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 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1193 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1194 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1197 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1199 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1200 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1203 /* Shared or private faults. */
1204 #define NR_NUMA_HINT_FAULT_TYPES 2
1206 /* Memory and CPU locality */
1207 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1209 /* Averaged statistics, and temporary buffers. */
1210 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1212 pid_t task_numa_group_id(struct task_struct *p)
1214 struct numa_group *ng;
1218 ng = rcu_dereference(p->numa_group);
1227 * The averaged statistics, shared & private, memory & CPU,
1228 * occupy the first half of the array. The second half of the
1229 * array is for current counters, which are averaged into the
1230 * first set by task_numa_placement.
1232 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1234 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1237 static inline unsigned long task_faults(struct task_struct *p, int nid)
1239 if (!p->numa_faults)
1242 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1243 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1246 static inline unsigned long group_faults(struct task_struct *p, int nid)
1248 struct numa_group *ng = deref_task_numa_group(p);
1253 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1254 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1257 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1259 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1260 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1263 static inline unsigned long group_faults_priv(struct numa_group *ng)
1265 unsigned long faults = 0;
1268 for_each_online_node(node) {
1269 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1275 static inline unsigned long group_faults_shared(struct numa_group *ng)
1277 unsigned long faults = 0;
1280 for_each_online_node(node) {
1281 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1288 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1289 * considered part of a numa group's pseudo-interleaving set. Migrations
1290 * between these nodes are slowed down, to allow things to settle down.
1292 #define ACTIVE_NODE_FRACTION 3
1294 static bool numa_is_active_node(int nid, struct numa_group *ng)
1296 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1299 /* Handle placement on systems where not all nodes are directly connected. */
1300 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1301 int maxdist, bool task)
1303 unsigned long score = 0;
1307 * All nodes are directly connected, and the same distance
1308 * from each other. No need for fancy placement algorithms.
1310 if (sched_numa_topology_type == NUMA_DIRECT)
1314 * This code is called for each node, introducing N^2 complexity,
1315 * which should be ok given the number of nodes rarely exceeds 8.
1317 for_each_online_node(node) {
1318 unsigned long faults;
1319 int dist = node_distance(nid, node);
1322 * The furthest away nodes in the system are not interesting
1323 * for placement; nid was already counted.
1325 if (dist == sched_max_numa_distance || node == nid)
1329 * On systems with a backplane NUMA topology, compare groups
1330 * of nodes, and move tasks towards the group with the most
1331 * memory accesses. When comparing two nodes at distance
1332 * "hoplimit", only nodes closer by than "hoplimit" are part
1333 * of each group. Skip other nodes.
1335 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1339 /* Add up the faults from nearby nodes. */
1341 faults = task_faults(p, node);
1343 faults = group_faults(p, node);
1346 * On systems with a glueless mesh NUMA topology, there are
1347 * no fixed "groups of nodes". Instead, nodes that are not
1348 * directly connected bounce traffic through intermediate
1349 * nodes; a numa_group can occupy any set of nodes.
1350 * The further away a node is, the less the faults count.
1351 * This seems to result in good task placement.
1353 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1354 faults *= (sched_max_numa_distance - dist);
1355 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1365 * These return the fraction of accesses done by a particular task, or
1366 * task group, on a particular numa node. The group weight is given a
1367 * larger multiplier, in order to group tasks together that are almost
1368 * evenly spread out between numa nodes.
1370 static inline unsigned long task_weight(struct task_struct *p, int nid,
1373 unsigned long faults, total_faults;
1375 if (!p->numa_faults)
1378 total_faults = p->total_numa_faults;
1383 faults = task_faults(p, nid);
1384 faults += score_nearby_nodes(p, nid, dist, true);
1386 return 1000 * faults / total_faults;
1389 static inline unsigned long group_weight(struct task_struct *p, int nid,
1392 struct numa_group *ng = deref_task_numa_group(p);
1393 unsigned long faults, total_faults;
1398 total_faults = ng->total_faults;
1403 faults = group_faults(p, nid);
1404 faults += score_nearby_nodes(p, nid, dist, false);
1406 return 1000 * faults / total_faults;
1409 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1410 int src_nid, int dst_cpu)
1412 struct numa_group *ng = deref_curr_numa_group(p);
1413 int dst_nid = cpu_to_node(dst_cpu);
1414 int last_cpupid, this_cpupid;
1416 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1417 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1420 * Allow first faults or private faults to migrate immediately early in
1421 * the lifetime of a task. The magic number 4 is based on waiting for
1422 * two full passes of the "multi-stage node selection" test that is
1425 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1426 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1430 * Multi-stage node selection is used in conjunction with a periodic
1431 * migration fault to build a temporal task<->page relation. By using
1432 * a two-stage filter we remove short/unlikely relations.
1434 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1435 * a task's usage of a particular page (n_p) per total usage of this
1436 * page (n_t) (in a given time-span) to a probability.
1438 * Our periodic faults will sample this probability and getting the
1439 * same result twice in a row, given these samples are fully
1440 * independent, is then given by P(n)^2, provided our sample period
1441 * is sufficiently short compared to the usage pattern.
1443 * This quadric squishes small probabilities, making it less likely we
1444 * act on an unlikely task<->page relation.
1446 if (!cpupid_pid_unset(last_cpupid) &&
1447 cpupid_to_nid(last_cpupid) != dst_nid)
1450 /* Always allow migrate on private faults */
1451 if (cpupid_match_pid(p, last_cpupid))
1454 /* A shared fault, but p->numa_group has not been set up yet. */
1459 * Destination node is much more heavily used than the source
1460 * node? Allow migration.
1462 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1463 ACTIVE_NODE_FRACTION)
1467 * Distribute memory according to CPU & memory use on each node,
1468 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1470 * faults_cpu(dst) 3 faults_cpu(src)
1471 * --------------- * - > ---------------
1472 * faults_mem(dst) 4 faults_mem(src)
1474 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1475 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1478 static unsigned long cpu_runnable_load(struct rq *rq);
1480 /* Cached statistics for all CPUs within a node */
1484 /* Total compute capacity of CPUs on a node */
1485 unsigned long compute_capacity;
1489 * XXX borrowed from update_sg_lb_stats
1491 static void update_numa_stats(struct numa_stats *ns, int nid)
1495 memset(ns, 0, sizeof(*ns));
1496 for_each_cpu(cpu, cpumask_of_node(nid)) {
1497 struct rq *rq = cpu_rq(cpu);
1499 ns->load += cpu_runnable_load(rq);
1500 ns->compute_capacity += capacity_of(cpu);
1505 struct task_numa_env {
1506 struct task_struct *p;
1508 int src_cpu, src_nid;
1509 int dst_cpu, dst_nid;
1511 struct numa_stats src_stats, dst_stats;
1516 struct task_struct *best_task;
1521 static void task_numa_assign(struct task_numa_env *env,
1522 struct task_struct *p, long imp)
1524 struct rq *rq = cpu_rq(env->dst_cpu);
1526 /* Bail out if run-queue part of active NUMA balance. */
1527 if (xchg(&rq->numa_migrate_on, 1))
1531 * Clear previous best_cpu/rq numa-migrate flag, since task now
1532 * found a better CPU to move/swap.
1534 if (env->best_cpu != -1) {
1535 rq = cpu_rq(env->best_cpu);
1536 WRITE_ONCE(rq->numa_migrate_on, 0);
1540 put_task_struct(env->best_task);
1545 env->best_imp = imp;
1546 env->best_cpu = env->dst_cpu;
1549 static bool load_too_imbalanced(long src_load, long dst_load,
1550 struct task_numa_env *env)
1553 long orig_src_load, orig_dst_load;
1554 long src_capacity, dst_capacity;
1557 * The load is corrected for the CPU capacity available on each node.
1560 * ------------ vs ---------
1561 * src_capacity dst_capacity
1563 src_capacity = env->src_stats.compute_capacity;
1564 dst_capacity = env->dst_stats.compute_capacity;
1566 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1568 orig_src_load = env->src_stats.load;
1569 orig_dst_load = env->dst_stats.load;
1571 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1573 /* Would this change make things worse? */
1574 return (imb > old_imb);
1578 * Maximum NUMA importance can be 1998 (2*999);
1579 * SMALLIMP @ 30 would be close to 1998/64.
1580 * Used to deter task migration.
1585 * This checks if the overall compute and NUMA accesses of the system would
1586 * be improved if the source tasks was migrated to the target dst_cpu taking
1587 * into account that it might be best if task running on the dst_cpu should
1588 * be exchanged with the source task
1590 static void task_numa_compare(struct task_numa_env *env,
1591 long taskimp, long groupimp, bool maymove)
1593 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
1594 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1595 long imp = p_ng ? groupimp : taskimp;
1596 struct task_struct *cur;
1597 long src_load, dst_load;
1598 int dist = env->dist;
1602 if (READ_ONCE(dst_rq->numa_migrate_on))
1606 cur = task_rcu_dereference(&dst_rq->curr);
1607 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1611 * Because we have preemption enabled we can get migrated around and
1612 * end try selecting ourselves (current == env->p) as a swap candidate.
1618 if (maymove && moveimp >= env->best_imp)
1625 * "imp" is the fault differential for the source task between the
1626 * source and destination node. Calculate the total differential for
1627 * the source task and potential destination task. The more negative
1628 * the value is, the more remote accesses that would be expected to
1629 * be incurred if the tasks were swapped.
1631 /* Skip this swap candidate if cannot move to the source cpu */
1632 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
1636 * If dst and source tasks are in the same NUMA group, or not
1637 * in any group then look only at task weights.
1639 cur_ng = rcu_dereference(cur->numa_group);
1640 if (cur_ng == p_ng) {
1641 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1642 task_weight(cur, env->dst_nid, dist);
1644 * Add some hysteresis to prevent swapping the
1645 * tasks within a group over tiny differences.
1651 * Compare the group weights. If a task is all by itself
1652 * (not part of a group), use the task weight instead.
1655 imp += group_weight(cur, env->src_nid, dist) -
1656 group_weight(cur, env->dst_nid, dist);
1658 imp += task_weight(cur, env->src_nid, dist) -
1659 task_weight(cur, env->dst_nid, dist);
1662 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1669 * If the NUMA importance is less than SMALLIMP,
1670 * task migration might only result in ping pong
1671 * of tasks and also hurt performance due to cache
1674 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1678 * In the overloaded case, try and keep the load balanced.
1680 load = task_h_load(env->p) - task_h_load(cur);
1684 dst_load = env->dst_stats.load + load;
1685 src_load = env->src_stats.load - load;
1687 if (load_too_imbalanced(src_load, dst_load, env))
1692 * One idle CPU per node is evaluated for a task numa move.
1693 * Call select_idle_sibling to maybe find a better one.
1697 * select_idle_siblings() uses an per-CPU cpumask that
1698 * can be used from IRQ context.
1700 local_irq_disable();
1701 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1706 task_numa_assign(env, cur, imp);
1711 static void task_numa_find_cpu(struct task_numa_env *env,
1712 long taskimp, long groupimp)
1714 long src_load, dst_load, load;
1715 bool maymove = false;
1718 load = task_h_load(env->p);
1719 dst_load = env->dst_stats.load + load;
1720 src_load = env->src_stats.load - load;
1723 * If the improvement from just moving env->p direction is better
1724 * than swapping tasks around, check if a move is possible.
1726 maymove = !load_too_imbalanced(src_load, dst_load, env);
1728 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1729 /* Skip this CPU if the source task cannot migrate */
1730 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
1734 task_numa_compare(env, taskimp, groupimp, maymove);
1738 static int task_numa_migrate(struct task_struct *p)
1740 struct task_numa_env env = {
1743 .src_cpu = task_cpu(p),
1744 .src_nid = task_node(p),
1746 .imbalance_pct = 112,
1752 unsigned long taskweight, groupweight;
1753 struct sched_domain *sd;
1754 long taskimp, groupimp;
1755 struct numa_group *ng;
1760 * Pick the lowest SD_NUMA domain, as that would have the smallest
1761 * imbalance and would be the first to start moving tasks about.
1763 * And we want to avoid any moving of tasks about, as that would create
1764 * random movement of tasks -- counter the numa conditions we're trying
1768 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1770 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1774 * Cpusets can break the scheduler domain tree into smaller
1775 * balance domains, some of which do not cross NUMA boundaries.
1776 * Tasks that are "trapped" in such domains cannot be migrated
1777 * elsewhere, so there is no point in (re)trying.
1779 if (unlikely(!sd)) {
1780 sched_setnuma(p, task_node(p));
1784 env.dst_nid = p->numa_preferred_nid;
1785 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1786 taskweight = task_weight(p, env.src_nid, dist);
1787 groupweight = group_weight(p, env.src_nid, dist);
1788 update_numa_stats(&env.src_stats, env.src_nid);
1789 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1790 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1791 update_numa_stats(&env.dst_stats, env.dst_nid);
1793 /* Try to find a spot on the preferred nid. */
1794 task_numa_find_cpu(&env, taskimp, groupimp);
1797 * Look at other nodes in these cases:
1798 * - there is no space available on the preferred_nid
1799 * - the task is part of a numa_group that is interleaved across
1800 * multiple NUMA nodes; in order to better consolidate the group,
1801 * we need to check other locations.
1803 ng = deref_curr_numa_group(p);
1804 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
1805 for_each_online_node(nid) {
1806 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1809 dist = node_distance(env.src_nid, env.dst_nid);
1810 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1812 taskweight = task_weight(p, env.src_nid, dist);
1813 groupweight = group_weight(p, env.src_nid, dist);
1816 /* Only consider nodes where both task and groups benefit */
1817 taskimp = task_weight(p, nid, dist) - taskweight;
1818 groupimp = group_weight(p, nid, dist) - groupweight;
1819 if (taskimp < 0 && groupimp < 0)
1824 update_numa_stats(&env.dst_stats, env.dst_nid);
1825 task_numa_find_cpu(&env, taskimp, groupimp);
1830 * If the task is part of a workload that spans multiple NUMA nodes,
1831 * and is migrating into one of the workload's active nodes, remember
1832 * this node as the task's preferred numa node, so the workload can
1834 * A task that migrated to a second choice node will be better off
1835 * trying for a better one later. Do not set the preferred node here.
1838 if (env.best_cpu == -1)
1841 nid = cpu_to_node(env.best_cpu);
1843 if (nid != p->numa_preferred_nid)
1844 sched_setnuma(p, nid);
1847 /* No better CPU than the current one was found. */
1848 if (env.best_cpu == -1)
1851 best_rq = cpu_rq(env.best_cpu);
1852 if (env.best_task == NULL) {
1853 ret = migrate_task_to(p, env.best_cpu);
1854 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1856 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1860 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
1861 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1864 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1865 put_task_struct(env.best_task);
1869 /* Attempt to migrate a task to a CPU on the preferred node. */
1870 static void numa_migrate_preferred(struct task_struct *p)
1872 unsigned long interval = HZ;
1874 /* This task has no NUMA fault statistics yet */
1875 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
1878 /* Periodically retry migrating the task to the preferred node */
1879 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1880 p->numa_migrate_retry = jiffies + interval;
1882 /* Success if task is already running on preferred CPU */
1883 if (task_node(p) == p->numa_preferred_nid)
1886 /* Otherwise, try migrate to a CPU on the preferred node */
1887 task_numa_migrate(p);
1891 * Find out how many nodes on the workload is actively running on. Do this by
1892 * tracking the nodes from which NUMA hinting faults are triggered. This can
1893 * be different from the set of nodes where the workload's memory is currently
1896 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1898 unsigned long faults, max_faults = 0;
1899 int nid, active_nodes = 0;
1901 for_each_online_node(nid) {
1902 faults = group_faults_cpu(numa_group, nid);
1903 if (faults > max_faults)
1904 max_faults = faults;
1907 for_each_online_node(nid) {
1908 faults = group_faults_cpu(numa_group, nid);
1909 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1913 numa_group->max_faults_cpu = max_faults;
1914 numa_group->active_nodes = active_nodes;
1918 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1919 * increments. The more local the fault statistics are, the higher the scan
1920 * period will be for the next scan window. If local/(local+remote) ratio is
1921 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1922 * the scan period will decrease. Aim for 70% local accesses.
1924 #define NUMA_PERIOD_SLOTS 10
1925 #define NUMA_PERIOD_THRESHOLD 7
1928 * Increase the scan period (slow down scanning) if the majority of
1929 * our memory is already on our local node, or if the majority of
1930 * the page accesses are shared with other processes.
1931 * Otherwise, decrease the scan period.
1933 static void update_task_scan_period(struct task_struct *p,
1934 unsigned long shared, unsigned long private)
1936 unsigned int period_slot;
1937 int lr_ratio, ps_ratio;
1940 unsigned long remote = p->numa_faults_locality[0];
1941 unsigned long local = p->numa_faults_locality[1];
1944 * If there were no record hinting faults then either the task is
1945 * completely idle or all activity is areas that are not of interest
1946 * to automatic numa balancing. Related to that, if there were failed
1947 * migration then it implies we are migrating too quickly or the local
1948 * node is overloaded. In either case, scan slower
1950 if (local + shared == 0 || p->numa_faults_locality[2]) {
1951 p->numa_scan_period = min(p->numa_scan_period_max,
1952 p->numa_scan_period << 1);
1954 p->mm->numa_next_scan = jiffies +
1955 msecs_to_jiffies(p->numa_scan_period);
1961 * Prepare to scale scan period relative to the current period.
1962 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1963 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1964 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1966 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1967 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1968 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1970 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1972 * Most memory accesses are local. There is no need to
1973 * do fast NUMA scanning, since memory is already local.
1975 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
1978 diff = slot * period_slot;
1979 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
1981 * Most memory accesses are shared with other tasks.
1982 * There is no point in continuing fast NUMA scanning,
1983 * since other tasks may just move the memory elsewhere.
1985 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
1988 diff = slot * period_slot;
1991 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
1992 * yet they are not on the local NUMA node. Speed up
1993 * NUMA scanning to get the memory moved over.
1995 int ratio = max(lr_ratio, ps_ratio);
1996 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1999 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2000 task_scan_min(p), task_scan_max(p));
2001 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2005 * Get the fraction of time the task has been running since the last
2006 * NUMA placement cycle. The scheduler keeps similar statistics, but
2007 * decays those on a 32ms period, which is orders of magnitude off
2008 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2009 * stats only if the task is so new there are no NUMA statistics yet.
2011 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2013 u64 runtime, delta, now;
2014 /* Use the start of this time slice to avoid calculations. */
2015 now = p->se.exec_start;
2016 runtime = p->se.sum_exec_runtime;
2018 if (p->last_task_numa_placement) {
2019 delta = runtime - p->last_sum_exec_runtime;
2020 *period = now - p->last_task_numa_placement;
2022 /* Avoid time going backwards, prevent potential divide error: */
2023 if (unlikely((s64)*period < 0))
2026 delta = p->se.avg.load_sum;
2027 *period = LOAD_AVG_MAX;
2030 p->last_sum_exec_runtime = runtime;
2031 p->last_task_numa_placement = now;
2037 * Determine the preferred nid for a task in a numa_group. This needs to
2038 * be done in a way that produces consistent results with group_weight,
2039 * otherwise workloads might not converge.
2041 static int preferred_group_nid(struct task_struct *p, int nid)
2046 /* Direct connections between all NUMA nodes. */
2047 if (sched_numa_topology_type == NUMA_DIRECT)
2051 * On a system with glueless mesh NUMA topology, group_weight
2052 * scores nodes according to the number of NUMA hinting faults on
2053 * both the node itself, and on nearby nodes.
2055 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2056 unsigned long score, max_score = 0;
2057 int node, max_node = nid;
2059 dist = sched_max_numa_distance;
2061 for_each_online_node(node) {
2062 score = group_weight(p, node, dist);
2063 if (score > max_score) {
2072 * Finding the preferred nid in a system with NUMA backplane
2073 * interconnect topology is more involved. The goal is to locate
2074 * tasks from numa_groups near each other in the system, and
2075 * untangle workloads from different sides of the system. This requires
2076 * searching down the hierarchy of node groups, recursively searching
2077 * inside the highest scoring group of nodes. The nodemask tricks
2078 * keep the complexity of the search down.
2080 nodes = node_online_map;
2081 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2082 unsigned long max_faults = 0;
2083 nodemask_t max_group = NODE_MASK_NONE;
2086 /* Are there nodes at this distance from each other? */
2087 if (!find_numa_distance(dist))
2090 for_each_node_mask(a, nodes) {
2091 unsigned long faults = 0;
2092 nodemask_t this_group;
2093 nodes_clear(this_group);
2095 /* Sum group's NUMA faults; includes a==b case. */
2096 for_each_node_mask(b, nodes) {
2097 if (node_distance(a, b) < dist) {
2098 faults += group_faults(p, b);
2099 node_set(b, this_group);
2100 node_clear(b, nodes);
2104 /* Remember the top group. */
2105 if (faults > max_faults) {
2106 max_faults = faults;
2107 max_group = this_group;
2109 * subtle: at the smallest distance there is
2110 * just one node left in each "group", the
2111 * winner is the preferred nid.
2116 /* Next round, evaluate the nodes within max_group. */
2124 static void task_numa_placement(struct task_struct *p)
2126 int seq, nid, max_nid = NUMA_NO_NODE;
2127 unsigned long max_faults = 0;
2128 unsigned long fault_types[2] = { 0, 0 };
2129 unsigned long total_faults;
2130 u64 runtime, period;
2131 spinlock_t *group_lock = NULL;
2132 struct numa_group *ng;
2135 * The p->mm->numa_scan_seq field gets updated without
2136 * exclusive access. Use READ_ONCE() here to ensure
2137 * that the field is read in a single access:
2139 seq = READ_ONCE(p->mm->numa_scan_seq);
2140 if (p->numa_scan_seq == seq)
2142 p->numa_scan_seq = seq;
2143 p->numa_scan_period_max = task_scan_max(p);
2145 total_faults = p->numa_faults_locality[0] +
2146 p->numa_faults_locality[1];
2147 runtime = numa_get_avg_runtime(p, &period);
2149 /* If the task is part of a group prevent parallel updates to group stats */
2150 ng = deref_curr_numa_group(p);
2152 group_lock = &ng->lock;
2153 spin_lock_irq(group_lock);
2156 /* Find the node with the highest number of faults */
2157 for_each_online_node(nid) {
2158 /* Keep track of the offsets in numa_faults array */
2159 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2160 unsigned long faults = 0, group_faults = 0;
2163 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2164 long diff, f_diff, f_weight;
2166 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2167 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2168 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2169 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2171 /* Decay existing window, copy faults since last scan */
2172 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2173 fault_types[priv] += p->numa_faults[membuf_idx];
2174 p->numa_faults[membuf_idx] = 0;
2177 * Normalize the faults_from, so all tasks in a group
2178 * count according to CPU use, instead of by the raw
2179 * number of faults. Tasks with little runtime have
2180 * little over-all impact on throughput, and thus their
2181 * faults are less important.
2183 f_weight = div64_u64(runtime << 16, period + 1);
2184 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2186 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2187 p->numa_faults[cpubuf_idx] = 0;
2189 p->numa_faults[mem_idx] += diff;
2190 p->numa_faults[cpu_idx] += f_diff;
2191 faults += p->numa_faults[mem_idx];
2192 p->total_numa_faults += diff;
2195 * safe because we can only change our own group
2197 * mem_idx represents the offset for a given
2198 * nid and priv in a specific region because it
2199 * is at the beginning of the numa_faults array.
2201 ng->faults[mem_idx] += diff;
2202 ng->faults_cpu[mem_idx] += f_diff;
2203 ng->total_faults += diff;
2204 group_faults += ng->faults[mem_idx];
2209 if (faults > max_faults) {
2210 max_faults = faults;
2213 } else if (group_faults > max_faults) {
2214 max_faults = group_faults;
2220 numa_group_count_active_nodes(ng);
2221 spin_unlock_irq(group_lock);
2222 max_nid = preferred_group_nid(p, max_nid);
2226 /* Set the new preferred node */
2227 if (max_nid != p->numa_preferred_nid)
2228 sched_setnuma(p, max_nid);
2231 update_task_scan_period(p, fault_types[0], fault_types[1]);
2234 static inline int get_numa_group(struct numa_group *grp)
2236 return refcount_inc_not_zero(&grp->refcount);
2239 static inline void put_numa_group(struct numa_group *grp)
2241 if (refcount_dec_and_test(&grp->refcount))
2242 kfree_rcu(grp, rcu);
2245 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2248 struct numa_group *grp, *my_grp;
2249 struct task_struct *tsk;
2251 int cpu = cpupid_to_cpu(cpupid);
2254 if (unlikely(!deref_curr_numa_group(p))) {
2255 unsigned int size = sizeof(struct numa_group) +
2256 4*nr_node_ids*sizeof(unsigned long);
2258 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2262 refcount_set(&grp->refcount, 1);
2263 grp->active_nodes = 1;
2264 grp->max_faults_cpu = 0;
2265 spin_lock_init(&grp->lock);
2267 /* Second half of the array tracks nids where faults happen */
2268 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2271 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2272 grp->faults[i] = p->numa_faults[i];
2274 grp->total_faults = p->total_numa_faults;
2277 rcu_assign_pointer(p->numa_group, grp);
2281 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2283 if (!cpupid_match_pid(tsk, cpupid))
2286 grp = rcu_dereference(tsk->numa_group);
2290 my_grp = deref_curr_numa_group(p);
2295 * Only join the other group if its bigger; if we're the bigger group,
2296 * the other task will join us.
2298 if (my_grp->nr_tasks > grp->nr_tasks)
2302 * Tie-break on the grp address.
