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.
103 #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024)
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 void attach_entity_cfs_rq(struct sched_entity *se);
755 * With new tasks being created, their initial util_avgs are extrapolated
756 * based on the cfs_rq's current util_avg:
758 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
760 * However, in many cases, the above util_avg does not give a desired
761 * value. Moreover, the sum of the util_avgs may be divergent, such
762 * as when the series is a harmonic series.
764 * To solve this problem, we also cap the util_avg of successive tasks to
765 * only 1/2 of the left utilization budget:
767 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
769 * where n denotes the nth task and cpu_scale the CPU capacity.
771 * For example, for a CPU with 1024 of capacity, a simplest series from
772 * the beginning would be like:
774 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
775 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
777 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
778 * if util_avg > util_avg_cap.
780 void post_init_entity_util_avg(struct task_struct *p)
782 struct sched_entity *se = &p->se;
783 struct cfs_rq *cfs_rq = cfs_rq_of(se);
784 struct sched_avg *sa = &se->avg;
785 long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
786 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
789 if (cfs_rq->avg.util_avg != 0) {
790 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
791 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
793 if (sa->util_avg > cap)
800 if (p->sched_class != &fair_sched_class) {
802 * For !fair tasks do:
804 update_cfs_rq_load_avg(now, cfs_rq);
805 attach_entity_load_avg(cfs_rq, se, 0);
806 switched_from_fair(rq, p);
808 * such that the next switched_to_fair() has the
811 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
815 attach_entity_cfs_rq(se);
818 #else /* !CONFIG_SMP */
819 void init_entity_runnable_average(struct sched_entity *se)
822 void post_init_entity_util_avg(struct task_struct *p)
825 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
828 #endif /* CONFIG_SMP */
831 * Update the current task's runtime statistics.
833 static void update_curr(struct cfs_rq *cfs_rq)
835 struct sched_entity *curr = cfs_rq->curr;
836 u64 now = rq_clock_task(rq_of(cfs_rq));
842 delta_exec = now - curr->exec_start;
843 if (unlikely((s64)delta_exec <= 0))
846 curr->exec_start = now;
848 schedstat_set(curr->statistics.exec_max,
849 max(delta_exec, curr->statistics.exec_max));
851 curr->sum_exec_runtime += delta_exec;
852 schedstat_add(cfs_rq->exec_clock, delta_exec);
854 curr->vruntime += calc_delta_fair(delta_exec, curr);
855 update_min_vruntime(cfs_rq);
857 if (entity_is_task(curr)) {
858 struct task_struct *curtask = task_of(curr);
860 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
861 cgroup_account_cputime(curtask, delta_exec);
862 account_group_exec_runtime(curtask, delta_exec);
865 account_cfs_rq_runtime(cfs_rq, delta_exec);
868 static void update_curr_fair(struct rq *rq)
870 update_curr(cfs_rq_of(&rq->curr->se));
874 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
876 u64 wait_start, prev_wait_start;
878 if (!schedstat_enabled())
881 wait_start = rq_clock(rq_of(cfs_rq));
882 prev_wait_start = schedstat_val(se->statistics.wait_start);
884 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
885 likely(wait_start > prev_wait_start))
886 wait_start -= prev_wait_start;
888 __schedstat_set(se->statistics.wait_start, wait_start);
892 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
894 struct task_struct *p;
897 if (!schedstat_enabled())
900 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
902 if (entity_is_task(se)) {
904 if (task_on_rq_migrating(p)) {
906 * Preserve migrating task's wait time so wait_start
907 * time stamp can be adjusted to accumulate wait time
908 * prior to migration.
910 __schedstat_set(se->statistics.wait_start, delta);
913 trace_sched_stat_wait(p, delta);
916 __schedstat_set(se->statistics.wait_max,
917 max(schedstat_val(se->statistics.wait_max), delta));
918 __schedstat_inc(se->statistics.wait_count);
919 __schedstat_add(se->statistics.wait_sum, delta);
920 __schedstat_set(se->statistics.wait_start, 0);
924 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
926 struct task_struct *tsk = NULL;
927 u64 sleep_start, block_start;
929 if (!schedstat_enabled())
932 sleep_start = schedstat_val(se->statistics.sleep_start);
933 block_start = schedstat_val(se->statistics.block_start);
935 if (entity_is_task(se))
939 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
944 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
945 __schedstat_set(se->statistics.sleep_max, delta);
947 __schedstat_set(se->statistics.sleep_start, 0);
948 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
951 account_scheduler_latency(tsk, delta >> 10, 1);
952 trace_sched_stat_sleep(tsk, delta);
956 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
961 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
962 __schedstat_set(se->statistics.block_max, delta);
964 __schedstat_set(se->statistics.block_start, 0);
965 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
968 if (tsk->in_iowait) {
969 __schedstat_add(se->statistics.iowait_sum, delta);
970 __schedstat_inc(se->statistics.iowait_count);
971 trace_sched_stat_iowait(tsk, delta);
974 trace_sched_stat_blocked(tsk, delta);
977 * Blocking time is in units of nanosecs, so shift by
978 * 20 to get a milliseconds-range estimation of the
979 * amount of time that the task spent sleeping:
981 if (unlikely(prof_on == SLEEP_PROFILING)) {
982 profile_hits(SLEEP_PROFILING,
983 (void *)get_wchan(tsk),
986 account_scheduler_latency(tsk, delta >> 10, 0);
992 * Task is being enqueued - update stats:
995 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
997 if (!schedstat_enabled())
1001 * Are we enqueueing a waiting task? (for current tasks
1002 * a dequeue/enqueue event is a NOP)
1004 if (se != cfs_rq->curr)
1005 update_stats_wait_start(cfs_rq, se);
1007 if (flags & ENQUEUE_WAKEUP)
1008 update_stats_enqueue_sleeper(cfs_rq, se);
1012 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1015 if (!schedstat_enabled())
1019 * Mark the end of the wait period if dequeueing a
1022 if (se != cfs_rq->curr)
1023 update_stats_wait_end(cfs_rq, se);
1025 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1026 struct task_struct *tsk = task_of(se);
1028 if (tsk->state & TASK_INTERRUPTIBLE)
1029 __schedstat_set(se->statistics.sleep_start,
1030 rq_clock(rq_of(cfs_rq)));
1031 if (tsk->state & TASK_UNINTERRUPTIBLE)
1032 __schedstat_set(se->statistics.block_start,
1033 rq_clock(rq_of(cfs_rq)));
1038 * We are picking a new current task - update its stats:
1041 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1044 * We are starting a new run period:
1046 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1049 /**************************************************
1050 * Scheduling class queueing methods:
1053 #ifdef CONFIG_NUMA_BALANCING
1055 * Approximate time to scan a full NUMA task in ms. The task scan period is
1056 * calculated based on the tasks virtual memory size and
1057 * numa_balancing_scan_size.
1059 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1060 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1062 /* Portion of address space to scan in MB */
1063 unsigned int sysctl_numa_balancing_scan_size = 256;
1065 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1066 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1069 refcount_t refcount;
1071 spinlock_t lock; /* nr_tasks, tasks */
1076 struct rcu_head rcu;
1077 unsigned long total_faults;
1078 unsigned long max_faults_cpu;
1080 * Faults_cpu is used to decide whether memory should move
1081 * towards the CPU. As a consequence, these stats are weighted
1082 * more by CPU use than by memory faults.
