_cpu ? p : NULL;
6339 cpu_util = schedutil_cpu_util(cpu, util_cfs, cpu_cap,
6340 FREQUENCY_UTIL, tsk);
6341 max_util = max(max_util, cpu_util);
6344 energy += em_pd_energy(pd->em_pd, max_util, sum_util);
6351 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6352 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6353 * spare capacity in each performance domain and uses it as a potential
6354 * candidate to execute the task. Then, it uses the Energy Model to figure
6355 * out which of the CPU candidates is the most energy-efficient.
6357 * The rationale for this heuristic is as follows. In a performance domain,
6358 * all the most energy efficient CPU candidates (according to the Energy
6359 * Model) are those for which we'll request a low frequency. When there are
6360 * several CPUs for which the frequency request will be the same, we don't
6361 * have enough data to break the tie between them, because the Energy Model
6362 * only includes active power costs. With this model, if we assume that
6363 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6364 * the maximum spare capacity in a performance domain is guaranteed to be among
6365 * the best candidates of the performance domain.
6367 * In practice, it could be preferable from an energy standpoint to pack
6368 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6369 * but that could also hurt our chances to go cluster idle, and we have no
6370 * ways to tell with the current Energy Model if this is actually a good
6371 * idea or not. So, find_energy_efficient_cpu() basically favors
6372 * cluster-packing, and spreading inside a cluster. That should at least be
6373 * a good thing for latency, and this is consistent with the idea that most
6374 * of the energy savings of EAS come from the asymmetry of the system, and
6375 * not so much from breaking the tie between identical CPUs. That's also the
6376 * reason why EAS is enabled in the topology code only for systems where
6377 * SD_ASYM_CPUCAPACITY is set.
6379 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6380 * they don't have any useful utilization data yet and it's not possible to
6381 * forecast their impact on energy consumption. Consequently, they will be
6382 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6383 * to be energy-inefficient in some use-cases. The alternative would be to
6384 * bias new tasks towards specific types of CPUs first, or to try to infer
6385 * their util_avg from the parent task, but those heuristics could hurt
6386 * other use-cases too. So, until someone finds a better way to solve this,
6387 * let's keep things simple by re-using the existing slow path.
6390 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6392 unsigned long prev_energy = ULONG_MAX, best_energy = ULONG_MAX;
6393 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6394 int cpu, best_energy_cpu = prev_cpu;
6395 struct perf_domain *head, *pd;
6396 unsigned long cpu_cap, util;
6397 struct sched_domain *sd;
6400 pd = rcu_dereference(rd->pd);
6401 if (!pd || READ_ONCE(rd->overutilized))
6406 * Energy-aware wake-up happens on the lowest sched_domain starting
6407 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6409 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6410 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6415 sync_entity_load_avg(&p->se);
6416 if (!task_util_est(p))
6419 for (; pd; pd = pd->next) {
6420 unsigned long cur_energy, spare_cap, max_spare_cap = 0;
6421 int max_spare_cap_cpu = -1;
6423 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
6424 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6427 /* Skip CPUs that will be overutilized. */
6428 util = cpu_util_next(cpu, p, cpu);
6429 cpu_cap = capacity_of(cpu);
6430 if (cpu_cap * 1024 < util * capacity_margin)
6433 /* Always use prev_cpu as a candidate. */
6434 if (cpu == prev_cpu) {
6435 prev_energy = compute_energy(p, prev_cpu, head);
6436 best_energy = min(best_energy, prev_energy);
6441 * Find the CPU with the maximum spare capacity in
6442 * the performance domain
6444 spare_cap = cpu_cap - util;
6445 if (spare_cap > max_spare_cap) {
6446 max_spare_cap = spare_cap;
6447 max_spare_cap_cpu = cpu;
6451 /* Evaluate the energy impact of using this CPU. */
6452 if (max_spare_cap_cpu >= 0) {
6453 cur_energy = compute_energy(p, max_spare_cap_cpu, head);
6454 if (cur_energy < best_energy) {
6455 best_energy = cur_energy;
6456 best_energy_cpu = max_spare_cap_cpu;
6464 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6465 * least 6% of the energy used by prev_cpu.
6467 if (prev_energy == ULONG_MAX)
6468 return best_energy_cpu;
6470 if ((prev_energy - best_energy) > (prev_energy >> 4))
6471 return best_energy_cpu;
6482 * select_task_rq_fair: Select target runqueue for the waking task in domains
6483 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6484 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6486 * Balances load by selecting the idlest CPU in the idlest group, or under
6487 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6489 * Returns the target CPU number.
6491 * preempt must be disabled.
6494 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6496 struct sched_domain *tmp, *sd = NULL;
6497 int cpu = smp_processor_id();
6498 int new_cpu = prev_cpu;
6499 int want_affine = 0;
6500 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6502 if (sd_flag & SD_BALANCE_WAKE) {
6505 if (sched_energy_enabled()) {
6506 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
6512 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu) &&
6513 cpumask_test_cpu(cpu, p->cpus_ptr);
6517 for_each_domain(cpu, tmp) {
6518 if (!(tmp->flags & SD_LOAD_BALANCE))
6522 * If both 'cpu' and 'prev_cpu' are part of this domain,
6523 * cpu is a valid SD_WAKE_AFFINE target.
6525 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6526 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6527 if (cpu != prev_cpu)
6528 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6530 sd = NULL; /* Prefer wake_affine over balance flags */
6534 if (tmp->flags & sd_flag)
6536 else if (!want_affine)
6542 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6543 } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6546 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6549 current->recent_used_cpu = cpu;
6556 static void detach_entity_cfs_rq(struct sched_entity *se);
6559 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6560 * cfs_rq_of(p) references at time of call are still valid and identify the
6561 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6563 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6566 * As blocked tasks retain absolute vruntime the migration needs to
6567 * deal with this by subtracting the old and adding the new
6568 * min_vruntime -- the latter is done by enqueue_entity() when placing
6569 * the task on the new runqueue.
6571 if (p->state == TASK_WAKING) {
6572 struct sched_entity *se = &p->se;
6573 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6576 #ifndef CONFIG_64BIT
6577 u64 min_vruntime_copy;
6580 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6582 min_vruntime = cfs_rq->min_vruntime;
6583 } while (min_vruntime != min_vruntime_copy);
6585 min_vruntime = cfs_rq->min_vruntime;
6588 se->vruntime -= min_vruntime;
6591 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6593 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6594 * rq->lock and can modify state directly.
6596 lockdep_assert_held(&task_rq(p)->lock);
6597 detach_entity_cfs_rq(&p->se);
6601 * We are supposed to update the task to "current" time, then
6602 * its up to date and ready to go to new CPU/cfs_rq. But we
6603 * have difficulty in getting what current time is, so simply
6604 * throw away the out-of-date time. This will result in the
6605 * wakee task is less decayed, but giving the wakee more load
6608 remove_entity_load_avg(&p->se);
6611 /* Tell new CPU we are migrated */
6612 p->se.avg.last_update_time = 0;
6614 /* We have migrated, no longer consider this task hot */
6615 p->se.exec_start = 0;
6617 update_scan_period(p, new_cpu);
6620 static void task_dead_fair(struct task_struct *p)
6622 remove_entity_load_avg(&p->se);
6624 #endif /* CONFIG_SMP */
6626 static unsigned long wakeup_gran(struct sched_entity *se)
6628 unsigned long gran = sysctl_sched_wakeup_granularity;
6631 * Since its curr running now, convert the gran from real-time
6632 * to virtual-time in his units.
6634 * By using 'se' instead of 'curr' we penalize light tasks, so
6635 * they get preempted easier. That is, if 'se' < 'curr' then
6636 * the resulting gran will be larger, therefore penalizing the
6637 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6638 * be smaller, again penalizing the lighter task.
6640 * This is especially important for buddies when the leftmost
6641 * task is higher priority than the buddy.
6643 return calc_delta_fair(gran, se);
6647 * Should 'se' preempt 'curr'.
6661 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6663 s64 gran, vdiff = curr->vruntime - se->vruntime;
6668 gran = wakeup_gran(se);
6675 static void set_last_buddy(struct sched_entity *se)
6677 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6680 for_each_sched_entity(se) {
6681 if (SCHED_WARN_ON(!se->on_rq))
6683 cfs_rq_of(se)->last = se;
6687 static void set_next_buddy(struct sched_entity *se)
6689 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6692 for_each_sched_entity(se) {
6693 if (SCHED_WARN_ON(!se->on_rq))
6695 cfs_rq_of(se)->next = se;
6699 static void set_skip_buddy(struct sched_entity *se)
6701 for_each_sched_entity(se)
6702 cfs_rq_of(se)->skip = se;
6706 * Preempt the current task with a newly woken task if needed:
6708 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6710 struct task_struct *curr = rq->curr;
6711 struct sched_entity *se = &curr->se, *pse = &p->se;
6712 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6713 int scale = cfs_rq->nr_running >= sched_nr_latency;
6714 int next_buddy_marked = 0;
6716 if (unlikely(se == pse))
6720 * This is possible from callers such as attach_tasks(), in which we
6721 * unconditionally check_prempt_curr() after an enqueue (which may have
6722 * lead to a throttle). This both saves work and prevents false
6723 * next-buddy nomination below.
6725 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6728 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6729 set_next_buddy(pse);
6730 next_buddy_marked = 1;
6734 * We can come here with TIF_NEED_RESCHED already set from new task
6737 * Note: this also catches the edge-case of curr being in a throttled
6738 * group (e.g. via set_curr_task), since update_curr() (in the
6739 * enqueue of curr) will have resulted in resched being set. This
6740 * prevents us from potentially nominating it as a false LAST_BUDDY
6743 if (test_tsk_need_resched(curr))
6746 /* Idle tasks are by definition preempted by non-idle tasks. */
6747 if (unlikely(task_has_idle_policy(curr)) &&
6748 likely(!task_has_idle_policy(p)))
6752 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6753 * is driven by the tick):
6755 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6758 find_matching_se(&se, &pse);
6759 update_curr(cfs_rq_of(se));
6761 if (wakeup_preempt_entity(se, pse) == 1) {
6763 * Bias pick_next to pick the sched entity that is
6764 * triggering this preemption.
6766 if (!next_buddy_marked)
6767 set_next_buddy(pse);
6776 * Only set the backward buddy when the current task is still
6777 * on the rq. This can happen when a wakeup gets interleaved
6778 * with schedule on the ->pre_schedule() or idle_balance()
6779 * point, either of which can * drop the rq lock.
6781 * Also, during early boot the idle thread is in the fair class,
6782 * for obvious reasons its a bad idea to schedule back to it.
6784 if (unlikely(!se->on_rq || curr == rq->idle))
6787 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6791 static struct task_struct *
6792 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6794 struct cfs_rq *cfs_rq = &rq->cfs;
6795 struct sched_entity *se;
6796 struct task_struct *p;
6800 if (!cfs_rq->nr_running)
6803 #ifdef CONFIG_FAIR_GROUP_SCHED
6804 if (prev->sched_class != &fair_sched_class)
6808 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6809 * likely that a next task is from the same cgroup as the current.
6811 * Therefore attempt to avoid putting and setting the entire cgroup
6812 * hierarchy, only change the part that actually changes.
6816 struct sched_entity *curr = cfs_rq->curr;
6819 * Since we got here without doing put_prev_entity() we also
6820 * have to consider cfs_rq->curr. If it is still a runnable
6821 * entity, update_curr() will update its vruntime, otherwise
6822 * forget we've ever seen it.
6826 update_curr(cfs_rq);
6831 * This call to check_cfs_rq_runtime() will do the
6832 * throttle and dequeue its entity in the parent(s).
6833 * Therefore the nr_running test will indeed
6836 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6839 if (!cfs_rq->nr_running)
6846 se = pick_next_entity(cfs_rq, curr);
6847 cfs_rq = group_cfs_rq(se);
6853 * Since we haven't yet done put_prev_entity and if the selected task
6854 * is a different task than we started out with, try and touch the
6855 * least amount of cfs_rqs.