2304 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2307 /* Always join threads in the same process. */
2308 if (tsk->mm == current->mm)
2311 /* Simple filter to avoid false positives due to PID collisions */
2312 if (flags & TNF_SHARED)
2315 /* Update priv based on whether false sharing was detected */
2318 if (join && !get_numa_group(grp))
2326 BUG_ON(irqs_disabled());
2327 double_lock_irq(&my_grp->lock, &grp->lock);
2329 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2330 my_grp->faults[i] -= p->numa_faults[i];
2331 grp->faults[i] += p->numa_faults[i];
2333 my_grp->total_faults -= p->total_numa_faults;
2334 grp->total_faults += p->total_numa_faults;
2339 spin_unlock(&my_grp->lock);
2340 spin_unlock_irq(&grp->lock);
2342 rcu_assign_pointer(p->numa_group, grp);
2344 put_numa_group(my_grp);
2353 * Get rid of NUMA staticstics associated with a task (either current or dead).
2354 * If @final is set, the task is dead and has reached refcount zero, so we can
2355 * safely free all relevant data structures. Otherwise, there might be
2356 * concurrent reads from places like load balancing and procfs, and we should
2357 * reset the data back to default state without freeing ->numa_faults.
2359 void task_numa_free(struct task_struct *p, bool final)
2361 /* safe: p either is current or is being freed by current */
2362 struct numa_group *grp = rcu_dereference_raw(p->numa_group);
2363 unsigned long *numa_faults = p->numa_faults;
2364 unsigned long flags;
2371 spin_lock_irqsave(&grp->lock, flags);
2372 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2373 grp->faults[i] -= p->numa_faults[i];
2374 grp->total_faults -= p->total_numa_faults;
2377 spin_unlock_irqrestore(&grp->lock, flags);
2378 RCU_INIT_POINTER(p->numa_group, NULL);
2379 put_numa_group(grp);
2383 p->numa_faults = NULL;
2386 p->total_numa_faults = 0;
2387 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2393 * Got a PROT_NONE fault for a page on @node.
2395 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2397 struct task_struct *p = current;
2398 bool migrated = flags & TNF_MIGRATED;
2399 int cpu_node = task_node(current);
2400 int local = !!(flags & TNF_FAULT_LOCAL);
2401 struct numa_group *ng;
2404 if (!static_branch_likely(&sched_numa_balancing))
2407 /* for example, ksmd faulting in a user's mm */
2411 /* Allocate buffer to track faults on a per-node basis */
2412 if (unlikely(!p->numa_faults)) {
2413 int size = sizeof(*p->numa_faults) *
2414 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2416 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2417 if (!p->numa_faults)
2420 p->total_numa_faults = 0;
2421 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2425 * First accesses are treated as private, otherwise consider accesses
2426 * to be private if the accessing pid has not changed
2428 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2431 priv = cpupid_match_pid(p, last_cpupid);
2432 if (!priv && !(flags & TNF_NO_GROUP))
2433 task_numa_group(p, last_cpupid, flags, &priv);
2437 * If a workload spans multiple NUMA nodes, a shared fault that
2438 * occurs wholly within the set of nodes that the workload is
2439 * actively using should be counted as local. This allows the
2440 * scan rate to slow down when a workload has settled down.
2442 ng = deref_curr_numa_group(p);
2443 if (!priv && !local && ng && ng->active_nodes > 1 &&
2444 numa_is_active_node(cpu_node, ng) &&
2445 numa_is_active_node(mem_node, ng))
2449 * Retry to migrate task to preferred node periodically, in case it
2450 * previously failed, or the scheduler moved us.
2452 if (time_after(jiffies, p->numa_migrate_retry)) {
2453 task_numa_placement(p);
2454 numa_migrate_preferred(p);
2458 p->numa_pages_migrated += pages;
2459 if (flags & TNF_MIGRATE_FAIL)
2460 p->numa_faults_locality[2] += pages;
2462 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2463 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2464 p->numa_faults_locality[local] += pages;
2467 static void reset_ptenuma_scan(struct task_struct *p)
2470 * We only did a read acquisition of the mmap sem, so
2471 * p->mm->numa_scan_seq is written to without exclusive access
2472 * and the update is not guaranteed to be atomic. That's not
2473 * much of an issue though, since this is just used for
2474 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2475 * expensive, to avoid any form of compiler optimizations:
2477 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2478 p->mm->numa_scan_offset = 0;
2482 * The expensive part of numa migration is done from task_work context.
2483 * Triggered from task_tick_numa().
2485 static void task_numa_work(struct callback_head *work)
2487 unsigned long migrate, next_scan, now = jiffies;
2488 struct task_struct *p = current;
2489 struct mm_struct *mm = p->mm;
2490 u64 runtime = p->se.sum_exec_runtime;
2491 struct vm_area_struct *vma;
2492 unsigned long start, end;
2493 unsigned long nr_pte_updates = 0;
2494 long pages, virtpages;
2496 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2500 * Who cares about NUMA placement when they're dying.
2502 * NOTE: make sure not to dereference p->mm before this check,
2503 * exit_task_work() happens _after_ exit_mm() so we could be called
2504 * without p->mm even though we still had it when we enqueued this
2507 if (p->flags & PF_EXITING)
2510 if (!mm->numa_next_scan) {
2511 mm->numa_next_scan = now +
2512 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2516 * Enforce maximal scan/migration frequency..
2518 migrate = mm->numa_next_scan;
2519 if (time_before(now, migrate))
2522 if (p->numa_scan_period == 0) {
2523 p->numa_scan_period_max = task_scan_max(p);
2524 p->numa_scan_period = task_scan_start(p);
2527 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2528 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2532 * Delay this task enough that another task of this mm will likely win
2533 * the next time around.
2535 p->node_stamp += 2 * TICK_NSEC;
2537 start = mm->numa_scan_offset;
2538 pages = sysctl_numa_balancing_scan_size;
2539 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2540 virtpages = pages * 8; /* Scan up to this much virtual space */
2545 if (!down_read_trylock(&mm->mmap_sem))
2547 vma = find_vma(mm, start);
2549 reset_ptenuma_scan(p);
2553 for (; vma; vma = vma->vm_next) {
2554 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2555 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2560 * Shared library pages mapped by multiple processes are not
2561 * migrated as it is expected they are cache replicated. Avoid
2562 * hinting faults in read-only file-backed mappings or the vdso
2563 * as migrating the pages will be of marginal benefit.
2566 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2570 * Skip inaccessible VMAs to avoid any confusion between
2571 * PROT_NONE and NUMA hinting ptes
2573 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2577 start = max(start, vma->vm_start);
2578 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2579 end = min(end, vma->vm_end);
2580 nr_pte_updates = change_prot_numa(vma, start, end);
2583 * Try to scan sysctl_numa_balancing_size worth of
2584 * hpages that have at least one present PTE that
2585 * is not already pte-numa. If the VMA contains
2586 * areas that are unused or already full of prot_numa
2587 * PTEs, scan up to virtpages, to skip through those
2591 pages -= (end - start) >> PAGE_SHIFT;
2592 virtpages -= (end - start) >> PAGE_SHIFT;
2595 if (pages <= 0 || virtpages <= 0)
2599 } while (end != vma->vm_end);
2604 * It is possible to reach the end of the VMA list but the last few
2605 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2606 * would find the !migratable VMA on the next scan but not reset the
2607 * scanner to the start so check it now.
2610 mm->numa_scan_offset = start;
2612 reset_ptenuma_scan(p);
2613 up_read(&mm->mmap_sem);
2616 * Make sure tasks use at least 32x as much time to run other code
2617 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2618 * Usually update_task_scan_period slows down scanning enough; on an
2619 * overloaded system we need to limit overhead on a per task basis.
2621 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2622 u64 diff = p->se.sum_exec_runtime - runtime;
2623 p->node_stamp += 32 * diff;
2627 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
2630 struct mm_struct *mm = p->mm;
2633 mm_users = atomic_read(&mm->mm_users);
2634 if (mm_users == 1) {
2635 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2636 mm->numa_scan_seq = 0;
2640 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
2641 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2642 /* Protect against double add, see task_tick_numa and task_numa_work */
2643 p->numa_work.next = &p->numa_work;
2644 p->numa_faults = NULL;
2645 RCU_INIT_POINTER(p->numa_group, NULL);
2646 p->last_task_numa_placement = 0;
2647 p->last_sum_exec_runtime = 0;
2649 init_task_work(&p->numa_work, task_numa_work);
2651 /* New address space, reset the preferred nid */
2652 if (!(clone_flags & CLONE_VM)) {
2653 p->numa_preferred_nid = NUMA_NO_NODE;
2658 * New thread, keep existing numa_preferred_nid which should be copied
2659 * already by arch_dup_task_struct but stagger when scans start.
2664 delay = min_t(unsigned int, task_scan_max(current),
2665 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
2666 delay += 2 * TICK_NSEC;
2667 p->node_stamp = delay;
2672 * Drive the periodic memory faults..
2674 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2676 struct callback_head *work = &curr->numa_work;
2680 * We don't care about NUMA placement if we don't have memory.
2682 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2686 * Using runtime rather than walltime has the dual advantage that
2687 * we (mostly) drive the selection from busy threads and that the
2688 * task needs to have done some actual work before we bother with
2691 now = curr->se.sum_exec_runtime;
2692 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2694 if (now > curr->node_stamp + period) {
2695 if (!curr->node_stamp)
2696 curr->numa_scan_period = task_scan_start(curr);
2697 curr->node_stamp += period;
2699 if (!time_before(jiffies, curr->mm->numa_next_scan))
2700 task_work_add(curr, work, true);
2704 static void update_scan_period(struct task_struct *p, int new_cpu)
2706 int src_nid = cpu_to_node(task_cpu(p));
2707 int dst_nid = cpu_to_node(new_cpu);
2709 if (!static_branch_likely(&sched_numa_balancing))
2712 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2715 if (src_nid == dst_nid)
2719 * Allow resets if faults have been trapped before one scan
2720 * has completed. This is most likely due to a new task that
2721 * is pulled cross-node due to wakeups or load balancing.
2723 if (p->numa_scan_seq) {
2725 * Avoid scan adjustments if moving to the preferred
2726 * node or if the task was not previously running on
2727 * the preferred node.
2729 if (dst_nid == p->numa_preferred_nid ||
2730 (p->numa_preferred_nid != NUMA_NO_NODE &&
2731 src_nid != p->numa_preferred_nid))
2735 p->numa_scan_period = task_scan_start(p);
2739 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2743 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2747 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2751 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2755 #endif /* CONFIG_NUMA_BALANCING */
2758 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2760 update_load_add(&cfs_rq->load, se->load.weight);
2762 if (entity_is_task(se)) {
2763 struct rq *rq = rq_of(cfs_rq);
2765 account_numa_enqueue(rq, task_of(se));
2766 list_add(&se->group_node, &rq->cfs_tasks);
2769 cfs_rq->nr_running++;
2773 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2775 update_load_sub(&cfs_rq->load, se->load.weight);
2777 if (entity_is_task(se)) {
2778 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2779 list_del_init(&se->group_node);
2782 cfs_rq->nr_running--;
2786 * Signed add and clamp on underflow.
2788 * Explicitly do a load-store to ensure the intermediate value never hits
2789 * memory. This allows lockless observations without ever seeing the negative
2792 #define add_positive(_ptr, _val) do { \
2793 typeof(_ptr) ptr = (_ptr); \
2794 typeof(_val) val = (_val); \
2795 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2799 if (val < 0 && res > var) \
2802 WRITE_ONCE(*ptr, res); \
2806 * Unsigned subtract and clamp on underflow.
2808 * Explicitly do a load-store to ensure the intermediate value never hits
2809 * memory. This allows lockless observations without ever seeing the negative
2812 #define sub_positive(_ptr, _val) do { \
2813 typeof(_ptr) ptr = (_ptr); \
2814 typeof(*ptr) val = (_val); \
2815 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2819 WRITE_ONCE(*ptr, res); \
2823 * Remove and clamp on negative, from a local variable.
2825 * A variant of sub_positive(), which does not use explicit load-store
2826 * and is thus optimized for local variable updates.
2828 #define lsub_positive(_ptr, _val) do { \
2829 typeof(_ptr) ptr = (_ptr); \
2830 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
2835 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2837 cfs_rq->runnable_weight += se->runnable_weight;
2839 cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2840 cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2844 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2846 cfs_rq->runnable_weight -= se->runnable_weight;
2848 sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2849 sub_positive(&cfs_rq->avg.runnable_load_sum,
2850 se_runnable(se) * se->avg.runnable_load_sum);
2854 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2856 cfs_rq->avg.load_avg += se->avg.load_avg;
2857 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2861 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2863 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2864 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2868 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2870 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2872 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2874 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2877 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2878 unsigned long weight, unsigned long runnable)
2881 /* commit outstanding execution time */
2882 if (cfs_rq->curr == se)
2883 update_curr(cfs_rq);
2884 account_entity_dequeue(cfs_rq, se);
2885 dequeue_runnable_load_avg(cfs_rq, se);
2887 dequeue_load_avg(cfs_rq, se);
2889 se->runnable_weight = runnable;
2890 update_load_set(&se->load, weight);
2894 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2896 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2897 se->avg.runnable_load_avg =
2898 div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2902 enqueue_load_avg(cfs_rq, se);
2904 account_entity_enqueue(cfs_rq, se);
2905 enqueue_runnable_load_avg(cfs_rq, se);
2909 void reweight_task(struct task_struct *p, int prio)
2911 struct sched_entity *se = &p->se;
2912 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2913 struct load_weight *load = &se->load;
2914 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2916 reweight_entity(cfs_rq, se, weight, weight);
2917 load->inv_weight = sched_prio_to_wmult[prio];
2920 #ifdef CONFIG_FAIR_GROUP_SCHED
2923 * All this does is approximate the hierarchical proportion which includes that
2924 * global sum we all love to hate.
2926 * That is, the weight of a group entity, is the proportional share of the
2927 * group weight based on the group runqueue weights. That is:
2929 * tg->weight * grq->load.weight
2930 * ge->load.weight = ----------------------------- (1)
2931 * \Sum grq->load.weight
2933 * Now, because computing that sum is prohibitively expensive to compute (been
2934 * there, done that) we approximate it with this average stuff. The average
2935 * moves slower and therefore the approximation is cheaper and more stable.
2937 * So instead of the above, we substitute:
2939 * grq->load.weight -> grq->avg.load_avg (2)
2941 * which yields the following:
2943 * tg->weight * grq->avg.load_avg
2944 * ge->load.weight = ------------------------------ (3)
2947 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2949 * That is shares_avg, and it is right (given the approximation (2)).
2951 * The problem with it is that because the average is slow -- it was designed
2952 * to be exactly that of course -- this leads to transients in boundary
2953 * conditions. In specific, the case where the group was idle and we start the
2954 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2955 * yielding bad latency etc..
2957 * Now, in that special case (1) reduces to:
2959 * tg->weight * grq->load.weight
2960 * ge->load.weight = ----------------------------- = tg->weight (4)
2963 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2965 * So what we do is modify our approximation (3) to approach (4) in the (near)
2970 * tg->weight * grq->load.weight
2971 * --------------------------------------------------- (5)
2972 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2974 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2975 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2978 * tg->weight * grq->load.weight
2979 * ge->load.weight = ----------------------------- (6)
2984 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2985 * max(grq->load.weight, grq->avg.load_avg)
2987 * And that is shares_weight and is icky. In the (near) UP case it approaches
2988 * (4) while in the normal case it approaches (3). It consistently
2989 * overestimates the ge->load.weight and therefore:
2991 * \Sum ge->load.weight >= tg->weight
2995 static long calc_group_shares(struct cfs_rq *cfs_rq)
2997 long tg_weight, tg_shares, load, shares;
2998 struct task_group *tg = cfs_rq->tg;
3000 tg_shares = READ_ONCE(tg->shares);
3002 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3004 tg_weight = atomic_long_read(&tg->load_avg);
3006 /* Ensure tg_weight >= load */
3007 tg_weight -= cfs_rq->tg_load_avg_contrib;
3010 shares = (tg_shares * load);
3012 shares /= tg_weight;
3015 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3016 * of a group with small tg->shares value. It is a floor value which is
3017 * assigned as a minimum load.weight to the sched_entity representing
3018 * the group on a CPU.
3020 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3021 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3022 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3023 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3026 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3030 * This calculates the effective runnable weight for a group entity based on
3031 * the group entity weight calculated above.
3033 * Because of the above approximation (2), our group entity weight is
3034 * an load_avg based ratio (3). This means that it includes blocked load and
3035 * does not represent the runnable weight.
3037 * Approximate the group entity's runnable weight per ratio from the group
3040 * grq->avg.runnable_load_avg
3041 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
3044 * However, analogous to above, since the avg numbers are slow, this leads to
3045 * transients in the from-idle case. Instead we use:
3047 * ge->runnable_weight = ge->load.weight *
3049 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
3050 * ----------------------------------------------------- (8)
3051 * max(grq->avg.load_avg, grq->load.weight)
3053 * Where these max() serve both to use the 'instant' values to fix the slow
3054 * from-idle and avoid the /0 on to-idle, similar to (6).
3056 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
3058 long runnable, load_avg;
3060 load_avg = max(cfs_rq->avg.load_avg,
3061 scale_load_down(cfs_rq->load.weight));
3063 runnable = max(cfs_rq->avg.runnable_load_avg,
3064 scale_load_down(cfs_rq->runnable_weight));
3068 runnable /= load_avg;
3070 return clamp_t(long, runnable, MIN_SHARES, shares);
3072 #endif /* CONFIG_SMP */
3074 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3077 * Recomputes the group entity based on the current state of its group
3080 static void update_cfs_group(struct sched_entity *se)
3082 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3083 long shares, runnable;
3088 if (throttled_hierarchy(gcfs_rq))
3092 runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
3094 if (likely(se->load.weight == shares))
3097 shares = calc_group_shares(gcfs_rq);
3098 runnable = calc_group_runnable(gcfs_rq, shares);
3101 reweight_entity(cfs_rq_of(se), se, shares, runnable);
3104 #else /* CONFIG_FAIR_GROUP_SCHED */
3105 static inline void update_cfs_group(struct sched_entity *se)
3108 #endif /* CONFIG_FAIR_GROUP_SCHED */
3110 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3112 struct rq *rq = rq_of(cfs_rq);
3114 if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) {
3116 * There are a few boundary cases this might miss but it should
3117 * get called often enough that that should (hopefully) not be
3120 * It will not get called when we go idle, because the idle
3121 * thread is a different class (!fair), nor will the utilization
3122 * number include things like RT tasks.
3124 * As is, the util number is not freq-invariant (we'd have to
3125 * implement arch_scale_freq_capacity() for that).
3129 cpufreq_update_util(rq, flags);
3134 #ifdef CONFIG_FAIR_GROUP_SCHED
3136 * update_tg_load_avg - update the tg's load avg
3137 * @cfs_rq: the cfs_rq whose avg changed
3138 * @force: update regardless of how small the difference
3140 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3141 * However, because tg->load_avg is a global value there are performance
3144 * In order to avoid having to look at the other cfs_rq's, we use a
3145 * differential update where we store the last value we propagated. This in
3146 * turn allows skipping updates if the differential is 'small'.
3148 * Updating tg's load_avg is necessary before update_cfs_share().
3150 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3152 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3155 * No need to update load_avg for root_task_group as it is not used.
3157 if (cfs_rq->tg == &root_task_group)
3160 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3161 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3162 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3167 * Called within set_task_rq() right before setting a task's CPU. The
3168 * caller only guarantees p->pi_lock is held; no other assumptions,
3169 * including the state of rq->lock, should be made.
3171 void set_task_rq_fair(struct sched_entity *se,
3172 struct cfs_rq *prev, struct cfs_rq *next)
3174 u64 p_last_update_time;
3175 u64 n_last_update_time;
3177 if (!sched_feat(ATTACH_AGE_LOAD))
3181 * We are supposed to update the task to "current" time, then its up to
3182 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3183 * getting what current time is, so simply throw away the out-of-date
3184 * time. This will result in the wakee task is less decayed, but giving
3185 * the wakee more load sounds not bad.
3187 if (!(se->avg.last_update_time && prev))
3190 #ifndef CONFIG_64BIT
3192 u64 p_last_update_time_copy;
3193 u64 n_last_update_time_copy;
3196 p_last_update_time_copy = prev->load_last_update_time_copy;
3197 n_last_update_time_copy = next->load_last_update_time_copy;
3201 p_last_update_time = prev->avg.last_update_time;
3202 n_last_update_time = next->avg.last_update_time;
3204 } while (p_last_update_time != p_last_update_time_copy ||
3205 n_last_update_time != n_last_update_time_copy);
3208 p_last_update_time = prev->avg.last_update_time;
3209 n_last_update_time = next->avg.last_update_time;
3211 __update_load_avg_blocked_se(p_last_update_time, se);
3212 se->avg.last_update_time = n_last_update_time;
3217 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3218 * propagate its contribution. The key to this propagation is the invariant
3219 * that for each group:
3221 * ge->avg == grq->avg (1)
3223 * _IFF_ we look at the pure running and runnable sums. Because they
3224 * represent the very same entity, just at different points in the hierarchy.
3226 * Per the above update_tg_cfs_util() is trivial and simply copies the running
3227 * sum over (but still wrong, because the group entity and group rq do not have
3228 * their PELT windows aligned).
3230 * However, update_tg_cfs_runnable() is more complex. So we have:
3232 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3234 * And since, like util, the runnable part should be directly transferable,
3235 * the following would _appear_ to be the straight forward approach:
3237 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3239 * And per (1) we have:
3241 * ge->avg.runnable_avg == grq->avg.runnable_avg
3245 * ge->load.weight * grq->avg.load_avg
3246 * ge->avg.load_avg = ----------------------------------- (4)
3249 * Except that is wrong!
3251 * Because while for entities historical weight is not important and we
3252 * really only care about our future and therefore can consider a pure
3253 * runnable sum, runqueues can NOT do this.
3255 * We specifically want runqueues to have a load_avg that includes
3256 * historical weights. Those represent the blocked load, the load we expect
3257 * to (shortly) return to us. This only works by keeping the weights as
3258 * integral part of the sum. We therefore cannot decompose as per (3).
3260 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3261 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3262 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3263 * runnable section of these tasks overlap (or not). If they were to perfectly
3264 * align the rq as a whole would be runnable 2/3 of the time. If however we
3265 * always have at least 1 runnable task, the rq as a whole is always runnable.
3267 * So we'll have to approximate.. :/
3269 * Given the constraint:
3271 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3273 * We can construct a rule that adds runnable to a rq by assuming minimal
3276 * On removal, we'll assume each task is equally runnable; which yields:
3278 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3280 * XXX: only do this for the part of runnable > running ?
3285 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3287 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3289 /* Nothing to update */
3294 * The relation between sum and avg is:
3296 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3298 * however, the PELT windows are not aligned between grq and gse.
3301 /* Set new sched_entity's utilization */
3302 se->avg.util_avg = gcfs_rq->avg.util_avg;
3303 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3305 /* Update parent cfs_rq utilization */
3306 add_positive(&cfs_rq->avg.util_avg, delta);
3307 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3311 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3313 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3314 unsigned long runnable_load_avg, load_avg;
3315 u64 runnable_load_sum, load_sum = 0;
3321 gcfs_rq->prop_runnable_sum = 0;
3323 if (runnable_sum >= 0) {
3325 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3326 * the CPU is saturated running == runnable.
3328 runnable_sum += se->avg.load_sum;
3329 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3332 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3333 * assuming all tasks are equally runnable.
3335 if (scale_load_down(gcfs_rq->load.weight)) {
3336 load_sum = div_s64(gcfs_rq->avg.load_sum,
3337 scale_load_down(gcfs_rq->load.weight));
3340 /* But make sure to not inflate se's runnable */
3341 runnable_sum = min(se->avg.load_sum, load_sum);
3345 * runnable_sum can't be lower than running_sum
3346 * Rescale running sum to be in the same range as runnable sum
3347 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3348 * runnable_sum is in [0 : LOAD_AVG_MAX]
3350 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3351 runnable_sum = max(runnable_sum, running_sum);
3353 load_sum = (s64)se_weight(se) * runnable_sum;
3354 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3356 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3357 delta_avg = load_avg - se->avg.load_avg;
3359 se->avg.load_sum = runnable_sum;
3360 se->avg.load_avg = load_avg;
3361 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3362 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3364 runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3365 runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3366 delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3367 delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3369 se->avg.runnable_load_sum = runnable_sum;
3370 se->avg.runnable_load_avg = runnable_load_avg;
3373 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3374 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3378 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3380 cfs_rq->propagate = 1;
3381 cfs_rq->prop_runnable_sum += runnable_sum;
3384 /* Update task and its cfs_rq load average */
3385 static inline int propagate_entity_load_avg(struct sched_entity *se)
3387 struct cfs_rq *cfs_rq, *gcfs_rq;
3389 if (entity_is_task(se))
3392 gcfs_rq = group_cfs_rq(se);
3393 if (!gcfs_rq->propagate)
3396 gcfs_rq->propagate = 0;
3398 cfs_rq = cfs_rq_of(se);
3400 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3402 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3403 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3405 trace_pelt_cfs_tp(cfs_rq);
3406 trace_pelt_se_tp(se);
3412 * Check if we need to update the load and the utilization of a blocked
3415 static inline bool skip_blocked_update(struct sched_entity *se)
3417 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3420 * If sched_entity still have not zero load or utilization, we have to
3423 if (se->avg.load_avg || se->avg.util_avg)
3427 * If there is a pending propagation, we have to update the load and
3428 * the utilization of the sched_entity:
3430 if (gcfs_rq->propagate)
3434 * Otherwise, the load and the utilization of the sched_entity is
3435 * already zero and there is no pending propagation, so it will be a
3436 * waste of time to try to decay it:
3441 #else /* CONFIG_FAIR_GROUP_SCHED */
3443 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3445 static inline int propagate_entity_load_avg(struct sched_entity *se)
3450 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3452 #endif /* CONFIG_FAIR_GROUP_SCHED */
3455 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3456 * @now: current time, as per cfs_rq_clock_pelt()
3457 * @cfs_rq: cfs_rq to update
3459 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3460 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3461 * post_init_entity_util_avg().