1084 unsigned long *faults_cpu;
1085 unsigned long faults[0];
1089 * For functions that can be called in multiple contexts that permit reading
1090 * ->numa_group (see struct task_struct for locking rules).
1092 static struct numa_group *deref_task_numa_group(struct task_struct *p)
1094 return rcu_dereference_check(p->numa_group, p == current ||
1095 (lockdep_is_held(&task_rq(p)->lock) && !READ_ONCE(p->on_cpu)));
1098 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1100 return rcu_dereference_protected(p->numa_group, p == current);
1103 static inline unsigned long group_faults_priv(struct numa_group *ng);
1104 static inline unsigned long group_faults_shared(struct numa_group *ng);
1106 static unsigned int task_nr_scan_windows(struct task_struct *p)
1108 unsigned long rss = 0;
1109 unsigned long nr_scan_pages;
1112 * Calculations based on RSS as non-present and empty pages are skipped
1113 * by the PTE scanner and NUMA hinting faults should be trapped based
1116 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1117 rss = get_mm_rss(p->mm);
1119 rss = nr_scan_pages;
1121 rss = round_up(rss, nr_scan_pages);
1122 return rss / nr_scan_pages;
1125 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1126 #define MAX_SCAN_WINDOW 2560
1128 static unsigned int task_scan_min(struct task_struct *p)
1130 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1131 unsigned int scan, floor;
1132 unsigned int windows = 1;
1134 if (scan_size < MAX_SCAN_WINDOW)
1135 windows = MAX_SCAN_WINDOW / scan_size;
1136 floor = 1000 / windows;
1138 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1139 return max_t(unsigned int, floor, scan);
1142 static unsigned int task_scan_start(struct task_struct *p)
1144 unsigned long smin = task_scan_min(p);
1145 unsigned long period = smin;
1146 struct numa_group *ng;
1148 /* Scale the maximum scan period with the amount of shared memory. */
1150 ng = rcu_dereference(p->numa_group);
1152 unsigned long shared = group_faults_shared(ng);
1153 unsigned long private = group_faults_priv(ng);
1155 period *= refcount_read(&ng->refcount);
1156 period *= shared + 1;
1157 period /= private + shared + 1;
1161 return max(smin, period);
1164 static unsigned int task_scan_max(struct task_struct *p)
1166 unsigned long smin = task_scan_min(p);
1168 struct numa_group *ng;
1170 /* Watch for min being lower than max due to floor calculations */
1171 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1173 /* Scale the maximum scan period with the amount of shared memory. */
1174 ng = deref_curr_numa_group(p);
1176 unsigned long shared = group_faults_shared(ng);
1177 unsigned long private = group_faults_priv(ng);
1178 unsigned long period = smax;
1180 period *= refcount_read(&ng->refcount);
1181 period *= shared + 1;
1182 period /= private + shared + 1;
1184 smax = max(smax, period);
1187 return max(smin, smax);
1190 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1192 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1193 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1196 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1198 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1199 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1202 /* Shared or private faults. */
1203 #define NR_NUMA_HINT_FAULT_TYPES 2
1205 /* Memory and CPU locality */
1206 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1208 /* Averaged statistics, and temporary buffers. */
1209 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1211 pid_t task_numa_group_id(struct task_struct *p)
1213 struct numa_group *ng;
1217 ng = rcu_dereference(p->numa_group);
1226 * The averaged statistics, shared & private, memory & CPU,
1227 * occupy the first half of the array. The second half of the
1228 * array is for current counters, which are averaged into the
1229 * first set by task_numa_placement.
1231 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1233 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1236 static inline unsigned long task_faults(struct task_struct *p, int nid)
1238 if (!p->numa_faults)
1241 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1242 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1245 static inline unsigned long group_faults(struct task_struct *p, int nid)
1247 struct numa_group *ng = deref_task_numa_group(p);
1252 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1253 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1256 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1258 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1259 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1262 static inline unsigned long group_faults_priv(struct numa_group *ng)
1264 unsigned long faults = 0;
1267 for_each_online_node(node) {
1268 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1274 static inline unsigned long group_faults_shared(struct numa_group *ng)
1276 unsigned long faults = 0;
1279 for_each_online_node(node) {
1280 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1287 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1288 * considered part of a numa group's pseudo-interleaving set. Migrations
1289 * between these nodes are slowed down, to allow things to settle down.
1291 #define ACTIVE_NODE_FRACTION 3
1293 static bool numa_is_active_node(int nid, struct numa_group *ng)
1295 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1298 /* Handle placement on systems where not all nodes are directly connected. */
1299 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1300 int maxdist, bool task)
1302 unsigned long score = 0;
1306 * All nodes are directly connected, and the same distance
1307 * from each other. No need for fancy placement algorithms.
1309 if (sched_numa_topology_type == NUMA_DIRECT)
1313 * This code is called for each node, introducing N^2 complexity,
1314 * which should be ok given the number of nodes rarely exceeds 8.
1316 for_each_online_node(node) {
1317 unsigned long faults;
1318 int dist = node_distance(nid, node);
1321 * The furthest away nodes in the system are not interesting
1322 * for placement; nid was already counted.
1324 if (dist == sched_max_numa_distance || node == nid)
1328 * On systems with a backplane NUMA topology, compare groups
1329 * of nodes, and move tasks towards the group with the most
1330 * memory accesses. When comparing two nodes at distance
1331 * "hoplimit", only nodes closer by than "hoplimit" are part
1332 * of each group. Skip other nodes.
1334 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1338 /* Add up the faults from nearby nodes. */
1340 faults = task_faults(p, node);
1342 faults = group_faults(p, node);
1345 * On systems with a glueless mesh NUMA topology, there are
1346 * no fixed "groups of nodes". Instead, nodes that are not
1347 * directly connected bounce traffic through intermediate
1348 * nodes; a numa_group can occupy any set of nodes.
1349 * The further away a node is, the less the faults count.
1350 * This seems to result in good task placement.
1352 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1353 faults *= (sched_max_numa_distance - dist);
1354 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1364 * These return the fraction of accesses done by a particular task, or
1365 * task group, on a particular numa node. The group weight is given a
1366 * larger multiplier, in order to group tasks together that are almost
1367 * evenly spread out between numa nodes.
1369 static inline unsigned long task_weight(struct task_struct *p, int nid,
1372 unsigned long faults, total_faults;
1374 if (!p->numa_faults)
1377 total_faults = p->total_numa_faults;
1382 faults = task_faults(p, nid);
1383 faults += score_nearby_nodes(p, nid, dist, true);
1385 return 1000 * faults / total_faults;
1388 static inline unsigned long group_weight(struct task_struct *p, int nid,
1391 struct numa_group *ng = deref_task_numa_group(p);
1392 unsigned long faults, total_faults;
1397 total_faults = ng->total_faults;
1402 faults = group_faults(p, nid);
1403 faults += score_nearby_nodes(p, nid, dist, false);
1405 return 1000 * faults / total_faults;
1408 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1409 int src_nid, int dst_cpu)
1411 struct numa_group *ng = deref_curr_numa_group(p);
1412 int dst_nid = cpu_to_node(dst_cpu);
1413 int last_cpupid, this_cpupid;
1415 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1416 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1419 * Allow first faults or private faults to migrate immediately early in
1420 * the lifetime of a task. The magic number 4 is based on waiting for
1421 * two full passes of the "multi-stage node selection" test that is
1424 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1425 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1429 * Multi-stage node selection is used in conjunction with a periodic
1430 * migration fault to build a temporal task<->page relation. By using
1431 * a two-stage filter we remove short/unlikely relations.