6858 struct sched_entity *pse = &prev->se;
6860 while (!(cfs_rq = is_same_group(se, pse))) {
6861 int se_depth = se->depth;
6862 int pse_depth = pse->depth;
6864 if (se_depth <= pse_depth) {
6865 put_prev_entity(cfs_rq_of(pse), pse);
6866 pse = parent_entity(pse);
6868 if (se_depth >= pse_depth) {
6869 set_next_entity(cfs_rq_of(se), se);
6870 se = parent_entity(se);
6874 put_prev_entity(cfs_rq, pse);
6875 set_next_entity(cfs_rq, se);
6882 put_prev_task(rq, prev);
6885 se = pick_next_entity(cfs_rq, NULL);
6886 set_next_entity(cfs_rq, se);
6887 cfs_rq = group_cfs_rq(se);
6892 done: __maybe_unused;
6895 * Move the next running task to the front of
6896 * the list, so our cfs_tasks list becomes MRU
6899 list_move(&p->se.group_node, &rq->cfs_tasks);
6902 if (hrtick_enabled(rq))
6903 hrtick_start_fair(rq, p);
6905 update_misfit_status(p, rq);
6910 update_misfit_status(NULL, rq);
6911 new_tasks = idle_balance(rq, rf);
6914 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6915 * possible for any higher priority task to appear. In that case we
6916 * must re-start the pick_next_entity() loop.
6925 * rq is about to be idle, check if we need to update the
6926 * lost_idle_time of clock_pelt
6928 update_idle_rq_clock_pelt(rq);
6934 * Account for a descheduled task:
6936 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6938 struct sched_entity *se = &prev->se;
6939 struct cfs_rq *cfs_rq;
6941 for_each_sched_entity(se) {
6942 cfs_rq = cfs_rq_of(se);
6943 put_prev_entity(cfs_rq, se);
6948 * sched_yield() is very simple
6950 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6952 static void yield_task_fair(struct rq *rq)
6954 struct task_struct *curr = rq->curr;
6955 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6956 struct sched_entity *se = &curr->se;
6959 * Are we the only task in the tree?
6961 if (unlikely(rq->nr_running == 1))
6964 clear_buddies(cfs_rq, se);
6966 if (curr->policy != SCHED_BATCH) {
6967 update_rq_clock(rq);
6969 * Update run-time statistics of the 'current'.
6971 update_curr(cfs_rq);
6973 * Tell update_rq_clock() that we've just updated,
6974 * so we don't do microscopic update in schedule()
6975 * and double the fastpath cost.
6977 rq_clock_skip_update(rq);
6983 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6985 struct sched_entity *se = &p->se;
6987 /* throttled hierarchies are not runnable */
6988 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6991 /* Tell the scheduler that we'd really like pse to run next. */
6994 yield_task_fair(rq);
7000 /**************************************************
7001 * Fair scheduling class load-balancing methods.
7005 * The purpose of load-balancing is to achieve the same basic fairness the
7006 * per-CPU scheduler provides, namely provide a proportional amount of compute
7007 * time to each task. This is expressed in the following equation:
7009 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7011 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7012 * W_i,0 is defined as:
7014 * W_i,0 = \Sum_j w_i,j (2)
7016 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7017 * is derived from the nice value as per sched_prio_to_weight[].
7019 * The weight average is an exponential decay average of the instantaneous
7022 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7024 * C_i is the compute capacity of CPU i, typically it is the
7025 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7026 * can also include other factors [XXX].
7028 * To achieve this balance we define a measure of imbalance which follows
7029 * directly from (1):
7031 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7033 * We them move tasks around to minimize the imbalance. In the continuous
7034 * function space it is obvious this converges, in the discrete case we get
7035 * a few fun cases generally called infeasible weight scenarios.
7038 * - infeasible weights;
7039 * - local vs global optima in the discrete case. ]
7044 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7045 * for all i,j solution, we create a tree of CPUs that follows the hardware
7046 * topology where each level pairs two lower groups (or better). This results
7047 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7048 * tree to only the first of the previous level and we decrease the frequency
7049 * of load-balance at each level inv. proportional to the number of CPUs in
7055 * \Sum { --- * --- * 2^i } = O(n) (5)
7057 * `- size of each group
7058 * | | `- number of CPUs doing load-balance
7060 * `- sum over all levels
7062 * Coupled with a limit on how many tasks we can migrate every balance pass,
7063 * this makes (5) the runtime complexity of the balancer.
7065 * An important property here is that each CPU is still (indirectly) connected
7066 * to every other CPU in at most O(log n) steps:
7068 * The adjacency matrix of the resulting graph is given by:
7071 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7074 * And you'll find that:
7076 * A^(log_2 n)_i,j != 0 for all i,j (7)
7078 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7079 * The task movement gives a factor of O(m), giving a convergence complexity
7082 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7087 * In order to avoid CPUs going idle while there's still work to do, new idle
7088 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7089 * tree itself instead of relying on other CPUs to bring it work.
7091 * This adds some complexity to both (5) and (8) but it reduces the total idle
7099 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7102 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7107 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7109 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7111 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7114 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7115 * rewrite all of this once again.]
7118 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7120 enum fbq_type { regular, remote, all };
7129 #define LBF_ALL_PINNED 0x01
7130 #define LBF_NEED_BREAK 0x02
7131 #define LBF_DST_PINNED 0x04
7132 #define LBF_SOME_PINNED 0x08
7133 #define LBF_NOHZ_STATS 0x10
7134 #define LBF_NOHZ_AGAIN 0x20
7137 struct sched_domain *sd;
7145 struct cpumask *dst_grpmask;
7147 enum cpu_idle_type idle;
7149 /* The set of CPUs under consideration for load-balancing */
7150 struct cpumask *cpus;
7155 unsigned int loop_break;
7156 unsigned int loop_max;
7158 enum fbq_type fbq_type;
7159 enum group_type src_grp_type;
7160 struct list_head tasks;
7164 * Is this task likely cache-hot:
7166 static int task_hot(struct task_struct *p, struct lb_env *env)
7170 lockdep_assert_held(&env->src_rq->lock);
7172 if (p->sched_class != &fair_sched_class)
7175 if (unlikely(task_has_idle_policy(p)))
7179 * Buddy candidates are cache hot:
7181 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7182 (&p->se == cfs_rq_of(&p->se)->next ||
7183 &p->se == cfs_rq_of(&p->se)->last))
7186 if (sysctl_sched_migration_cost == -1)
7188 if (sysctl_sched_migration_cost == 0)
7191 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7193 return delta < (s64)sysctl_sched_migration_cost;
7196 #ifdef CONFIG_NUMA_BALANCING
7198 * Returns 1, if task migration degrades locality
7199 * Returns 0, if task migration improves locality i.e migration preferred.
7200 * Returns -1, if task migration is not affected by locality.
7202 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7204 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7205 unsigned long src_weight, dst_weight;
7206 int src_nid, dst_nid, dist;
7208 if (!static_branch_likely(&sched_numa_balancing))
7211 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7214 src_nid = cpu_to_node(env->src_cpu);
7215 dst_nid = cpu_to_node(env->dst_cpu);
7217 if (src_nid == dst_nid)
7220 /* Migrating away from the preferred node is always bad. */
7221 if (src_nid == p->numa_preferred_nid) {
7222 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7228 /* Encourage migration to the preferred node. */
7229 if (dst_nid == p->numa_preferred_nid)
7232 /* Leaving a core idle is often worse than degrading locality. */
7233 if (env->idle == CPU_IDLE)
7236 dist = node_distance(src_nid, dst_nid);
7238 src_weight = group_weight(p, src_nid, dist);
7239 dst_weight = group_weight(p, dst_nid, dist);
7241 src_weight = task_weight(p, src_nid, dist);
7242 dst_weight = task_weight(p, dst_nid, dist);
7245 return dst_weight < src_weight;
7249 static inline int migrate_degrades_locality(struct task_struct *p,
7257 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7260 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7264 lockdep_assert_held(&env->src_rq->lock);
7267 * We do not migrate tasks that are:
7268 * 1) throttled_lb_pair, or
7269 * 2) cannot be migrated to this CPU due to cpus_ptr, or
7270 * 3) running (obviously), or
7271 * 4) are cache-hot on their current CPU.
7273 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7276 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
7279 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7281 env->flags |= LBF_SOME_PINNED;
7284 * Remember if this task can be migrated to any other CPU in
7285 * our sched_group. We may want to revisit it if we couldn't
7286 * meet load balance goals by pulling other tasks on src_cpu.
7288 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7289 * already computed one in current iteration.
7291 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7294 /* Prevent to re-select dst_cpu via env's CPUs: */
7295 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7296 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
7297 env->flags |= LBF_DST_PINNED;
7298 env->new_dst_cpu = cpu;
7306 /* Record that we found atleast one task that could run on dst_cpu */
7307 env->flags &= ~LBF_ALL_PINNED;
7309 if (task_running(env->src_rq, p)) {
7310 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7315 * Aggressive migration if:
7316 * 1) destination numa is preferred
7317 * 2) task is cache cold, or
7318 * 3) too many balance attempts have failed.
7320 tsk_cache_hot = migrate_degrades_locality(p, env);
7321 if (tsk_cache_hot == -1)
7322 tsk_cache_hot = task_hot(p, env);
7324 if (tsk_cache_hot <= 0 ||
7325 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7326 if (tsk_cache_hot == 1) {
7327 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7328 schedstat_inc(p->se.statistics.nr_forced_migrations);
7333 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7338 * detach_task() -- detach the task for the migration specified in env
7340 static void detach_task(struct task_struct *p, struct lb_env *env)
7342 lockdep_assert_held(&env->src_rq->lock);
7344 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7345 set_task_cpu(p, env->dst_cpu);
7349 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7350 * part of active balancing operations within "domain".
7352 * Returns a task if successful and NULL otherwise.
7354 static struct task_struct *detach_one_task(struct lb_env *env)
7356 struct task_struct *p;
7358 lockdep_assert_held(&env->src_rq->lock);
7360 list_for_each_entry_reverse(p,
7361 &env->src_rq->cfs_tasks, se.group_node) {
7362 if (!can_migrate_task(p, env))
7365 detach_task(p, env);
7368 * Right now, this is only the second place where
7369 * lb_gained[env->idle] is updated (other is detach_tasks)
7370 * so we can safely collect stats here rather than
7371 * inside detach_tasks().
7373 schedstat_inc(env->sd->lb_gained[env->idle]);
7379 static const unsigned int sched_nr_migrate_break = 32;
7382 * detach_tasks() -- tries to detach up to imbalance runnable load from
7383 * busiest_rq, as part of a balancing operation within domain "sd".
7385 * Returns number of detached tasks if successful and 0 otherwise.
7387 static int detach_tasks(struct lb_env *env)
7389 struct list_head *tasks = &env->src_rq->cfs_tasks;
7390 struct task_struct *p;
7394 lockdep_assert_held(&env->src_rq->lock);
7396 if (env->imbalance <= 0)
7399 while (!list_empty(tasks)) {
7401 * We don't want to steal all, otherwise we may be treated likewise,
7402 * which could at worst lead to a livelock crash.
7404 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7407 p = list_last_entry(tasks, struct task_struct, se.group_node);
7410 /* We've more or less seen every task there is, call it quits */
7411 if (env->loop > env->loop_max)
7414 /* take a breather every nr_migrate tasks */
7415 if (env->loop > env->loop_break) {
7416 env->loop_break += sched_nr_migrate_break;
7417 env->flags |= LBF_NEED_BREAK;
7421 if (!can_migrate_task(p, env))
7424 load = task_h_load(p);
7426 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7429 if ((load / 2) > env->imbalance)
7432 detach_task(p, env);
7433 list_add(&p->se.group_node, &env->tasks);
7436 env->imbalance -= load;
7438 #ifdef CONFIG_PREEMPT
7440 * NEWIDLE balancing is a source of latency, so preemptible
7441 * kernels will stop after the first task is detached to minimize
7442 * the critical section.
7444 if (env->idle == CPU_NEWLY_IDLE)
7449 * We only want to steal up to the prescribed amount of
7452 if (env->imbalance <= 0)
7457 list_move(&p->se.group_node, tasks);
7461 * Right now, this is one of only two places we collect this stat
7462 * so we can safely collect detach_one_task() stats here rather
7463 * than inside detach_one_task().
7465 schedstat_add(env->sd->lb_gained[env->idle], detached);
7471 * attach_task() -- attach the task detached by detach_task() to its new rq.