3463 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3465 * Returns true if the load decayed or we removed load.
3467 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3468 * call update_tg_load_avg() when this function returns true.
3471 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3473 unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3474 struct sched_avg *sa = &cfs_rq->avg;
3477 if (cfs_rq->removed.nr) {
3479 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3481 raw_spin_lock(&cfs_rq->removed.lock);
3482 swap(cfs_rq->removed.util_avg, removed_util);
3483 swap(cfs_rq->removed.load_avg, removed_load);
3484 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3485 cfs_rq->removed.nr = 0;
3486 raw_spin_unlock(&cfs_rq->removed.lock);
3489 sub_positive(&sa->load_avg, r);
3490 sub_positive(&sa->load_sum, r * divider);
3493 sub_positive(&sa->util_avg, r);
3494 sub_positive(&sa->util_sum, r * divider);
3496 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3501 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
3503 #ifndef CONFIG_64BIT
3505 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3509 cfs_rq_util_change(cfs_rq, 0);
3515 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3516 * @cfs_rq: cfs_rq to attach to
3517 * @se: sched_entity to attach
3518 * @flags: migration hints
3520 * Must call update_cfs_rq_load_avg() before this, since we rely on
3521 * cfs_rq->avg.last_update_time being current.
3523 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3525 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3528 * When we attach the @se to the @cfs_rq, we must align the decay
3529 * window because without that, really weird and wonderful things can
3534 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3535 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3538 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3539 * period_contrib. This isn't strictly correct, but since we're
3540 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3543 se->avg.util_sum = se->avg.util_avg * divider;
3545 se->avg.load_sum = divider;
3546 if (se_weight(se)) {
3548 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3551 se->avg.runnable_load_sum = se->avg.load_sum;
3553 enqueue_load_avg(cfs_rq, se);
3554 cfs_rq->avg.util_avg += se->avg.util_avg;
3555 cfs_rq->avg.util_sum += se->avg.util_sum;
3557 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3559 cfs_rq_util_change(cfs_rq, flags);
3561 trace_pelt_cfs_tp(cfs_rq);
3565 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3566 * @cfs_rq: cfs_rq to detach from
3567 * @se: sched_entity to detach
3569 * Must call update_cfs_rq_load_avg() before this, since we rely on
3570 * cfs_rq->avg.last_update_time being current.
3572 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3574 dequeue_load_avg(cfs_rq, se);
3575 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3576 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3578 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3580 cfs_rq_util_change(cfs_rq, 0);
3582 trace_pelt_cfs_tp(cfs_rq);
3586 * Optional action to be done while updating the load average
3588 #define UPDATE_TG 0x1
3589 #define SKIP_AGE_LOAD 0x2
3590 #define DO_ATTACH 0x4
3592 /* Update task and its cfs_rq load average */
3593 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3595 u64 now = cfs_rq_clock_pelt(cfs_rq);
3599 * Track task load average for carrying it to new CPU after migrated, and
3600 * track group sched_entity load average for task_h_load calc in migration
3602 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3603 __update_load_avg_se(now, cfs_rq, se);
3605 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3606 decayed |= propagate_entity_load_avg(se);
3608 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3611 * DO_ATTACH means we're here from enqueue_entity().
3612 * !last_update_time means we've passed through
3613 * migrate_task_rq_fair() indicating we migrated.
3615 * IOW we're enqueueing a task on a new CPU.
3617 attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
3618 update_tg_load_avg(cfs_rq, 0);
3620 } else if (decayed && (flags & UPDATE_TG))
3621 update_tg_load_avg(cfs_rq, 0);
3624 #ifndef CONFIG_64BIT
3625 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3627 u64 last_update_time_copy;
3628 u64 last_update_time;
3631 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3633 last_update_time = cfs_rq->avg.last_update_time;
3634 } while (last_update_time != last_update_time_copy);
3636 return last_update_time;
3639 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3641 return cfs_rq->avg.last_update_time;
3646 * Synchronize entity load avg of dequeued entity without locking
3649 static void sync_entity_load_avg(struct sched_entity *se)
3651 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3652 u64 last_update_time;
3654 last_update_time = cfs_rq_last_update_time(cfs_rq);
3655 __update_load_avg_blocked_se(last_update_time, se);
3659 * Task first catches up with cfs_rq, and then subtract
3660 * itself from the cfs_rq (task must be off the queue now).
3662 static void remove_entity_load_avg(struct sched_entity *se)
3664 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3665 unsigned long flags;
3668 * tasks cannot exit without having gone through wake_up_new_task() ->
3669 * post_init_entity_util_avg() which will have added things to the
3670 * cfs_rq, so we can remove unconditionally.
3673 sync_entity_load_avg(se);
3675 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3676 ++cfs_rq->removed.nr;
3677 cfs_rq->removed.util_avg += se->avg.util_avg;
3678 cfs_rq->removed.load_avg += se->avg.load_avg;
3679 cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */
3680 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3683 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3685 return cfs_rq->avg.runnable_load_avg;
3688 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3690 return cfs_rq->avg.load_avg;
3693 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3695 static inline unsigned long task_util(struct task_struct *p)
3697 return READ_ONCE(p->se.avg.util_avg);
3700 static inline unsigned long _task_util_est(struct task_struct *p)
3702 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3704 return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED);
3707 static inline unsigned long task_util_est(struct task_struct *p)
3709 return max(task_util(p), _task_util_est(p));
3712 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3713 struct task_struct *p)
3715 unsigned int enqueued;
3717 if (!sched_feat(UTIL_EST))
3720 /* Update root cfs_rq's estimated utilization */
3721 enqueued = cfs_rq->avg.util_est.enqueued;
3722 enqueued += _task_util_est(p);
3723 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3727 * Check if a (signed) value is within a specified (unsigned) margin,
3728 * based on the observation that:
3730 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3732 * NOTE: this only works when value + maring < INT_MAX.
3734 static inline bool within_margin(int value, int margin)
3736 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3740 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3742 long last_ewma_diff;
3746 if (!sched_feat(UTIL_EST))
3749 /* Update root cfs_rq's estimated utilization */
3750 ue.enqueued = cfs_rq->avg.util_est.enqueued;
3751 ue.enqueued -= min_t(unsigned int, ue.enqueued, _task_util_est(p));
3752 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3755 * Skip update of task's estimated utilization when the task has not
3756 * yet completed an activation, e.g. being migrated.
3762 * If the PELT values haven't changed since enqueue time,
3763 * skip the util_est update.
3765 ue = p->se.avg.util_est;
3766 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3770 * Skip update of task's estimated utilization when its EWMA is
3771 * already ~1% close to its last activation value.
3773 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3774 last_ewma_diff = ue.enqueued - ue.ewma;
3775 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3779 * To avoid overestimation of actual task utilization, skip updates if
3780 * we cannot grant there is idle time in this CPU.
3782 cpu = cpu_of(rq_of(cfs_rq));
3783 if (task_util(p) > capacity_orig_of(cpu))
3787 * Update Task's estimated utilization
3789 * When *p completes an activation we can consolidate another sample
3790 * of the task size. This is done by storing the current PELT value
3791 * as ue.enqueued and by using this value to update the Exponential
3792 * Weighted Moving Average (EWMA):
3794 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
3795 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
3796 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
3797 * = w * ( last_ewma_diff ) + ewma(t-1)
3798 * = w * (last_ewma_diff + ewma(t-1) / w)
3800 * Where 'w' is the weight of new samples, which is configured to be
3801 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3803 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
3804 ue.ewma += last_ewma_diff;
3805 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
3806 WRITE_ONCE(p->se.avg.util_est, ue);
3809 static inline int task_fits_capacity(struct task_struct *p, long capacity)
3811 return capacity * 1024 > task_util_est(p) * capacity_margin;
3814 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
3816 if (!static_branch_unlikely(&sched_asym_cpucapacity))
3820 rq->misfit_task_load = 0;
3824 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
3825 rq->misfit_task_load = 0;
3829 rq->misfit_task_load = task_h_load(p);
3832 #else /* CONFIG_SMP */
3834 #define UPDATE_TG 0x0
3835 #define SKIP_AGE_LOAD 0x0
3836 #define DO_ATTACH 0x0
3838 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3840 cfs_rq_util_change(cfs_rq, 0);
3843 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3846 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
3848 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3850 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3856 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
3859 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
3861 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
3863 #endif /* CONFIG_SMP */
3865 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3867 #ifdef CONFIG_SCHED_DEBUG
3868 s64 d = se->vruntime - cfs_rq->min_vruntime;
3873 if (d > 3*sysctl_sched_latency)
3874 schedstat_inc(cfs_rq->nr_spread_over);
3879 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3881 u64 vruntime = cfs_rq->min_vruntime;
3884 * The 'current' period is already promised to the current tasks,
3885 * however the extra weight of the new task will slow them down a
3886 * little, place the new task so that it fits in the slot that
3887 * stays open at the end.
3889 if (initial && sched_feat(START_DEBIT))
3890 vruntime += sched_vslice(cfs_rq, se);
3892 /* sleeps up to a single latency don't count. */
3894 unsigned long thresh = sysctl_sched_latency;
3897 * Halve their sleep time's effect, to allow
3898 * for a gentler effect of sleepers:
3900 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3906 /* ensure we never gain time by being placed backwards. */
3907 se->vruntime = max_vruntime(se->vruntime, vruntime);
3910 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3912 static inline void check_schedstat_required(void)
3914 #ifdef CONFIG_SCHEDSTATS
3915 if (schedstat_enabled())
3918 /* Force schedstat enabled if a dependent tracepoint is active */
3919 if (trace_sched_stat_wait_enabled() ||
3920 trace_sched_stat_sleep_enabled() ||
3921 trace_sched_stat_iowait_enabled() ||
3922 trace_sched_stat_blocked_enabled() ||
3923 trace_sched_stat_runtime_enabled()) {
3924 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3925 "stat_blocked and stat_runtime require the "
3926 "kernel parameter schedstats=enable or "
3927 "kernel.sched_schedstats=1\n");
3938 * update_min_vruntime()
3939 * vruntime -= min_vruntime
3943 * update_min_vruntime()
3944 * vruntime += min_vruntime
3946 * this way the vruntime transition between RQs is done when both
3947 * min_vruntime are up-to-date.
3951 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3952 * vruntime -= min_vruntime
3956 * update_min_vruntime()
3957 * vruntime += min_vruntime
3959 * this way we don't have the most up-to-date min_vruntime on the originating
3960 * CPU and an up-to-date min_vruntime on the destination CPU.
3964 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3966 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3967 bool curr = cfs_rq->curr == se;
3970 * If we're the current task, we must renormalise before calling
3974 se->vruntime += cfs_rq->min_vruntime;
3976 update_curr(cfs_rq);
3979 * Otherwise, renormalise after, such that we're placed at the current
3980 * moment in time, instead of some random moment in the past. Being
3981 * placed in the past could significantly boost this task to the
3982 * fairness detriment of existing tasks.
3984 if (renorm && !curr)
3985 se->vruntime += cfs_rq->min_vruntime;
3988 * When enqueuing a sched_entity, we must:
3989 * - Update loads to have both entity and cfs_rq synced with now.
3990 * - Add its load to cfs_rq->runnable_avg
3991 * - For group_entity, update its weight to reflect the new share of
3993 * - Add its new weight to cfs_rq->load.weight
3995 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
3996 update_cfs_group(se);
3997 enqueue_runnable_load_avg(cfs_rq, se);
3998 account_entity_enqueue(cfs_rq, se);
4000 if (flags & ENQUEUE_WAKEUP)
4001 place_entity(cfs_rq, se, 0);
4003 check_schedstat_required();
4004 update_stats_enqueue(cfs_rq, se, flags);
4005 check_spread(cfs_rq, se);
4007 __enqueue_entity(cfs_rq, se);
4010 if (cfs_rq->nr_running == 1) {
4011 list_add_leaf_cfs_rq(cfs_rq);
4012 check_enqueue_throttle(cfs_rq);
4016 static void __clear_buddies_last(struct sched_entity *se)
4018 for_each_sched_entity(se) {
4019 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4020 if (cfs_rq->last != se)
4023 cfs_rq->last = NULL;
4027 static void __clear_buddies_next(struct sched_entity *se)
4029 for_each_sched_entity(se) {
4030 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4031 if (cfs_rq->next != se)
4034 cfs_rq->next = NULL;
4038 static void __clear_buddies_skip(struct sched_entity *se)
4040 for_each_sched_entity(se) {
4041 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4042 if (cfs_rq->skip != se)
4045 cfs_rq->skip = NULL;
4049 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4051 if (cfs_rq->last == se)
4052 __clear_buddies_last(se);
4054 if (cfs_rq->next == se)
4055 __clear_buddies_next(se);
4057 if (cfs_rq->skip == se)
4058 __clear_buddies_skip(se);
4061 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4064 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4067 * Update run-time statistics of the 'current'.
4069 update_curr(cfs_rq);
4072 * When dequeuing a sched_entity, we must:
4073 * - Update loads to have both entity and cfs_rq synced with now.
4074 * - Subtract its load from the cfs_rq->runnable_avg.
4075 * - Subtract its previous weight from cfs_rq->load.weight.
4076 * - For group entity, update its weight to reflect the new share
4077 * of its group cfs_rq.
4079 update_load_avg(cfs_rq, se, UPDATE_TG);
4080 dequeue_runnable_load_avg(cfs_rq, se);
4082 update_stats_dequeue(cfs_rq, se, flags);
4084 clear_buddies(cfs_rq, se);
4086 if (se != cfs_rq->curr)
4087 __dequeue_entity(cfs_rq, se);
4089 account_entity_dequeue(cfs_rq, se);
4092 * Normalize after update_curr(); which will also have moved
4093 * min_vruntime if @se is the one holding it back. But before doing
4094 * update_min_vruntime() again, which will discount @se's position and
4095 * can move min_vruntime forward still more.
4097 if (!(flags & DEQUEUE_SLEEP))
4098 se->vruntime -= cfs_rq->min_vruntime;
4100 /* return excess runtime on last dequeue */
4101 return_cfs_rq_runtime(cfs_rq);
4103 update_cfs_group(se);
4106 * Now advance min_vruntime if @se was the entity holding it back,
4107 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4108 * put back on, and if we advance min_vruntime, we'll be placed back
4109 * further than we started -- ie. we'll be penalized.
4111 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4112 update_min_vruntime(cfs_rq);
4116 * Preempt the current task with a newly woken task if needed:
4119 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4121 unsigned long ideal_runtime, delta_exec;
4122 struct sched_entity *se;
4125 ideal_runtime = sched_slice(cfs_rq, curr);
4126 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4127 if (delta_exec > ideal_runtime) {
4128 resched_curr(rq_of(cfs_rq));
4130 * The current task ran long enough, ensure it doesn't get
4131 * re-elected due to buddy favours.
4133 clear_buddies(cfs_rq, curr);
4138 * Ensure that a task that missed wakeup preemption by a
4139 * narrow margin doesn't have to wait for a full slice.
4140 * This also mitigates buddy induced latencies under load.
4142 if (delta_exec < sysctl_sched_min_granularity)
4145 se = __pick_first_entity(cfs_rq);
4146 delta = curr->vruntime - se->vruntime;
4151 if (delta > ideal_runtime)
4152 resched_curr(rq_of(cfs_rq));
4156 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4158 /* 'current' is not kept within the tree. */
4161 * Any task has to be enqueued before it get to execute on
4162 * a CPU. So account for the time it spent waiting on the
4165 update_stats_wait_end(cfs_rq, se);
4166 __dequeue_entity(cfs_rq, se);
4167 update_load_avg(cfs_rq, se, UPDATE_TG);
4170 update_stats_curr_start(cfs_rq, se);
4174 * Track our maximum slice length, if the CPU's load is at
4175 * least twice that of our own weight (i.e. dont track it
4176 * when there are only lesser-weight tasks around):
4178 if (schedstat_enabled() &&
4179 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
4180 schedstat_set(se->statistics.slice_max,
4181 max((u64)schedstat_val(se->statistics.slice_max),
4182 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4185 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4189 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4192 * Pick the next process, keeping these things in mind, in this order:
4193 * 1) keep things fair between processes/task groups
4194 * 2) pick the "next" process, since someone really wants that to run
4195 * 3) pick the "last" process, for cache locality
4196 * 4) do not run the "skip" process, if something else is available
4198 static struct sched_entity *
4199 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4201 struct sched_entity *left = __pick_first_entity(cfs_rq);
4202 struct sched_entity *se;
4205 * If curr is set we have to see if its left of the leftmost entity
4206 * still in the tree, provided there was anything in the tree at all.
4208 if (!left || (curr && entity_before(curr, left)))
4211 se = left; /* ideally we run the leftmost entity */
4214 * Avoid running the skip buddy, if running something else can
4215 * be done without getting too unfair.
4217 if (cfs_rq->skip == se) {
4218 struct sched_entity *second;
4221 second = __pick_first_entity(cfs_rq);
4223 second = __pick_next_entity(se);
4224 if (!second || (curr && entity_before(curr, second)))
4228 if (second && wakeup_preempt_entity(second, left) < 1)
4233 * Prefer last buddy, try to return the CPU to a preempted task.
4235 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4239 * Someone really wants this to run. If it's not unfair, run it.
4241 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4244 clear_buddies(cfs_rq, se);
4249 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4251 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4254 * If still on the runqueue then deactivate_task()
4255 * was not called and update_curr() has to be done:
4258 update_curr(cfs_rq);
4260 /* throttle cfs_rqs exceeding runtime */
4261 check_cfs_rq_runtime(cfs_rq);
4263 check_spread(cfs_rq, prev);
4266 update_stats_wait_start(cfs_rq, prev);
4267 /* Put 'current' back into the tree. */
4268 __enqueue_entity(cfs_rq, prev);
4269 /* in !on_rq case, update occurred at dequeue */
4270 update_load_avg(cfs_rq, prev, 0);
4272 cfs_rq->curr = NULL;
4276 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4279 * Update run-time statistics of the 'current'.
4281 update_curr(cfs_rq);
4284 * Ensure that runnable average is periodically updated.
4286 update_load_avg(cfs_rq, curr, UPDATE_TG);
4287 update_cfs_group(curr);
4289 #ifdef CONFIG_SCHED_HRTICK
4291 * queued ticks are scheduled to match the slice, so don't bother
4292 * validating it and just reschedule.
4295 resched_curr(rq_of(cfs_rq));
4299 * don't let the period tick interfere with the hrtick preemption
4301 if (!sched_feat(DOUBLE_TICK) &&
4302 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4306 if (cfs_rq->nr_running > 1)
4307 check_preempt_tick(cfs_rq, curr);
4311 /**************************************************
4312 * CFS bandwidth control machinery
4315 #ifdef CONFIG_CFS_BANDWIDTH
4317 #ifdef CONFIG_JUMP_LABEL
4318 static struct static_key __cfs_bandwidth_used;
4320 static inline bool cfs_bandwidth_used(void)
4322 return static_key_false(&__cfs_bandwidth_used);
4325 void cfs_bandwidth_usage_inc(void)
4327 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4330 void cfs_bandwidth_usage_dec(void)
4332 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4334 #else /* CONFIG_JUMP_LABEL */
4335 static bool cfs_bandwidth_used(void)
4340 void cfs_bandwidth_usage_inc(void) {}
4341 void cfs_bandwidth_usage_dec(void) {}
4342 #endif /* CONFIG_JUMP_LABEL */
4345 * default period for cfs group bandwidth.
4346 * default: 0.1s, units: nanoseconds
4348 static inline u64 default_cfs_period(void)
4350 return 100000000ULL;
4353 static inline u64 sched_cfs_bandwidth_slice(void)
4355 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4359 * Replenish runtime according to assigned quota and update expiration time.
4360 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4361 * additional synchronization around rq->lock.
4363 * requires cfs_b->lock
4365 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4369 if (cfs_b->quota == RUNTIME_INF)
4372 now = sched_clock_cpu(smp_processor_id());
4373 cfs_b->runtime = cfs_b->quota;
4374 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4375 cfs_b->expires_seq++;
4378 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4380 return &tg->cfs_bandwidth;
4383 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4384 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4386 if (unlikely(cfs_rq->throttle_count))
4387 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4389 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4392 /* returns 0 on failure to allocate runtime */
4393 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4395 struct task_group *tg = cfs_rq->tg;
4396 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4397 u64 amount = 0, min_amount, expires;
4400 /* note: this is a positive sum as runtime_remaining <= 0 */
4401 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4403 raw_spin_lock(&cfs_b->lock);
4404 if (cfs_b->quota == RUNTIME_INF)
4405 amount = min_amount;
4407 start_cfs_bandwidth(cfs_b);
4409 if (cfs_b->runtime > 0) {
4410 amount = min(cfs_b->runtime, min_amount);
4411 cfs_b->runtime -= amount;
4415 expires_seq = cfs_b->expires_seq;
4416 expires = cfs_b->runtime_expires;
4417 raw_spin_unlock(&cfs_b->lock);
4419 cfs_rq->runtime_remaining += amount;
4421 * we may have advanced our local expiration to account for allowed
4422 * spread between our sched_clock and the one on which runtime was
4425 if (cfs_rq->expires_seq != expires_seq) {
4426 cfs_rq->expires_seq = expires_seq;
4427 cfs_rq->runtime_expires = expires;
4430 return cfs_rq->runtime_remaining > 0;
4434 * Note: This depends on the synchronization provided by sched_clock and the
4435 * fact that rq->clock snapshots this value.
4437 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4439 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4441 /* if the deadline is ahead of our clock, nothing to do */
4442 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4445 if (cfs_rq->runtime_remaining < 0)
4449 * If the local deadline has passed we have to consider the
4450 * possibility that our sched_clock is 'fast' and the global deadline
4451 * has not truly expired.
4453 * Fortunately we can check determine whether this the case by checking
4454 * whether the global deadline(cfs_b->expires_seq) has advanced.
4456 if (cfs_rq->expires_seq == cfs_b->expires_seq) {
4457 /* extend local deadline, drift is bounded above by 2 ticks */
4458 cfs_rq->runtime_expires += TICK_NSEC;
4460 /* global deadline is ahead, expiration has passed */
4461 cfs_rq->runtime_remaining = 0;
4465 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4467 /* dock delta_exec before expiring quota (as it could span periods) */
4468 cfs_rq->runtime_remaining -= delta_exec;
4469 expire_cfs_rq_runtime(cfs_rq);
4471 if (likely(cfs_rq->runtime_remaining > 0))
4475 * if we're unable to extend our runtime we resched so that the active
4476 * hierarchy can be throttled
4478 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4479 resched_curr(rq_of(cfs_rq));
4482 static __always_inline
4483 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4485 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4488 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4491 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4493 return cfs_bandwidth_used() && cfs_rq->throttled;
4496 /* check whether cfs_rq, or any parent, is throttled */
4497 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4499 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4503 * Ensure that neither of the group entities corresponding to src_cpu or
4504 * dest_cpu are members of a throttled hierarchy when performing group
4505 * load-balance operations.
4507 static inline int throttled_lb_pair(struct task_group *tg,
4508 int src_cpu, int dest_cpu)
4510 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4512 src_cfs_rq = tg->cfs_rq[src_cpu];
4513 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4515 return throttled_hierarchy(src_cfs_rq) ||
4516 throttled_hierarchy(dest_cfs_rq);
4519 static int tg_unthrottle_up(struct task_group *tg, void *data)
4521 struct rq *rq = data;
4522 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4524 cfs_rq->throttle_count--;
4525 if (!cfs_rq->throttle_count) {
4526 /* adjust cfs_rq_clock_task() */
4527 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4528 cfs_rq->throttled_clock_task;
4530 /* Add cfs_rq with already running entity in the list */
4531 if (cfs_rq->nr_running >= 1)
4532 list_add_leaf_cfs_rq(cfs_rq);
4538 static int tg_throttle_down(struct task_group *tg, void *data)
4540 struct rq *rq = data;
4541 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4543 /* group is entering throttled state, stop time */
4544 if (!cfs_rq->throttle_count) {
4545 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4546 list_del_leaf_cfs_rq(cfs_rq);
4548 cfs_rq->throttle_count++;
4553 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4555 struct rq *rq = rq_of(cfs_rq);
4556 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4557 struct sched_entity *se;
4558 long task_delta, idle_task_delta, dequeue = 1;
4561 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4563 /* freeze hierarchy runnable averages while throttled */
4565 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4568 task_delta = cfs_rq->h_nr_running;
4569 idle_task_delta = cfs_rq->idle_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;
4579 qcfs_rq->idle_h_nr_running -= idle_task_delta;
4581 if (qcfs_rq->load.weight)
4586 sub_nr_running(rq, task_delta);
4588 cfs_rq->throttled = 1;
4589 cfs_rq->throttled_clock = rq_clock(rq);
4590 raw_spin_lock(&cfs_b->lock);
4591 empty = list_empty(&cfs_b->throttled_cfs_rq);
4594 * Add to the _head_ of the list, so that an already-started
4595 * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
4596 * not running add to the tail so that later runqueues don't get starved.