1433 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1434 * a task's usage of a particular page (n_p) per total usage of this
1435 * page (n_t) (in a given time-span) to a probability.
1437 * Our periodic faults will sample this probability and getting the
1438 * same result twice in a row, given these samples are fully
1439 * independent, is then given by P(n)^2, provided our sample period
1440 * is sufficiently short compared to the usage pattern.
1442 * This quadric squishes small probabilities, making it less likely we
1443 * act on an unlikely task<->page relation.
1445 if (!cpupid_pid_unset(last_cpupid) &&
1446 cpupid_to_nid(last_cpupid) != dst_nid)
1449 /* Always allow migrate on private faults */
1450 if (cpupid_match_pid(p, last_cpupid))
1453 /* A shared fault, but p->numa_group has not been set up yet. */
1458 * Destination node is much more heavily used than the source
1459 * node? Allow migration.
1461 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1462 ACTIVE_NODE_FRACTION)
1466 * Distribute memory according to CPU & memory use on each node,
1467 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1469 * faults_cpu(dst) 3 faults_cpu(src)
1470 * --------------- * - > ---------------
1471 * faults_mem(dst) 4 faults_mem(src)
1473 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1474 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1477 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq);
1479 static unsigned long cpu_runnable_load(struct rq *rq)
1481 return cfs_rq_runnable_load_avg(&rq->cfs);
1484 /* Cached statistics for all CPUs within a node */
1488 /* Total compute capacity of CPUs on a node */
1489 unsigned long compute_capacity;
1493 * XXX borrowed from update_sg_lb_stats
1495 static void update_numa_stats(struct numa_stats *ns, int nid)
1499 memset(ns, 0, sizeof(*ns));
1500 for_each_cpu(cpu, cpumask_of_node(nid)) {
1501 struct rq *rq = cpu_rq(cpu);
1503 ns->load += cpu_runnable_load(rq);
1504 ns->compute_capacity += capacity_of(cpu);
1509 struct task_numa_env {
1510 struct task_struct *p;
1512 int src_cpu, src_nid;
1513 int dst_cpu, dst_nid;
1515 struct numa_stats src_stats, dst_stats;
1520 struct task_struct *best_task;
1525 static void task_numa_assign(struct task_numa_env *env,
1526 struct task_struct *p, long imp)
1528 struct rq *rq = cpu_rq(env->dst_cpu);
1530 /* Bail out if run-queue part of active NUMA balance. */
1531 if (xchg(&rq->numa_migrate_on, 1))
1535 * Clear previous best_cpu/rq numa-migrate flag, since task now
1536 * found a better CPU to move/swap.
1538 if (env->best_cpu != -1) {
1539 rq = cpu_rq(env->best_cpu);
1540 WRITE_ONCE(rq->numa_migrate_on, 0);
1544 put_task_struct(env->best_task);
1549 env->best_imp = imp;
1550 env->best_cpu = env->dst_cpu;
1553 static bool load_too_imbalanced(long src_load, long dst_load,
1554 struct task_numa_env *env)
1557 long orig_src_load, orig_dst_load;
1558 long src_capacity, dst_capacity;
1561 * The load is corrected for the CPU capacity available on each node.
1564 * ------------ vs ---------
1565 * src_capacity dst_capacity
1567 src_capacity = env->src_stats.compute_capacity;
1568 dst_capacity = env->dst_stats.compute_capacity;
1570 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1572 orig_src_load = env->src_stats.load;
1573 orig_dst_load = env->dst_stats.load;
1575 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1577 /* Would this change make things worse? */
1578 return (imb > old_imb);
1582 * Maximum NUMA importance can be 1998 (2*999);
1583 * SMALLIMP @ 30 would be close to 1998/64.
1584 * Used to deter task migration.
1589 * This checks if the overall compute and NUMA accesses of the system would
1590 * be improved if the source tasks was migrated to the target dst_cpu taking
1591 * into account that it might be best if task running on the dst_cpu should
1592 * be exchanged with the source task
1594 static void task_numa_compare(struct task_numa_env *env,
1595 long taskimp, long groupimp, bool maymove)
1597 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
1598 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1599 long imp = p_ng ? groupimp : taskimp;
1600 struct task_struct *cur;
1601 long src_load, dst_load;
1602 int dist = env->dist;
1606 if (READ_ONCE(dst_rq->numa_migrate_on))
1610 cur = rcu_dereference(dst_rq->curr);
1611 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1615 * Because we have preemption enabled we can get migrated around and
1616 * end try selecting ourselves (current == env->p) as a swap candidate.
1622 if (maymove && moveimp >= env->best_imp)
1629 * "imp" is the fault differential for the source task between the
1630 * source and destination node. Calculate the total differential for
1631 * the source task and potential destination task. The more negative
1632 * the value is, the more remote accesses that would be expected to
1633 * be incurred if the tasks were swapped.
1635 /* Skip this swap candidate if cannot move to the source cpu */
1636 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
1640 * If dst and source tasks are in the same NUMA group, or not
1641 * in any group then look only at task weights.
1643 cur_ng = rcu_dereference(cur->numa_group);
1644 if (cur_ng == p_ng) {
1645 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1646 task_weight(cur, env->dst_nid, dist);
1648 * Add some hysteresis to prevent swapping the
1649 * tasks within a group over tiny differences.
1655 * Compare the group weights. If a task is all by itself
1656 * (not part of a group), use the task weight instead.
1659 imp += group_weight(cur, env->src_nid, dist) -
1660 group_weight(cur, env->dst_nid, dist);
1662 imp += task_weight(cur, env->src_nid, dist) -
1663 task_weight(cur, env->dst_nid, dist);
1666 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1673 * If the NUMA importance is less than SMALLIMP,
1674 * task migration might only result in ping pong
1675 * of tasks and also hurt performance due to cache
1678 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1682 * In the overloaded case, try and keep the load balanced.
1684 load = task_h_load(env->p) - task_h_load(cur);
1688 dst_load = env->dst_stats.load + load;
1689 src_load = env->src_stats.load - load;
1691 if (load_too_imbalanced(src_load, dst_load, env))
1696 * One idle CPU per node is evaluated for a task numa move.
1697 * Call select_idle_sibling to maybe find a better one.
1701 * select_idle_siblings() uses an per-CPU cpumask that
1702 * can be used from IRQ context.
1704 local_irq_disable();
1705 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1710 task_numa_assign(env, cur, imp);
1715 static void task_numa_find_cpu(struct task_numa_env *env,
1716 long taskimp, long groupimp)
1718 long src_load, dst_load, load;
1719 bool maymove = false;
1722 load = task_h_load(env->p);
1723 dst_load = env->dst_stats.load + load;
1724 src_load = env->src_stats.load - load;
1727 * If the improvement from just moving env->p direction is better
1728 * than swapping tasks around, check if a move is possible.