7473 static void attach_task(struct rq *rq, struct task_struct *p)
7475 lockdep_assert_held(&rq->lock);
7477 BUG_ON(task_rq(p) != rq);
7478 activate_task(rq, p, ENQUEUE_NOCLOCK);
7479 check_preempt_curr(rq, p, 0);
7483 * attach_one_task() -- attaches the task returned from detach_one_task() to
7486 static void attach_one_task(struct rq *rq, struct task_struct *p)
7491 update_rq_clock(rq);
7497 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7500 static void attach_tasks(struct lb_env *env)
7502 struct list_head *tasks = &env->tasks;
7503 struct task_struct *p;
7506 rq_lock(env->dst_rq, &rf);
7507 update_rq_clock(env->dst_rq);
7509 while (!list_empty(tasks)) {
7510 p = list_first_entry(tasks, struct task_struct, se.group_node);
7511 list_del_init(&p->se.group_node);
7513 attach_task(env->dst_rq, p);
7516 rq_unlock(env->dst_rq, &rf);
7519 #ifdef CONFIG_NO_HZ_COMMON
7520 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7522 if (cfs_rq->avg.load_avg)
7525 if (cfs_rq->avg.util_avg)
7531 static inline bool others_have_blocked(struct rq *rq)
7533 if (READ_ONCE(rq->avg_rt.util_avg))
7536 if (READ_ONCE(rq->avg_dl.util_avg))
7539 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7540 if (READ_ONCE(rq->avg_irq.util_avg))
7547 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
7549 rq->last_blocked_load_update_tick = jiffies;
7552 rq->has_blocked_load = 0;
7555 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
7556 static inline bool others_have_blocked(struct rq *rq) { return false; }
7557 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
7560 #ifdef CONFIG_FAIR_GROUP_SCHED
7562 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7564 if (cfs_rq->load.weight)
7567 if (cfs_rq->avg.load_sum)
7570 if (cfs_rq->avg.util_sum)
7573 if (cfs_rq->avg.runnable_load_sum)
7579 static void update_blocked_averages(int cpu)
7581 struct rq *rq = cpu_rq(cpu);
7582 struct cfs_rq *cfs_rq, *pos;
7583 const struct sched_class *curr_class;
7587 rq_lock_irqsave(rq, &rf);
7588 update_rq_clock(rq);
7591 * Iterates the task_group tree in a bottom up fashion, see
7592 * list_add_leaf_cfs_rq() for details.
7594 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7595 struct sched_entity *se;
7597 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq))
7598 update_tg_load_avg(cfs_rq, 0);
7600 /* Propagate pending load changes to the parent, if any: */
7601 se = cfs_rq->tg->se[cpu];
7602 if (se && !skip_blocked_update(se))
7603 update_load_avg(cfs_rq_of(se), se, 0);
7606 * There can be a lot of idle CPU cgroups. Don't let fully
7607 * decayed cfs_rqs linger on the list.
7609 if (cfs_rq_is_decayed(cfs_rq))
7610 list_del_leaf_cfs_rq(cfs_rq);
7612 /* Don't need periodic decay once load/util_avg are null */
7613 if (cfs_rq_has_blocked(cfs_rq))
7617 curr_class = rq->curr->sched_class;
7618 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class);
7619 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class);
7620 update_irq_load_avg(rq, 0);
7621 /* Don't need periodic decay once load/util_avg are null */
7622 if (others_have_blocked(rq))
7625 update_blocked_load_status(rq, !done);
7626 rq_unlock_irqrestore(rq, &rf);
7630 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7631 * This needs to be done in a top-down fashion because the load of a child
7632 * group is a fraction of its parents load.
7634 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7636 struct rq *rq = rq_of(cfs_rq);
7637 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7638 unsigned long now = jiffies;
7641 if (cfs_rq->last_h_load_update == now)
7644 WRITE_ONCE(cfs_rq->h_load_next, NULL);
7645 for_each_sched_entity(se) {
7646 cfs_rq = cfs_rq_of(se);
7647 WRITE_ONCE(cfs_rq->h_load_next, se);
7648 if (cfs_rq->last_h_load_update == now)
7653 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7654 cfs_rq->last_h_load_update = now;
7657 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
7658 load = cfs_rq->h_load;
7659 load = div64_ul(load * se->avg.load_avg,
7660 cfs_rq_load_avg(cfs_rq) + 1);
7661 cfs_rq = group_cfs_rq(se);
7662 cfs_rq->h_load = load;
7663 cfs_rq->last_h_load_update = now;
7667 static unsigned long task_h_load(struct task_struct *p)
7669 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7671 update_cfs_rq_h_load(cfs_rq);
7672 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7673 cfs_rq_load_avg(cfs_rq) + 1);
7676 static inline void update_blocked_averages(int cpu)
7678 struct rq *rq = cpu_rq(cpu);
7679 struct cfs_rq *cfs_rq = &rq->cfs;
7680 const struct sched_class *curr_class;
7683 rq_lock_irqsave(rq, &rf);
7684 update_rq_clock(rq);
7685 update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
7687 curr_class = rq->curr->sched_class;
7688 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class);
7689 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class);
7690 update_irq_load_avg(rq, 0);
7691 update_blocked_load_status(rq, cfs_rq_has_blocked(cfs_rq) || others_have_blocked(rq));
7692 rq_unlock_irqrestore(rq, &rf);
7695 static unsigned long task_h_load(struct task_struct *p)
7697 return p->se.avg.load_avg;
7701 /********** Helpers for find_busiest_group ************************/
7704 * sg_lb_stats - stats of a sched_group required for load_balancing
7706 struct sg_lb_stats {
7707 unsigned long avg_load; /*Avg load across the CPUs of the group */
7708 unsigned long group_load; /* Total load over the CPUs of the group */
7709 unsigned long load_per_task;
7710 unsigned long group_capacity;
7711 unsigned long group_util; /* Total utilization of the group */
7712 unsigned int sum_nr_running; /* Nr tasks running in the group */
7713 unsigned int idle_cpus;
7714 unsigned int group_weight;
7715 enum group_type group_type;
7716 int group_no_capacity;
7717 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
7718 #ifdef CONFIG_NUMA_BALANCING
7719 unsigned int nr_numa_running;
7720 unsigned int nr_preferred_running;
7725 * sd_lb_stats - Structure to store the statistics of a sched_domain
7726 * during load balancing.
7728 struct sd_lb_stats {
7729 struct sched_group *busiest; /* Busiest group in this sd */
7730 struct sched_group *local; /* Local group in this sd */
7731 unsigned long total_running;
7732 unsigned long total_load; /* Total load of all groups in sd */
7733 unsigned long total_capacity; /* Total capacity of all groups in sd */
7734 unsigned long avg_load; /* Average load across all groups in sd */
7736 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7737 struct sg_lb_stats local_stat; /* Statistics of the local group */
7740 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7743 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7744 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7745 * We must however clear busiest_stat::avg_load because
7746 * update_sd_pick_busiest() reads this before assignment.
7748 *sds = (struct sd_lb_stats){
7751 .total_running = 0UL,
7753 .total_capacity = 0UL,
7756 .sum_nr_running = 0,
7757 .group_type = group_other,
7762 static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu)
7764 struct rq *rq = cpu_rq(cpu);
7765 unsigned long max = arch_scale_cpu_capacity(cpu);
7766 unsigned long used, free;
7769 irq = cpu_util_irq(rq);
7771 if (unlikely(irq >= max))
7774 used = READ_ONCE(rq->avg_rt.util_avg);
7775 used += READ_ONCE(rq->avg_dl.util_avg);
7777 if (unlikely(used >= max))
7782 return scale_irq_capacity(free, irq, max);
7785 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7787 unsigned long capacity = scale_rt_capacity(sd, cpu);
7788 struct sched_group *sdg = sd->groups;
7790 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
7795 cpu_rq(cpu)->cpu_capacity = capacity;
7796 sdg->sgc->capacity = capacity;
7797 sdg->sgc->min_capacity = capacity;
7798 sdg->sgc->max_capacity = capacity;
7801 void update_group_capacity(struct sched_domain *sd, int cpu)
7803 struct sched_domain *child = sd->child;
7804 struct sched_group *group, *sdg = sd->groups;
7805 unsigned long capacity, min_capacity, max_capacity;
7806 unsigned long interval;
7808 interval = msecs_to_jiffies(sd->balance_interval);
7809 interval = clamp(interval, 1UL, max_load_balance_interval);
7810 sdg->sgc->next_update = jiffies + interval;
7813 update_cpu_capacity(sd, cpu);
7818 min_capacity = ULONG_MAX;
7821 if (child->flags & SD_OVERLAP) {
7823 * SD_OVERLAP domains cannot assume that child groups
7824 * span the current group.
7827 for_each_cpu(cpu, sched_group_span(sdg)) {
7828 struct sched_group_capacity *sgc;
7829 struct rq *rq = cpu_rq(cpu);
7832 * build_sched_domains() -> init_sched_groups_capacity()
7833 * gets here before we've attached the domains to the
7836 * Use capacity_of(), which is set irrespective of domains
7837 * in update_cpu_capacity().
7839 * This avoids capacity from being 0 and
7840 * causing divide-by-zero issues on boot.
7842 if (unlikely(!rq->sd)) {
7843 capacity += capacity_of(cpu);
7845 sgc = rq->sd->groups->sgc;
7846 capacity += sgc->capacity;
7849 min_capacity = min(capacity, min_capacity);
7850 max_capacity = max(capacity, max_capacity);
7854 * !SD_OVERLAP domains can assume that child groups
7855 * span the current group.
7858 group = child->groups;
7860 struct sched_group_capacity *sgc = group->sgc;
7862 capacity += sgc->capacity;
7863 min_capacity = min(sgc->min_capacity, min_capacity);
7864 max_capacity = max(sgc->max_capacity, max_capacity);
7865 group = group->next;
7866 } while (group != child->groups);
7869 sdg->sgc->capacity = capacity;
7870 sdg->sgc->min_capacity = min_capacity;
7871 sdg->sgc->max_capacity = max_capacity;
7875 * Check whether the capacity of the rq has been noticeably reduced by side
7876 * activity. The imbalance_pct is used for the threshold.
7877 * Return true is the capacity is reduced
7880 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7882 return ((rq->cpu_capacity * sd->imbalance_pct) <
7883 (rq->cpu_capacity_orig * 100));
7887 * Check whether a rq has a misfit task and if it looks like we can actually
7888 * help that task: we can migrate the task to a CPU of higher capacity, or
7889 * the task's current CPU is heavily pressured.
7891 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
7893 return rq->misfit_task_load &&
7894 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
7895 check_cpu_capacity(rq, sd));
7899 * Group imbalance indicates (and tries to solve) the problem where balancing
7900 * groups is inadequate due to ->cpus_ptr constraints.
7902 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
7903 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
7906 * { 0 1 2 3 } { 4 5 6 7 }
7909 * If we were to balance group-wise we'd place two tasks in the first group and
7910 * two tasks in the second group. Clearly this is undesired as it will overload
7911 * cpu 3 and leave one of the CPUs in the second group unused.
7913 * The current solution to this issue is detecting the skew in the first group
7914 * by noticing the lower domain failed to reach balance and had difficulty
7915 * moving tasks due to affinity constraints.
7917 * When this is so detected; this group becomes a candidate for busiest; see
7918 * update_sd_pick_busiest(). And calculate_imbalance() and
7919 * find_busiest_group() avoid some of the usual balance conditions to allow it
7920 * to create an effective group imbalance.
7922 * This is a somewhat tricky proposition since the next run might not find the
7923 * group imbalance and decide the groups need to be balanced again. A most
7924 * subtle and fragile situation.
7927 static inline int sg_imbalanced(struct sched_group *group)
7929 return group->sgc->imbalance;
7933 * group_has_capacity returns true if the group has spare capacity that could
7934 * be used by some tasks.
7935 * We consider that a group has spare capacity if the * number of task is
7936 * smaller than the number of CPUs or if the utilization is lower than the
7937 * available capacity for CFS tasks.
7938 * For the latter, we use a threshold to stabilize the state, to take into
7939 * account the variance of the tasks' load and to return true if the available
7940 * capacity in meaningful for the load balancer.
7941 * As an example, an available capacity of 1% can appear but it doesn't make
7942 * any benefit for the load balance.
7945 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7947 if (sgs->sum_nr_running < sgs->group_weight)
7950 if ((sgs->group_capacity * 100) >
7951 (sgs->group_util * env->sd->imbalance_pct))
7958 * group_is_overloaded returns true if the group has more tasks than it can
7960 * group_is_overloaded is not equals to !group_has_capacity because a group
7961 * with the exact right number of tasks, has no more spare capacity but is not
7962 * overloaded so both group_has_capacity and group_is_overloaded return
7966 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7968 if (sgs->sum_nr_running <= sgs->group_weight)
7971 if ((sgs->group_capacity * 100) <
7972 (sgs->group_util * env->sd->imbalance_pct))
7979 * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller
7980 * per-CPU capacity than sched_group ref.