4598 if (cfs_b->distribute_running)
4599 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4601 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4604 * If we're the first throttled task, make sure the bandwidth
4608 start_cfs_bandwidth(cfs_b);
4610 raw_spin_unlock(&cfs_b->lock);
4613 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4615 struct rq *rq = rq_of(cfs_rq);
4616 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4617 struct sched_entity *se;
4619 long task_delta, idle_task_delta;
4621 se = cfs_rq->tg->se[cpu_of(rq)];
4623 cfs_rq->throttled = 0;
4625 update_rq_clock(rq);
4627 raw_spin_lock(&cfs_b->lock);
4628 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4629 list_del_rcu(&cfs_rq->throttled_list);
4630 raw_spin_unlock(&cfs_b->lock);
4632 /* update hierarchical throttle state */
4633 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4635 if (!cfs_rq->load.weight)
4638 task_delta = cfs_rq->h_nr_running;
4639 idle_task_delta = cfs_rq->idle_h_nr_running;
4640 for_each_sched_entity(se) {
4644 cfs_rq = cfs_rq_of(se);
4646 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4647 cfs_rq->h_nr_running += task_delta;
4648 cfs_rq->idle_h_nr_running += idle_task_delta;
4650 if (cfs_rq_throttled(cfs_rq))
4654 assert_list_leaf_cfs_rq(rq);
4657 add_nr_running(rq, task_delta);
4659 /* Determine whether we need to wake up potentially idle CPU: */
4660 if (rq->curr == rq->idle && rq->cfs.nr_running)
4664 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4665 u64 remaining, u64 expires)
4667 struct cfs_rq *cfs_rq;
4669 u64 starting_runtime = remaining;
4672 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4674 struct rq *rq = rq_of(cfs_rq);
4677 rq_lock_irqsave(rq, &rf);
4678 if (!cfs_rq_throttled(cfs_rq))
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;
5262 int idle_h_nr_running = task_has_idle_policy(p);
5265 * The code below (indirectly) updates schedutil which looks at
5266 * the cfs_rq utilization to select a frequency.
5267 * Let's add the task's estimated utilization to the cfs_rq's
5268 * estimated utilization, before we update schedutil.
5270 util_est_enqueue(&rq->cfs, p);
5273 * If in_iowait is set, the code below may not trigger any cpufreq
5274 * utilization updates, so do it here explicitly with the IOWAIT flag
5278 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5280 for_each_sched_entity(se) {
5283 cfs_rq = cfs_rq_of(se);
5284 enqueue_entity(cfs_rq, se, flags);
5287 * end evaluation on encountering a throttled cfs_rq
5289 * note: in the case of encountering a throttled cfs_rq we will
5290 * post the final h_nr_running increment below.
5292 if (cfs_rq_throttled(cfs_rq))
5294 cfs_rq->h_nr_running++;
5295 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5297 flags = ENQUEUE_WAKEUP;
5300 for_each_sched_entity(se) {
5301 cfs_rq = cfs_rq_of(se);
5302 cfs_rq->h_nr_running++;
5303 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5305 if (cfs_rq_throttled(cfs_rq))
5308 update_load_avg(cfs_rq, se, UPDATE_TG);
5309 update_cfs_group(se);
5313 add_nr_running(rq, 1);
5315 * Since new tasks are assigned an initial util_avg equal to
5316 * half of the spare capacity of their CPU, tiny tasks have the
5317 * ability to cross the overutilized threshold, which will
5318 * result in the load balancer ruining all the task placement
5319 * done by EAS. As a way to mitigate that effect, do not account
5320 * for the first enqueue operation of new tasks during the
5321 * overutilized flag detection.
5323 * A better way of solving this problem would be to wait for
5324 * the PELT signals of tasks to converge before taking them
5325 * into account, but that is not straightforward to implement,
5326 * and the following generally works well enough in practice.
5328 if (flags & ENQUEUE_WAKEUP)
5329 update_overutilized_status(rq);
5333 if (cfs_bandwidth_used()) {
5335 * When bandwidth control is enabled; the cfs_rq_throttled()
5336 * breaks in the above iteration can result in incomplete
5337 * leaf list maintenance, resulting in triggering the assertion
5340 for_each_sched_entity(se) {
5341 cfs_rq = cfs_rq_of(se);
5343 if (list_add_leaf_cfs_rq(cfs_rq))
5348 assert_list_leaf_cfs_rq(rq);
5353 static void set_next_buddy(struct sched_entity *se);
5356 * The dequeue_task method is called before nr_running is
5357 * decreased. We remove the task from the rbtree and
5358 * update the fair scheduling stats:
5360 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5362 struct cfs_rq *cfs_rq;
5363 struct sched_entity *se = &p->se;
5364 int task_sleep = flags & DEQUEUE_SLEEP;
5365 int idle_h_nr_running = task_has_idle_policy(p);
5367 for_each_sched_entity(se) {
5368 cfs_rq = cfs_rq_of(se);
5369 dequeue_entity(cfs_rq, se, flags);
5372 * end evaluation on encountering a throttled cfs_rq
5374 * note: in the case of encountering a throttled cfs_rq we will
5375 * post the final h_nr_running decrement below.
5377 if (cfs_rq_throttled(cfs_rq))
5379 cfs_rq->h_nr_running--;
5380 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5382 /* Don't dequeue parent if it has other entities besides us */
5383 if (cfs_rq->load.weight) {
5384 /* Avoid re-evaluating load for this entity: */
5385 se = parent_entity(se);
5387 * Bias pick_next to pick a task from this cfs_rq, as
5388 * p is sleeping when it is within its sched_slice.
5390 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5394 flags |= DEQUEUE_SLEEP;
5397 for_each_sched_entity(se) {
5398 cfs_rq = cfs_rq_of(se);
5399 cfs_rq->h_nr_running--;
5400 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5402 if (cfs_rq_throttled(cfs_rq))
5405 update_load_avg(cfs_rq, se, UPDATE_TG);
5406 update_cfs_group(se);
5410 sub_nr_running(rq, 1);
5412 util_est_dequeue(&rq->cfs, p, task_sleep);
5418 /* Working cpumask for: load_balance, load_balance_newidle. */
5419 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5420 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5422 #ifdef CONFIG_NO_HZ_COMMON
5425 cpumask_var_t idle_cpus_mask;
5427 int has_blocked; /* Idle CPUS has blocked load */
5428 unsigned long next_balance; /* in jiffy units */
5429 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5430 } nohz ____cacheline_aligned;
5432 #endif /* CONFIG_NO_HZ_COMMON */
5434 /* CPU only has SCHED_IDLE tasks enqueued */
5435 static int sched_idle_cpu(int cpu)
5437 struct rq *rq = cpu_rq(cpu);
5439 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
5443 static unsigned long cpu_runnable_load(struct rq *rq)
5445 return cfs_rq_runnable_load_avg(&rq->cfs);
5448 static unsigned long capacity_of(int cpu)
5450 return cpu_rq(cpu)->cpu_capacity;
5453 static unsigned long cpu_avg_load_per_task(int cpu)
5455 struct rq *rq = cpu_rq(cpu);
5456 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5457 unsigned long load_avg = cpu_runnable_load(rq);
5460 return load_avg / nr_running;
5465 static void record_wakee(struct task_struct *p)
5468 * Only decay a single time; tasks that have less then 1 wakeup per
5469 * jiffy will not have built up many flips.
5471 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5472 current->wakee_flips >>= 1;
5473 current->wakee_flip_decay_ts = jiffies;
5476 if (current->last_wakee != p) {
5477 current->last_wakee = p;
5478 current->wakee_flips++;
5483 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5485 * A waker of many should wake a different task than the one last awakened
5486 * at a frequency roughly N times higher than one of its wakees.
5488 * In order to determine whether we should let the load spread vs consolidating
5489 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5490 * partner, and a factor of lls_size higher frequency in the other.
5492 * With both conditions met, we can be relatively sure that the relationship is
5493 * non-monogamous, with partner count exceeding socket size.
5495 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5496 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5499 static int wake_wide(struct task_struct *p)
5501 unsigned int master = current->wakee_flips;
5502 unsigned int slave = p->wakee_flips;
5503 int factor = this_cpu_read(sd_llc_size);
5506 swap(master, slave);
5507 if (slave < factor || master < slave * factor)
5513 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5514 * soonest. For the purpose of speed we only consider the waking and previous
5517 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5518 * cache-affine and is (or will be) idle.
5520 * wake_affine_weight() - considers the weight to reflect the average
5521 * scheduling latency of the CPUs. This seems to work
5522 * for the overloaded case.
5525 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5528 * If this_cpu is idle, it implies the wakeup is from interrupt
5529 * context. Only allow the move if cache is shared. Otherwise an
5530 * interrupt intensive workload could force all tasks onto one
5531 * node depending on the IO topology or IRQ affinity settings.
5533 * If the prev_cpu is idle and cache affine then avoid a migration.
5534 * There is no guarantee that the cache hot data from an interrupt
5535 * is more important than cache hot data on the prev_cpu and from
5536 * a cpufreq perspective, it's better to have higher utilisation
5539 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5540 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5542 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5545 return nr_cpumask_bits;
5549 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5550 int this_cpu, int prev_cpu, int sync)
5552 s64 this_eff_load, prev_eff_load;
5553 unsigned long task_load;
5555 this_eff_load = cpu_runnable_load(cpu_rq(this_cpu));
5558 unsigned long current_load = task_h_load(current);
5560 if (current_load > this_eff_load)
5563 this_eff_load -= current_load;
5566 task_load = task_h_load(p);
5568 this_eff_load += task_load;
5569 if (sched_feat(WA_BIAS))
5570 this_eff_load *= 100;
5571 this_eff_load *= capacity_of(prev_cpu);
5573 prev_eff_load = cpu_runnable_load(cpu_rq(prev_cpu));
5574 prev_eff_load -= task_load;
5575 if (sched_feat(WA_BIAS))
5576 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5577 prev_eff_load *= capacity_of(this_cpu);
5580 * If sync, adjust the weight of prev_eff_load such that if
5581 * prev_eff == this_eff that select_idle_sibling() will consider
5582 * stacking the wakee on top of the waker if no other CPU is
5588 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5591 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5592 int this_cpu, int prev_cpu, int sync)
5594 int target = nr_cpumask_bits;
5596 if (sched_feat(WA_IDLE))
5597 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5599 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5600 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5602 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5603 if (target == nr_cpumask_bits)
5606 schedstat_inc(sd->ttwu_move_affine);
5607 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5611 static unsigned long cpu_util_without(int cpu, struct task_struct *p);
5613 static unsigned long capacity_spare_without(int cpu, struct task_struct *p)
5615 return max_t(long, capacity_of(cpu) - cpu_util_without(cpu, p), 0);
5619 * find_idlest_group finds and returns the least busy CPU group within the
5622 * Assumes p is allowed on at least one CPU in sd.
5624 static struct sched_group *
5625 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5626 int this_cpu, int sd_flag)
5628 struct sched_group *idlest = NULL, *group = sd->groups;
5629 struct sched_group *most_spare_sg = NULL;
5630 unsigned long min_runnable_load = ULONG_MAX;
5631 unsigned long this_runnable_load = ULONG_MAX;
5632 unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5633 unsigned long most_spare = 0, this_spare = 0;
5634 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5635 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5636 (sd->imbalance_pct-100) / 100;
5639 unsigned long load, avg_load, runnable_load;
5640 unsigned long spare_cap, max_spare_cap;
5644 /* Skip over this group if it has no CPUs allowed */
5645 if (!cpumask_intersects(sched_group_span(group),
5649 local_group = cpumask_test_cpu(this_cpu,
5650 sched_group_span(group));
5653 * Tally up the load of all CPUs in the group and find
5654 * the group containing the CPU with most spare capacity.
5660 for_each_cpu(i, sched_group_span(group)) {
5661 load = cpu_runnable_load(cpu_rq(i));
5662 runnable_load += load;
5664 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5666 spare_cap = capacity_spare_without(i, p);
5668 if (spare_cap > max_spare_cap)
5669 max_spare_cap = spare_cap;
5672 /* Adjust by relative CPU capacity of the group */
5673 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5674 group->sgc->capacity;
5675 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5676 group->sgc->capacity;
5679 this_runnable_load = runnable_load;
5680 this_avg_load = avg_load;
5681 this_spare = max_spare_cap;
5683 if (min_runnable_load > (runnable_load + imbalance)) {
5685 * The runnable load is significantly smaller
5686 * so we can pick this new CPU:
5688 min_runnable_load = runnable_load;
5689 min_avg_load = avg_load;
5691 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5692 (100*min_avg_load > imbalance_scale*avg_load)) {
5694 * The runnable loads are close so take the
5695 * blocked load into account through avg_load:
5697 min_avg_load = avg_load;
5701 if (most_spare < max_spare_cap) {
5702 most_spare = max_spare_cap;
5703 most_spare_sg = group;
5706 } while (group = group->next, group != sd->groups);
5709 * The cross-over point between using spare capacity or least load
5710 * is too conservative for high utilization tasks on partially
5711 * utilized systems if we require spare_capacity > task_util(p),
5712 * so we allow for some task stuffing by using
5713 * spare_capacity > task_util(p)/2.
5715 * Spare capacity can't be used for fork because the utilization has
5716 * not been set yet, we must first select a rq to compute the initial
5719 if (sd_flag & SD_BALANCE_FORK)
5722 if (this_spare > task_util(p) / 2 &&
5723 imbalance_scale*this_spare > 100*most_spare)
5726 if (most_spare > task_util(p) / 2)
5727 return most_spare_sg;
5734 * When comparing groups across NUMA domains, it's possible for the
5735 * local domain to be very lightly loaded relative to the remote
5736 * domains but "imbalance" skews the comparison making remote CPUs
5737 * look much more favourable. When considering cross-domain, add
5738 * imbalance to the runnable load on the remote node and consider
5741 if ((sd->flags & SD_NUMA) &&
5742 min_runnable_load + imbalance >= this_runnable_load)
5745 if (min_runnable_load > (this_runnable_load + imbalance))
5748 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5749 (100*this_avg_load < imbalance_scale*min_avg_load))
5756 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5759 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5761 unsigned long load, min_load = ULONG_MAX;
5762 unsigned int min_exit_latency = UINT_MAX;
5763 u64 latest_idle_timestamp = 0;
5764 int least_loaded_cpu = this_cpu;
5765 int shallowest_idle_cpu = -1, si_cpu = -1;
5768 /* Check if we have any choice: */
5769 if (group->group_weight == 1)
5770 return cpumask_first(sched_group_span(group));
5772 /* Traverse only the allowed CPUs */
5773 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
5774 if (available_idle_cpu(i)) {
5775 struct rq *rq = cpu_rq(i);
5776 struct cpuidle_state *idle = idle_get_state(rq);
5777 if (idle && idle->exit_latency < min_exit_latency) {
5779 * We give priority to a CPU whose idle state
5780 * has the smallest exit latency irrespective
5781 * of any idle timestamp.
5783 min_exit_latency = idle->exit_latency;
5784 latest_idle_timestamp = rq->idle_stamp;
5785 shallowest_idle_cpu = i;
5786 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5787 rq->idle_stamp > latest_idle_timestamp) {
5789 * If equal or no active idle state, then
5790 * the most recently idled CPU might have
5793 latest_idle_timestamp = rq->idle_stamp;
5794 shallowest_idle_cpu = i;
5796 } else if (shallowest_idle_cpu == -1 && si_cpu == -1) {
5797 if (sched_idle_cpu(i)) {
5802 load = cpu_runnable_load(cpu_rq(i));
5803 if (load < min_load) {
5805 least_loaded_cpu = i;
5810 if (shallowest_idle_cpu != -1)
5811 return shallowest_idle_cpu;
5814 return least_loaded_cpu;
5817 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5818 int cpu, int prev_cpu, int sd_flag)
5822 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
5826 * We need task's util for capacity_spare_without, sync it up to
5827 * prev_cpu's last_update_time.
5829 if (!(sd_flag & SD_BALANCE_FORK))
5830 sync_entity_load_avg(&p->se);
5833 struct sched_group *group;
5834 struct sched_domain *tmp;
5837 if (!(sd->flags & sd_flag)) {
5842 group = find_idlest_group(sd, p, cpu, sd_flag);
5848 new_cpu = find_idlest_group_cpu(group, p, cpu);
5849 if (new_cpu == cpu) {
5850 /* Now try balancing at a lower domain level of 'cpu': */
5855 /* Now try balancing at a lower domain level of 'new_cpu': */
5857 weight = sd->span_weight;
5859 for_each_domain(cpu, tmp) {
5860 if (weight <= tmp->span_weight)
5862 if (tmp->flags & sd_flag)
5870 #ifdef CONFIG_SCHED_SMT
5871 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5872 EXPORT_SYMBOL_GPL(sched_smt_present);
5874 static inline void set_idle_cores(int cpu, int val)
5876 struct sched_domain_shared *sds;
5878 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5880 WRITE_ONCE(sds->has_idle_cores, val);
5883 static inline bool test_idle_cores(int cpu, bool def)
5885 struct sched_domain_shared *sds;
5887 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5889 return READ_ONCE(sds->has_idle_cores);
5895 * Scans the local SMT mask to see if the entire core is idle, and records this
5896 * information in sd_llc_shared->has_idle_cores.
5898 * Since SMT siblings share all cache levels, inspecting this limited remote
5899 * state should be fairly cheap.
5901 void __update_idle_core(struct rq *rq)
5903 int core = cpu_of(rq);
5907 if (test_idle_cores(core, true))
5910 for_each_cpu(cpu, cpu_smt_mask(core)) {
5914 if (!available_idle_cpu(cpu))
5918 set_idle_cores(core, 1);
5924 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5925 * there are no idle cores left in the system; tracked through
5926 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5928 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5930 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5933 if (!static_branch_likely(&sched_smt_present))
5936 if (!test_idle_cores(target, false))
5939 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
5941 for_each_cpu_wrap(core, cpus, target) {
5944 for_each_cpu(cpu, cpu_smt_mask(core)) {
5945 __cpumask_clear_cpu(cpu, cpus);
5946 if (!available_idle_cpu(cpu))
5955 * Failed to find an idle core; stop looking for one.
5957 set_idle_cores(target, 0);
5963 * Scan the local SMT mask for idle CPUs.
5965 static int select_idle_smt(struct task_struct *p, int target)
5967 int cpu, si_cpu = -1;
5969 if (!static_branch_likely(&sched_smt_present))
5972 for_each_cpu(cpu, cpu_smt_mask(target)) {
5973 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
5975 if (available_idle_cpu(cpu))
5977 if (si_cpu == -1 && sched_idle_cpu(cpu))
5984 #else /* CONFIG_SCHED_SMT */
5986 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5991 static inline int select_idle_smt(struct task_struct *p, int target)
5996 #endif /* CONFIG_SCHED_SMT */
5999 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6000 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6001 * average idle time for this rq (as found in rq->avg_idle).
6003 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
6005 struct sched_domain *this_sd;
6006 u64 avg_cost, avg_idle;
6009 int this = smp_processor_id();
6010 int cpu, nr = INT_MAX, si_cpu = -1;
6012 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6017 * Due to large variance we need a large fuzz factor; hackbench in
6018 * particularly is sensitive here.
6020 avg_idle = this_rq()->avg_idle / 512;
6021 avg_cost = this_sd->avg_scan_cost + 1;
6023 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
6026 if (sched_feat(SIS_PROP)) {
6027 u64 span_avg = sd->span_weight * avg_idle;
6028 if (span_avg > 4*avg_cost)
6029 nr = div_u64(span_avg, avg_cost);
6034 time = cpu_clock(this);
6036 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
6039 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6041 if (available_idle_cpu(cpu))
6043 if (si_cpu == -1 && sched_idle_cpu(cpu))
6047 time = cpu_clock(this) - time;
6048 cost = this_sd->avg_scan_cost;
6049 delta = (s64)(time - cost) / 8;
6050 this_sd->avg_scan_cost += delta;
6056 * Try and locate an idle core/thread in the LLC cache domain.
6058 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6060 struct sched_domain *sd;
6061 int i, recent_used_cpu;
6063 if (available_idle_cpu(target) || sched_idle_cpu(target))
6067 * If the previous CPU is cache affine and idle, don't be stupid:
6069 if (prev != target && cpus_share_cache(prev, target) &&
6070 (available_idle_cpu(prev) || sched_idle_cpu(prev)))
6073 /* Check a recently used CPU as a potential idle candidate: */
6074 recent_used_cpu = p->recent_used_cpu;
6075 if (recent_used_cpu != prev &&
6076 recent_used_cpu != target &&
6077 cpus_share_cache(recent_used_cpu, target) &&
6078 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
6079 cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr)) {
6081 * Replace recent_used_cpu with prev as it is a potential
6082 * candidate for the next wake:
6084 p->recent_used_cpu = prev;
6085 return recent_used_cpu;
6088 sd = rcu_dereference(per_cpu(sd_llc, target));
6092 i = select_idle_core(p, sd, target);
6093 if ((unsigned)i < nr_cpumask_bits)
6096 i = select_idle_cpu(p, sd, target);
6097 if ((unsigned)i < nr_cpumask_bits)
6100 i = select_idle_smt(p, target);
6101 if ((unsigned)i < nr_cpumask_bits)
6108 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6109 * @cpu: the CPU to get the utilization of
6111 * The unit of the return value must be the one of capacity so we can compare
6112 * the utilization with the capacity of the CPU that is available for CFS task
6113 * (ie cpu_capacity).
6115 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6116 * recent utilization of currently non-runnable tasks on a CPU. It represents
6117 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6118 * capacity_orig is the cpu_capacity available at the highest frequency
6119 * (arch_scale_freq_capacity()).
6120 * The utilization of a CPU converges towards a sum equal to or less than the
6121 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6122 * the running time on this CPU scaled by capacity_curr.
6124 * The estimated utilization of a CPU is defined to be the maximum between its
6125 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6126 * currently RUNNABLE on that CPU.
6127 * This allows to properly represent the expected utilization of a CPU which
6128 * has just got a big task running since a long sleep period. At the same time
6129 * however it preserves the benefits of the "blocked utilization" in
6130 * describing the potential for other tasks waking up on the same CPU.
6132 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6133 * higher than capacity_orig because of unfortunate rounding in
6134 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6135 * the average stabilizes with the new running time. We need to check that the
6136 * utilization stays within the range of [0..capacity_orig] and cap it if
6137 * necessary. Without utilization capping, a group could be seen as overloaded
6138 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6139 * available capacity. We allow utilization to overshoot capacity_curr (but not
6140 * capacity_orig) as it useful for predicting the capacity required after task
6141 * migrations (scheduler-driven DVFS).
6143 * Return: the (estimated) utilization for the specified CPU
6145 static inline unsigned long cpu_util(int cpu)
6147 struct cfs_rq *cfs_rq;
6150 cfs_rq = &cpu_rq(cpu)->cfs;
6151 util = READ_ONCE(cfs_rq->avg.util_avg);
6153 if (sched_feat(UTIL_EST))
6154 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6156 return min_t(unsigned long, util, capacity_orig_of(cpu));
6160 * cpu_util_without: compute cpu utilization without any contributions from *p
6161 * @cpu: the CPU which utilization is requested
6162 * @p: the task which utilization should be discounted
6164 * The utilization of a CPU is defined by the utilization of tasks currently
6165 * enqueued on that CPU as well as tasks which are currently sleeping after an
6166 * execution on that CPU.
6168 * This method returns the utilization of the specified CPU by discounting the
6169 * utilization of the specified task, whenever the task is currently
6170 * contributing to the CPU utilization.
6172 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6174 struct cfs_rq *cfs_rq;
6177 /* Task has no contribution or is new */
6178 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6179 return cpu_util(cpu);
6181 cfs_rq = &cpu_rq(cpu)->cfs;
6182 util = READ_ONCE(cfs_rq->avg.util_avg);
6184 /* Discount task's util from CPU's util */
6185 lsub_positive(&util, task_util(p));
6190 * a) if *p is the only task sleeping on this CPU, then:
6191 * cpu_util (== task_util) > util_est (== 0)
6192 * and thus we return:
6193 * cpu_util_without = (cpu_util - task_util) = 0
6195 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6197 * cpu_util >= task_util
6198 * cpu_util > util_est (== 0)
6199 * and thus we discount *p's blocked utilization to return:
6200 * cpu_util_without = (cpu_util - task_util) >= 0
6202 * c) if other tasks are RUNNABLE on that CPU and
6203 * util_est > cpu_util
6204 * then we use util_est since it returns a more restrictive
6205 * estimation of the spare capacity on that CPU, by just
6206 * considering the expected utilization of tasks already
6207 * runnable on that CPU.
6209 * Cases a) and b) are covered by the above code, while case c) is
6210 * covered by the following code when estimated utilization is
6213 if (sched_feat(UTIL_EST)) {
6214 unsigned int estimated =
6215 READ_ONCE(cfs_rq->avg.util_est.enqueued);
6218 * Despite the following checks we still have a small window
6219 * for a possible race, when an execl's select_task_rq_fair()
6220 * races with LB's detach_task():
6223 * p->on_rq = TASK_ON_RQ_MIGRATING;
6224 * ---------------------------------- A
6225 * deactivate_task() \
6226 * dequeue_task() + RaceTime
6227 * util_est_dequeue() /
6228 * ---------------------------------- B
6230 * The additional check on "current == p" it's required to
6231 * properly fix the execl regression and it helps in further
6232 * reducing the chances for the above race.
6234 if (unlikely(task_on_rq_queued(p) || current == p))
6235 lsub_positive(&estimated, _task_util_est(p));
6237 util = max(util, estimated);
6241 * Utilization (estimated) can exceed the CPU capacity, thus let's
6242 * clamp to the maximum CPU capacity to ensure consistency with
6243 * the cpu_util call.