1730 maymove = !load_too_imbalanced(src_load, dst_load, env);
1732 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1733 /* Skip this CPU if the source task cannot migrate */
1734 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
1738 task_numa_compare(env, taskimp, groupimp, maymove);
1742 static int task_numa_migrate(struct task_struct *p)
1744 struct task_numa_env env = {
1747 .src_cpu = task_cpu(p),
1748 .src_nid = task_node(p),
1750 .imbalance_pct = 112,
1756 unsigned long taskweight, groupweight;
1757 struct sched_domain *sd;
1758 long taskimp, groupimp;
1759 struct numa_group *ng;
1764 * Pick the lowest SD_NUMA domain, as that would have the smallest
1765 * imbalance and would be the first to start moving tasks about.
1767 * And we want to avoid any moving of tasks about, as that would create
1768 * random movement of tasks -- counter the numa conditions we're trying
1772 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1774 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1778 * Cpusets can break the scheduler domain tree into smaller
1779 * balance domains, some of which do not cross NUMA boundaries.
1780 * Tasks that are "trapped" in such domains cannot be migrated
1781 * elsewhere, so there is no point in (re)trying.
1783 if (unlikely(!sd)) {
1784 sched_setnuma(p, task_node(p));
1788 env.dst_nid = p->numa_preferred_nid;
1789 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1790 taskweight = task_weight(p, env.src_nid, dist);
1791 groupweight = group_weight(p, env.src_nid, dist);
1792 update_numa_stats(&env.src_stats, env.src_nid);
1793 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1794 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1795 update_numa_stats(&env.dst_stats, env.dst_nid);
1797 /* Try to find a spot on the preferred nid. */
1798 task_numa_find_cpu(&env, taskimp, groupimp);
1801 * Look at other nodes in these cases:
1802 * - there is no space available on the preferred_nid
1803 * - the task is part of a numa_group that is interleaved across
1804 * multiple NUMA nodes; in order to better consolidate the group,
1805 * we need to check other locations.
1807 ng = deref_curr_numa_group(p);
1808 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
1809 for_each_online_node(nid) {
1810 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1813 dist = node_distance(env.src_nid, env.dst_nid);
1814 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1816 taskweight = task_weight(p, env.src_nid, dist);
1817 groupweight = group_weight(p, env.src_nid, dist);
1820 /* Only consider nodes where both task and groups benefit */
1821 taskimp = task_weight(p, nid, dist) - taskweight;
1822 groupimp = group_weight(p, nid, dist) - groupweight;
1823 if (taskimp < 0 && groupimp < 0)
1828 update_numa_stats(&env.dst_stats, env.dst_nid);
1829 task_numa_find_cpu(&env, taskimp, groupimp);
1834 * If the task is part of a workload that spans multiple NUMA nodes,
1835 * and is migrating into one of the workload's active nodes, remember
1836 * this node as the task's preferred numa node, so the workload can
1838 * A task that migrated to a second choice node will be better off
1839 * trying for a better one later. Do not set the preferred node here.
1842 if (env.best_cpu == -1)
1845 nid = cpu_to_node(env.best_cpu);
1847 if (nid != p->numa_preferred_nid)
1848 sched_setnuma(p, nid);
1851 /* No better CPU than the current one was found. */
1852 if (env.best_cpu == -1)
1855 best_rq = cpu_rq(env.best_cpu);
1856 if (env.best_task == NULL) {
1857 ret = migrate_task_to(p, env.best_cpu);
1858 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1860 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1864 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
1865 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1868 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1869 put_task_struct(env.best_task);
1873 /* Attempt to migrate a task to a CPU on the preferred node. */
1874 static void numa_migrate_preferred(struct task_struct *p)
1876 unsigned long interval = HZ;
1878 /* This task has no NUMA fault statistics yet */
1879 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
1882 /* Periodically retry migrating the task to the preferred node */
1883 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1884 p->numa_migrate_retry = jiffies + interval;
1886 /* Success if task is already running on preferred CPU */
1887 if (task_node(p) == p->numa_preferred_nid)
1890 /* Otherwise, try migrate to a CPU on the preferred node */
1891 task_numa_migrate(p);
1895 * Find out how many nodes on the workload is actively running on. Do this by
1896 * tracking the nodes from which NUMA hinting faults are triggered. This can
1897 * be different from the set of nodes where the workload's memory is currently
1900 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1902 unsigned long faults, max_faults = 0;
1903 int nid, active_nodes = 0;
1905 for_each_online_node(nid) {
1906 faults = group_faults_cpu(numa_group, nid);
1907 if (faults > max_faults)
1908 max_faults = faults;
1911 for_each_online_node(nid) {
1912 faults = group_faults_cpu(numa_group, nid);
1913 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1917 numa_group->max_faults_cpu = max_faults;
1918 numa_group->active_nodes = active_nodes;
1922 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1923 * increments. The more local the fault statistics are, the higher the scan
1924 * period will be for the next scan window. If local/(local+remote) ratio is
1925 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1926 * the scan period will decrease. Aim for 70% local accesses.
1928 #define NUMA_PERIOD_SLOTS 10
1929 #define NUMA_PERIOD_THRESHOLD 7
1932 * Increase the scan period (slow down scanning) if the majority of
1933 * our memory is already on our local node, or if the majority of
1934 * the page accesses are shared with other processes.
1935 * Otherwise, decrease the scan period.
1937 static void update_task_scan_period(struct task_struct *p,
1938 unsigned long shared, unsigned long private)
1940 unsigned int period_slot;
1941 int lr_ratio, ps_ratio;
1944 unsigned long remote = p->numa_faults_locality[0];
1945 unsigned long local = p->numa_faults_locality[1];
1948 * If there were no record hinting faults then either the task is
1949 * completely idle or all activity is areas that are not of interest
1950 * to automatic numa balancing. Related to that, if there were failed
1951 * migration then it implies we are migrating too quickly or the local
1952 * node is overloaded. In either case, scan slower
1954 if (local + shared == 0 || p->numa_faults_locality[2]) {
1955 p->numa_scan_period = min(p->numa_scan_period_max,
1956 p->numa_scan_period << 1);
1958 p->mm->numa_next_scan = jiffies +
1959 msecs_to_jiffies(p->numa_scan_period);
1965 * Prepare to scale scan period relative to the current period.
1966 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1967 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1968 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1970 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1971 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1972 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1974 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1976 * Most memory accesses are local. There is no need to
1977 * do fast NUMA scanning, since memory is already local.
1979 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
1982 diff = slot * period_slot;
1983 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
1985 * Most memory accesses are shared with other tasks.
1986 * There is no point in continuing fast NUMA scanning,
1987 * since other tasks may just move the memory elsewhere.
1989 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
1992 diff = slot * period_slot;
1995 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
1996 * yet they are not on the local NUMA node. Speed up
1997 * NUMA scanning to get the memory moved over.
1999 int ratio = max(lr_ratio, ps_ratio);
2000 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2003 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2004 task_scan_min(p), task_scan_max(p));
2005 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2009 * Get the fraction of time the task has been running since the last
2010 * NUMA placement cycle. The scheduler keeps similar statistics, but
2011 * decays those on a 32ms period, which is orders of magnitude off
2012 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2013 * stats only if the task is so new there are no NUMA statistics yet.