7983 group_smaller_min_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7985 return sg->sgc->min_capacity * capacity_margin <
7986 ref->sgc->min_capacity * 1024;
7990 * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller
7991 * per-CPU capacity_orig than sched_group ref.
7994 group_smaller_max_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7996 return sg->sgc->max_capacity * capacity_margin <
7997 ref->sgc->max_capacity * 1024;
8001 group_type group_classify(struct sched_group *group,
8002 struct sg_lb_stats *sgs)
8004 if (sgs->group_no_capacity)
8005 return group_overloaded;
8007 if (sg_imbalanced(group))
8008 return group_imbalanced;
8010 if (sgs->group_misfit_task_load)
8011 return group_misfit_task;
8016 static bool update_nohz_stats(struct rq *rq, bool force)
8018 #ifdef CONFIG_NO_HZ_COMMON
8019 unsigned int cpu = rq->cpu;
8021 if (!rq->has_blocked_load)
8024 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
8027 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
8030 update_blocked_averages(cpu);
8032 return rq->has_blocked_load;
8039 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8040 * @env: The load balancing environment.
8041 * @group: sched_group whose statistics are to be updated.
8042 * @sgs: variable to hold the statistics for this group.
8043 * @sg_status: Holds flag indicating the status of the sched_group
8045 static inline void update_sg_lb_stats(struct lb_env *env,
8046 struct sched_group *group,
8047 struct sg_lb_stats *sgs,
8052 memset(sgs, 0, sizeof(*sgs));
8054 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8055 struct rq *rq = cpu_rq(i);
8057 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
8058 env->flags |= LBF_NOHZ_AGAIN;
8060 sgs->group_load += cpu_runnable_load(rq);
8061 sgs->group_util += cpu_util(i);
8062 sgs->sum_nr_running += rq->cfs.h_nr_running;
8064 nr_running = rq->nr_running;
8066 *sg_status |= SG_OVERLOAD;
8068 if (cpu_overutilized(i))
8069 *sg_status |= SG_OVERUTILIZED;
8071 #ifdef CONFIG_NUMA_BALANCING
8072 sgs->nr_numa_running += rq->nr_numa_running;
8073 sgs->nr_preferred_running += rq->nr_preferred_running;
8076 * No need to call idle_cpu() if nr_running is not 0
8078 if (!nr_running && idle_cpu(i))
8081 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8082 sgs->group_misfit_task_load < rq->misfit_task_load) {
8083 sgs->group_misfit_task_load = rq->misfit_task_load;
8084 *sg_status |= SG_OVERLOAD;
8088 /* Adjust by relative CPU capacity of the group */
8089 sgs->group_capacity = group->sgc->capacity;
8090 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
8092 if (sgs->sum_nr_running)
8093 sgs->load_per_task = sgs->group_load / sgs->sum_nr_running;
8095 sgs->group_weight = group->group_weight;
8097 sgs->group_no_capacity = group_is_overloaded(env, sgs);
8098 sgs->group_type = group_classify(group, sgs);
8102 * update_sd_pick_busiest - return 1 on busiest group
8103 * @env: The load balancing environment.
8104 * @sds: sched_domain statistics
8105 * @sg: sched_group candidate to be checked for being the busiest
8106 * @sgs: sched_group statistics
8108 * Determine if @sg is a busier group than the previously selected
8111 * Return: %true if @sg is a busier group than the previously selected
8112 * busiest group. %false otherwise.
8114 static bool update_sd_pick_busiest(struct lb_env *env,
8115 struct sd_lb_stats *sds,
8116 struct sched_group *sg,
8117 struct sg_lb_stats *sgs)
8119 struct sg_lb_stats *busiest = &sds->busiest_stat;
8122 * Don't try to pull misfit tasks we can't help.
8123 * We can use max_capacity here as reduction in capacity on some
8124 * CPUs in the group should either be possible to resolve
8125 * internally or be covered by avg_load imbalance (eventually).
8127 if (sgs->group_type == group_misfit_task &&
8128 (!group_smaller_max_cpu_capacity(sg, sds->local) ||
8129 !group_has_capacity(env, &sds->local_stat)))
8132 if (sgs->group_type > busiest->group_type)
8135 if (sgs->group_type < busiest->group_type)
8138 if (sgs->avg_load <= busiest->avg_load)
8141 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
8145 * Candidate sg has no more than one task per CPU and
8146 * has higher per-CPU capacity. Migrating tasks to less
8147 * capable CPUs may harm throughput. Maximize throughput,
8148 * power/energy consequences are not considered.
8150 if (sgs->sum_nr_running <= sgs->group_weight &&
8151 group_smaller_min_cpu_capacity(sds->local, sg))
8155 * If we have more than one misfit sg go with the biggest misfit.
8157 if (sgs->group_type == group_misfit_task &&
8158 sgs->group_misfit_task_load < busiest->group_misfit_task_load)
8162 /* This is the busiest node in its class. */
8163 if (!(env->sd->flags & SD_ASYM_PACKING))
8166 /* No ASYM_PACKING if target CPU is already busy */
8167 if (env->idle == CPU_NOT_IDLE)
8170 * ASYM_PACKING needs to move all the work to the highest
8171 * prority CPUs in the group, therefore mark all groups
8172 * of lower priority than ourself as busy.
8174 if (sgs->sum_nr_running &&
8175 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
8179 /* Prefer to move from lowest priority CPU's work */
8180 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
8181 sg->asym_prefer_cpu))
8188 #ifdef CONFIG_NUMA_BALANCING
8189 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8191 if (sgs->sum_nr_running > sgs->nr_numa_running)
8193 if (sgs->sum_nr_running > sgs->nr_preferred_running)
8198 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8200 if (rq->nr_running > rq->nr_numa_running)
8202 if (rq->nr_running > rq->nr_preferred_running)
8207 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8212 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8216 #endif /* CONFIG_NUMA_BALANCING */
8219 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8220 * @env: The load balancing environment.
8221 * @sds: variable to hold the statistics for this sched_domain.
8223 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8225 struct sched_domain *child = env->sd->child;
8226 struct sched_group *sg = env->sd->groups;
8227 struct sg_lb_stats *local = &sds->local_stat;
8228 struct sg_lb_stats tmp_sgs;
8229 bool prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
8232 #ifdef CONFIG_NO_HZ_COMMON
8233 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
8234 env->flags |= LBF_NOHZ_STATS;
8238 struct sg_lb_stats *sgs = &tmp_sgs;
8241 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
8246 if (env->idle != CPU_NEWLY_IDLE ||
8247 time_after_eq(jiffies, sg->sgc->next_update))
8248 update_group_capacity(env->sd, env->dst_cpu);
8251 update_sg_lb_stats(env, sg, sgs, &sg_status);
8257 * In case the child domain prefers tasks go to siblings
8258 * first, lower the sg capacity so that we'll try
8259 * and move all the excess tasks away. We lower the capacity
8260 * of a group only if the local group has the capacity to fit
8261 * these excess tasks. The extra check prevents the case where
8262 * you always pull from the heaviest group when it is already
8263 * under-utilized (possible with a large weight task outweighs
8264 * the tasks on the system).
8266 if (prefer_sibling && sds->local &&
8267 group_has_capacity(env, local) &&
8268 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
8269 sgs->group_no_capacity = 1;
8270 sgs->group_type = group_classify(sg, sgs);
8273 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8275 sds->busiest_stat = *sgs;
8279 /* Now, start updating sd_lb_stats */
8280 sds->total_running += sgs->sum_nr_running;
8281 sds->total_load += sgs->group_load;
8282 sds->total_capacity += sgs->group_capacity;
8285 } while (sg != env->sd->groups);
8287 #ifdef CONFIG_NO_HZ_COMMON
8288 if ((env->flags & LBF_NOHZ_AGAIN) &&
8289 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8291 WRITE_ONCE(nohz.next_blocked,
8292 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8296 if (env->sd->flags & SD_NUMA)
8297 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8299 if (!env->sd->parent) {
8300 struct root_domain *rd = env->dst_rq->rd;
8302 /* update overload indicator if we are at root domain */
8303 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
8305 /* Update over-utilization (tipping point, U >= 0) indicator */
8306 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
8307 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
8308 } else if (sg_status & SG_OVERUTILIZED) {
8309 struct root_domain *rd = env->dst_rq->rd;
8311 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
8312 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
8317 * check_asym_packing - Check to see if the group is packed into the
8320 * This is primarily intended to used at the sibling level. Some
8321 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8322 * case of POWER7, it can move to lower SMT modes only when higher
8323 * threads are idle. When in lower SMT modes, the threads will
8324 * perform better since they share less core resources. Hence when we
8325 * have idle threads, we want them to be the higher ones.
8327 * This packing function is run on idle threads. It checks to see if
8328 * the busiest CPU in this domain (core in the P7 case) has a higher
8329 * CPU number than the packing function is being run on. Here we are
8330 * assuming lower CPU number will be equivalent to lower a SMT thread
8333 * Return: 1 when packing is required and a task should be moved to
8334 * this CPU. The amount of the imbalance is returned in env->imbalance.
8336 * @env: The load balancing environment.
8337 * @sds: Statistics of the sched_domain which is to be packed
8339 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8343 if (!(env->sd->flags & SD_ASYM_PACKING))
8346 if (env->idle == CPU_NOT_IDLE)
8352 busiest_cpu = sds->busiest->asym_prefer_cpu;
8353 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
8356 env->imbalance = sds->busiest_stat.group_load;
8362 * fix_small_imbalance - Calculate the minor imbalance that exists
8363 * amongst the groups of a sched_domain, during
8365 * @env: The load balancing environment.
8366 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8369 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8371 unsigned long tmp, capa_now = 0, capa_move = 0;
8372 unsigned int imbn = 2;
8373 unsigned long scaled_busy_load_per_task;
8374 struct sg_lb_stats *local, *busiest;
8376 local = &sds->local_stat;
8377 busiest = &sds->busiest_stat;
8379 if (!local->sum_nr_running)
8380 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8381 else if (busiest->load_per_task > local->load_per_task)
8384 scaled_busy_load_per_task =
8385 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8386 busiest->group_capacity;
8388 if (busiest->avg_load + scaled_busy_load_per_task >=
8389 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8390 env->imbalance = busiest->load_per_task;
8395 * OK, we don't have enough imbalance to justify moving tasks,
8396 * however we may be able to increase total CPU capacity used by
8400 capa_now += busiest->group_capacity *
8401 min(busiest->load_per_task, busiest->avg_load);
8402 capa_now += local->group_capacity *
8403 min(local->load_per_task, local->avg_load);
8404 capa_now /= SCHED_CAPACITY_SCALE;
8406 /* Amount of load we'd subtract */
8407 if (busiest->avg_load > scaled_busy_load_per_task) {
8408 capa_move += busiest->group_capacity *
8409 min(busiest->load_per_task,
8410 busiest->avg_load - scaled_busy_load_per_task);
8413 /* Amount of load we'd add */
8414 if (busiest->avg_load * busiest->group_capacity <
8415 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8416 tmp = (busiest->avg_load * busiest->group_capacity) /
8417 local->group_capacity;
8419 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8420 local->group_capacity;
8422 capa_move += local->group_capacity *
8423 min(local->load_per_task, local->avg_load + tmp);
8424 capa_move /= SCHED_CAPACITY_SCALE;
8426 /* Move if we gain throughput */
8427 if (capa_move > capa_now)
8428 env->imbalance = busiest->load_per_task;
8432 * calculate_imbalance - Calculate the amount of imbalance present within the
8433 * groups of a given sched_domain during load balance.
8434 * @env: load balance environment
8435 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8437 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8439 unsigned long max_pull, load_above_capacity = ~0UL;
8440 struct sg_lb_stats *local, *busiest;
8442 local = &sds->local_stat;
8443 busiest = &sds->busiest_stat;
8445 if (busiest->group_type == group_imbalanced) {
8447 * In the group_imb case we cannot rely on group-wide averages
8448 * to ensure CPU-load equilibrium, look at wider averages. XXX
8450 busiest->load_per_task =
8451 min(busiest->load_per_task, sds->avg_load);
8455 * Avg load of busiest sg can be less and avg load of local sg can
8456 * be greater than avg load across all sgs of sd because avg load
8457 * factors in sg capacity and sgs with smaller group_type are
8458 * skipped when updating the busiest sg:
8460 if (busiest->group_type != group_misfit_task &&
8461 (busiest->avg_load <= sds->avg_load ||
8462 local->avg_load >= sds->avg_load)) {
8464 return fix_small_imbalance(env, sds);
8468 * If there aren't any idle CPUs, avoid creating some.