6245 return min_t(unsigned long, util, capacity_orig_of(cpu));
6249 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6250 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6252 * In that case WAKE_AFFINE doesn't make sense and we'll let
6253 * BALANCE_WAKE sort things out.
6255 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6257 long min_cap, max_cap;
6259 if (!static_branch_unlikely(&sched_asym_cpucapacity))
6262 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6263 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6265 /* Minimum capacity is close to max, no need to abort wake_affine */
6266 if (max_cap - min_cap < max_cap >> 3)
6269 /* Bring task utilization in sync with prev_cpu */
6270 sync_entity_load_avg(&p->se);
6272 return !task_fits_capacity(p, min_cap);
6276 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6279 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6281 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6282 unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
6285 * If @p migrates from @cpu to another, remove its contribution. Or,
6286 * if @p migrates from another CPU to @cpu, add its contribution. In
6287 * the other cases, @cpu is not impacted by the migration, so the
6288 * util_avg should already be correct.
6290 if (task_cpu(p) == cpu && dst_cpu != cpu)
6291 sub_positive(&util, task_util(p));
6292 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6293 util += task_util(p);
6295 if (sched_feat(UTIL_EST)) {
6296 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6299 * During wake-up, the task isn't enqueued yet and doesn't
6300 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6301 * so just add it (if needed) to "simulate" what will be
6302 * cpu_util() after the task has been enqueued.
6305 util_est += _task_util_est(p);
6307 util = max(util, util_est);
6310 return min(util, capacity_orig_of(cpu));
6314 * compute_energy(): Estimates the energy that would be consumed if @p was
6315 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6316 * landscape of the * CPUs after the task migration, and uses the Energy Model
6317 * to compute what would be the energy if we decided to actually migrate that
6321 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
6323 unsigned int max_util, util_cfs, cpu_util, cpu_cap;
6324 unsigned long sum_util, energy = 0;
6325 struct task_struct *tsk;
6328 for (; pd; pd = pd->next) {
6329 struct cpumask *pd_mask = perf_domain_span(pd);
6332 * The energy model mandates all the CPUs of a performance
6333 * domain have the same capacity.
6335 cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask));
6336 max_util = sum_util = 0;
6339 * The capacity state of CPUs of the current rd can be driven by
6340 * CPUs of another rd if they belong to the same performance
6341 * domain. So, account for the utilization of these CPUs too
6342 * by masking pd with cpu_online_mask instead of the rd span.
6344 * If an entire performance domain is outside of the current rd,
6345 * it will not appear in its pd list and will not be accounted
6346 * by compute_energy().
6348 for_each_cpu_and(cpu, pd_mask, cpu_online_mask) {
6349 util_cfs = cpu_util_next(cpu, p, dst_cpu);
6352 * Busy time computation: utilization clamping is not
6353 * required since the ratio (sum_util / cpu_capacity)
6354 * is already enough to scale the EM reported power
6355 * consumption at the (eventually clamped) cpu_capacity.
6357 sum_util += schedutil_cpu_util(cpu, util_cfs, cpu_cap,
6361 * Performance domain frequency: utilization clamping
6362 * must be considered since it affects the selection
6363 * of the performance domain frequency.
6364 * NOTE: in case RT tasks are running, by default the
6365 * FREQUENCY_UTIL's utilization can be max OPP.
6367 tsk = cpu == dst_cpu ? p : NULL;
6368 cpu_util = schedutil_cpu_util(cpu, util_cfs, cpu_cap,
6369 FREQUENCY_UTIL, tsk);
6370 max_util = max(max_util, cpu_util);
6373 energy += em_pd_energy(pd->em_pd, max_util, sum_util);
6380 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6381 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6382 * spare capacity in each performance domain and uses it as a potential
6383 * candidate to execute the task. Then, it uses the Energy Model to figure
6384 * out which of the CPU candidates is the most energy-efficient.
6386 * The rationale for this heuristic is as follows. In a performance domain,
6387 * all the most energy efficient CPU candidates (according to the Energy
6388 * Model) are those for which we'll request a low frequency. When there are
6389 * several CPUs for which the frequency request will be the same, we don't
6390 * have enough data to break the tie between them, because the Energy Model
6391 * only includes active power costs. With this model, if we assume that
6392 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6393 * the maximum spare capacity in a performance domain is guaranteed to be among
6394 * the best candidates of the performance domain.
6396 * In practice, it could be preferable from an energy standpoint to pack
6397 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6398 * but that could also hurt our chances to go cluster idle, and we have no
6399 * ways to tell with the current Energy Model if this is actually a good
6400 * idea or not. So, find_energy_efficient_cpu() basically favors
6401 * cluster-packing, and spreading inside a cluster. That should at least be
6402 * a good thing for latency, and this is consistent with the idea that most
6403 * of the energy savings of EAS come from the asymmetry of the system, and
6404 * not so much from breaking the tie between identical CPUs. That's also the
6405 * reason why EAS is enabled in the topology code only for systems where
6406 * SD_ASYM_CPUCAPACITY is set.
6408 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6409 * they don't have any useful utilization data yet and it's not possible to
6410 * forecast their impact on energy consumption. Consequently, they will be
6411 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6412 * to be energy-inefficient in some use-cases. The alternative would be to
6413 * bias new tasks towards specific types of CPUs first, or to try to infer
6414 * their util_avg from the parent task, but those heuristics could hurt
6415 * other use-cases too. So, until someone finds a better way to solve this,
6416 * let's keep things simple by re-using the existing slow path.
6419 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6421 unsigned long prev_energy = ULONG_MAX, best_energy = ULONG_MAX;
6422 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6423 int cpu, best_energy_cpu = prev_cpu;
6424 struct perf_domain *head, *pd;
6425 unsigned long cpu_cap, util;
6426 struct sched_domain *sd;
6429 pd = rcu_dereference(rd->pd);
6430 if (!pd || READ_ONCE(rd->overutilized))
6435 * Energy-aware wake-up happens on the lowest sched_domain starting
6436 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6438 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6439 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6444 sync_entity_load_avg(&p->se);
6445 if (!task_util_est(p))
6448 for (; pd; pd = pd->next) {
6449 unsigned long cur_energy, spare_cap, max_spare_cap = 0;
6450 int max_spare_cap_cpu = -1;
6452 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
6453 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6456 /* Skip CPUs that will be overutilized. */
6457 util = cpu_util_next(cpu, p, cpu);
6458 cpu_cap = capacity_of(cpu);
6459 if (cpu_cap * 1024 < util * capacity_margin)
6462 /* Always use prev_cpu as a candidate. */
6463 if (cpu == prev_cpu) {
6464 prev_energy = compute_energy(p, prev_cpu, head);
6465 best_energy = min(best_energy, prev_energy);
6470 * Find the CPU with the maximum spare capacity in
6471 * the performance domain
6473 spare_cap = cpu_cap - util;
6474 if (spare_cap > max_spare_cap) {
6475 max_spare_cap = spare_cap;
6476 max_spare_cap_cpu = cpu;
6480 /* Evaluate the energy impact of using this CPU. */
6481 if (max_spare_cap_cpu >= 0) {
6482 cur_energy = compute_energy(p, max_spare_cap_cpu, head);
6483 if (cur_energy < best_energy) {
6484 best_energy = cur_energy;
6485 best_energy_cpu = max_spare_cap_cpu;
6493 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6494 * least 6% of the energy used by prev_cpu.
6496 if (prev_energy == ULONG_MAX)
6497 return best_energy_cpu;
6499 if ((prev_energy - best_energy) > (prev_energy >> 4))
6500 return best_energy_cpu;
6511 * select_task_rq_fair: Select target runqueue for the waking task in domains
6512 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6513 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6515 * Balances load by selecting the idlest CPU in the idlest group, or under
6516 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6518 * Returns the target CPU number.
6520 * preempt must be disabled.
6523 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6525 struct sched_domain *tmp, *sd = NULL;
6526 int cpu = smp_processor_id();
6527 int new_cpu = prev_cpu;
6528 int want_affine = 0;
6529 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6531 if (sd_flag & SD_BALANCE_WAKE) {
6534 if (sched_energy_enabled()) {
6535 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
6541 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu) &&
6542 cpumask_test_cpu(cpu, p->cpus_ptr);
6546 for_each_domain(cpu, tmp) {
6547 if (!(tmp->flags & SD_LOAD_BALANCE))
6551 * If both 'cpu' and 'prev_cpu' are part of this domain,
6552 * cpu is a valid SD_WAKE_AFFINE target.
6554 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6555 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6556 if (cpu != prev_cpu)
6557 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6559 sd = NULL; /* Prefer wake_affine over balance flags */
6563 if (tmp->flags & sd_flag)
6565 else if (!want_affine)
6571 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6572 } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6575 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6578 current->recent_used_cpu = cpu;
6585 static void detach_entity_cfs_rq(struct sched_entity *se);
6588 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6589 * cfs_rq_of(p) references at time of call are still valid and identify the
6590 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6592 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6595 * As blocked tasks retain absolute vruntime the migration needs to
6596 * deal with this by subtracting the old and adding the new
6597 * min_vruntime -- the latter is done by enqueue_entity() when placing
6598 * the task on the new runqueue.
6600 if (p->state == TASK_WAKING) {
6601 struct sched_entity *se = &p->se;
6602 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6605 #ifndef CONFIG_64BIT
6606 u64 min_vruntime_copy;
6609 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6611 min_vruntime = cfs_rq->min_vruntime;
6612 } while (min_vruntime != min_vruntime_copy);
6614 min_vruntime = cfs_rq->min_vruntime;
6617 se->vruntime -= min_vruntime;
6620 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6622 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6623 * rq->lock and can modify state directly.
6625 lockdep_assert_held(&task_rq(p)->lock);
6626 detach_entity_cfs_rq(&p->se);
6630 * We are supposed to update the task to "current" time, then
6631 * its up to date and ready to go to new CPU/cfs_rq. But we
6632 * have difficulty in getting what current time is, so simply
6633 * throw away the out-of-date time. This will result in the
6634 * wakee task is less decayed, but giving the wakee more load
6637 remove_entity_load_avg(&p->se);
6640 /* Tell new CPU we are migrated */
6641 p->se.avg.last_update_time = 0;
6643 /* We have migrated, no longer consider this task hot */
6644 p->se.exec_start = 0;
6646 update_scan_period(p, new_cpu);
6649 static void task_dead_fair(struct task_struct *p)
6651 remove_entity_load_avg(&p->se);
6653 #endif /* CONFIG_SMP */
6655 static unsigned long wakeup_gran(struct sched_entity *se)
6657 unsigned long gran = sysctl_sched_wakeup_granularity;
6660 * Since its curr running now, convert the gran from real-time
6661 * to virtual-time in his units.
6663 * By using 'se' instead of 'curr' we penalize light tasks, so
6664 * they get preempted easier. That is, if 'se' < 'curr' then
6665 * the resulting gran will be larger, therefore penalizing the
6666 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6667 * be smaller, again penalizing the lighter task.
6669 * This is especially important for buddies when the leftmost
6670 * task is higher priority than the buddy.
6672 return calc_delta_fair(gran, se);
6676 * Should 'se' preempt 'curr'.
6690 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6692 s64 gran, vdiff = curr->vruntime - se->vruntime;
6697 gran = wakeup_gran(se);
6704 static void set_last_buddy(struct sched_entity *se)
6706 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6709 for_each_sched_entity(se) {
6710 if (SCHED_WARN_ON(!se->on_rq))
6712 cfs_rq_of(se)->last = se;
6716 static void set_next_buddy(struct sched_entity *se)
6718 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6721 for_each_sched_entity(se) {
6722 if (SCHED_WARN_ON(!se->on_rq))
6724 cfs_rq_of(se)->next = se;
6728 static void set_skip_buddy(struct sched_entity *se)
6730 for_each_sched_entity(se)
6731 cfs_rq_of(se)->skip = se;
6735 * Preempt the current task with a newly woken task if needed:
6737 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6739 struct task_struct *curr = rq->curr;
6740 struct sched_entity *se = &curr->se, *pse = &p->se;
6741 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6742 int scale = cfs_rq->nr_running >= sched_nr_latency;
6743 int next_buddy_marked = 0;
6745 if (unlikely(se == pse))
6749 * This is possible from callers such as attach_tasks(), in which we
6750 * unconditionally check_prempt_curr() after an enqueue (which may have
6751 * lead to a throttle). This both saves work and prevents false
6752 * next-buddy nomination below.
6754 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6757 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6758 set_next_buddy(pse);
6759 next_buddy_marked = 1;
6763 * We can come here with TIF_NEED_RESCHED already set from new task
6766 * Note: this also catches the edge-case of curr being in a throttled
6767 * group (e.g. via set_curr_task), since update_curr() (in the
6768 * enqueue of curr) will have resulted in resched being set. This
6769 * prevents us from potentially nominating it as a false LAST_BUDDY
6772 if (test_tsk_need_resched(curr))
6775 /* Idle tasks are by definition preempted by non-idle tasks. */
6776 if (unlikely(task_has_idle_policy(curr)) &&
6777 likely(!task_has_idle_policy(p)))
6781 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6782 * is driven by the tick):
6784 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6787 find_matching_se(&se, &pse);
6788 update_curr(cfs_rq_of(se));
6790 if (wakeup_preempt_entity(se, pse) == 1) {
6792 * Bias pick_next to pick the sched entity that is
6793 * triggering this preemption.
6795 if (!next_buddy_marked)
6796 set_next_buddy(pse);
6805 * Only set the backward buddy when the current task is still
6806 * on the rq. This can happen when a wakeup gets interleaved
6807 * with schedule on the ->pre_schedule() or idle_balance()
6808 * point, either of which can * drop the rq lock.
6810 * Also, during early boot the idle thread is in the fair class,
6811 * for obvious reasons its a bad idea to schedule back to it.
6813 if (unlikely(!se->on_rq || curr == rq->idle))
6816 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6820 static struct task_struct *
6821 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6823 struct cfs_rq *cfs_rq = &rq->cfs;
6824 struct sched_entity *se;
6825 struct task_struct *p;
6829 if (!cfs_rq->nr_running)
6832 #ifdef CONFIG_FAIR_GROUP_SCHED
6833 if (prev->sched_class != &fair_sched_class)
6837 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6838 * likely that a next task is from the same cgroup as the current.
6840 * Therefore attempt to avoid putting and setting the entire cgroup
6841 * hierarchy, only change the part that actually changes.
6845 struct sched_entity *curr = cfs_rq->curr;
6848 * Since we got here without doing put_prev_entity() we also
6849 * have to consider cfs_rq->curr. If it is still a runnable
6850 * entity, update_curr() will update its vruntime, otherwise
6851 * forget we've ever seen it.
6855 update_curr(cfs_rq);
6860 * This call to check_cfs_rq_runtime() will do the
6861 * throttle and dequeue its entity in the parent(s).
6862 * Therefore the nr_running test will indeed
6865 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6868 if (!cfs_rq->nr_running)
6875 se = pick_next_entity(cfs_rq, curr);
6876 cfs_rq = group_cfs_rq(se);
6882 * Since we haven't yet done put_prev_entity and if the selected task
6883 * is a different task than we started out with, try and touch the
6884 * least amount of cfs_rqs.
6887 struct sched_entity *pse = &prev->se;
6889 while (!(cfs_rq = is_same_group(se, pse))) {
6890 int se_depth = se->depth;
6891 int pse_depth = pse->depth;
6893 if (se_depth <= pse_depth) {
6894 put_prev_entity(cfs_rq_of(pse), pse);
6895 pse = parent_entity(pse);
6897 if (se_depth >= pse_depth) {
6898 set_next_entity(cfs_rq_of(se), se);
6899 se = parent_entity(se);
6903 put_prev_entity(cfs_rq, pse);
6904 set_next_entity(cfs_rq, se);
6911 put_prev_task(rq, prev);
6914 se = pick_next_entity(cfs_rq, NULL);
6915 set_next_entity(cfs_rq, se);
6916 cfs_rq = group_cfs_rq(se);
6921 done: __maybe_unused;
6924 * Move the next running task to the front of
6925 * the list, so our cfs_tasks list becomes MRU
6928 list_move(&p->se.group_node, &rq->cfs_tasks);
6931 if (hrtick_enabled(rq))
6932 hrtick_start_fair(rq, p);
6934 update_misfit_status(p, rq);
6939 update_misfit_status(NULL, rq);
6940 new_tasks = idle_balance(rq, rf);
6943 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6944 * possible for any higher priority task to appear. In that case we
6945 * must re-start the pick_next_entity() loop.
6954 * rq is about to be idle, check if we need to update the
6955 * lost_idle_time of clock_pelt
6957 update_idle_rq_clock_pelt(rq);
6963 * Account for a descheduled task:
6965 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6967 struct sched_entity *se = &prev->se;
6968 struct cfs_rq *cfs_rq;
6970 for_each_sched_entity(se) {
6971 cfs_rq = cfs_rq_of(se);
6972 put_prev_entity(cfs_rq, se);
6977 * sched_yield() is very simple
6979 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6981 static void yield_task_fair(struct rq *rq)
6983 struct task_struct *curr = rq->curr;
6984 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6985 struct sched_entity *se = &curr->se;
6988 * Are we the only task in the tree?
6990 if (unlikely(rq->nr_running == 1))
6993 clear_buddies(cfs_rq, se);
6995 if (curr->policy != SCHED_BATCH) {
6996 update_rq_clock(rq);
6998 * Update run-time statistics of the 'current'.
7000 update_curr(cfs_rq);
7002 * Tell update_rq_clock() that we've just updated,
7003 * so we don't do microscopic update in schedule()
7004 * and double the fastpath cost.
7006 rq_clock_skip_update(rq);
7012 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
7014 struct sched_entity *se = &p->se;
7016 /* throttled hierarchies are not runnable */
7017 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
7020 /* Tell the scheduler that we'd really like pse to run next. */
7023 yield_task_fair(rq);
7029 /**************************************************
7030 * Fair scheduling class load-balancing methods.
7034 * The purpose of load-balancing is to achieve the same basic fairness the
7035 * per-CPU scheduler provides, namely provide a proportional amount of compute
7036 * time to each task. This is expressed in the following equation:
7038 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7040 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7041 * W_i,0 is defined as:
7043 * W_i,0 = \Sum_j w_i,j (2)
7045 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7046 * is derived from the nice value as per sched_prio_to_weight[].
7048 * The weight average is an exponential decay average of the instantaneous
7051 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7053 * C_i is the compute capacity of CPU i, typically it is the
7054 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7055 * can also include other factors [XXX].
7057 * To achieve this balance we define a measure of imbalance which follows
7058 * directly from (1):
7060 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7062 * We them move tasks around to minimize the imbalance. In the continuous
7063 * function space it is obvious this converges, in the discrete case we get
7064 * a few fun cases generally called infeasible weight scenarios.
7067 * - infeasible weights;
7068 * - local vs global optima in the discrete case. ]
7073 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7074 * for all i,j solution, we create a tree of CPUs that follows the hardware
7075 * topology where each level pairs two lower groups (or better). This results
7076 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7077 * tree to only the first of the previous level and we decrease the frequency
7078 * of load-balance at each level inv. proportional to the number of CPUs in
7084 * \Sum { --- * --- * 2^i } = O(n) (5)
7086 * `- size of each group
7087 * | | `- number of CPUs doing load-balance
7089 * `- sum over all levels
7091 * Coupled with a limit on how many tasks we can migrate every balance pass,
7092 * this makes (5) the runtime complexity of the balancer.
7094 * An important property here is that each CPU is still (indirectly) connected
7095 * to every other CPU in at most O(log n) steps:
7097 * The adjacency matrix of the resulting graph is given by:
7100 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7103 * And you'll find that:
7105 * A^(log_2 n)_i,j != 0 for all i,j (7)
7107 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7108 * The task movement gives a factor of O(m), giving a convergence complexity
7111 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7116 * In order to avoid CPUs going idle while there's still work to do, new idle
7117 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7118 * tree itself instead of relying on other CPUs to bring it work.
7120 * This adds some complexity to both (5) and (8) but it reduces the total idle
7128 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7131 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7136 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7138 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7140 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7143 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7144 * rewrite all of this once again.]
7147 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7149 enum fbq_type { regular, remote, all };
7158 #define LBF_ALL_PINNED 0x01
7159 #define LBF_NEED_BREAK 0x02
7160 #define LBF_DST_PINNED 0x04
7161 #define LBF_SOME_PINNED 0x08
7162 #define LBF_NOHZ_STATS 0x10
7163 #define LBF_NOHZ_AGAIN 0x20
7166 struct sched_domain *sd;
7174 struct cpumask *dst_grpmask;
7176 enum cpu_idle_type idle;
7178 /* The set of CPUs under consideration for load-balancing */
7179 struct cpumask *cpus;
7184 unsigned int loop_break;
7185 unsigned int loop_max;
7187 enum fbq_type fbq_type;
7188 enum group_type src_grp_type;
7189 struct list_head tasks;
7193 * Is this task likely cache-hot:
7195 static int task_hot(struct task_struct *p, struct lb_env *env)
7199 lockdep_assert_held(&env->src_rq->lock);
7201 if (p->sched_class != &fair_sched_class)
7204 if (unlikely(task_has_idle_policy(p)))
7208 * Buddy candidates are cache hot:
7210 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7211 (&p->se == cfs_rq_of(&p->se)->next ||
7212 &p->se == cfs_rq_of(&p->se)->last))
7215 if (sysctl_sched_migration_cost == -1)
7217 if (sysctl_sched_migration_cost == 0)
7220 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7222 return delta < (s64)sysctl_sched_migration_cost;
7225 #ifdef CONFIG_NUMA_BALANCING
7227 * Returns 1, if task migration degrades locality
7228 * Returns 0, if task migration improves locality i.e migration preferred.
7229 * Returns -1, if task migration is not affected by locality.
7231 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7233 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7234 unsigned long src_weight, dst_weight;
7235 int src_nid, dst_nid, dist;
7237 if (!static_branch_likely(&sched_numa_balancing))
7240 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7243 src_nid = cpu_to_node(env->src_cpu);
7244 dst_nid = cpu_to_node(env->dst_cpu);
7246 if (src_nid == dst_nid)
7249 /* Migrating away from the preferred node is always bad. */
7250 if (src_nid == p->numa_preferred_nid) {
7251 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7257 /* Encourage migration to the preferred node. */
7258 if (dst_nid == p->numa_preferred_nid)
7261 /* Leaving a core idle is often worse than degrading locality. */
7262 if (env->idle == CPU_IDLE)
7265 dist = node_distance(src_nid, dst_nid);
7267 src_weight = group_weight(p, src_nid, dist);
7268 dst_weight = group_weight(p, dst_nid, dist);
7270 src_weight = task_weight(p, src_nid, dist);
7271 dst_weight = task_weight(p, dst_nid, dist);
7274 return dst_weight < src_weight;
7278 static inline int migrate_degrades_locality(struct task_struct *p,
7286 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7289 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7293 lockdep_assert_held(&env->src_rq->lock);
7296 * We do not migrate tasks that are:
7297 * 1) throttled_lb_pair, or
7298 * 2) cannot be migrated to this CPU due to cpus_ptr, or
7299 * 3) running (obviously), or
7300 * 4) are cache-hot on their current CPU.
7302 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7305 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
7308 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7310 env->flags |= LBF_SOME_PINNED;
7313 * Remember if this task can be migrated to any other CPU in
7314 * our sched_group. We may want to revisit it if we couldn't
7315 * meet load balance goals by pulling other tasks on src_cpu.
7317 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7318 * already computed one in current iteration.
7320 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7323 /* Prevent to re-select dst_cpu via env's CPUs: */
7324 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7325 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
7326 env->flags |= LBF_DST_PINNED;
7327 env->new_dst_cpu = cpu;
7335 /* Record that we found atleast one task that could run on dst_cpu */
7336 env->flags &= ~LBF_ALL_PINNED;
7338 if (task_running(env->src_rq, p)) {
7339 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7344 * Aggressive migration if:
7345 * 1) destination numa is preferred
7346 * 2) task is cache cold, or
7347 * 3) too many balance attempts have failed.
7349 tsk_cache_hot = migrate_degrades_locality(p, env);
7350 if (tsk_cache_hot == -1)
7351 tsk_cache_hot = task_hot(p, env);
7353 if (tsk_cache_hot <= 0 ||
7354 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7355 if (tsk_cache_hot == 1) {
7356 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7357 schedstat_inc(p->se.statistics.nr_forced_migrations);
7362 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7367 * detach_task() -- detach the task for the migration specified in env
7369 static void detach_task(struct task_struct *p, struct lb_env *env)
7371 lockdep_assert_held(&env->src_rq->lock);
7373 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7374 set_task_cpu(p, env->dst_cpu);
7378 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7379 * part of active balancing operations within "domain".
7381 * Returns a task if successful and NULL otherwise.
7383 static struct task_struct *detach_one_task(struct lb_env *env)
7385 struct task_struct *p;
7387 lockdep_assert_held(&env->src_rq->lock);
7389 list_for_each_entry_reverse(p,
7390 &env->src_rq->cfs_tasks, se.group_node) {
7391 if (!can_migrate_task(p, env))
7394 detach_task(p, env);
7397 * Right now, this is only the second place where
7398 * lb_gained[env->idle] is updated (other is detach_tasks)
7399 * so we can safely collect stats here rather than
7400 * inside detach_tasks().
7402 schedstat_inc(env->sd->lb_gained[env->idle]);
7408 static const unsigned int sched_nr_migrate_break = 32;
7411 * detach_tasks() -- tries to detach up to imbalance runnable load from
7412 * busiest_rq, as part of a balancing operation within domain "sd".