2015 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2017 u64 runtime, delta, now;
2018 /* Use the start of this time slice to avoid calculations. */
2019 now = p->se.exec_start;
2020 runtime = p->se.sum_exec_runtime;
2022 if (p->last_task_numa_placement) {
2023 delta = runtime - p->last_sum_exec_runtime;
2024 *period = now - p->last_task_numa_placement;
2026 /* Avoid time going backwards, prevent potential divide error: */
2027 if (unlikely((s64)*period < 0))
2030 delta = p->se.avg.load_sum;
2031 *period = LOAD_AVG_MAX;
2034 p->last_sum_exec_runtime = runtime;
2035 p->last_task_numa_placement = now;
2041 * Determine the preferred nid for a task in a numa_group. This needs to
2042 * be done in a way that produces consistent results with group_weight,
2043 * otherwise workloads might not converge.
2045 static int preferred_group_nid(struct task_struct *p, int nid)
2050 /* Direct connections between all NUMA nodes. */
2051 if (sched_numa_topology_type == NUMA_DIRECT)
2055 * On a system with glueless mesh NUMA topology, group_weight
2056 * scores nodes according to the number of NUMA hinting faults on
2057 * both the node itself, and on nearby nodes.
2059 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2060 unsigned long score, max_score = 0;
2061 int node, max_node = nid;
2063 dist = sched_max_numa_distance;
2065 for_each_online_node(node) {
2066 score = group_weight(p, node, dist);
2067 if (score > max_score) {
2076 * Finding the preferred nid in a system with NUMA backplane
2077 * interconnect topology is more involved. The goal is to locate
2078 * tasks from numa_groups near each other in the system, and
2079 * untangle workloads from different sides of the system. This requires
2080 * searching down the hierarchy of node groups, recursively searching
2081 * inside the highest scoring group of nodes. The nodemask tricks
2082 * keep the complexity of the search down.
2084 nodes = node_online_map;
2085 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2086 unsigned long max_faults = 0;
2087 nodemask_t max_group = NODE_MASK_NONE;
2090 /* Are there nodes at this distance from each other? */
2091 if (!find_numa_distance(dist))
2094 for_each_node_mask(a, nodes) {
2095 unsigned long faults = 0;
2096 nodemask_t this_group;
2097 nodes_clear(this_group);
2099 /* Sum group's NUMA faults; includes a==b case. */
2100 for_each_node_mask(b, nodes) {
2101 if (node_distance(a, b) < dist) {
2102 faults += group_faults(p, b);
2103 node_set(b, this_group);
2104 node_clear(b, nodes);
2108 /* Remember the top group. */
2109 if (faults > max_faults) {
2110 max_faults = faults;
2111 max_group = this_group;
2113 * subtle: at the smallest distance there is
2114 * just one node left in each "group", the
2115 * winner is the preferred nid.
2120 /* Next round, evaluate the nodes within max_group. */
2128 static void task_numa_placement(struct task_struct *p)
2130 int seq, nid, max_nid = NUMA_NO_NODE;
2131 unsigned long max_faults = 0;
2132 unsigned long fault_types[2] = { 0, 0 };
2133 unsigned long total_faults;
2134 u64 runtime, period;
2135 spinlock_t *group_lock = NULL;
2136 struct numa_group *ng;
2139 * The p->mm->numa_scan_seq field gets updated without
2140 * exclusive access. Use READ_ONCE() here to ensure
2141 * that the field is read in a single access:
2143 seq = READ_ONCE(p->mm->numa_scan_seq);
2144 if (p->numa_scan_seq == seq)
2146 p->numa_scan_seq = seq;
2147 p->numa_scan_period_max = task_scan_max(p);
2149 total_faults = p->numa_faults_locality[0] +
2150 p->numa_faults_locality[1];
2151 runtime = numa_get_avg_runtime(p, &period);
2153 /* If the task is part of a group prevent parallel updates to group stats */
2154 ng = deref_curr_numa_group(p);
2156 group_lock = &ng->lock;
2157 spin_lock_irq(group_lock);
2160 /* Find the node with the highest number of faults */
2161 for_each_online_node(nid) {
2162 /* Keep track of the offsets in numa_faults array */
2163 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2164 unsigned long faults = 0, group_faults = 0;
2167 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2168 long diff, f_diff, f_weight;
2170 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2171 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2172 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2173 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2175 /* Decay existing window, copy faults since last scan */
2176 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2177 fault_types[priv] += p->numa_faults[membuf_idx];
2178 p->numa_faults[membuf_idx] = 0;
2181 * Normalize the faults_from, so all tasks in a group
2182 * count according to CPU use, instead of by the raw
2183 * number of faults. Tasks with little runtime have
2184 * little over-all impact on throughput, and thus their
2185 * faults are less important.
2187 f_weight = div64_u64(runtime << 16, period + 1);
2188 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2190 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2191 p->numa_faults[cpubuf_idx] = 0;
2193 p->numa_faults[mem_idx] += diff;
2194 p->numa_faults[cpu_idx] += f_diff;
2195 faults += p->numa_faults[mem_idx];
2196 p->total_numa_faults += diff;
2199 * safe because we can only change our own group
2201 * mem_idx represents the offset for a given
2202 * nid and priv in a specific region because it
2203 * is at the beginning of the numa_faults array.
2205 ng->faults[mem_idx] += diff;
2206 ng->faults_cpu[mem_idx] += f_diff;
2207 ng->total_faults += diff;
2208 group_faults += ng->faults[mem_idx];
2213 if (faults > max_faults) {
2214 max_faults = faults;
2217 } else if (group_faults > max_faults) {
2218 max_faults = group_faults;
2224 numa_group_count_active_nodes(ng);
2225 spin_unlock_irq(group_lock);
2226 max_nid = preferred_group_nid(p, max_nid);
2230 /* Set the new preferred node */
2231 if (max_nid != p->numa_preferred_nid)
2232 sched_setnuma(p, max_nid);
2235 update_task_scan_period(p, fault_types[0], fault_types[1]);
2238 static inline int get_numa_group(struct numa_group *grp)
2240 return refcount_inc_not_zero(&grp->refcount);
2243 static inline void put_numa_group(struct numa_group *grp)
2245 if (refcount_dec_and_test(&grp->refcount))
2246 kfree_rcu(grp, rcu);
2249 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2252 struct numa_group *grp, *my_grp;
2253 struct task_struct *tsk;
2255 int cpu = cpupid_to_cpu(cpupid);
2258 if (unlikely(!deref_curr_numa_group(p))) {
2259 unsigned int size = sizeof(struct numa_group) +
2260 4*nr_node_ids*sizeof(unsigned long);
2262 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2266 refcount_set(&grp->refcount, 1);
2267 grp->active_nodes = 1;
2268 grp->max_faults_cpu = 0;
2269 spin_lock_init(&grp->lock);
2271 /* Second half of the array tracks nids where faults happen */
2272 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2275 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2276 grp->faults[i] = p->numa_faults[i];
2278 grp->total_faults = p->total_numa_faults;
2281 rcu_assign_pointer(p->numa_group, grp);
2285 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2287 if (!cpupid_match_pid(tsk, cpupid))
2290 grp = rcu_dereference(tsk->numa_group);
2294 my_grp = deref_curr_numa_group(p);
2299 * Only join the other group if its bigger; if we're the bigger group,
2300 * the other task will join us.