8470 if (busiest->group_type == group_overloaded &&
8471 local->group_type == group_overloaded) {
8472 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
8473 if (load_above_capacity > busiest->group_capacity) {
8474 load_above_capacity -= busiest->group_capacity;
8475 load_above_capacity *= scale_load_down(NICE_0_LOAD);
8476 load_above_capacity /= busiest->group_capacity;
8478 load_above_capacity = ~0UL;
8482 * We're trying to get all the CPUs to the average_load, so we don't
8483 * want to push ourselves above the average load, nor do we wish to
8484 * reduce the max loaded CPU below the average load. At the same time,
8485 * we also don't want to reduce the group load below the group
8486 * capacity. Thus we look for the minimum possible imbalance.
8488 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8490 /* How much load to actually move to equalise the imbalance */
8491 env->imbalance = min(
8492 max_pull * busiest->group_capacity,
8493 (sds->avg_load - local->avg_load) * local->group_capacity
8494 ) / SCHED_CAPACITY_SCALE;
8496 /* Boost imbalance to allow misfit task to be balanced. */
8497 if (busiest->group_type == group_misfit_task) {
8498 env->imbalance = max_t(long, env->imbalance,
8499 busiest->group_misfit_task_load);
8503 * if *imbalance is less than the average load per runnable task
8504 * there is no guarantee that any tasks will be moved so we'll have
8505 * a think about bumping its value to force at least one task to be
8508 if (env->imbalance < busiest->load_per_task)
8509 return fix_small_imbalance(env, sds);
8512 /******* find_busiest_group() helpers end here *********************/
8515 * find_busiest_group - Returns the busiest group within the sched_domain
8516 * if there is an imbalance.
8518 * Also calculates the amount of runnable load which should be moved
8519 * to restore balance.
8521 * @env: The load balancing environment.
8523 * Return: - The busiest group if imbalance exists.
8525 static struct sched_group *find_busiest_group(struct lb_env *env)
8527 struct sg_lb_stats *local, *busiest;
8528 struct sd_lb_stats sds;
8530 init_sd_lb_stats(&sds);
8533 * Compute the various statistics relavent for load balancing at
8536 update_sd_lb_stats(env, &sds);
8538 if (sched_energy_enabled()) {
8539 struct root_domain *rd = env->dst_rq->rd;
8541 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
8545 local = &sds.local_stat;
8546 busiest = &sds.busiest_stat;
8548 /* ASYM feature bypasses nice load balance check */
8549 if (check_asym_packing(env, &sds))
8552 /* There is no busy sibling group to pull tasks from */
8553 if (!sds.busiest || busiest->sum_nr_running == 0)
8556 /* XXX broken for overlapping NUMA groups */
8557 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8558 / sds.total_capacity;
8561 * If the busiest group is imbalanced the below checks don't
8562 * work because they assume all things are equal, which typically
8563 * isn't true due to cpus_ptr constraints and the like.
8565 if (busiest->group_type == group_imbalanced)
8569 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8570 * capacities from resulting in underutilization due to avg_load.
8572 if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8573 busiest->group_no_capacity)
8576 /* Misfit tasks should be dealt with regardless of the avg load */
8577 if (busiest->group_type == group_misfit_task)
8581 * If the local group is busier than the selected busiest group
8582 * don't try and pull any tasks.
8584 if (local->avg_load >= busiest->avg_load)
8588 * Don't pull any tasks if this group is already above the domain
8591 if (local->avg_load >= sds.avg_load)
8594 if (env->idle == CPU_IDLE) {
8596 * This CPU is idle. If the busiest group is not overloaded
8597 * and there is no imbalance between this and busiest group
8598 * wrt idle CPUs, it is balanced. The imbalance becomes
8599 * significant if the diff is greater than 1 otherwise we
8600 * might end up to just move the imbalance on another group
8602 if ((busiest->group_type != group_overloaded) &&
8603 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8607 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8608 * imbalance_pct to be conservative.
8610 if (100 * busiest->avg_load <=
8611 env->sd->imbalance_pct * local->avg_load)
8616 /* Looks like there is an imbalance. Compute it */
8617 env->src_grp_type = busiest->group_type;
8618 calculate_imbalance(env, &sds);
8619 return env->imbalance ? sds.busiest : NULL;
8627 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
8629 static struct rq *find_busiest_queue(struct lb_env *env,
8630 struct sched_group *group)
8632 struct rq *busiest = NULL, *rq;
8633 unsigned long busiest_load = 0, busiest_capacity = 1;
8636 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8637 unsigned long capacity, load;
8641 rt = fbq_classify_rq(rq);
8644 * We classify groups/runqueues into three groups:
8645 * - regular: there are !numa tasks
8646 * - remote: there are numa tasks that run on the 'wrong' node
8647 * - all: there is no distinction
8649 * In order to avoid migrating ideally placed numa tasks,
8650 * ignore those when there's better options.
8652 * If we ignore the actual busiest queue to migrate another
8653 * task, the next balance pass can still reduce the busiest
8654 * queue by moving tasks around inside the node.
8656 * If we cannot move enough load due to this classification
8657 * the next pass will adjust the group classification and
8658 * allow migration of more tasks.
8660 * Both cases only affect the total convergence complexity.
8662 if (rt > env->fbq_type)
8666 * For ASYM_CPUCAPACITY domains with misfit tasks we simply
8667 * seek the "biggest" misfit task.
8669 if (env->src_grp_type == group_misfit_task) {
8670 if (rq->misfit_task_load > busiest_load) {
8671 busiest_load = rq->misfit_task_load;
8678 capacity = capacity_of(i);
8681 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
8682 * eventually lead to active_balancing high->low capacity.
8683 * Higher per-CPU capacity is considered better than balancing
8686 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8687 capacity_of(env->dst_cpu) < capacity &&
8688 rq->nr_running == 1)
8691 load = cpu_runnable_load(rq);
8694 * When comparing with imbalance, use cpu_runnable_load()
8695 * which is not scaled with the CPU capacity.
8698 if (rq->nr_running == 1 && load > env->imbalance &&
8699 !check_cpu_capacity(rq, env->sd))
8703 * For the load comparisons with the other CPU's, consider
8704 * the cpu_runnable_load() scaled with the CPU capacity, so
8705 * that the load can be moved away from the CPU that is
8706 * potentially running at a lower capacity.
8708 * Thus we're looking for max(load_i / capacity_i), crosswise
8709 * multiplication to rid ourselves of the division works out
8710 * to: load_i * capacity_j > load_j * capacity_i; where j is
8711 * our previous maximum.
8713 if (load * busiest_capacity > busiest_load * capacity) {
8714 busiest_load = load;
8715 busiest_capacity = capacity;
8724 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8725 * so long as it is large enough.
8727 #define MAX_PINNED_INTERVAL 512
8730 asym_active_balance(struct lb_env *env)
8733 * ASYM_PACKING needs to force migrate tasks from busy but
8734 * lower priority CPUs in order to pack all tasks in the
8735 * highest priority CPUs.
8737 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
8738 sched_asym_prefer(env->dst_cpu, env->src_cpu);
8742 voluntary_active_balance(struct lb_env *env)
8744 struct sched_domain *sd = env->sd;
8746 if (asym_active_balance(env))
8750 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8751 * It's worth migrating the task if the src_cpu's capacity is reduced
8752 * because of other sched_class or IRQs if more capacity stays
8753 * available on dst_cpu.
8755 if ((env->idle != CPU_NOT_IDLE) &&
8756 (env->src_rq->cfs.h_nr_running == 1)) {
8757 if ((check_cpu_capacity(env->src_rq, sd)) &&
8758 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8762 if (env->src_grp_type == group_misfit_task)
8768 static int need_active_balance(struct lb_env *env)
8770 struct sched_domain *sd = env->sd;
8772 if (voluntary_active_balance(env))
8775 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8778 static int active_load_balance_cpu_stop(void *data);
8780 static int should_we_balance(struct lb_env *env)
8782 struct sched_group *sg = env->sd->groups;
8783 int cpu, balance_cpu = -1;
8786 * Ensure the balancing environment is consistent; can happen
8787 * when the softirq triggers 'during' hotplug.
8789 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8793 * In the newly idle case, we will allow all the CPUs
8794 * to do the newly idle load balance.
8796 if (env->idle == CPU_NEWLY_IDLE)
8799 /* Try to find first idle CPU */
8800 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8808 if (balance_cpu == -1)
8809 balance_cpu = group_balance_cpu(sg);
8812 * First idle CPU or the first CPU(busiest) in this sched group
8813 * is eligible for doing load balancing at this and above domains.
8815 return balance_cpu == env->dst_cpu;
8819 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8820 * tasks if there is an imbalance.
8822 static int load_balance(int this_cpu, struct rq *this_rq,
8823 struct sched_domain *sd, enum cpu_idle_type idle,
8824 int *continue_balancing)
8826 int ld_moved, cur_ld_moved, active_balance = 0;
8827 struct sched_domain *sd_parent = sd->parent;
8828 struct sched_group *group;
8831 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8833 struct lb_env env = {
8835 .dst_cpu = this_cpu,
8837 .dst_grpmask = sched_group_span(sd->groups),
8839 .loop_break = sched_nr_migrate_break,
8842 .tasks = LIST_HEAD_INIT(env.tasks),
8845 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8847 schedstat_inc(sd->lb_count[idle]);
8850 if (!should_we_balance(&env)) {
8851 *continue_balancing = 0;
8855 group = find_busiest_group(&env);
8857 schedstat_inc(sd->lb_nobusyg[idle]);
8861 busiest = find_busiest_queue(&env, group);
8863 schedstat_inc(sd->lb_nobusyq[idle]);
8867 BUG_ON(busiest == env.dst_rq);
8869 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8871 env.src_cpu = busiest->cpu;
8872 env.src_rq = busiest;
8875 if (busiest->nr_running > 1) {
8877 * Attempt to move tasks. If find_busiest_group has found
8878 * an imbalance but busiest->nr_running <= 1, the group is
8879 * still unbalanced. ld_moved simply stays zero, so it is
8880 * correctly treated as an imbalance.
8882 env.flags |= LBF_ALL_PINNED;
8883 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8886 rq_lock_irqsave(busiest, &rf);
8887 update_rq_clock(busiest);
8890 * cur_ld_moved - load moved in current iteration
8891 * ld_moved - cumulative load moved across iterations
8893 cur_ld_moved = detach_tasks(&env);
8896 * We've detached some tasks from busiest_rq. Every
8897 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8898 * unlock busiest->lock, and we are able to be sure
8899 * that nobody can manipulate the tasks in parallel.
8900 * See task_rq_lock() family for the details.
8903 rq_unlock(busiest, &rf);
8907 ld_moved += cur_ld_moved;
8910 local_irq_restore(rf.flags);
8912 if (env.flags & LBF_NEED_BREAK) {
8913 env.flags &= ~LBF_NEED_BREAK;
8918 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8919 * us and move them to an alternate dst_cpu in our sched_group
8920 * where they can run. The upper limit on how many times we
8921 * iterate on same src_cpu is dependent on number of CPUs in our
8924 * This changes load balance semantics a bit on who can move
8925 * load to a given_cpu. In addition to the given_cpu itself
8926 * (or a ilb_cpu acting on its behalf where given_cpu is
8927 * nohz-idle), we now have balance_cpu in a position to move
8928 * load to given_cpu. In rare situations, this may cause
8929 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8930 * _independently_ and at _same_ time to move some load to
8931 * given_cpu) causing exceess load to be moved to given_cpu.
8932 * This however should not happen so much in practice and
8933 * moreover subsequent load balance cycles should correct the
8934 * excess load moved.
8936 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8938 /* Prevent to re-select dst_cpu via env's CPUs */
8939 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
8941 env.dst_rq = cpu_rq(env.new_dst_cpu);
8942 env.dst_cpu = env.new_dst_cpu;
8943 env.flags &= ~LBF_DST_PINNED;
8945 env.loop_break = sched_nr_migrate_break;
8948 * Go back to "more_balance" rather than "redo" since we
8949 * need to continue with same src_cpu.
8955 * We failed to reach balance because of affinity.
8958 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8960 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8961 *group_imbalance = 1;
8964 /* All tasks on this runqueue were pinned by CPU affinity */
8965 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8966 __cpumask_clear_cpu(cpu_of(busiest), cpus);
8968 * Attempting to continue load balancing at the current
8969 * sched_domain level only makes sense if there are
8970 * active CPUs remaining as possible busiest CPUs to
8971 * pull load from which are not contained within the
8972 * destination group that is receiving any migrated
8975 if (!cpumask_subset(cpus, env.dst_grpmask)) {
8977 env.loop_break = sched_nr_migrate_break;
8980 goto out_all_pinned;
8985 schedstat_inc(sd->lb_failed[idle]);
8987 * Increment the failure counter only on periodic balance.