7414 * Returns number of detached tasks if successful and 0 otherwise.
7416 static int detach_tasks(struct lb_env *env)
7418 struct list_head *tasks = &env->src_rq->cfs_tasks;
7419 struct task_struct *p;
7423 lockdep_assert_held(&env->src_rq->lock);
7425 if (env->imbalance <= 0)
7428 while (!list_empty(tasks)) {
7430 * We don't want to steal all, otherwise we may be treated likewise,
7431 * which could at worst lead to a livelock crash.
7433 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7436 p = list_last_entry(tasks, struct task_struct, se.group_node);
7439 /* We've more or less seen every task there is, call it quits */
7440 if (env->loop > env->loop_max)
7443 /* take a breather every nr_migrate tasks */
7444 if (env->loop > env->loop_break) {
7445 env->loop_break += sched_nr_migrate_break;
7446 env->flags |= LBF_NEED_BREAK;
7450 if (!can_migrate_task(p, env))
7453 load = task_h_load(p);
7455 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7458 if ((load / 2) > env->imbalance)
7461 detach_task(p, env);
7462 list_add(&p->se.group_node, &env->tasks);
7465 env->imbalance -= load;
7467 #ifdef CONFIG_PREEMPT
7469 * NEWIDLE balancing is a source of latency, so preemptible
7470 * kernels will stop after the first task is detached to minimize
7471 * the critical section.
7473 if (env->idle == CPU_NEWLY_IDLE)
7478 * We only want to steal up to the prescribed amount of
7481 if (env->imbalance <= 0)
7486 list_move(&p->se.group_node, tasks);
7490 * Right now, this is one of only two places we collect this stat
7491 * so we can safely collect detach_one_task() stats here rather
7492 * than inside detach_one_task().
7494 schedstat_add(env->sd->lb_gained[env->idle], detached);
7500 * attach_task() -- attach the task detached by detach_task() to its new rq.
7502 static void attach_task(struct rq *rq, struct task_struct *p)
7504 lockdep_assert_held(&rq->lock);
7506 BUG_ON(task_rq(p) != rq);
7507 activate_task(rq, p, ENQUEUE_NOCLOCK);
7508 check_preempt_curr(rq, p, 0);
7512 * attach_one_task() -- attaches the task returned from detach_one_task() to
7515 static void attach_one_task(struct rq *rq, struct task_struct *p)
7520 update_rq_clock(rq);
7526 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7529 static void attach_tasks(struct lb_env *env)
7531 struct list_head *tasks = &env->tasks;
7532 struct task_struct *p;
7535 rq_lock(env->dst_rq, &rf);
7536 update_rq_clock(env->dst_rq);
7538 while (!list_empty(tasks)) {
7539 p = list_first_entry(tasks, struct task_struct, se.group_node);
7540 list_del_init(&p->se.group_node);
7542 attach_task(env->dst_rq, p);
7545 rq_unlock(env->dst_rq, &rf);
7548 #ifdef CONFIG_NO_HZ_COMMON
7549 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7551 if (cfs_rq->avg.load_avg)
7554 if (cfs_rq->avg.util_avg)
7560 static inline bool others_have_blocked(struct rq *rq)
7562 if (READ_ONCE(rq->avg_rt.util_avg))
7565 if (READ_ONCE(rq->avg_dl.util_avg))
7568 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7569 if (READ_ONCE(rq->avg_irq.util_avg))
7576 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
7578 rq->last_blocked_load_update_tick = jiffies;
7581 rq->has_blocked_load = 0;
7584 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
7585 static inline bool others_have_blocked(struct rq *rq) { return false; }
7586 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
7589 #ifdef CONFIG_FAIR_GROUP_SCHED
7591 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7593 if (cfs_rq->load.weight)
7596 if (cfs_rq->avg.load_sum)
7599 if (cfs_rq->avg.util_sum)
7602 if (cfs_rq->avg.runnable_load_sum)
7608 static void update_blocked_averages(int cpu)
7610 struct rq *rq = cpu_rq(cpu);
7611 struct cfs_rq *cfs_rq, *pos;
7612 const struct sched_class *curr_class;
7616 rq_lock_irqsave(rq, &rf);
7617 update_rq_clock(rq);
7620 * Iterates the task_group tree in a bottom up fashion, see
7621 * list_add_leaf_cfs_rq() for details.
7623 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7624 struct sched_entity *se;
7626 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq))
7627 update_tg_load_avg(cfs_rq, 0);
7629 /* Propagate pending load changes to the parent, if any: */
7630 se = cfs_rq->tg->se[cpu];
7631 if (se && !skip_blocked_update(se))
7632 update_load_avg(cfs_rq_of(se), se, 0);
7635 * There can be a lot of idle CPU cgroups. Don't let fully
7636 * decayed cfs_rqs linger on the list.
7638 if (cfs_rq_is_decayed(cfs_rq))
7639 list_del_leaf_cfs_rq(cfs_rq);
7641 /* Don't need periodic decay once load/util_avg are null */
7642 if (cfs_rq_has_blocked(cfs_rq))
7646 curr_class = rq->curr->sched_class;
7647 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class);
7648 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class);
7649 update_irq_load_avg(rq, 0);
7650 /* Don't need periodic decay once load/util_avg are null */
7651 if (others_have_blocked(rq))
7654 update_blocked_load_status(rq, !done);
7655 rq_unlock_irqrestore(rq, &rf);
7659 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7660 * This needs to be done in a top-down fashion because the load of a child
7661 * group is a fraction of its parents load.
7663 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7665 struct rq *rq = rq_of(cfs_rq);
7666 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7667 unsigned long now = jiffies;
7670 if (cfs_rq->last_h_load_update == now)
7673 WRITE_ONCE(cfs_rq->h_load_next, NULL);
7674 for_each_sched_entity(se) {
7675 cfs_rq = cfs_rq_of(se);
7676 WRITE_ONCE(cfs_rq->h_load_next, se);
7677 if (cfs_rq->last_h_load_update == now)
7682 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7683 cfs_rq->last_h_load_update = now;
7686 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
7687 load = cfs_rq->h_load;
7688 load = div64_ul(load * se->avg.load_avg,
7689 cfs_rq_load_avg(cfs_rq) + 1);
7690 cfs_rq = group_cfs_rq(se);
7691 cfs_rq->h_load = load;
7692 cfs_rq->last_h_load_update = now;
7696 static unsigned long task_h_load(struct task_struct *p)
7698 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7700 update_cfs_rq_h_load(cfs_rq);
7701 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7702 cfs_rq_load_avg(cfs_rq) + 1);
7705 static inline void update_blocked_averages(int cpu)
7707 struct rq *rq = cpu_rq(cpu);
7708 struct cfs_rq *cfs_rq = &rq->cfs;
7709 const struct sched_class *curr_class;
7712 rq_lock_irqsave(rq, &rf);
7713 update_rq_clock(rq);
7714 update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
7716 curr_class = rq->curr->sched_class;
7717 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class);
7718 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class);
7719 update_irq_load_avg(rq, 0);
7720 update_blocked_load_status(rq, cfs_rq_has_blocked(cfs_rq) || others_have_blocked(rq));
7721 rq_unlock_irqrestore(rq, &rf);
7724 static unsigned long task_h_load(struct task_struct *p)
7726 return p->se.avg.load_avg;
7730 /********** Helpers for find_busiest_group ************************/
7733 * sg_lb_stats - stats of a sched_group required for load_balancing
7735 struct sg_lb_stats {
7736 unsigned long avg_load; /*Avg load across the CPUs of the group */
7737 unsigned long group_load; /* Total load over the CPUs of the group */
7738 unsigned long load_per_task;
7739 unsigned long group_capacity;
7740 unsigned long group_util; /* Total utilization of the group */
7741 unsigned int sum_nr_running; /* Nr tasks running in the group */
7742 unsigned int idle_cpus;
7743 unsigned int group_weight;
7744 enum group_type group_type;
7745 int group_no_capacity;
7746 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
7747 #ifdef CONFIG_NUMA_BALANCING
7748 unsigned int nr_numa_running;
7749 unsigned int nr_preferred_running;
7754 * sd_lb_stats - Structure to store the statistics of a sched_domain
7755 * during load balancing.
7757 struct sd_lb_stats {
7758 struct sched_group *busiest; /* Busiest group in this sd */
7759 struct sched_group *local; /* Local group in this sd */
7760 unsigned long total_running;
7761 unsigned long total_load; /* Total load of all groups in sd */
7762 unsigned long total_capacity; /* Total capacity of all groups in sd */
7763 unsigned long avg_load; /* Average load across all groups in sd */
7765 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7766 struct sg_lb_stats local_stat; /* Statistics of the local group */
7769 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7772 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7773 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7774 * We must however clear busiest_stat::avg_load because
7775 * update_sd_pick_busiest() reads this before assignment.
7777 *sds = (struct sd_lb_stats){
7780 .total_running = 0UL,
7782 .total_capacity = 0UL,
7785 .sum_nr_running = 0,
7786 .group_type = group_other,
7791 static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu)
7793 struct rq *rq = cpu_rq(cpu);
7794 unsigned long max = arch_scale_cpu_capacity(cpu);
7795 unsigned long used, free;
7798 irq = cpu_util_irq(rq);
7800 if (unlikely(irq >= max))
7803 used = READ_ONCE(rq->avg_rt.util_avg);
7804 used += READ_ONCE(rq->avg_dl.util_avg);
7806 if (unlikely(used >= max))
7811 return scale_irq_capacity(free, irq, max);
7814 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7816 unsigned long capacity = scale_rt_capacity(sd, cpu);
7817 struct sched_group *sdg = sd->groups;
7819 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
7824 cpu_rq(cpu)->cpu_capacity = capacity;
7825 sdg->sgc->capacity = capacity;
7826 sdg->sgc->min_capacity = capacity;
7827 sdg->sgc->max_capacity = capacity;
7830 void update_group_capacity(struct sched_domain *sd, int cpu)
7832 struct sched_domain *child = sd->child;
7833 struct sched_group *group, *sdg = sd->groups;
7834 unsigned long capacity, min_capacity, max_capacity;
7835 unsigned long interval;
7837 interval = msecs_to_jiffies(sd->balance_interval);
7838 interval = clamp(interval, 1UL, max_load_balance_interval);
7839 sdg->sgc->next_update = jiffies + interval;
7842 update_cpu_capacity(sd, cpu);
7847 min_capacity = ULONG_MAX;
7850 if (child->flags & SD_OVERLAP) {
7852 * SD_OVERLAP domains cannot assume that child groups
7853 * span the current group.
7856 for_each_cpu(cpu, sched_group_span(sdg)) {
7857 struct sched_group_capacity *sgc;
7858 struct rq *rq = cpu_rq(cpu);
7861 * build_sched_domains() -> init_sched_groups_capacity()
7862 * gets here before we've attached the domains to the
7865 * Use capacity_of(), which is set irrespective of domains
7866 * in update_cpu_capacity().
7868 * This avoids capacity from being 0 and
7869 * causing divide-by-zero issues on boot.
7871 if (unlikely(!rq->sd)) {
7872 capacity += capacity_of(cpu);
7874 sgc = rq->sd->groups->sgc;
7875 capacity += sgc->capacity;
7878 min_capacity = min(capacity, min_capacity);
7879 max_capacity = max(capacity, max_capacity);
7883 * !SD_OVERLAP domains can assume that child groups
7884 * span the current group.
7887 group = child->groups;
7889 struct sched_group_capacity *sgc = group->sgc;
7891 capacity += sgc->capacity;
7892 min_capacity = min(sgc->min_capacity, min_capacity);
7893 max_capacity = max(sgc->max_capacity, max_capacity);
7894 group = group->next;
7895 } while (group != child->groups);
7898 sdg->sgc->capacity = capacity;
7899 sdg->sgc->min_capacity = min_capacity;
7900 sdg->sgc->max_capacity = max_capacity;
7904 * Check whether the capacity of the rq has been noticeably reduced by side
7905 * activity. The imbalance_pct is used for the threshold.
7906 * Return true is the capacity is reduced
7909 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7911 return ((rq->cpu_capacity * sd->imbalance_pct) <
7912 (rq->cpu_capacity_orig * 100));
7916 * Check whether a rq has a misfit task and if it looks like we can actually
7917 * help that task: we can migrate the task to a CPU of higher capacity, or
7918 * the task's current CPU is heavily pressured.
7920 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
7922 return rq->misfit_task_load &&
7923 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
7924 check_cpu_capacity(rq, sd));
7928 * Group imbalance indicates (and tries to solve) the problem where balancing
7929 * groups is inadequate due to ->cpus_ptr constraints.
7931 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
7932 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
7935 * { 0 1 2 3 } { 4 5 6 7 }
7938 * If we were to balance group-wise we'd place two tasks in the first group and
7939 * two tasks in the second group. Clearly this is undesired as it will overload
7940 * cpu 3 and leave one of the CPUs in the second group unused.
7942 * The current solution to this issue is detecting the skew in the first group
7943 * by noticing the lower domain failed to reach balance and had difficulty
7944 * moving tasks due to affinity constraints.
7946 * When this is so detected; this group becomes a candidate for busiest; see
7947 * update_sd_pick_busiest(). And calculate_imbalance() and
7948 * find_busiest_group() avoid some of the usual balance conditions to allow it
7949 * to create an effective group imbalance.
7951 * This is a somewhat tricky proposition since the next run might not find the
7952 * group imbalance and decide the groups need to be balanced again. A most
7953 * subtle and fragile situation.
7956 static inline int sg_imbalanced(struct sched_group *group)
7958 return group->sgc->imbalance;
7962 * group_has_capacity returns true if the group has spare capacity that could
7963 * be used by some tasks.
7964 * We consider that a group has spare capacity if the * number of task is
7965 * smaller than the number of CPUs or if the utilization is lower than the
7966 * available capacity for CFS tasks.
7967 * For the latter, we use a threshold to stabilize the state, to take into
7968 * account the variance of the tasks' load and to return true if the available
7969 * capacity in meaningful for the load balancer.
7970 * As an example, an available capacity of 1% can appear but it doesn't make
7971 * any benefit for the load balance.
7974 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7976 if (sgs->sum_nr_running < sgs->group_weight)
7979 if ((sgs->group_capacity * 100) >
7980 (sgs->group_util * env->sd->imbalance_pct))
7987 * group_is_overloaded returns true if the group has more tasks than it can
7989 * group_is_overloaded is not equals to !group_has_capacity because a group
7990 * with the exact right number of tasks, has no more spare capacity but is not
7991 * overloaded so both group_has_capacity and group_is_overloaded return
7995 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7997 if (sgs->sum_nr_running <= sgs->group_weight)
8000 if ((sgs->group_capacity * 100) <
8001 (sgs->group_util * env->sd->imbalance_pct))
8008 * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller
8009 * per-CPU capacity than sched_group ref.
8012 group_smaller_min_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
8014 return sg->sgc->min_capacity * capacity_margin <
8015 ref->sgc->min_capacity * 1024;
8019 * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller
8020 * per-CPU capacity_orig than sched_group ref.
8023 group_smaller_max_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
8025 return sg->sgc->max_capacity * capacity_margin <
8026 ref->sgc->max_capacity * 1024;
8030 group_type group_classify(struct sched_group *group,
8031 struct sg_lb_stats *sgs)
8033 if (sgs->group_no_capacity)
8034 return group_overloaded;
8036 if (sg_imbalanced(group))
8037 return group_imbalanced;
8039 if (sgs->group_misfit_task_load)
8040 return group_misfit_task;
8045 static bool update_nohz_stats(struct rq *rq, bool force)
8047 #ifdef CONFIG_NO_HZ_COMMON
8048 unsigned int cpu = rq->cpu;
8050 if (!rq->has_blocked_load)
8053 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
8056 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
8059 update_blocked_averages(cpu);
8061 return rq->has_blocked_load;
8068 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8069 * @env: The load balancing environment.
8070 * @group: sched_group whose statistics are to be updated.
8071 * @sgs: variable to hold the statistics for this group.
8072 * @sg_status: Holds flag indicating the status of the sched_group
8074 static inline void update_sg_lb_stats(struct lb_env *env,
8075 struct sched_group *group,
8076 struct sg_lb_stats *sgs,
8081 memset(sgs, 0, sizeof(*sgs));
8083 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8084 struct rq *rq = cpu_rq(i);
8086 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
8087 env->flags |= LBF_NOHZ_AGAIN;
8089 sgs->group_load += cpu_runnable_load(rq);
8090 sgs->group_util += cpu_util(i);
8091 sgs->sum_nr_running += rq->cfs.h_nr_running;
8093 nr_running = rq->nr_running;
8095 *sg_status |= SG_OVERLOAD;
8097 if (cpu_overutilized(i))
8098 *sg_status |= SG_OVERUTILIZED;
8100 #ifdef CONFIG_NUMA_BALANCING
8101 sgs->nr_numa_running += rq->nr_numa_running;
8102 sgs->nr_preferred_running += rq->nr_preferred_running;
8105 * No need to call idle_cpu() if nr_running is not 0
8107 if (!nr_running && idle_cpu(i))
8110 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8111 sgs->group_misfit_task_load < rq->misfit_task_load) {
8112 sgs->group_misfit_task_load = rq->misfit_task_load;
8113 *sg_status |= SG_OVERLOAD;
8117 /* Adjust by relative CPU capacity of the group */
8118 sgs->group_capacity = group->sgc->capacity;
8119 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
8121 if (sgs->sum_nr_running)
8122 sgs->load_per_task = sgs->group_load / sgs->sum_nr_running;
8124 sgs->group_weight = group->group_weight;
8126 sgs->group_no_capacity = group_is_overloaded(env, sgs);
8127 sgs->group_type = group_classify(group, sgs);
8131 * update_sd_pick_busiest - return 1 on busiest group
8132 * @env: The load balancing environment.
8133 * @sds: sched_domain statistics
8134 * @sg: sched_group candidate to be checked for being the busiest
8135 * @sgs: sched_group statistics
8137 * Determine if @sg is a busier group than the previously selected
8140 * Return: %true if @sg is a busier group than the previously selected
8141 * busiest group. %false otherwise.
8143 static bool update_sd_pick_busiest(struct lb_env *env,
8144 struct sd_lb_stats *sds,
8145 struct sched_group *sg,
8146 struct sg_lb_stats *sgs)
8148 struct sg_lb_stats *busiest = &sds->busiest_stat;
8151 * Don't try to pull misfit tasks we can't help.
8152 * We can use max_capacity here as reduction in capacity on some
8153 * CPUs in the group should either be possible to resolve
8154 * internally or be covered by avg_load imbalance (eventually).
8156 if (sgs->group_type == group_misfit_task &&
8157 (!group_smaller_max_cpu_capacity(sg, sds->local) ||
8158 !group_has_capacity(env, &sds->local_stat)))
8161 if (sgs->group_type > busiest->group_type)
8164 if (sgs->group_type < busiest->group_type)
8167 if (sgs->avg_load <= busiest->avg_load)
8170 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
8174 * Candidate sg has no more than one task per CPU and
8175 * has higher per-CPU capacity. Migrating tasks to less
8176 * capable CPUs may harm throughput. Maximize throughput,
8177 * power/energy consequences are not considered.
8179 if (sgs->sum_nr_running <= sgs->group_weight &&
8180 group_smaller_min_cpu_capacity(sds->local, sg))
8184 * If we have more than one misfit sg go with the biggest misfit.
8186 if (sgs->group_type == group_misfit_task &&
8187 sgs->group_misfit_task_load < busiest->group_misfit_task_load)
8191 /* This is the busiest node in its class. */
8192 if (!(env->sd->flags & SD_ASYM_PACKING))
8195 /* No ASYM_PACKING if target CPU is already busy */
8196 if (env->idle == CPU_NOT_IDLE)
8199 * ASYM_PACKING needs to move all the work to the highest
8200 * prority CPUs in the group, therefore mark all groups
8201 * of lower priority than ourself as busy.
8203 if (sgs->sum_nr_running &&
8204 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
8208 /* Prefer to move from lowest priority CPU's work */
8209 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
8210 sg->asym_prefer_cpu))
8217 #ifdef CONFIG_NUMA_BALANCING
8218 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8220 if (sgs->sum_nr_running > sgs->nr_numa_running)
8222 if (sgs->sum_nr_running > sgs->nr_preferred_running)
8227 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8229 if (rq->nr_running > rq->nr_numa_running)
8231 if (rq->nr_running > rq->nr_preferred_running)
8236 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8241 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8245 #endif /* CONFIG_NUMA_BALANCING */
8248 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8249 * @env: The load balancing environment.
8250 * @sds: variable to hold the statistics for this sched_domain.
8252 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8254 struct sched_domain *child = env->sd->child;
8255 struct sched_group *sg = env->sd->groups;
8256 struct sg_lb_stats *local = &sds->local_stat;
8257 struct sg_lb_stats tmp_sgs;
8258 bool prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
8261 #ifdef CONFIG_NO_HZ_COMMON
8262 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
8263 env->flags |= LBF_NOHZ_STATS;
8267 struct sg_lb_stats *sgs = &tmp_sgs;
8270 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
8275 if (env->idle != CPU_NEWLY_IDLE ||
8276 time_after_eq(jiffies, sg->sgc->next_update))
8277 update_group_capacity(env->sd, env->dst_cpu);
8280 update_sg_lb_stats(env, sg, sgs, &sg_status);
8286 * In case the child domain prefers tasks go to siblings
8287 * first, lower the sg capacity so that we'll try
8288 * and move all the excess tasks away. We lower the capacity
8289 * of a group only if the local group has the capacity to fit
8290 * these excess tasks. The extra check prevents the case where
8291 * you always pull from the heaviest group when it is already
8292 * under-utilized (possible with a large weight task outweighs
8293 * the tasks on the system).
8295 if (prefer_sibling && sds->local &&
8296 group_has_capacity(env, local) &&
8297 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
8298 sgs->group_no_capacity = 1;
8299 sgs->group_type = group_classify(sg, sgs);
8302 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8304 sds->busiest_stat = *sgs;
8308 /* Now, start updating sd_lb_stats */
8309 sds->total_running += sgs->sum_nr_running;
8310 sds->total_load += sgs->group_load;
8311 sds->total_capacity += sgs->group_capacity;
8314 } while (sg != env->sd->groups);
8316 #ifdef CONFIG_NO_HZ_COMMON
8317 if ((env->flags & LBF_NOHZ_AGAIN) &&
8318 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8320 WRITE_ONCE(nohz.next_blocked,
8321 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8325 if (env->sd->flags & SD_NUMA)
8326 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8328 if (!env->sd->parent) {
8329 struct root_domain *rd = env->dst_rq->rd;
8331 /* update overload indicator if we are at root domain */
8332 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
8334 /* Update over-utilization (tipping point, U >= 0) indicator */
8335 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
8336 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
8337 } else if (sg_status & SG_OVERUTILIZED) {
8338 struct root_domain *rd = env->dst_rq->rd;
8340 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
8341 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
8346 * check_asym_packing - Check to see if the group is packed into the
8349 * This is primarily intended to used at the sibling level. Some
8350 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8351 * case of POWER7, it can move to lower SMT modes only when higher
8352 * threads are idle. When in lower SMT modes, the threads will
8353 * perform better since they share less core resources. Hence when we
8354 * have idle threads, we want them to be the higher ones.
8356 * This packing function is run on idle threads. It checks to see if
8357 * the busiest CPU in this domain (core in the P7 case) has a higher
8358 * CPU number than the packing function is being run on. Here we are
8359 * assuming lower CPU number will be equivalent to lower a SMT thread
8362 * Return: 1 when packing is required and a task should be moved to
8363 * this CPU. The amount of the imbalance is returned in env->imbalance.
8365 * @env: The load balancing environment.
8366 * @sds: Statistics of the sched_domain which is to be packed
8368 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8372 if (!(env->sd->flags & SD_ASYM_PACKING))
8375 if (env->idle == CPU_NOT_IDLE)
8381 busiest_cpu = sds->busiest->asym_prefer_cpu;
8382 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
8385 env->imbalance = sds->busiest_stat.group_load;
8391 * fix_small_imbalance - Calculate the minor imbalance that exists
8392 * amongst the groups of a sched_domain, during
8394 * @env: The load balancing environment.
8395 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8398 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8400 unsigned long tmp, capa_now = 0, capa_move = 0;
8401 unsigned int imbn = 2;
8402 unsigned long scaled_busy_load_per_task;
8403 struct sg_lb_stats *local, *busiest;
8405 local = &sds->local_stat;
8406 busiest = &sds->busiest_stat;
8408 if (!local->sum_nr_running)
8409 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8410 else if (busiest->load_per_task > local->load_per_task)
8413 scaled_busy_load_per_task =
8414 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8415 busiest->group_capacity;
8417 if (busiest->avg_load + scaled_busy_load_per_task >=
8418 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8419 env->imbalance = busiest->load_per_task;
8424 * OK, we don't have enough imbalance to justify moving tasks,
8425 * however we may be able to increase total CPU capacity used by
8429 capa_now += busiest->group_capacity *
8430 min(busiest->load_per_task, busiest->avg_load);
8431 capa_now += local->group_capacity *
8432 min(local->load_per_task, local->avg_load);
8433 capa_now /= SCHED_CAPACITY_SCALE;
8435 /* Amount of load we'd subtract */
8436 if (busiest->avg_load > scaled_busy_load_per_task) {
8437 capa_move += busiest->group_capacity *
8438 min(busiest->load_per_task,
8439 busiest->avg_load - scaled_busy_load_per_task);
8442 /* Amount of load we'd add */
8443 if (busiest->avg_load * busiest->group_capacity <
8444 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8445 tmp = (busiest->avg_load * busiest->group_capacity) /
8446 local->group_capacity;
8448 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8449 local->group_capacity;
8451 capa_move += local->group_capacity *
8452 min(local->load_per_task, local->avg_load + tmp);
8453 capa_move /= SCHED_CAPACITY_SCALE;
8455 /* Move if we gain throughput */
8456 if (capa_move > capa_now)
8457 env->imbalance = busiest->load_per_task;
8461 * calculate_imbalance - Calculate the amount of imbalance present within the
8462 * groups of a given sched_domain during load balance.