2302 if (my_grp->nr_tasks > grp->nr_tasks)
2306 * Tie-break on the grp address.
2308 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2311 /* Always join threads in the same process. */
2312 if (tsk->mm == current->mm)
2315 /* Simple filter to avoid false positives due to PID collisions */
2316 if (flags & TNF_SHARED)
2319 /* Update priv based on whether false sharing was detected */
2322 if (join && !get_numa_group(grp))
2330 BUG_ON(irqs_disabled());
2331 double_lock_irq(&my_grp->lock, &grp->lock);
2333 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2334 my_grp->faults[i] -= p->numa_faults[i];
2335 grp->faults[i] += p->numa_faults[i];
2337 my_grp->total_faults -= p->total_numa_faults;
2338 grp->total_faults += p->total_numa_faults;
2343 spin_unlock(&my_grp->lock);
2344 spin_unlock_irq(&grp->lock);
2346 rcu_assign_pointer(p->numa_group, grp);
2348 put_numa_group(my_grp);
2357 * Get rid of NUMA staticstics associated with a task (either current or dead).
2358 * If @final is set, the task is dead and has reached refcount zero, so we can
2359 * safely free all relevant data structures. Otherwise, there might be
2360 * concurrent reads from places like load balancing and procfs, and we should
2361 * reset the data back to default state without freeing ->numa_faults.
2363 void task_numa_free(struct task_struct *p, bool final)
2365 /* safe: p either is current or is being freed by current */
2366 struct numa_group *grp = rcu_dereference_raw(p->numa_group);
2367 unsigned long *numa_faults = p->numa_faults;
2368 unsigned long flags;
2375 spin_lock_irqsave(&grp->lock, flags);
2376 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2377 grp->faults[i] -= p->numa_faults[i];
2378 grp->total_faults -= p->total_numa_faults;
2381 spin_unlock_irqrestore(&grp->lock, flags);
2382 RCU_INIT_POINTER(p->numa_group, NULL);
2383 put_numa_group(grp);
2387 p->numa_faults = NULL;
2390 p->total_numa_faults = 0;
2391 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2397 * Got a PROT_NONE fault for a page on @node.
2399 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2401 struct task_struct *p = current;
2402 bool migrated = flags & TNF_MIGRATED;
2403 int cpu_node = task_node(current);
2404 int local = !!(flags & TNF_FAULT_LOCAL);
2405 struct numa_group *ng;
2408 if (!static_branch_likely(&sched_numa_balancing))
2411 /* for example, ksmd faulting in a user's mm */
2415 /* Allocate buffer to track faults on a per-node basis */
2416 if (unlikely(!p->numa_faults)) {
2417 int size = sizeof(*p->numa_faults) *
2418 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2420 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2421 if (!p->numa_faults)
2424 p->total_numa_faults = 0;
2425 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2429 * First accesses are treated as private, otherwise consider accesses
2430 * to be private if the accessing pid has not changed
2432 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2435 priv = cpupid_match_pid(p, last_cpupid);
2436 if (!priv && !(flags & TNF_NO_GROUP))
2437 task_numa_group(p, last_cpupid, flags, &priv);
2441 * If a workload spans multiple NUMA nodes, a shared fault that
2442 * occurs wholly within the set of nodes that the workload is
2443 * actively using should be counted as local. This allows the
2444 * scan rate to slow down when a workload has settled down.
2446 ng = deref_curr_numa_group(p);
2447 if (!priv && !local && ng && ng->active_nodes > 1 &&
2448 numa_is_active_node(cpu_node, ng) &&
2449 numa_is_active_node(mem_node, ng))
2453 * Retry to migrate task to preferred node periodically, in case it
2454 * previously failed, or the scheduler moved us.
2456 if (time_after(jiffies, p->numa_migrate_retry)) {
2457 task_numa_placement(p);
2458 numa_migrate_preferred(p);
2462 p->numa_pages_migrated += pages;
2463 if (flags & TNF_MIGRATE_FAIL)
2464 p->numa_faults_locality[2] += pages;
2466 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2467 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2468 p->numa_faults_locality[local] += pages;
2471 static void reset_ptenuma_scan(struct task_struct *p)
2474 * We only did a read acquisition of the mmap sem, so
2475 * p->mm->numa_scan_seq is written to without exclusive access
2476 * and the update is not guaranteed to be atomic. That's not
2477 * much of an issue though, since this is just used for
2478 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2479 * expensive, to avoid any form of compiler optimizations:
2481 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2482 p->mm->numa_scan_offset = 0;
2486 * The expensive part of numa migration is done from task_work context.
2487 * Triggered from task_tick_numa().
2489 static void task_numa_work(struct callback_head *work)
2491 unsigned long migrate, next_scan, now = jiffies;
2492 struct task_struct *p = current;
2493 struct mm_struct *mm = p->mm;
2494 u64 runtime = p->se.sum_exec_runtime;
2495 struct vm_area_struct *vma;
2496 unsigned long start, end;
2497 unsigned long nr_pte_updates = 0;
2498 long pages, virtpages;
2500 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2504 * Who cares about NUMA placement when they're dying.
2506 * NOTE: make sure not to dereference p->mm before this check,
2507 * exit_task_work() happens _after_ exit_mm() so we could be called
2508 * without p->mm even though we still had it when we enqueued this
2511 if (p->flags & PF_EXITING)
2514 if (!mm->numa_next_scan) {
2515 mm->numa_next_scan = now +
2516 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2520 * Enforce maximal scan/migration frequency..
2522 migrate = mm->numa_next_scan;
2523 if (time_before(now, migrate))
2526 if (p->numa_scan_period == 0) {
2527 p->numa_scan_period_max = task_scan_max(p);
2528 p->numa_scan_period = task_scan_start(p);
2531 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2532 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2536 * Delay this task enough that another task of this mm will likely win
2537 * the next time around.
2539 p->node_stamp += 2 * TICK_NSEC;
2541 start = mm->numa_scan_offset;
2542 pages = sysctl_numa_balancing_scan_size;
2543 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2544 virtpages = pages * 8; /* Scan up to this much virtual space */
2549 if (!down_read_trylock(&mm->mmap_sem))
2551 vma = find_vma(mm, start);
2553 reset_ptenuma_scan(p);
2557 for (; vma; vma = vma->vm_next) {
2558 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2559 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2564 * Shared library pages mapped by multiple processes are not
2565 * migrated as it is expected they are cache replicated. Avoid
2566 * hinting faults in read-only file-backed mappings or the vdso
2567 * as migrating the pages will be of marginal benefit.
2570 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2574 * Skip inaccessible VMAs to avoid any confusion between
2575 * PROT_NONE and NUMA hinting ptes
2577 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2581 start = max(start, vma->vm_start);
2582 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2583 end = min(end, vma->vm_end);
2584 nr_pte_updates = change_prot_numa(vma, start, end);
2587 * Try to scan sysctl_numa_balancing_size worth of
2588 * hpages that have at least one present PTE that
2589 * is not already pte-numa. If the VMA contains
2590 * areas that are unused or already full of prot_numa
2591 * PTEs, scan up to virtpages, to skip through those
2595 pages -= (end - start) >> PAGE_SHIFT;
2596 virtpages -= (end - start) >> PAGE_SHIFT;
2599 if (pages <= 0 || virtpages <= 0)
2603 } while (end != vma->vm_end);
2608 * It is possible to reach the end of the VMA list but the last few
2609 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2610 * would find the !migratable VMA on the next scan but not reset the
2611 * scanner to the start so check it now.