8988 * We do not want newidle balance, which can be very
8989 * frequent, pollute the failure counter causing
8990 * excessive cache_hot migrations and active balances.
8992 if (idle != CPU_NEWLY_IDLE)
8993 sd->nr_balance_failed++;
8995 if (need_active_balance(&env)) {
8996 unsigned long flags;
8998 raw_spin_lock_irqsave(&busiest->lock, flags);
9001 * Don't kick the active_load_balance_cpu_stop,
9002 * if the curr task on busiest CPU can't be
9003 * moved to this_cpu:
9005 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
9006 raw_spin_unlock_irqrestore(&busiest->lock,
9008 env.flags |= LBF_ALL_PINNED;
9009 goto out_one_pinned;
9013 * ->active_balance synchronizes accesses to
9014 * ->active_balance_work. Once set, it's cleared
9015 * only after active load balance is finished.
9017 if (!busiest->active_balance) {
9018 busiest->active_balance = 1;
9019 busiest->push_cpu = this_cpu;
9022 raw_spin_unlock_irqrestore(&busiest->lock, flags);
9024 if (active_balance) {
9025 stop_one_cpu_nowait(cpu_of(busiest),
9026 active_load_balance_cpu_stop, busiest,
9027 &busiest->active_balance_work);
9030 /* We've kicked active balancing, force task migration. */
9031 sd->nr_balance_failed = sd->cache_nice_tries+1;
9034 sd->nr_balance_failed = 0;
9036 if (likely(!active_balance) || voluntary_active_balance(&env)) {
9037 /* We were unbalanced, so reset the balancing interval */
9038 sd->balance_interval = sd->min_interval;
9041 * If we've begun active balancing, start to back off. This
9042 * case may not be covered by the all_pinned logic if there
9043 * is only 1 task on the busy runqueue (because we don't call
9046 if (sd->balance_interval < sd->max_interval)
9047 sd->balance_interval *= 2;
9054 * We reach balance although we may have faced some affinity
9055 * constraints. Clear the imbalance flag if it was set.
9058 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9060 if (*group_imbalance)
9061 *group_imbalance = 0;
9066 * We reach balance because all tasks are pinned at this level so
9067 * we can't migrate them. Let the imbalance flag set so parent level
9068 * can try to migrate them.
9070 schedstat_inc(sd->lb_balanced[idle]);
9072 sd->nr_balance_failed = 0;
9078 * idle_balance() disregards balance intervals, so we could repeatedly
9079 * reach this code, which would lead to balance_interval skyrocketting
9080 * in a short amount of time. Skip the balance_interval increase logic
9083 if (env.idle == CPU_NEWLY_IDLE)
9086 /* tune up the balancing interval */
9087 if ((env.flags & LBF_ALL_PINNED &&
9088 sd->balance_interval < MAX_PINNED_INTERVAL) ||
9089 sd->balance_interval < sd->max_interval)
9090 sd->balance_interval *= 2;
9095 static inline unsigned long
9096 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
9098 unsigned long interval = sd->balance_interval;
9101 interval *= sd->busy_factor;
9103 /* scale ms to jiffies */
9104 interval = msecs_to_jiffies(interval);
9105 interval = clamp(interval, 1UL, max_load_balance_interval);
9111 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
9113 unsigned long interval, next;
9115 /* used by idle balance, so cpu_busy = 0 */
9116 interval = get_sd_balance_interval(sd, 0);
9117 next = sd->last_balance + interval;
9119 if (time_after(*next_balance, next))
9120 *next_balance = next;
9124 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9125 * running tasks off the busiest CPU onto idle CPUs. It requires at
9126 * least 1 task to be running on each physical CPU where possible, and
9127 * avoids physical / logical imbalances.
9129 static int active_load_balance_cpu_stop(void *data)
9131 struct rq *busiest_rq = data;
9132 int busiest_cpu = cpu_of(busiest_rq);
9133 int target_cpu = busiest_rq->push_cpu;
9134 struct rq *target_rq = cpu_rq(target_cpu);
9135 struct sched_domain *sd;
9136 struct task_struct *p = NULL;
9139 rq_lock_irq(busiest_rq, &rf);
9141 * Between queueing the stop-work and running it is a hole in which
9142 * CPUs can become inactive. We should not move tasks from or to
9145 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
9148 /* Make sure the requested CPU hasn't gone down in the meantime: */
9149 if (unlikely(busiest_cpu != smp_processor_id() ||
9150 !busiest_rq->active_balance))
9153 /* Is there any task to move? */
9154 if (busiest_rq->nr_running <= 1)
9158 * This condition is "impossible", if it occurs
9159 * we need to fix it. Originally reported by
9160 * Bjorn Helgaas on a 128-CPU setup.
9162 BUG_ON(busiest_rq == target_rq);
9164 /* Search for an sd spanning us and the target CPU. */
9166 for_each_domain(target_cpu, sd) {
9167 if ((sd->flags & SD_LOAD_BALANCE) &&
9168 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9173 struct lb_env env = {
9175 .dst_cpu = target_cpu,
9176 .dst_rq = target_rq,
9177 .src_cpu = busiest_rq->cpu,
9178 .src_rq = busiest_rq,
9181 * can_migrate_task() doesn't need to compute new_dst_cpu
9182 * for active balancing. Since we have CPU_IDLE, but no
9183 * @dst_grpmask we need to make that test go away with lying
9186 .flags = LBF_DST_PINNED,
9189 schedstat_inc(sd->alb_count);
9190 update_rq_clock(busiest_rq);
9192 p = detach_one_task(&env);
9194 schedstat_inc(sd->alb_pushed);
9195 /* Active balancing done, reset the failure counter. */
9196 sd->nr_balance_failed = 0;
9198 schedstat_inc(sd->alb_failed);
9203 busiest_rq->active_balance = 0;
9204 rq_unlock(busiest_rq, &rf);
9207 attach_one_task(target_rq, p);
9214 static DEFINE_SPINLOCK(balancing);
9217 * Scale the max load_balance interval with the number of CPUs in the system.
9218 * This trades load-balance latency on larger machines for less cross talk.
9220 void update_max_interval(void)
9222 max_load_balance_interval = HZ*num_online_cpus()/10;
9226 * It checks each scheduling domain to see if it is due to be balanced,
9227 * and initiates a balancing operation if so.
9229 * Balancing parameters are set up in init_sched_domains.
9231 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9233 int continue_balancing = 1;
9235 unsigned long interval;
9236 struct sched_domain *sd;
9237 /* Earliest time when we have to do rebalance again */
9238 unsigned long next_balance = jiffies + 60*HZ;
9239 int update_next_balance = 0;
9240 int need_serialize, need_decay = 0;
9244 for_each_domain(cpu, sd) {
9246 * Decay the newidle max times here because this is a regular
9247 * visit to all the domains. Decay ~1% per second.
9249 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9250 sd->max_newidle_lb_cost =
9251 (sd->max_newidle_lb_cost * 253) / 256;
9252 sd->next_decay_max_lb_cost = jiffies + HZ;
9255 max_cost += sd->max_newidle_lb_cost;
9257 if (!(sd->flags & SD_LOAD_BALANCE))
9261 * Stop the load balance at this level. There is another
9262 * CPU in our sched group which is doing load balancing more
9265 if (!continue_balancing) {
9271 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9273 need_serialize = sd->flags & SD_SERIALIZE;
9274 if (need_serialize) {
9275 if (!spin_trylock(&balancing))
9279 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9280 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9282 * The LBF_DST_PINNED logic could have changed
9283 * env->dst_cpu, so we can't know our idle
9284 * state even if we migrated tasks. Update it.
9286 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9288 sd->last_balance = jiffies;
9289 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9292 spin_unlock(&balancing);
9294 if (time_after(next_balance, sd->last_balance + interval)) {
9295 next_balance = sd->last_balance + interval;
9296 update_next_balance = 1;
9301 * Ensure the rq-wide value also decays but keep it at a
9302 * reasonable floor to avoid funnies with rq->avg_idle.
9304 rq->max_idle_balance_cost =
9305 max((u64)sysctl_sched_migration_cost, max_cost);
9310 * next_balance will be updated only when there is a need.
9311 * When the cpu is attached to null domain for ex, it will not be
9314 if (likely(update_next_balance)) {
9315 rq->next_balance = next_balance;
9317 #ifdef CONFIG_NO_HZ_COMMON
9319 * If this CPU has been elected to perform the nohz idle
9320 * balance. Other idle CPUs have already rebalanced with
9321 * nohz_idle_balance() and nohz.next_balance has been
9322 * updated accordingly. This CPU is now running the idle load
9323 * balance for itself and we need to update the
9324 * nohz.next_balance accordingly.
9326 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9327 nohz.next_balance = rq->next_balance;
9332 static inline int on_null_domain(struct rq *rq)
9334 return unlikely(!rcu_dereference_sched(rq->sd));
9337 #ifdef CONFIG_NO_HZ_COMMON
9339 * idle load balancing details
9340 * - When one of the busy CPUs notice that there may be an idle rebalancing
9341 * needed, they will kick the idle load balancer, which then does idle
9342 * load balancing for all the idle CPUs.
9343 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
9347 static inline int find_new_ilb(void)
9351 for_each_cpu_and(ilb, nohz.idle_cpus_mask,
9352 housekeeping_cpumask(HK_FLAG_MISC)) {
9361 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
9362 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
9364 static void kick_ilb(unsigned int flags)
9368 nohz.next_balance++;
9370 ilb_cpu = find_new_ilb();
9372 if (ilb_cpu >= nr_cpu_ids)
9375 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
9376 if (flags & NOHZ_KICK_MASK)
9380 * Use smp_send_reschedule() instead of resched_cpu().
9381 * This way we generate a sched IPI on the target CPU which
9382 * is idle. And the softirq performing nohz idle load balance
9383 * will be run before returning from the IPI.
9385 smp_send_reschedule(ilb_cpu);
9389 * Current decision point for kicking the idle load balancer in the presence
9390 * of idle CPUs in the system.
9392 static void nohz_balancer_kick(struct rq *rq)
9394 unsigned long now = jiffies;
9395 struct sched_domain_shared *sds;
9396 struct sched_domain *sd;
9397 int nr_busy, i, cpu = rq->cpu;
9398 unsigned int flags = 0;
9400 if (unlikely(rq->idle_balance))
9404 * We may be recently in ticked or tickless idle mode. At the first
9405 * busy tick after returning from idle, we will update the busy stats.
9407 nohz_balance_exit_idle(rq);
9410 * None are in tickless mode and hence no need for NOHZ idle load
9413 if (likely(!atomic_read(&nohz.nr_cpus)))
9416 if (READ_ONCE(nohz.has_blocked) &&
9417 time_after(now, READ_ONCE(nohz.next_blocked)))
9418 flags = NOHZ_STATS_KICK;
9420 if (time_before(now, nohz.next_balance))
9423 if (rq->nr_running >= 2) {
9424 flags = NOHZ_KICK_MASK;
9430 sd = rcu_dereference(rq->sd);
9433 * If there's a CFS task and the current CPU has reduced
9434 * capacity; kick the ILB to see if there's a better CPU to run
9437 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
9438 flags = NOHZ_KICK_MASK;
9443 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
9446 * When ASYM_PACKING; see if there's a more preferred CPU
9447 * currently idle; in which case, kick the ILB to move tasks
9450 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
9451 if (sched_asym_prefer(i, cpu)) {
9452 flags = NOHZ_KICK_MASK;
9458 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
9461 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
9462 * to run the misfit task on.
9464 if (check_misfit_status(rq, sd)) {
9465 flags = NOHZ_KICK_MASK;
9470 * For asymmetric systems, we do not want to nicely balance
9471 * cache use, instead we want to embrace asymmetry and only
9472 * ensure tasks have enough CPU capacity.
9474 * Skip the LLC logic because it's not relevant in that case.
9479 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9482 * If there is an imbalance between LLC domains (IOW we could
9483 * increase the overall cache use), we need some less-loaded LLC
9484 * domain to pull some load. Likewise, we may need to spread
9485 * load within the current LLC domain (e.g. packed SMT cores but
9486 * other CPUs are idle). We can't really know from here how busy
9487 * the others are - so just get a nohz balance going if it looks
9488 * like this LLC domain has tasks we could move.