8463 * @env: load balance environment
8464 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8466 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8468 unsigned long max_pull, load_above_capacity = ~0UL;
8469 struct sg_lb_stats *local, *busiest;
8471 local = &sds->local_stat;
8472 busiest = &sds->busiest_stat;
8474 if (busiest->group_type == group_imbalanced) {
8476 * In the group_imb case we cannot rely on group-wide averages
8477 * to ensure CPU-load equilibrium, look at wider averages. XXX
8479 busiest->load_per_task =
8480 min(busiest->load_per_task, sds->avg_load);
8484 * Avg load of busiest sg can be less and avg load of local sg can
8485 * be greater than avg load across all sgs of sd because avg load
8486 * factors in sg capacity and sgs with smaller group_type are
8487 * skipped when updating the busiest sg:
8489 if (busiest->group_type != group_misfit_task &&
8490 (busiest->avg_load <= sds->avg_load ||
8491 local->avg_load >= sds->avg_load)) {
8493 return fix_small_imbalance(env, sds);
8497 * If there aren't any idle CPUs, avoid creating some.
8499 if (busiest->group_type == group_overloaded &&
8500 local->group_type == group_overloaded) {
8501 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
8502 if (load_above_capacity > busiest->group_capacity) {
8503 load_above_capacity -= busiest->group_capacity;
8504 load_above_capacity *= scale_load_down(NICE_0_LOAD);
8505 load_above_capacity /= busiest->group_capacity;
8507 load_above_capacity = ~0UL;
8511 * We're trying to get all the CPUs to the average_load, so we don't
8512 * want to push ourselves above the average load, nor do we wish to
8513 * reduce the max loaded CPU below the average load. At the same time,
8514 * we also don't want to reduce the group load below the group
8515 * capacity. Thus we look for the minimum possible imbalance.
8517 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8519 /* How much load to actually move to equalise the imbalance */
8520 env->imbalance = min(
8521 max_pull * busiest->group_capacity,
8522 (sds->avg_load - local->avg_load) * local->group_capacity
8523 ) / SCHED_CAPACITY_SCALE;
8525 /* Boost imbalance to allow misfit task to be balanced. */
8526 if (busiest->group_type == group_misfit_task) {
8527 env->imbalance = max_t(long, env->imbalance,
8528 busiest->group_misfit_task_load);
8532 * if *imbalance is less than the average load per runnable task
8533 * there is no guarantee that any tasks will be moved so we'll have
8534 * a think about bumping its value to force at least one task to be
8537 if (env->imbalance < busiest->load_per_task)
8538 return fix_small_imbalance(env, sds);
8541 /******* find_busiest_group() helpers end here *********************/
8544 * find_busiest_group - Returns the busiest group within the sched_domain
8545 * if there is an imbalance.
8547 * Also calculates the amount of runnable load which should be moved
8548 * to restore balance.
8550 * @env: The load balancing environment.
8552 * Return: - The busiest group if imbalance exists.
8554 static struct sched_group *find_busiest_group(struct lb_env *env)
8556 struct sg_lb_stats *local, *busiest;
8557 struct sd_lb_stats sds;
8559 init_sd_lb_stats(&sds);
8562 * Compute the various statistics relavent for load balancing at
8565 update_sd_lb_stats(env, &sds);
8567 if (sched_energy_enabled()) {
8568 struct root_domain *rd = env->dst_rq->rd;
8570 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
8574 local = &sds.local_stat;
8575 busiest = &sds.busiest_stat;
8577 /* ASYM feature bypasses nice load balance check */
8578 if (check_asym_packing(env, &sds))
8581 /* There is no busy sibling group to pull tasks from */
8582 if (!sds.busiest || busiest->sum_nr_running == 0)
8585 /* XXX broken for overlapping NUMA groups */
8586 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8587 / sds.total_capacity;
8590 * If the busiest group is imbalanced the below checks don't
8591 * work because they assume all things are equal, which typically
8592 * isn't true due to cpus_ptr constraints and the like.
8594 if (busiest->group_type == group_imbalanced)
8598 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8599 * capacities from resulting in underutilization due to avg_load.
8601 if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8602 busiest->group_no_capacity)
8605 /* Misfit tasks should be dealt with regardless of the avg load */
8606 if (busiest->group_type == group_misfit_task)
8610 * If the local group is busier than the selected busiest group
8611 * don't try and pull any tasks.
8613 if (local->avg_load >= busiest->avg_load)
8617 * Don't pull any tasks if this group is already above the domain
8620 if (local->avg_load >= sds.avg_load)
8623 if (env->idle == CPU_IDLE) {
8625 * This CPU is idle. If the busiest group is not overloaded
8626 * and there is no imbalance between this and busiest group
8627 * wrt idle CPUs, it is balanced. The imbalance becomes
8628 * significant if the diff is greater than 1 otherwise we
8629 * might end up to just move the imbalance on another group
8631 if ((busiest->group_type != group_overloaded) &&
8632 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8636 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8637 * imbalance_pct to be conservative.
8639 if (100 * busiest->avg_load <=
8640 env->sd->imbalance_pct * local->avg_load)
8645 /* Looks like there is an imbalance. Compute it */
8646 env->src_grp_type = busiest->group_type;
8647 calculate_imbalance(env, &sds);
8648 return env->imbalance ? sds.busiest : NULL;
8656 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
8658 static struct rq *find_busiest_queue(struct lb_env *env,
8659 struct sched_group *group)
8661 struct rq *busiest = NULL, *rq;
8662 unsigned long busiest_load = 0, busiest_capacity = 1;
8665 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8666 unsigned long capacity, load;
8670 rt = fbq_classify_rq(rq);
8673 * We classify groups/runqueues into three groups:
8674 * - regular: there are !numa tasks
8675 * - remote: there are numa tasks that run on the 'wrong' node
8676 * - all: there is no distinction
8678 * In order to avoid migrating ideally placed numa tasks,
8679 * ignore those when there's better options.
8681 * If we ignore the actual busiest queue to migrate another
8682 * task, the next balance pass can still reduce the busiest
8683 * queue by moving tasks around inside the node.
8685 * If we cannot move enough load due to this classification
8686 * the next pass will adjust the group classification and
8687 * allow migration of more tasks.
8689 * Both cases only affect the total convergence complexity.
8691 if (rt > env->fbq_type)
8695 * For ASYM_CPUCAPACITY domains with misfit tasks we simply
8696 * seek the "biggest" misfit task.
8698 if (env->src_grp_type == group_misfit_task) {
8699 if (rq->misfit_task_load > busiest_load) {
8700 busiest_load = rq->misfit_task_load;
8707 capacity = capacity_of(i);
8710 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
8711 * eventually lead to active_balancing high->low capacity.
8712 * Higher per-CPU capacity is considered better than balancing
8715 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8716 capacity_of(env->dst_cpu) < capacity &&
8717 rq->nr_running == 1)
8720 load = cpu_runnable_load(rq);
8723 * When comparing with imbalance, use cpu_runnable_load()
8724 * which is not scaled with the CPU capacity.
8727 if (rq->nr_running == 1 && load > env->imbalance &&
8728 !check_cpu_capacity(rq, env->sd))
8732 * For the load comparisons with the other CPU's, consider
8733 * the cpu_runnable_load() scaled with the CPU capacity, so
8734 * that the load can be moved away from the CPU that is
8735 * potentially running at a lower capacity.
8737 * Thus we're looking for max(load_i / capacity_i), crosswise
8738 * multiplication to rid ourselves of the division works out
8739 * to: load_i * capacity_j > load_j * capacity_i; where j is
8740 * our previous maximum.
8742 if (load * busiest_capacity > busiest_load * capacity) {
8743 busiest_load = load;
8744 busiest_capacity = capacity;
8753 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8754 * so long as it is large enough.
8756 #define MAX_PINNED_INTERVAL 512
8759 asym_active_balance(struct lb_env *env)
8762 * ASYM_PACKING needs to force migrate tasks from busy but
8763 * lower priority CPUs in order to pack all tasks in the
8764 * highest priority CPUs.
8766 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
8767 sched_asym_prefer(env->dst_cpu, env->src_cpu);
8771 voluntary_active_balance(struct lb_env *env)
8773 struct sched_domain *sd = env->sd;
8775 if (asym_active_balance(env))
8779 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8780 * It's worth migrating the task if the src_cpu's capacity is reduced
8781 * because of other sched_class or IRQs if more capacity stays
8782 * available on dst_cpu.
8784 if ((env->idle != CPU_NOT_IDLE) &&
8785 (env->src_rq->cfs.h_nr_running == 1)) {
8786 if ((check_cpu_capacity(env->src_rq, sd)) &&
8787 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8791 if (env->src_grp_type == group_misfit_task)
8797 static int need_active_balance(struct lb_env *env)
8799 struct sched_domain *sd = env->sd;
8801 if (voluntary_active_balance(env))
8804 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8807 static int active_load_balance_cpu_stop(void *data);
8809 static int should_we_balance(struct lb_env *env)
8811 struct sched_group *sg = env->sd->groups;
8812 int cpu, balance_cpu = -1;
8815 * Ensure the balancing environment is consistent; can happen
8816 * when the softirq triggers 'during' hotplug.
8818 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8822 * In the newly idle case, we will allow all the CPUs
8823 * to do the newly idle load balance.
8825 if (env->idle == CPU_NEWLY_IDLE)
8828 /* Try to find first idle CPU */
8829 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8837 if (balance_cpu == -1)
8838 balance_cpu = group_balance_cpu(sg);
8841 * First idle CPU or the first CPU(busiest) in this sched group
8842 * is eligible for doing load balancing at this and above domains.
8844 return balance_cpu == env->dst_cpu;
8848 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8849 * tasks if there is an imbalance.
8851 static int load_balance(int this_cpu, struct rq *this_rq,
8852 struct sched_domain *sd, enum cpu_idle_type idle,
8853 int *continue_balancing)
8855 int ld_moved, cur_ld_moved, active_balance = 0;
8856 struct sched_domain *sd_parent = sd->parent;
8857 struct sched_group *group;
8860 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8862 struct lb_env env = {
8864 .dst_cpu = this_cpu,
8866 .dst_grpmask = sched_group_span(sd->groups),
8868 .loop_break = sched_nr_migrate_break,
8871 .tasks = LIST_HEAD_INIT(env.tasks),
8874 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8876 schedstat_inc(sd->lb_count[idle]);
8879 if (!should_we_balance(&env)) {
8880 *continue_balancing = 0;
8884 group = find_busiest_group(&env);
8886 schedstat_inc(sd->lb_nobusyg[idle]);
8890 busiest = find_busiest_queue(&env, group);
8892 schedstat_inc(sd->lb_nobusyq[idle]);
8896 BUG_ON(busiest == env.dst_rq);
8898 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8900 env.src_cpu = busiest->cpu;
8901 env.src_rq = busiest;
8904 if (busiest->nr_running > 1) {
8906 * Attempt to move tasks. If find_busiest_group has found
8907 * an imbalance but busiest->nr_running <= 1, the group is
8908 * still unbalanced. ld_moved simply stays zero, so it is
8909 * correctly treated as an imbalance.
8911 env.flags |= LBF_ALL_PINNED;
8912 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8915 rq_lock_irqsave(busiest, &rf);
8916 update_rq_clock(busiest);
8919 * cur_ld_moved - load moved in current iteration
8920 * ld_moved - cumulative load moved across iterations
8922 cur_ld_moved = detach_tasks(&env);
8925 * We've detached some tasks from busiest_rq. Every
8926 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8927 * unlock busiest->lock, and we are able to be sure
8928 * that nobody can manipulate the tasks in parallel.
8929 * See task_rq_lock() family for the details.
8932 rq_unlock(busiest, &rf);
8936 ld_moved += cur_ld_moved;
8939 local_irq_restore(rf.flags);
8941 if (env.flags & LBF_NEED_BREAK) {
8942 env.flags &= ~LBF_NEED_BREAK;
8947 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8948 * us and move them to an alternate dst_cpu in our sched_group
8949 * where they can run. The upper limit on how many times we
8950 * iterate on same src_cpu is dependent on number of CPUs in our
8953 * This changes load balance semantics a bit on who can move
8954 * load to a given_cpu. In addition to the given_cpu itself
8955 * (or a ilb_cpu acting on its behalf where given_cpu is
8956 * nohz-idle), we now have balance_cpu in a position to move
8957 * load to given_cpu. In rare situations, this may cause
8958 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8959 * _independently_ and at _same_ time to move some load to
8960 * given_cpu) causing exceess load to be moved to given_cpu.
8961 * This however should not happen so much in practice and
8962 * moreover subsequent load balance cycles should correct the
8963 * excess load moved.
8965 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8967 /* Prevent to re-select dst_cpu via env's CPUs */
8968 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
8970 env.dst_rq = cpu_rq(env.new_dst_cpu);
8971 env.dst_cpu = env.new_dst_cpu;
8972 env.flags &= ~LBF_DST_PINNED;
8974 env.loop_break = sched_nr_migrate_break;
8977 * Go back to "more_balance" rather than "redo" since we
8978 * need to continue with same src_cpu.
8984 * We failed to reach balance because of affinity.
8987 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8989 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8990 *group_imbalance = 1;
8993 /* All tasks on this runqueue were pinned by CPU affinity */
8994 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8995 __cpumask_clear_cpu(cpu_of(busiest), cpus);
8997 * Attempting to continue load balancing at the current
8998 * sched_domain level only makes sense if there are
8999 * active CPUs remaining as possible busiest CPUs to
9000 * pull load from which are not contained within the
9001 * destination group that is receiving any migrated
9004 if (!cpumask_subset(cpus, env.dst_grpmask)) {
9006 env.loop_break = sched_nr_migrate_break;
9009 goto out_all_pinned;
9014 schedstat_inc(sd->lb_failed[idle]);
9016 * Increment the failure counter only on periodic balance.
9017 * We do not want newidle balance, which can be very
9018 * frequent, pollute the failure counter causing
9019 * excessive cache_hot migrations and active balances.
9021 if (idle != CPU_NEWLY_IDLE)
9022 sd->nr_balance_failed++;
9024 if (need_active_balance(&env)) {
9025 unsigned long flags;
9027 raw_spin_lock_irqsave(&busiest->lock, flags);
9030 * Don't kick the active_load_balance_cpu_stop,
9031 * if the curr task on busiest CPU can't be
9032 * moved to this_cpu:
9034 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
9035 raw_spin_unlock_irqrestore(&busiest->lock,
9037 env.flags |= LBF_ALL_PINNED;
9038 goto out_one_pinned;
9042 * ->active_balance synchronizes accesses to
9043 * ->active_balance_work. Once set, it's cleared
9044 * only after active load balance is finished.
9046 if (!busiest->active_balance) {
9047 busiest->active_balance = 1;
9048 busiest->push_cpu = this_cpu;
9051 raw_spin_unlock_irqrestore(&busiest->lock, flags);
9053 if (active_balance) {
9054 stop_one_cpu_nowait(cpu_of(busiest),
9055 active_load_balance_cpu_stop, busiest,
9056 &busiest->active_balance_work);
9059 /* We've kicked active balancing, force task migration. */
9060 sd->nr_balance_failed = sd->cache_nice_tries+1;
9063 sd->nr_balance_failed = 0;
9065 if (likely(!active_balance) || voluntary_active_balance(&env)) {
9066 /* We were unbalanced, so reset the balancing interval */
9067 sd->balance_interval = sd->min_interval;
9070 * If we've begun active balancing, start to back off. This
9071 * case may not be covered by the all_pinned logic if there
9072 * is only 1 task on the busy runqueue (because we don't call
9075 if (sd->balance_interval < sd->max_interval)
9076 sd->balance_interval *= 2;
9083 * We reach balance although we may have faced some affinity
9084 * constraints. Clear the imbalance flag only if other tasks got
9085 * a chance to move and fix the imbalance.
9087 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
9088 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9090 if (*group_imbalance)
9091 *group_imbalance = 0;
9096 * We reach balance because all tasks are pinned at this level so
9097 * we can't migrate them. Let the imbalance flag set so parent level
9098 * can try to migrate them.
9100 schedstat_inc(sd->lb_balanced[idle]);
9102 sd->nr_balance_failed = 0;
9108 * idle_balance() disregards balance intervals, so we could repeatedly
9109 * reach this code, which would lead to balance_interval skyrocketting
9110 * in a short amount of time. Skip the balance_interval increase logic
9113 if (env.idle == CPU_NEWLY_IDLE)
9116 /* tune up the balancing interval */
9117 if ((env.flags & LBF_ALL_PINNED &&
9118 sd->balance_interval < MAX_PINNED_INTERVAL) ||
9119 sd->balance_interval < sd->max_interval)
9120 sd->balance_interval *= 2;
9125 static inline unsigned long
9126 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
9128 unsigned long interval = sd->balance_interval;
9131 interval *= sd->busy_factor;
9133 /* scale ms to jiffies */
9134 interval = msecs_to_jiffies(interval);
9135 interval = clamp(interval, 1UL, max_load_balance_interval);
9141 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
9143 unsigned long interval, next;
9145 /* used by idle balance, so cpu_busy = 0 */
9146 interval = get_sd_balance_interval(sd, 0);
9147 next = sd->last_balance + interval;
9149 if (time_after(*next_balance, next))
9150 *next_balance = next;
9154 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9155 * running tasks off the busiest CPU onto idle CPUs. It requires at
9156 * least 1 task to be running on each physical CPU where possible, and
9157 * avoids physical / logical imbalances.
9159 static int active_load_balance_cpu_stop(void *data)
9161 struct rq *busiest_rq = data;
9162 int busiest_cpu = cpu_of(busiest_rq);
9163 int target_cpu = busiest_rq->push_cpu;
9164 struct rq *target_rq = cpu_rq(target_cpu);
9165 struct sched_domain *sd;
9166 struct task_struct *p = NULL;
9169 rq_lock_irq(busiest_rq, &rf);
9171 * Between queueing the stop-work and running it is a hole in which
9172 * CPUs can become inactive. We should not move tasks from or to
9175 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
9178 /* Make sure the requested CPU hasn't gone down in the meantime: */
9179 if (unlikely(busiest_cpu != smp_processor_id() ||
9180 !busiest_rq->active_balance))
9183 /* Is there any task to move? */
9184 if (busiest_rq->nr_running <= 1)
9188 * This condition is "impossible", if it occurs
9189 * we need to fix it. Originally reported by
9190 * Bjorn Helgaas on a 128-CPU setup.
9192 BUG_ON(busiest_rq == target_rq);
9194 /* Search for an sd spanning us and the target CPU. */
9196 for_each_domain(target_cpu, sd) {
9197 if ((sd->flags & SD_LOAD_BALANCE) &&
9198 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9203 struct lb_env env = {
9205 .dst_cpu = target_cpu,
9206 .dst_rq = target_rq,
9207 .src_cpu = busiest_rq->cpu,
9208 .src_rq = busiest_rq,
9211 * can_migrate_task() doesn't need to compute new_dst_cpu
9212 * for active balancing. Since we have CPU_IDLE, but no
9213 * @dst_grpmask we need to make that test go away with lying
9216 .flags = LBF_DST_PINNED,
9219 schedstat_inc(sd->alb_count);
9220 update_rq_clock(busiest_rq);
9222 p = detach_one_task(&env);
9224 schedstat_inc(sd->alb_pushed);
9225 /* Active balancing done, reset the failure counter. */
9226 sd->nr_balance_failed = 0;
9228 schedstat_inc(sd->alb_failed);
9233 busiest_rq->active_balance = 0;
9234 rq_unlock(busiest_rq, &rf);
9237 attach_one_task(target_rq, p);
9244 static DEFINE_SPINLOCK(balancing);
9247 * Scale the max load_balance interval with the number of CPUs in the system.
9248 * This trades load-balance latency on larger machines for less cross talk.
9250 void update_max_interval(void)
9252 max_load_balance_interval = HZ*num_online_cpus()/10;
9256 * It checks each scheduling domain to see if it is due to be balanced,
9257 * and initiates a balancing operation if so.
9259 * Balancing parameters are set up in init_sched_domains.
9261 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9263 int continue_balancing = 1;
9265 unsigned long interval;
9266 struct sched_domain *sd;
9267 /* Earliest time when we have to do rebalance again */
9268 unsigned long next_balance = jiffies + 60*HZ;
9269 int update_next_balance = 0;
9270 int need_serialize, need_decay = 0;
9274 for_each_domain(cpu, sd) {
9276 * Decay the newidle max times here because this is a regular
9277 * visit to all the domains. Decay ~1% per second.
9279 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9280 sd->max_newidle_lb_cost =
9281 (sd->max_newidle_lb_cost * 253) / 256;
9282 sd->next_decay_max_lb_cost = jiffies + HZ;
9285 max_cost += sd->max_newidle_lb_cost;
9287 if (!(sd->flags & SD_LOAD_BALANCE))
9291 * Stop the load balance at this level. There is another
9292 * CPU in our sched group which is doing load balancing more
9295 if (!continue_balancing) {
9301 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9303 need_serialize = sd->flags & SD_SERIALIZE;
9304 if (need_serialize) {
9305 if (!spin_trylock(&balancing))
9309 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9310 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9312 * The LBF_DST_PINNED logic could have changed
9313 * env->dst_cpu, so we can't know our idle
9314 * state even if we migrated tasks. Update it.
9316 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9318 sd->last_balance = jiffies;
9319 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9322 spin_unlock(&balancing);
9324 if (time_after(next_balance, sd->last_balance + interval)) {
9325 next_balance = sd->last_balance + interval;
9326 update_next_balance = 1;
9331 * Ensure the rq-wide value also decays but keep it at a
9332 * reasonable floor to avoid funnies with rq->avg_idle.
9334 rq->max_idle_balance_cost =
9335 max((u64)sysctl_sched_migration_cost, max_cost);
9340 * next_balance will be updated only when there is a need.
9341 * When the cpu is attached to null domain for ex, it will not be
9344 if (likely(update_next_balance)) {
9345 rq->next_balance = next_balance;
9347 #ifdef CONFIG_NO_HZ_COMMON
9349 * If this CPU has been elected to perform the nohz idle
9350 * balance. Other idle CPUs have already rebalanced with
9351 * nohz_idle_balance() and nohz.next_balance has been
9352 * updated accordingly. This CPU is now running the idle load
9353 * balance for itself and we need to update the
9354 * nohz.next_balance accordingly.
9356 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9357 nohz.next_balance = rq->next_balance;
9362 static inline int on_null_domain(struct rq *rq)
9364 return unlikely(!rcu_dereference_sched(rq->sd));
9367 #ifdef CONFIG_NO_HZ_COMMON
9369 * idle load balancing details
9370 * - When one of the busy CPUs notice that there may be an idle rebalancing
9371 * needed, they will kick the idle load balancer, which then does idle
9372 * load balancing for all the idle CPUs.
9373 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
9377 static inline int find_new_ilb(void)
9381 for_each_cpu_and(ilb, nohz.idle_cpus_mask,
9382 housekeeping_cpumask(HK_FLAG_MISC)) {
9391 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
9392 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
9394 static void kick_ilb(unsigned int flags)
9398 nohz.next_balance++;
9400 ilb_cpu = find_new_ilb();
9402 if (ilb_cpu >= nr_cpu_ids)
9405 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
9406 if (flags & NOHZ_KICK_MASK)
9410 * Use smp_send_reschedule() instead of resched_cpu().
9411 * This way we generate a sched IPI on the target CPU which
9412 * is idle. And the softirq performing nohz idle load balance
9413 * will be run before returning from the IPI.
9415 smp_send_reschedule(ilb_cpu);
9419 * Current decision point for kicking the idle load balancer in the presence
9420 * of idle CPUs in the system.
9422 static void nohz_balancer_kick(struct rq *rq)
9424 unsigned long now = jiffies;
9425 struct sched_domain_shared *sds;
9426 struct sched_domain *sd;
9427 int nr_busy, i, cpu = rq->cpu;
9428 unsigned int flags = 0;
9430 if (unlikely(rq->idle_balance))
9434 * We may be recently in ticked or tickless idle mode. At the first
9435 * busy tick after returning from idle, we will update the busy stats.
9437 nohz_balance_exit_idle(rq);
9440 * None are in tickless mode and hence no need for NOHZ idle load
9443 if (likely(!atomic_read(&nohz.nr_cpus)))
9446 if (READ_ONCE(nohz.has_blocked) &&
9447 time_after(now, READ_ONCE(nohz.next_blocked)))
9448 flags = NOHZ_STATS_KICK;
9450 if (time_before(now, nohz.next_balance))
9453 if (rq->nr_running >= 2) {
9454 flags = NOHZ_KICK_MASK;
9460 sd = rcu_dereference(rq->sd);
9463 * If there's a CFS task and the current CPU has reduced
9464 * capacity; kick the ILB to see if there's a better CPU to run
9467 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
9468 flags = NOHZ_KICK_MASK;
9473 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
9476 * When ASYM_PACKING; see if there's a more preferred CPU
9477 * currently idle; in which case, kick the ILB to move tasks
9480 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
9481 if (sched_asym_prefer(i, cpu)) {
9482 flags = NOHZ_KICK_MASK;
9488 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
9491 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
9492 * to run the misfit task on.