2614 mm->numa_scan_offset = start;
2616 reset_ptenuma_scan(p);
2617 up_read(&mm->mmap_sem);
2620 * Make sure tasks use at least 32x as much time to run other code
2621 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2622 * Usually update_task_scan_period slows down scanning enough; on an
2623 * overloaded system we need to limit overhead on a per task basis.
2625 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2626 u64 diff = p->se.sum_exec_runtime - runtime;
2627 p->node_stamp += 32 * diff;
2631 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
2634 struct mm_struct *mm = p->mm;
2637 mm_users = atomic_read(&mm->mm_users);
2638 if (mm_users == 1) {
2639 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2640 mm->numa_scan_seq = 0;
2644 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
2645 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2646 /* Protect against double add, see task_tick_numa and task_numa_work */
2647 p->numa_work.next = &p->numa_work;
2648 p->numa_faults = NULL;
2649 RCU_INIT_POINTER(p->numa_group, NULL);
2650 p->last_task_numa_placement = 0;
2651 p->last_sum_exec_runtime = 0;
2653 init_task_work(&p->numa_work, task_numa_work);
2655 /* New address space, reset the preferred nid */
2656 if (!(clone_flags & CLONE_VM)) {
2657 p->numa_preferred_nid = NUMA_NO_NODE;
2662 * New thread, keep existing numa_preferred_nid which should be copied
2663 * already by arch_dup_task_struct but stagger when scans start.
2668 delay = min_t(unsigned int, task_scan_max(current),
2669 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
2670 delay += 2 * TICK_NSEC;
2671 p->node_stamp = delay;
2676 * Drive the periodic memory faults..
2678 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2680 struct callback_head *work = &curr->numa_work;
2684 * We don't care about NUMA placement if we don't have memory.
2686 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2690 * Using runtime rather than walltime has the dual advantage that
2691 * we (mostly) drive the selection from busy threads and that the
2692 * task needs to have done some actual work before we bother with
2695 now = curr->se.sum_exec_runtime;
2696 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2698 if (now > curr->node_stamp + period) {
2699 if (!curr->node_stamp)
2700 curr->numa_scan_period = task_scan_start(curr);
2701 curr->node_stamp += period;
2703 if (!time_before(jiffies, curr->mm->numa_next_scan))
2704 task_work_add(curr, work, true);
2708 static void update_scan_period(struct task_struct *p, int new_cpu)
2710 int src_nid = cpu_to_node(task_cpu(p));
2711 int dst_nid = cpu_to_node(new_cpu);
2713 if (!static_branch_likely(&sched_numa_balancing))
2716 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2719 if (src_nid == dst_nid)
2723 * Allow resets if faults have been trapped before one scan
2724 * has completed. This is most likely due to a new task that
2725 * is pulled cross-node due to wakeups or load balancing.
2727 if (p->numa_scan_seq) {
2729 * Avoid scan adjustments if moving to the preferred
2730 * node or if the task was not previously running on
2731 * the preferred node.
2733 if (dst_nid == p->numa_preferred_nid ||
2734 (p->numa_preferred_nid != NUMA_NO_NODE &&
2735 src_nid != p->numa_preferred_nid))
2739 p->numa_scan_period = task_scan_start(p);
2743 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2747 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2751 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2755 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2759 #endif /* CONFIG_NUMA_BALANCING */
2762 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2764 update_load_add(&cfs_rq->load, se->load.weight);
2766 if (entity_is_task(se)) {
2767 struct rq *rq = rq_of(cfs_rq);
2769 account_numa_enqueue(rq, task_of(se));
2770 list_add(&se->group_node, &rq->cfs_tasks);
2773 cfs_rq->nr_running++;
2777 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2779 update_load_sub(&cfs_rq->load, se->load.weight);
2781 if (entity_is_task(se)) {
2782 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2783 list_del_init(&se->group_node);
2786 cfs_rq->nr_running--;
2790 * Signed add and clamp on underflow.
2792 * Explicitly do a load-store to ensure the intermediate value never hits
2793 * memory. This allows lockless observations without ever seeing the negative
2796 #define add_positive(_ptr, _val) do { \
2797 typeof(_ptr) ptr = (_ptr); \
2798 typeof(_val) val = (_val); \
2799 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2803 if (val < 0 && res > var) \
2806 WRITE_ONCE(*ptr, res); \
2810 * Unsigned subtract and clamp on underflow.
2812 * Explicitly do a load-store to ensure the intermediate value never hits
2813 * memory. This allows lockless observations without ever seeing the negative
2816 #define sub_positive(_ptr, _val) do { \
2817 typeof(_ptr) ptr = (_ptr); \
2818 typeof(*ptr) val = (_val); \
2819 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2823 WRITE_ONCE(*ptr, res); \
2827 * Remove and clamp on negative, from a local variable.
2829 * A variant of sub_positive(), which does not use explicit load-store
2830 * and is thus optimized for local variable updates.
2832 #define lsub_positive(_ptr, _val) do { \
2833 typeof(_ptr) ptr = (_ptr); \
2834 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
2839 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2841 cfs_rq->runnable_weight += se->runnable_weight;
2843 cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2844 cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2848 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2850 cfs_rq->runnable_weight -= se->runnable_weight;
2852 sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2853 sub_positive(&cfs_rq->avg.runnable_load_sum,
2854 se_runnable(se) * se->avg.runnable_load_sum);
2858 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2860 cfs_rq->avg.load_avg += se->avg.load_avg;
2861 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2865 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2867 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2868 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2872 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2874 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2876 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2878 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2881 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2882 unsigned long weight, unsigned long runnable)
2885 /* commit outstanding execution time */
2886 if (cfs_rq->curr == se)
2887 update_curr(cfs_rq);
2888 account_entity_dequeue(cfs_rq, se);
2889 dequeue_runnable_load_avg(cfs_rq, se);
2891 dequeue_load_avg(cfs_rq, se);
2893 se->runnable_weight = runnable;
2894 update_load_set(&se->load, weight);
2898 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2900 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2901 se->avg.runnable_load_avg =
2902 div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2906 enqueue_load_avg(cfs_rq, se);
2908 account_entity_enqueue(cfs_rq, se);
2909 enqueue_runnable_load_avg(cfs_rq, se);
2913 void reweight_task(struct task_struct *p, int prio)
2915 struct sched_entity *se = &p->se;
2916 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2917 struct load_weight *load = &se->load;
2918 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2920 reweight_entity(cfs_rq, se, weight, weight);
2921 load->inv_weight = sched_prio_to_wmult[prio];
2924 #ifdef CONFIG_FAIR_GROUP_SCHED
2927 * All this does is approximate the hierarchical proportion which includes that
2928 * global sum we all love to hate.