9490 nr_busy = atomic_read(&sds->nr_busy_cpus);
9492 flags = NOHZ_KICK_MASK;
9503 static void set_cpu_sd_state_busy(int cpu)
9505 struct sched_domain *sd;
9508 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9510 if (!sd || !sd->nohz_idle)
9514 atomic_inc(&sd->shared->nr_busy_cpus);
9519 void nohz_balance_exit_idle(struct rq *rq)
9521 SCHED_WARN_ON(rq != this_rq());
9523 if (likely(!rq->nohz_tick_stopped))
9526 rq->nohz_tick_stopped = 0;
9527 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
9528 atomic_dec(&nohz.nr_cpus);
9530 set_cpu_sd_state_busy(rq->cpu);
9533 static void set_cpu_sd_state_idle(int cpu)
9535 struct sched_domain *sd;
9538 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9540 if (!sd || sd->nohz_idle)
9544 atomic_dec(&sd->shared->nr_busy_cpus);
9550 * This routine will record that the CPU is going idle with tick stopped.
9551 * This info will be used in performing idle load balancing in the future.
9553 void nohz_balance_enter_idle(int cpu)
9555 struct rq *rq = cpu_rq(cpu);
9557 SCHED_WARN_ON(cpu != smp_processor_id());
9559 /* If this CPU is going down, then nothing needs to be done: */
9560 if (!cpu_active(cpu))
9563 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9564 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9568 * Can be set safely without rq->lock held
9569 * If a clear happens, it will have evaluated last additions because
9570 * rq->lock is held during the check and the clear
9572 rq->has_blocked_load = 1;
9575 * The tick is still stopped but load could have been added in the
9576 * meantime. We set the nohz.has_blocked flag to trig a check of the
9577 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
9578 * of nohz.has_blocked can only happen after checking the new load
9580 if (rq->nohz_tick_stopped)
9583 /* If we're a completely isolated CPU, we don't play: */
9584 if (on_null_domain(rq))
9587 rq->nohz_tick_stopped = 1;
9589 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9590 atomic_inc(&nohz.nr_cpus);
9593 * Ensures that if nohz_idle_balance() fails to observe our
9594 * @idle_cpus_mask store, it must observe the @has_blocked
9597 smp_mb__after_atomic();
9599 set_cpu_sd_state_idle(cpu);
9603 * Each time a cpu enter idle, we assume that it has blocked load and
9604 * enable the periodic update of the load of idle cpus
9606 WRITE_ONCE(nohz.has_blocked, 1);
9610 * Internal function that runs load balance for all idle cpus. The load balance
9611 * can be a simple update of blocked load or a complete load balance with
9612 * tasks movement depending of flags.
9613 * The function returns false if the loop has stopped before running
9614 * through all idle CPUs.
9616 static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
9617 enum cpu_idle_type idle)
9619 /* Earliest time when we have to do rebalance again */
9620 unsigned long now = jiffies;
9621 unsigned long next_balance = now + 60*HZ;
9622 bool has_blocked_load = false;
9623 int update_next_balance = 0;
9624 int this_cpu = this_rq->cpu;
9629 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
9632 * We assume there will be no idle load after this update and clear
9633 * the has_blocked flag. If a cpu enters idle in the mean time, it will
9634 * set the has_blocked flag and trig another update of idle load.
9635 * Because a cpu that becomes idle, is added to idle_cpus_mask before
9636 * setting the flag, we are sure to not clear the state and not
9637 * check the load of an idle cpu.
9639 WRITE_ONCE(nohz.has_blocked, 0);
9642 * Ensures that if we miss the CPU, we must see the has_blocked
9643 * store from nohz_balance_enter_idle().
9647 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9648 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9652 * If this CPU gets work to do, stop the load balancing
9653 * work being done for other CPUs. Next load
9654 * balancing owner will pick it up.
9656 if (need_resched()) {
9657 has_blocked_load = true;
9661 rq = cpu_rq(balance_cpu);
9663 has_blocked_load |= update_nohz_stats(rq, true);
9666 * If time for next balance is due,
9669 if (time_after_eq(jiffies, rq->next_balance)) {
9672 rq_lock_irqsave(rq, &rf);
9673 update_rq_clock(rq);
9674 rq_unlock_irqrestore(rq, &rf);
9676 if (flags & NOHZ_BALANCE_KICK)
9677 rebalance_domains(rq, CPU_IDLE);
9680 if (time_after(next_balance, rq->next_balance)) {
9681 next_balance = rq->next_balance;
9682 update_next_balance = 1;
9686 /* Newly idle CPU doesn't need an update */
9687 if (idle != CPU_NEWLY_IDLE) {
9688 update_blocked_averages(this_cpu);
9689 has_blocked_load |= this_rq->has_blocked_load;
9692 if (flags & NOHZ_BALANCE_KICK)
9693 rebalance_domains(this_rq, CPU_IDLE);
9695 WRITE_ONCE(nohz.next_blocked,
9696 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
9698 /* The full idle balance loop has been done */
9702 /* There is still blocked load, enable periodic update */
9703 if (has_blocked_load)
9704 WRITE_ONCE(nohz.has_blocked, 1);
9707 * next_balance will be updated only when there is a need.
9708 * When the CPU is attached to null domain for ex, it will not be
9711 if (likely(update_next_balance))
9712 nohz.next_balance = next_balance;
9718 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9719 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9721 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9723 int this_cpu = this_rq->cpu;
9726 if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
9729 if (idle != CPU_IDLE) {
9730 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9734 /* could be _relaxed() */
9735 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9736 if (!(flags & NOHZ_KICK_MASK))
9739 _nohz_idle_balance(this_rq, flags, idle);
9744 static void nohz_newidle_balance(struct rq *this_rq)
9746 int this_cpu = this_rq->cpu;
9749 * This CPU doesn't want to be disturbed by scheduler
9752 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
9755 /* Will wake up very soon. No time for doing anything else*/
9756 if (this_rq->avg_idle < sysctl_sched_migration_cost)
9759 /* Don't need to update blocked load of idle CPUs*/
9760 if (!READ_ONCE(nohz.has_blocked) ||
9761 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
9764 raw_spin_unlock(&this_rq->lock);
9766 * This CPU is going to be idle and blocked load of idle CPUs
9767 * need to be updated. Run the ilb locally as it is a good
9768 * candidate for ilb instead of waking up another idle CPU.
9769 * Kick an normal ilb if we failed to do the update.
9771 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
9772 kick_ilb(NOHZ_STATS_KICK);
9773 raw_spin_lock(&this_rq->lock);
9776 #else /* !CONFIG_NO_HZ_COMMON */
9777 static inline void nohz_balancer_kick(struct rq *rq) { }
9779 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9784 static inline void nohz_newidle_balance(struct rq *this_rq) { }
9785 #endif /* CONFIG_NO_HZ_COMMON */
9788 * idle_balance is called by schedule() if this_cpu is about to become
9789 * idle. Attempts to pull tasks from other CPUs.
9791 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
9793 unsigned long next_balance = jiffies + HZ;
9794 int this_cpu = this_rq->cpu;
9795 struct sched_domain *sd;
9796 int pulled_task = 0;
9800 * We must set idle_stamp _before_ calling idle_balance(), such that we
9801 * measure the duration of idle_balance() as idle time.
9803 this_rq->idle_stamp = rq_clock(this_rq);
9806 * Do not pull tasks towards !active CPUs...
9808 if (!cpu_active(this_cpu))
9812 * This is OK, because current is on_cpu, which avoids it being picked
9813 * for load-balance and preemption/IRQs are still disabled avoiding
9814 * further scheduler activity on it and we're being very careful to
9815 * re-start the picking loop.
9817 rq_unpin_lock(this_rq, rf);
9819 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
9820 !READ_ONCE(this_rq->rd->overload)) {
9823 sd = rcu_dereference_check_sched_domain(this_rq->sd);
9825 update_next_balance(sd, &next_balance);
9828 nohz_newidle_balance(this_rq);
9833 raw_spin_unlock(&this_rq->lock);
9835 update_blocked_averages(this_cpu);
9837 for_each_domain(this_cpu, sd) {
9838 int continue_balancing = 1;
9839 u64 t0, domain_cost;
9841 if (!(sd->flags & SD_LOAD_BALANCE))
9844 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
9845 update_next_balance(sd, &next_balance);
9849 if (sd->flags & SD_BALANCE_NEWIDLE) {
9850 t0 = sched_clock_cpu(this_cpu);
9852 pulled_task = load_balance(this_cpu, this_rq,
9854 &continue_balancing);
9856 domain_cost = sched_clock_cpu(this_cpu) - t0;
9857 if (domain_cost > sd->max_newidle_lb_cost)
9858 sd->max_newidle_lb_cost = domain_cost;
9860 curr_cost += domain_cost;
9863 update_next_balance(sd, &next_balance);
9866 * Stop searching for tasks to pull if there are
9867 * now runnable tasks on this rq.
9869 if (pulled_task || this_rq->nr_running > 0)
9874 raw_spin_lock(&this_rq->lock);
9876 if (curr_cost > this_rq->max_idle_balance_cost)
9877 this_rq->max_idle_balance_cost = curr_cost;
9881 * While browsing the domains, we released the rq lock, a task could
9882 * have been enqueued in the meantime. Since we're not going idle,
9883 * pretend we pulled a task.
9885 if (this_rq->cfs.h_nr_running && !pulled_task)
9888 /* Move the next balance forward */
9889 if (time_after(this_rq->next_balance, next_balance))
9890 this_rq->next_balance = next_balance;
9892 /* Is there a task of a high priority class? */
9893 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
9897 this_rq->idle_stamp = 0;
9899 rq_repin_lock(this_rq, rf);
9905 * run_rebalance_domains is triggered when needed from the scheduler tick.
9906 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9908 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9910 struct rq *this_rq = this_rq();
9911 enum cpu_idle_type idle = this_rq->idle_balance ?
9912 CPU_IDLE : CPU_NOT_IDLE;
9915 * If this CPU has a pending nohz_balance_kick, then do the
9916 * balancing on behalf of the other idle CPUs whose ticks are
9917 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9918 * give the idle CPUs a chance to load balance. Else we may
9919 * load balance only within the local sched_domain hierarchy
9920 * and abort nohz_idle_balance altogether if we pull some load.
9922 if (nohz_idle_balance(this_rq, idle))
9925 /* normal load balance */
9926 update_blocked_averages(this_rq->cpu);
9927 rebalance_domains(this_rq, idle);
9931 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9933 void trigger_load_balance(struct rq *rq)
9935 /* Don't need to rebalance while attached to NULL domain */
9936 if (unlikely(on_null_domain(rq)))
9939 if (time_after_eq(jiffies, rq->next_balance))
9940 raise_softirq(SCHED_SOFTIRQ);
9942 nohz_balancer_kick(rq);
9945 static void rq_online_fair(struct rq *rq)
9949 update_runtime_enabled(rq);
9952 static void rq_offline_fair(struct rq *rq)
9956 /* Ensure any throttled groups are reachable by pick_next_task */
9957 unthrottle_offline_cfs_rqs(rq);
9960 #endif /* CONFIG_SMP */
9963 * scheduler tick hitting a task of our scheduling class.
9965 * NOTE: This function can be called remotely by the tick offload that
9966 * goes along full dynticks. Therefore no local assumption can be made
9967 * and everything must be accessed through the @rq and @curr passed in
9970 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9972 struct cfs_rq *cfs_rq;
9973 struct sched_entity *se = &curr->se;
9975 for_each_sched_entity(se) {
9976 cfs_rq = cfs_rq_of(se);
9977 entity_tick(cfs_rq, se, queued);
9980 if (static_branch_unlikely(&sched_numa_balancing))
9981 task_tick_numa(rq, curr);
9983 update_misfit_status(curr, rq);
9984 update_overutilized_status(task_rq(curr));
9988 * called on fork with the child task as argument from the parent's context
9989 * - child not yet on the tasklist
9990 * - preemption disabled
9992 static void task_fork_fair(struct task_struct *p)
9994 struct cfs_rq *cfs_rq;
9995 struct sched_entity *se = &p->se, *curr;
9996 struct rq *rq = this_rq();
10000 update_rq_clock(rq);
10002 cfs_rq = task_cfs_rq(current);
10003 curr = cfs_rq->curr;
10005 update_curr(cfs_rq);
10006 se->vruntime = curr->vruntime;
10008 place_entity(cfs_rq, se, 1);
10010 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
10012 * Upon rescheduling, sched_class::put_prev_task() will place
10013 * 'current' within the tree based on its new key value.