9494 if (check_misfit_status(rq, sd)) {
9495 flags = NOHZ_KICK_MASK;
9500 * For asymmetric systems, we do not want to nicely balance
9501 * cache use, instead we want to embrace asymmetry and only
9502 * ensure tasks have enough CPU capacity.
9504 * Skip the LLC logic because it's not relevant in that case.
9509 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9512 * If there is an imbalance between LLC domains (IOW we could
9513 * increase the overall cache use), we need some less-loaded LLC
9514 * domain to pull some load. Likewise, we may need to spread
9515 * load within the current LLC domain (e.g. packed SMT cores but
9516 * other CPUs are idle). We can't really know from here how busy
9517 * the others are - so just get a nohz balance going if it looks
9518 * like this LLC domain has tasks we could move.
9520 nr_busy = atomic_read(&sds->nr_busy_cpus);
9522 flags = NOHZ_KICK_MASK;
9533 static void set_cpu_sd_state_busy(int cpu)
9535 struct sched_domain *sd;
9538 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9540 if (!sd || !sd->nohz_idle)
9544 atomic_inc(&sd->shared->nr_busy_cpus);
9549 void nohz_balance_exit_idle(struct rq *rq)
9551 SCHED_WARN_ON(rq != this_rq());
9553 if (likely(!rq->nohz_tick_stopped))
9556 rq->nohz_tick_stopped = 0;
9557 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
9558 atomic_dec(&nohz.nr_cpus);
9560 set_cpu_sd_state_busy(rq->cpu);
9563 static void set_cpu_sd_state_idle(int cpu)
9565 struct sched_domain *sd;
9568 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9570 if (!sd || sd->nohz_idle)
9574 atomic_dec(&sd->shared->nr_busy_cpus);
9580 * This routine will record that the CPU is going idle with tick stopped.
9581 * This info will be used in performing idle load balancing in the future.
9583 void nohz_balance_enter_idle(int cpu)
9585 struct rq *rq = cpu_rq(cpu);
9587 SCHED_WARN_ON(cpu != smp_processor_id());
9589 /* If this CPU is going down, then nothing needs to be done: */
9590 if (!cpu_active(cpu))
9593 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9594 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9598 * Can be set safely without rq->lock held
9599 * If a clear happens, it will have evaluated last additions because
9600 * rq->lock is held during the check and the clear
9602 rq->has_blocked_load = 1;
9605 * The tick is still stopped but load could have been added in the
9606 * meantime. We set the nohz.has_blocked flag to trig a check of the
9607 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
9608 * of nohz.has_blocked can only happen after checking the new load
9610 if (rq->nohz_tick_stopped)
9613 /* If we're a completely isolated CPU, we don't play: */
9614 if (on_null_domain(rq))
9617 rq->nohz_tick_stopped = 1;
9619 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9620 atomic_inc(&nohz.nr_cpus);
9623 * Ensures that if nohz_idle_balance() fails to observe our
9624 * @idle_cpus_mask store, it must observe the @has_blocked
9627 smp_mb__after_atomic();
9629 set_cpu_sd_state_idle(cpu);
9633 * Each time a cpu enter idle, we assume that it has blocked load and
9634 * enable the periodic update of the load of idle cpus
9636 WRITE_ONCE(nohz.has_blocked, 1);
9640 * Internal function that runs load balance for all idle cpus. The load balance
9641 * can be a simple update of blocked load or a complete load balance with
9642 * tasks movement depending of flags.
9643 * The function returns false if the loop has stopped before running
9644 * through all idle CPUs.
9646 static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
9647 enum cpu_idle_type idle)
9649 /* Earliest time when we have to do rebalance again */
9650 unsigned long now = jiffies;
9651 unsigned long next_balance = now + 60*HZ;
9652 bool has_blocked_load = false;
9653 int update_next_balance = 0;
9654 int this_cpu = this_rq->cpu;
9659 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
9662 * We assume there will be no idle load after this update and clear
9663 * the has_blocked flag. If a cpu enters idle in the mean time, it will
9664 * set the has_blocked flag and trig another update of idle load.
9665 * Because a cpu that becomes idle, is added to idle_cpus_mask before
9666 * setting the flag, we are sure to not clear the state and not
9667 * check the load of an idle cpu.
9669 WRITE_ONCE(nohz.has_blocked, 0);
9672 * Ensures that if we miss the CPU, we must see the has_blocked
9673 * store from nohz_balance_enter_idle().
9677 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9678 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9682 * If this CPU gets work to do, stop the load balancing
9683 * work being done for other CPUs. Next load
9684 * balancing owner will pick it up.
9686 if (need_resched()) {
9687 has_blocked_load = true;
9691 rq = cpu_rq(balance_cpu);
9693 has_blocked_load |= update_nohz_stats(rq, true);
9696 * If time for next balance is due,
9699 if (time_after_eq(jiffies, rq->next_balance)) {
9702 rq_lock_irqsave(rq, &rf);
9703 update_rq_clock(rq);
9704 rq_unlock_irqrestore(rq, &rf);
9706 if (flags & NOHZ_BALANCE_KICK)
9707 rebalance_domains(rq, CPU_IDLE);
9710 if (time_after(next_balance, rq->next_balance)) {
9711 next_balance = rq->next_balance;
9712 update_next_balance = 1;
9716 /* Newly idle CPU doesn't need an update */
9717 if (idle != CPU_NEWLY_IDLE) {
9718 update_blocked_averages(this_cpu);
9719 has_blocked_load |= this_rq->has_blocked_load;
9722 if (flags & NOHZ_BALANCE_KICK)
9723 rebalance_domains(this_rq, CPU_IDLE);
9725 WRITE_ONCE(nohz.next_blocked,
9726 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
9728 /* The full idle balance loop has been done */
9732 /* There is still blocked load, enable periodic update */
9733 if (has_blocked_load)
9734 WRITE_ONCE(nohz.has_blocked, 1);
9737 * next_balance will be updated only when there is a need.
9738 * When the CPU is attached to null domain for ex, it will not be
9741 if (likely(update_next_balance))
9742 nohz.next_balance = next_balance;
9748 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9749 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9751 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9753 int this_cpu = this_rq->cpu;
9756 if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
9759 if (idle != CPU_IDLE) {
9760 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9764 /* could be _relaxed() */
9765 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9766 if (!(flags & NOHZ_KICK_MASK))
9769 _nohz_idle_balance(this_rq, flags, idle);
9774 static void nohz_newidle_balance(struct rq *this_rq)
9776 int this_cpu = this_rq->cpu;
9779 * This CPU doesn't want to be disturbed by scheduler
9782 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
9785 /* Will wake up very soon. No time for doing anything else*/
9786 if (this_rq->avg_idle < sysctl_sched_migration_cost)
9789 /* Don't need to update blocked load of idle CPUs*/
9790 if (!READ_ONCE(nohz.has_blocked) ||
9791 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
9794 raw_spin_unlock(&this_rq->lock);
9796 * This CPU is going to be idle and blocked load of idle CPUs
9797 * need to be updated. Run the ilb locally as it is a good
9798 * candidate for ilb instead of waking up another idle CPU.
9799 * Kick an normal ilb if we failed to do the update.
9801 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
9802 kick_ilb(NOHZ_STATS_KICK);
9803 raw_spin_lock(&this_rq->lock);
9806 #else /* !CONFIG_NO_HZ_COMMON */
9807 static inline void nohz_balancer_kick(struct rq *rq) { }
9809 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9814 static inline void nohz_newidle_balance(struct rq *this_rq) { }
9815 #endif /* CONFIG_NO_HZ_COMMON */
9818 * idle_balance is called by schedule() if this_cpu is about to become
9819 * idle. Attempts to pull tasks from other CPUs.
9821 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
9823 unsigned long next_balance = jiffies + HZ;
9824 int this_cpu = this_rq->cpu;
9825 struct sched_domain *sd;
9826 int pulled_task = 0;
9830 * We must set idle_stamp _before_ calling idle_balance(), such that we
9831 * measure the duration of idle_balance() as idle time.
9833 this_rq->idle_stamp = rq_clock(this_rq);
9836 * Do not pull tasks towards !active CPUs...
9838 if (!cpu_active(this_cpu))
9842 * This is OK, because current is on_cpu, which avoids it being picked
9843 * for load-balance and preemption/IRQs are still disabled avoiding
9844 * further scheduler activity on it and we're being very careful to
9845 * re-start the picking loop.
9847 rq_unpin_lock(this_rq, rf);
9849 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
9850 !READ_ONCE(this_rq->rd->overload)) {
9853 sd = rcu_dereference_check_sched_domain(this_rq->sd);
9855 update_next_balance(sd, &next_balance);
9858 nohz_newidle_balance(this_rq);
9863 raw_spin_unlock(&this_rq->lock);
9865 update_blocked_averages(this_cpu);
9867 for_each_domain(this_cpu, sd) {
9868 int continue_balancing = 1;
9869 u64 t0, domain_cost;
9871 if (!(sd->flags & SD_LOAD_BALANCE))
9874 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
9875 update_next_balance(sd, &next_balance);
9879 if (sd->flags & SD_BALANCE_NEWIDLE) {
9880 t0 = sched_clock_cpu(this_cpu);
9882 pulled_task = load_balance(this_cpu, this_rq,
9884 &continue_balancing);
9886 domain_cost = sched_clock_cpu(this_cpu) - t0;
9887 if (domain_cost > sd->max_newidle_lb_cost)
9888 sd->max_newidle_lb_cost = domain_cost;
9890 curr_cost += domain_cost;
9893 update_next_balance(sd, &next_balance);
9896 * Stop searching for tasks to pull if there are
9897 * now runnable tasks on this rq.
9899 if (pulled_task || this_rq->nr_running > 0)
9904 raw_spin_lock(&this_rq->lock);
9906 if (curr_cost > this_rq->max_idle_balance_cost)
9907 this_rq->max_idle_balance_cost = curr_cost;
9911 * While browsing the domains, we released the rq lock, a task could
9912 * have been enqueued in the meantime. Since we're not going idle,
9913 * pretend we pulled a task.
9915 if (this_rq->cfs.h_nr_running && !pulled_task)
9918 /* Move the next balance forward */
9919 if (time_after(this_rq->next_balance, next_balance))
9920 this_rq->next_balance = next_balance;
9922 /* Is there a task of a high priority class? */
9923 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
9927 this_rq->idle_stamp = 0;
9929 rq_repin_lock(this_rq, rf);
9935 * run_rebalance_domains is triggered when needed from the scheduler tick.
9936 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9938 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9940 struct rq *this_rq = this_rq();
9941 enum cpu_idle_type idle = this_rq->idle_balance ?
9942 CPU_IDLE : CPU_NOT_IDLE;
9945 * If this CPU has a pending nohz_balance_kick, then do the
9946 * balancing on behalf of the other idle CPUs whose ticks are
9947 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9948 * give the idle CPUs a chance to load balance. Else we may
9949 * load balance only within the local sched_domain hierarchy
9950 * and abort nohz_idle_balance altogether if we pull some load.
9952 if (nohz_idle_balance(this_rq, idle))
9955 /* normal load balance */
9956 update_blocked_averages(this_rq->cpu);
9957 rebalance_domains(this_rq, idle);
9961 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9963 void trigger_load_balance(struct rq *rq)
9965 /* Don't need to rebalance while attached to NULL domain */
9966 if (unlikely(on_null_domain(rq)))
9969 if (time_after_eq(jiffies, rq->next_balance))
9970 raise_softirq(SCHED_SOFTIRQ);
9972 nohz_balancer_kick(rq);
9975 static void rq_online_fair(struct rq *rq)
9979 update_runtime_enabled(rq);
9982 static void rq_offline_fair(struct rq *rq)
9986 /* Ensure any throttled groups are reachable by pick_next_task */
9987 unthrottle_offline_cfs_rqs(rq);
9990 #endif /* CONFIG_SMP */
9993 * scheduler tick hitting a task of our scheduling class.
9995 * NOTE: This function can be called remotely by the tick offload that
9996 * goes along full dynticks. Therefore no local assumption can be made
9997 * and everything must be accessed through the @rq and @curr passed in
10000 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
10002 struct cfs_rq *cfs_rq;
10003 struct sched_entity *se = &curr->se;
10005 for_each_sched_entity(se) {
10006 cfs_rq = cfs_rq_of(se);
10007 entity_tick(cfs_rq, se, queued);
10010 if (static_branch_unlikely(&sched_numa_balancing))
10011 task_tick_numa(rq, curr);
10013 update_misfit_status(curr, rq);
10014 update_overutilized_status(task_rq(curr));
10018 * called on fork with the child task as argument from the parent's context
10019 * - child not yet on the tasklist
10020 * - preemption disabled
10022 static void task_fork_fair(struct task_struct *p)
10024 struct cfs_rq *cfs_rq;
10025 struct sched_entity *se = &p->se, *curr;
10026 struct rq *rq = this_rq();
10027 struct rq_flags rf;
10030 update_rq_clock(rq);
10032 cfs_rq = task_cfs_rq(current);
10033 curr = cfs_rq->curr;
10035 update_curr(cfs_rq);
10036 se->vruntime = curr->vruntime;
10038 place_entity(cfs_rq, se, 1);
10040 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
10042 * Upon rescheduling, sched_class::put_prev_task() will place
10043 * 'current' within the tree based on its new key value.
10045 swap(curr->vruntime, se->vruntime);
10049 se->vruntime -= cfs_rq->min_vruntime;
10050 rq_unlock(rq, &rf);
10054 * Priority of the task has changed. Check to see if we preempt
10055 * the current task.
10058 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
10060 if (!task_on_rq_queued(p))
10064 * Reschedule if we are currently running on this runqueue and
10065 * our priority decreased, or if we are not currently running on
10066 * this runqueue and our priority is higher than the current's
10068 if (rq->curr == p) {
10069 if (p->prio > oldprio)
10072 check_preempt_curr(rq, p, 0);
10075 static inline bool vruntime_normalized(struct task_struct *p)
10077 struct sched_entity *se = &p->se;
10080 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
10081 * the dequeue_entity(.flags=0) will already have normalized the
10088 * When !on_rq, vruntime of the task has usually NOT been normalized.
10089 * But there are some cases where it has already been normalized:
10091 * - A forked child which is waiting for being woken up by
10092 * wake_up_new_task().
10093 * - A task which has been woken up by try_to_wake_up() and
10094 * waiting for actually being woken up by sched_ttwu_pending().
10096 if (!se->sum_exec_runtime ||
10097 (p->state == TASK_WAKING && p->sched_remote_wakeup))
10103 #ifdef CONFIG_FAIR_GROUP_SCHED
10105 * Propagate the changes of the sched_entity across the tg tree to make it
10106 * visible to the root
10108 static void propagate_entity_cfs_rq(struct sched_entity *se)
10110 struct cfs_rq *cfs_rq;
10112 /* Start to propagate at parent */
10115 for_each_sched_entity(se) {
10116 cfs_rq = cfs_rq_of(se);
10118 if (cfs_rq_throttled(cfs_rq))
10121 update_load_avg(cfs_rq, se, UPDATE_TG);
10125 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
10128 static void detach_entity_cfs_rq(struct sched_entity *se)
10130 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10132 /* Catch up with the cfs_rq and remove our load when we leave */
10133 update_load_avg(cfs_rq, se, 0);
10134 detach_entity_load_avg(cfs_rq, se);
10135 update_tg_load_avg(cfs_rq, false);
10136 propagate_entity_cfs_rq(se);
10139 static void attach_entity_cfs_rq(struct sched_entity *se)
10141 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10143 #ifdef CONFIG_FAIR_GROUP_SCHED
10145 * Since the real-depth could have been changed (only FAIR
10146 * class maintain depth value), reset depth properly.
10148 se->depth = se->parent ? se->parent->depth + 1 : 0;
10151 /* Synchronize entity with its cfs_rq */
10152 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
10153 attach_entity_load_avg(cfs_rq, se, 0);
10154 update_tg_load_avg(cfs_rq, false);
10155 propagate_entity_cfs_rq(se);
10158 static void detach_task_cfs_rq(struct task_struct *p)
10160 struct sched_entity *se = &p->se;
10161 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10163 if (!vruntime_normalized(p)) {
10165 * Fix up our vruntime so that the current sleep doesn't
10166 * cause 'unlimited' sleep bonus.
10168 place_entity(cfs_rq, se, 0);
10169 se->vruntime -= cfs_rq->min_vruntime;
10172 detach_entity_cfs_rq(se);
10175 static void attach_task_cfs_rq(struct task_struct *p)
10177 struct sched_entity *se = &p->se;
10178 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10180 attach_entity_cfs_rq(se);
10182 if (!vruntime_normalized(p))
10183 se->vruntime += cfs_rq->min_vruntime;
10186 static void switched_from_fair(struct rq *rq, struct task_struct *p)
10188 detach_task_cfs_rq(p);
10191 static void switched_to_fair(struct rq *rq, struct task_struct *p)
10193 attach_task_cfs_rq(p);
10195 if (task_on_rq_queued(p)) {
10197 * We were most likely switched from sched_rt, so
10198 * kick off the schedule if running, otherwise just see
10199 * if we can still preempt the current task.
10204 check_preempt_curr(rq, p, 0);
10208 /* Account for a task changing its policy or group.
10210 * This routine is mostly called to set cfs_rq->curr field when a task
10211 * migrates between groups/classes.
10213 static void set_curr_task_fair(struct rq *rq)
10215 struct sched_entity *se = &rq->curr->se;
10217 for_each_sched_entity(se) {
10218 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10220 set_next_entity(cfs_rq, se);
10221 /* ensure bandwidth has been allocated on our new cfs_rq */
10222 account_cfs_rq_runtime(cfs_rq, 0);
10226 void init_cfs_rq(struct cfs_rq *cfs_rq)
10228 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
10229 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
10230 #ifndef CONFIG_64BIT
10231 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
10234 raw_spin_lock_init(&cfs_rq->removed.lock);
10238 #ifdef CONFIG_FAIR_GROUP_SCHED
10239 static void task_set_group_fair(struct task_struct *p)
10241 struct sched_entity *se = &p->se;
10243 set_task_rq(p, task_cpu(p));
10244 se->depth = se->parent ? se->parent->depth + 1 : 0;
10247 static void task_move_group_fair(struct task_struct *p)
10249 detach_task_cfs_rq(p);
10250 set_task_rq(p, task_cpu(p));
10253 /* Tell se's cfs_rq has been changed -- migrated */
10254 p->se.avg.last_update_time = 0;
10256 attach_task_cfs_rq(p);
10259 static void task_change_group_fair(struct task_struct *p, int type)
10262 case TASK_SET_GROUP:
10263 task_set_group_fair(p);
10266 case TASK_MOVE_GROUP:
10267 task_move_group_fair(p);
10272 void free_fair_sched_group(struct task_group *tg)
10276 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
10278 for_each_possible_cpu(i) {
10280 kfree(tg->cfs_rq[i]);
10289 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10291 struct sched_entity *se;
10292 struct cfs_rq *cfs_rq;
10295 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
10298 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
10302 tg->shares = NICE_0_LOAD;
10304 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
10306 for_each_possible_cpu(i) {
10307 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
10308 GFP_KERNEL, cpu_to_node(i));
10312 se = kzalloc_node(sizeof(struct sched_entity),
10313 GFP_KERNEL, cpu_to_node(i));
10317 init_cfs_rq(cfs_rq);
10318 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
10319 init_entity_runnable_average(se);
10330 void online_fair_sched_group(struct task_group *tg)
10332 struct sched_entity *se;
10336 for_each_possible_cpu(i) {
10340 raw_spin_lock_irq(&rq->lock);
10341 update_rq_clock(rq);
10342 attach_entity_cfs_rq(se);
10343 sync_throttle(tg, i);
10344 raw_spin_unlock_irq(&rq->lock);
10348 void unregister_fair_sched_group(struct task_group *tg)
10350 unsigned long flags;
10354 for_each_possible_cpu(cpu) {
10356 remove_entity_load_avg(tg->se[cpu]);
10359 * Only empty task groups can be destroyed; so we can speculatively
10360 * check on_list without danger of it being re-added.
10362 if (!tg->cfs_rq[cpu]->on_list)
10367 raw_spin_lock_irqsave(&rq->lock, flags);
10368 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
10369 raw_spin_unlock_irqrestore(&rq->lock, flags);
10373 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
10374 struct sched_entity *se, int cpu,
10375 struct sched_entity *parent)
10377 struct rq *rq = cpu_rq(cpu);
10381 init_cfs_rq_runtime(cfs_rq);
10383 tg->cfs_rq[cpu] = cfs_rq;
10386 /* se could be NULL for root_task_group */
10391 se->cfs_rq = &rq->cfs;
10394 se->cfs_rq = parent->my_q;
10395 se->depth = parent->depth + 1;
10399 /* guarantee group entities always have weight */
10400 update_load_set(&se->load, NICE_0_LOAD);
10401 se->parent = parent;
10404 static DEFINE_MUTEX(shares_mutex);
10406 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10411 * We can't change the weight of the root cgroup.
10416 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10418 mutex_lock(&shares_mutex);
10419 if (tg->shares == shares)
10422 tg->shares = shares;
10423 for_each_possible_cpu(i) {
10424 struct rq *rq = cpu_rq(i);
10425 struct sched_entity *se = tg->se[i];
10426 struct rq_flags rf;
10428 /* Propagate contribution to hierarchy */
10429 rq_lock_irqsave(rq, &rf);
10430 update_rq_clock(rq);
10431 for_each_sched_entity(se) {
10432 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
10433 update_cfs_group(se);
10435 rq_unlock_irqrestore(rq, &rf);
10439 mutex_unlock(&shares_mutex);
10442 #else /* CONFIG_FAIR_GROUP_SCHED */
10444 void free_fair_sched_group(struct task_group *tg) { }
10446 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10451 void online_fair_sched_group(struct task_group *tg) { }
10453 void unregister_fair_sched_group(struct task_group *tg) { }
10455 #endif /* CONFIG_FAIR_GROUP_SCHED */
10458 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10460 struct sched_entity *se = &task->se;
10461 unsigned int rr_interval = 0;
10464 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10467 if (rq->cfs.load.weight)
10468 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10470 return rr_interval;
10474 * All the scheduling class methods:
10476 const struct sched_class fair_sched_class = {
10477 .next = &idle_sched_class,
10478 .enqueue_task = enqueue_task_fair,
10479 .dequeue_task = dequeue_task_fair,
10480 .yield_task = yield_task_fair,
10481 .yield_to_task = yield_to_task_fair,
10483 .check_preempt_curr = check_preempt_wakeup,
10485 .pick_next_task = pick_next_task_fair,
10486 .put_prev_task = put_prev_task_fair,
10489 .select_task_rq = select_task_rq_fair,
10490 .migrate_task_rq = migrate_task_rq_fair,
10492 .rq_online = rq_online_fair,
10493 .rq_offline = rq_offline_fair,
10495 .task_dead = task_dead_fair,
10496 .set_cpus_allowed = set_cpus_allowed_common,
10499 .set_curr_task = set_curr_task_fair,
10500 .task_tick = task_tick_fair,
10501 .task_fork = task_fork_fair,
10503 .prio_changed = prio_changed_fair,
10504 .switched_from = switched_from_fair,
10505 .switched_to = switched_to_fair,
10507 .get_rr_interval = get_rr_interval_fair,
10509 .update_curr = update_curr_fair,
10511 #ifdef CONFIG_FAIR_GROUP_SCHED
10512 .task_change_group = task_change_group_fair,
10515 #ifdef CONFIG_UCLAMP_TASK
10516 .uclamp_enabled = 1,
10520 #ifdef CONFIG_SCHED_DEBUG
10521 void print_cfs_stats(struct seq_file *m, int cpu)
10523 struct cfs_rq *cfs_rq, *pos;
10526 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
10527 print_cfs_rq(m, cpu, cfs_rq);
10531 #ifdef CONFIG_NUMA_BALANCING
10532 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10535 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10536 struct numa_group *ng;
10539 ng = rcu_dereference(p->numa_group);
10540 for_each_online_node(node) {
10541 if (p->numa_faults) {
10542 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10543 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10546 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
10547 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
10549 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10553 #endif /* CONFIG_NUMA_BALANCING */
10554 #endif /* CONFIG_SCHED_DEBUG */
10556 __init void init_sched_fair_class(void)
10559 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10561 #ifdef CONFIG_NO_HZ_COMMON
10562 nohz.next_balance = jiffies;
10563 nohz.next_blocked = jiffies;
10564 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
10571 * Helper functions to facilitate extracting info from tracepoints.
10574 const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq)
10577 return cfs_rq ? &cfs_rq->avg : NULL;
10582 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg);
10584 char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len)
10588 strlcpy(str, "(null)", len);
10593 cfs_rq_tg_path(cfs_rq, str, len);
10596 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path);
10598 int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq)
10600 return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1;
10602 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu);
10604 const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq)
10607 return rq ? &rq->avg_rt : NULL;
10612 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt);
10614 const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq)
10617 return rq ? &rq->avg_dl : NULL;
10622 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl);
10624 const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq)
10626 #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ)
10627 return rq ? &rq->avg_irq : NULL;
10632 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq);
10634 int sched_trace_rq_cpu(struct rq *rq)
10636 return rq ? cpu_of(rq) : -1;
10638 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu);
10640 const struct cpumask *sched_trace_rd_span(struct root_domain *rd)
10643 return rd ? rd->span : NULL;
10648 EXPORT_SYMBOL_GPL(sched_trace_rd_span);