10015 swap(curr->vruntime, se->vruntime);
10019 se->vruntime -= cfs_rq->min_vruntime;
10020 rq_unlock(rq, &rf);
10024 * Priority of the task has changed. Check to see if we preempt
10025 * the current task.
10028 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
10030 if (!task_on_rq_queued(p))
10034 * Reschedule if we are currently running on this runqueue and
10035 * our priority decreased, or if we are not currently running on
10036 * this runqueue and our priority is higher than the current's
10038 if (rq->curr == p) {
10039 if (p->prio > oldprio)
10042 check_preempt_curr(rq, p, 0);
10045 static inline bool vruntime_normalized(struct task_struct *p)
10047 struct sched_entity *se = &p->se;
10050 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
10051 * the dequeue_entity(.flags=0) will already have normalized the
10058 * When !on_rq, vruntime of the task has usually NOT been normalized.
10059 * But there are some cases where it has already been normalized:
10061 * - A forked child which is waiting for being woken up by
10062 * wake_up_new_task().
10063 * - A task which has been woken up by try_to_wake_up() and
10064 * waiting for actually being woken up by sched_ttwu_pending().
10066 if (!se->sum_exec_runtime ||
10067 (p->state == TASK_WAKING && p->sched_remote_wakeup))
10073 #ifdef CONFIG_FAIR_GROUP_SCHED
10075 * Propagate the changes of the sched_entity across the tg tree to make it
10076 * visible to the root
10078 static void propagate_entity_cfs_rq(struct sched_entity *se)
10080 struct cfs_rq *cfs_rq;
10082 /* Start to propagate at parent */
10085 for_each_sched_entity(se) {
10086 cfs_rq = cfs_rq_of(se);
10088 if (cfs_rq_throttled(cfs_rq))
10091 update_load_avg(cfs_rq, se, UPDATE_TG);
10095 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
10098 static void detach_entity_cfs_rq(struct sched_entity *se)
10100 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10102 /* Catch up with the cfs_rq and remove our load when we leave */
10103 update_load_avg(cfs_rq, se, 0);
10104 detach_entity_load_avg(cfs_rq, se);
10105 update_tg_load_avg(cfs_rq, false);
10106 propagate_entity_cfs_rq(se);
10109 static void attach_entity_cfs_rq(struct sched_entity *se)
10111 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10113 #ifdef CONFIG_FAIR_GROUP_SCHED
10115 * Since the real-depth could have been changed (only FAIR
10116 * class maintain depth value), reset depth properly.
10118 se->depth = se->parent ? se->parent->depth + 1 : 0;
10121 /* Synchronize entity with its cfs_rq */
10122 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
10123 attach_entity_load_avg(cfs_rq, se, 0);
10124 update_tg_load_avg(cfs_rq, false);
10125 propagate_entity_cfs_rq(se);
10128 static void detach_task_cfs_rq(struct task_struct *p)
10130 struct sched_entity *se = &p->se;
10131 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10133 if (!vruntime_normalized(p)) {
10135 * Fix up our vruntime so that the current sleep doesn't
10136 * cause 'unlimited' sleep bonus.
10138 place_entity(cfs_rq, se, 0);
10139 se->vruntime -= cfs_rq->min_vruntime;
10142 detach_entity_cfs_rq(se);
10145 static void attach_task_cfs_rq(struct task_struct *p)
10147 struct sched_entity *se = &p->se;
10148 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10150 attach_entity_cfs_rq(se);
10152 if (!vruntime_normalized(p))
10153 se->vruntime += cfs_rq->min_vruntime;
10156 static void switched_from_fair(struct rq *rq, struct task_struct *p)
10158 detach_task_cfs_rq(p);
10161 static void switched_to_fair(struct rq *rq, struct task_struct *p)
10163 attach_task_cfs_rq(p);
10165 if (task_on_rq_queued(p)) {
10167 * We were most likely switched from sched_rt, so
10168 * kick off the schedule if running, otherwise just see
10169 * if we can still preempt the current task.
10174 check_preempt_curr(rq, p, 0);
10178 /* Account for a task changing its policy or group.
10180 * This routine is mostly called to set cfs_rq->curr field when a task
10181 * migrates between groups/classes.
10183 static void set_curr_task_fair(struct rq *rq)
10185 struct sched_entity *se = &rq->curr->se;
10187 for_each_sched_entity(se) {
10188 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10190 set_next_entity(cfs_rq, se);
10191 /* ensure bandwidth has been allocated on our new cfs_rq */
10192 account_cfs_rq_runtime(cfs_rq, 0);
10196 void init_cfs_rq(struct cfs_rq *cfs_rq)
10198 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
10199 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
10200 #ifndef CONFIG_64BIT
10201 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
10204 raw_spin_lock_init(&cfs_rq->removed.lock);
10208 #ifdef CONFIG_FAIR_GROUP_SCHED
10209 static void task_set_group_fair(struct task_struct *p)
10211 struct sched_entity *se = &p->se;
10213 set_task_rq(p, task_cpu(p));
10214 se->depth = se->parent ? se->parent->depth + 1 : 0;
10217 static void task_move_group_fair(struct task_struct *p)
10219 detach_task_cfs_rq(p);
10220 set_task_rq(p, task_cpu(p));
10223 /* Tell se's cfs_rq has been changed -- migrated */
10224 p->se.avg.last_update_time = 0;
10226 attach_task_cfs_rq(p);
10229 static void task_change_group_fair(struct task_struct *p, int type)
10232 case TASK_SET_GROUP:
10233 task_set_group_fair(p);
10236 case TASK_MOVE_GROUP:
10237 task_move_group_fair(p);
10242 void free_fair_sched_group(struct task_group *tg)
10246 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
10248 for_each_possible_cpu(i) {
10250 kfree(tg->cfs_rq[i]);
10259 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10261 struct sched_entity *se;
10262 struct cfs_rq *cfs_rq;
10265 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
10268 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
10272 tg->shares = NICE_0_LOAD;
10274 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
10276 for_each_possible_cpu(i) {
10277 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
10278 GFP_KERNEL, cpu_to_node(i));
10282 se = kzalloc_node(sizeof(struct sched_entity),
10283 GFP_KERNEL, cpu_to_node(i));
10287 init_cfs_rq(cfs_rq);
10288 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
10289 init_entity_runnable_average(se);
10300 void online_fair_sched_group(struct task_group *tg)
10302 struct sched_entity *se;
10306 for_each_possible_cpu(i) {
10310 raw_spin_lock_irq(&rq->lock);
10311 update_rq_clock(rq);
10312 attach_entity_cfs_rq(se);
10313 sync_throttle(tg, i);
10314 raw_spin_unlock_irq(&rq->lock);
10318 void unregister_fair_sched_group(struct task_group *tg)
10320 unsigned long flags;
10324 for_each_possible_cpu(cpu) {
10326 remove_entity_load_avg(tg->se[cpu]);
10329 * Only empty task groups can be destroyed; so we can speculatively
10330 * check on_list without danger of it being re-added.
10332 if (!tg->cfs_rq[cpu]->on_list)
10337 raw_spin_lock_irqsave(&rq->lock, flags);
10338 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
10339 raw_spin_unlock_irqrestore(&rq->lock, flags);
10343 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
10344 struct sched_entity *se, int cpu,
10345 struct sched_entity *parent)
10347 struct rq *rq = cpu_rq(cpu);
10351 init_cfs_rq_runtime(cfs_rq);
10353 tg->cfs_rq[cpu] = cfs_rq;
10356 /* se could be NULL for root_task_group */
10361 se->cfs_rq = &rq->cfs;
10364 se->cfs_rq = parent->my_q;
10365 se->depth = parent->depth + 1;
10369 /* guarantee group entities always have weight */
10370 update_load_set(&se->load, NICE_0_LOAD);
10371 se->parent = parent;
10374 static DEFINE_MUTEX(shares_mutex);
10376 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10381 * We can't change the weight of the root cgroup.
10386 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10388 mutex_lock(&shares_mutex);
10389 if (tg->shares == shares)
10392 tg->shares = shares;
10393 for_each_possible_cpu(i) {
10394 struct rq *rq = cpu_rq(i);
10395 struct sched_entity *se = tg->se[i];
10396 struct rq_flags rf;
10398 /* Propagate contribution to hierarchy */
10399 rq_lock_irqsave(rq, &rf);
10400 update_rq_clock(rq);
10401 for_each_sched_entity(se) {
10402 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
10403 update_cfs_group(se);
10405 rq_unlock_irqrestore(rq, &rf);
10409 mutex_unlock(&shares_mutex);
10412 #else /* CONFIG_FAIR_GROUP_SCHED */
10414 void free_fair_sched_group(struct task_group *tg) { }
10416 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10421 void online_fair_sched_group(struct task_group *tg) { }
10423 void unregister_fair_sched_group(struct task_group *tg) { }
10425 #endif /* CONFIG_FAIR_GROUP_SCHED */
10428 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10430 struct sched_entity *se = &task->se;
10431 unsigned int rr_interval = 0;
10434 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10437 if (rq->cfs.load.weight)
10438 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10440 return rr_interval;
10444 * All the scheduling class methods:
10446 const struct sched_class fair_sched_class = {
10447 .next = &idle_sched_class,
10448 .enqueue_task = enqueue_task_fair,
10449 .dequeue_task = dequeue_task_fair,
10450 .yield_task = yield_task_fair,
10451 .yield_to_task = yield_to_task_fair,
10453 .check_preempt_curr = check_preempt_wakeup,
10455 .pick_next_task = pick_next_task_fair,
10456 .put_prev_task = put_prev_task_fair,
10459 .select_task_rq = select_task_rq_fair,
10460 .migrate_task_rq = migrate_task_rq_fair,
10462 .rq_online = rq_online_fair,
10463 .rq_offline = rq_offline_fair,
10465 .task_dead = task_dead_fair,
10466 .set_cpus_allowed = set_cpus_allowed_common,
10469 .set_curr_task = set_curr_task_fair,
10470 .task_tick = task_tick_fair,
10471 .task_fork = task_fork_fair,
10473 .prio_changed = prio_changed_fair,
10474 .switched_from = switched_from_fair,
10475 .switched_to = switched_to_fair,
10477 .get_rr_interval = get_rr_interval_fair,
10479 .update_curr = update_curr_fair,
10481 #ifdef CONFIG_FAIR_GROUP_SCHED
10482 .task_change_group = task_change_group_fair,
10485 #ifdef CONFIG_UCLAMP_TASK
10486 .uclamp_enabled = 1,
10490 #ifdef CONFIG_SCHED_DEBUG
10491 void print_cfs_stats(struct seq_file *m, int cpu)
10493 struct cfs_rq *cfs_rq, *pos;
10496 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
10497 print_cfs_rq(m, cpu, cfs_rq);
10501 #ifdef CONFIG_NUMA_BALANCING
10502 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10505 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10506 struct numa_group *ng;
10509 ng = rcu_dereference(p->numa_group);
10510 for_each_online_node(node) {
10511 if (p->numa_faults) {
10512 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10513 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10516 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
10517 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
10519 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10523 #endif /* CONFIG_NUMA_BALANCING */
10524 #endif /* CONFIG_SCHED_DEBUG */
10526 __init void init_sched_fair_class(void)
10529 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10531 #ifdef CONFIG_NO_HZ_COMMON
10532 nohz.next_balance = jiffies;
10533 nohz.next_blocked = jiffies;
10534 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
10541 * Helper functions to facilitate extracting info from tracepoints.
10544 const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq)
10547 return cfs_rq ? &cfs_rq->avg : NULL;
10552 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg);
10554 char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len)
10558 strlcpy(str, "(null)", len);
10563 cfs_rq_tg_path(cfs_rq, str, len);
10566 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path);
10568 int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq)
10570 return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1;
10572 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu);
10574 const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq)
10577 return rq ? &rq->avg_rt : NULL;
10582 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt);
10584 const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq)
10587 return rq ? &rq->avg_dl : NULL;
10592 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl);
10594 const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq)
10596 #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ)
10597 return rq ? &rq->avg_irq : NULL;
10602 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq);
10604 int sched_trace_rq_cpu(struct rq *rq)
10606 return rq ? cpu_of(rq) : -1;
10608 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu);
10610 const struct cpumask *sched_trace_rd_span(struct root_domain *rd)
10613 return rd ? rd->span : NULL;
10618 EXPORT_SYMBOL_GPL(sched_trace_rd_span);