2930 * That is, the weight of a group entity, is the proportional share of the
2931 * group weight based on the group runqueue weights. That is:
2933 * tg->weight * grq->load.weight
2934 * ge->load.weight = ----------------------------- (1)
2935 * \Sum grq->load.weight
2937 * Now, because computing that sum is prohibitively expensive to compute (been
2938 * there, done that) we approximate it with this average stuff. The average
2939 * moves slower and therefore the approximation is cheaper and more stable.
2941 * So instead of the above, we substitute:
2943 * grq->load.weight -> grq->avg.load_avg (2)
2945 * which yields the following:
2947 * tg->weight * grq->avg.load_avg
2948 * ge->load.weight = ------------------------------ (3)
2951 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2953 * That is shares_avg, and it is right (given the approximation (2)).
2955 * The problem with it is that because the average is slow -- it was designed
2956 * to be exactly that of course -- this leads to transients in boundary
2957 * conditions. In specific, the case where the group was idle and we start the
2958 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2959 * yielding bad latency etc..
2961 * Now, in that special case (1) reduces to:
2963 * tg->weight * grq->load.weight
2964 * ge->load.weight = ----------------------------- = tg->weight (4)
2967 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2969 * So what we do is modify our approximation (3) to approach (4) in the (near)
2974 * tg->weight * grq->load.weight
2975 * --------------------------------------------------- (5)
2976 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2978 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2979 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2982 * tg->weight * grq->load.weight
2983 * ge->load.weight = ----------------------------- (6)
2988 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2989 * max(grq->load.weight, grq->avg.load_avg)
2991 * And that is shares_weight and is icky. In the (near) UP case it approaches
2992 * (4) while in the normal case it approaches (3). It consistently
2993 * overestimates the ge->load.weight and therefore:
2995 * \Sum ge->load.weight >= tg->weight
2999 static long calc_group_shares(struct cfs_rq *cfs_rq)
3001 long tg_weight, tg_shares, load, shares;
3002 struct task_group *tg = cfs_rq->tg;
3004 tg_shares = READ_ONCE(tg->shares);
3006 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3008 tg_weight = atomic_long_read(&tg->load_avg);
3010 /* Ensure tg_weight >= load */
3011 tg_weight -= cfs_rq->tg_load_avg_contrib;
3014 shares = (tg_shares * load);
3016 shares /= tg_weight;
3019 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3020 * of a group with small tg->shares value. It is a floor value which is
3021 * assigned as a minimum load.weight to the sched_entity representing
3022 * the group on a CPU.
3024 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3025 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3026 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3027 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3030 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3034 * This calculates the effective runnable weight for a group entity based on
3035 * the group entity weight calculated above.
3037 * Because of the above approximation (2), our group entity weight is
3038 * an load_avg based ratio (3). This means that it includes blocked load and
3039 * does not represent the runnable weight.
3041 * Approximate the group entity's runnable weight per ratio from the group
3044 * grq->avg.runnable_load_avg
3045 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
3048 * However, analogous to above, since the avg numbers are slow, this leads to
3049 * transients in the from-idle case. Instead we use:
3051 * ge->runnable_weight = ge->load.weight *
3053 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
3054 * ----------------------------------------------------- (8)
3055 * max(grq->avg.load_avg, grq->load.weight)
3057 * Where these max() serve both to use the 'instant' values to fix the slow
3058 * from-idle and avoid the /0 on to-idle, similar to (6).
3060 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
3062 long runnable, load_avg;
3064 load_avg = max(cfs_rq->avg.load_avg,
3065 scale_load_down(cfs_rq->load.weight));
3067 runnable = max(cfs_rq->avg.runnable_load_avg,
3068 scale_load_down(cfs_rq->runnable_weight));
3072 runnable /= load_avg;
3074 return clamp_t(long, runnable, MIN_SHARES, shares);
3076 #endif /* CONFIG_SMP */
3078 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3081 * Recomputes the group entity based on the current state of its group
3084 static void update_cfs_group(struct sched_entity *se)
3086 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3087 long shares, runnable;
3092 if (throttled_hierarchy(gcfs_rq))
3096 runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
3098 if (likely(se->load.weight == shares))
3101 shares = calc_group_shares(gcfs_rq);
3102 runnable = calc_group_runnable(gcfs_rq, shares);
3105 reweight_entity(cfs_rq_of(se), se, shares, runnable);
3108 #else /* CONFIG_FAIR_GROUP_SCHED */
3109 static inline void update_cfs_group(struct sched_entity *se)
3112 #endif /* CONFIG_FAIR_GROUP_SCHED */
3114 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3116 struct rq *rq = rq_of(cfs_rq);
3118 if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) {
3120 * There are a few boundary cases this might miss but it should
3121 * get called often enough that that should (hopefully) not be
3124 * It will not get called when we go idle, because the idle
3125 * thread is a different class (!fair), nor will the utilization
3126 * number include things like RT tasks.
3128 * As is, the util number is not freq-invariant (we'd have to
3129 * implement arch_scale_freq_capacity() for that).
3133 cpufreq_update_util(rq, flags);
3138 #ifdef CONFIG_FAIR_GROUP_SCHED
3140 * update_tg_load_avg - update the tg's load avg
3141 * @cfs_rq: the cfs_rq whose avg changed
3142 * @force: update regardless of how small the difference
3144 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3145 * However, because tg->load_avg is a global value there are performance
3148 * In order to avoid having to look at the other cfs_rq's, we use a
3149 * differential update where we store the last value we propagated. This in
3150 * turn allows skipping updates if the differential is 'small'.
3152 * Updating tg's load_avg is necessary before update_cfs_share().
3154 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3156 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3159 * No need to update load_avg for root_task_group as it is not used.
3161 if (cfs_rq->tg == &root_task_group)
3164 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3165 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3166 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3171 * Called within set_task_rq() right before setting a task's CPU. The
3172 * caller only guarantees p->pi_lock is held; no other assumptions,
3173 * including the state of rq->lock, should be made.
3175 void set_task_rq_fair(struct sched_entity *se,
3176 struct cfs_rq *prev, struct cfs_rq *next)
3178 u64 p_last_update_time;
3179 u64 n_last_update_time;
3181 if (!sched_feat(ATTACH_AGE_LOAD))
3185 * We are supposed to update the task to "current" time, then its up to
3186 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3187 * getting what current time is, so simply throw away the out-of-date
3188 * time. This will result in the wakee task is less decayed, but giving
3189 * the wakee more load sounds not bad.
3191 if (!(se->avg.last_update_time && prev))
3194 #ifndef CONFIG_64BIT
3196 u64 p_last_update_time_copy;
3197 u64 n_last_update_time_copy;
3200 p_last_update_time_copy = prev->load_last_update_time_copy;
3201 n_last_update_time_copy = next->load_last_update_time_copy;
3205 p_last_update_time = prev->avg.last_update_time;
3206 n_last_update_time = next->avg.last_update_time;
3208 } while (p_last_update_time != p_last_update_time_copy ||
3209 n_last_update_time != n_last_update_time_copy);
3212 p_last_update_time = prev->avg.last_update_time;
3213 n_last_update_time = next->avg.last_update_time;
3215 __update_load_avg_blocked_se(p_last_update_time, se);
3216 se->avg.last_update_time = n_last_update_time;
3221 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3222 * propagate its contribution. The key to this propagation is the invariant
3223 * that for each group:
3225 * ge->avg == grq->avg (1)
3227 * _IFF_ we look at the pure running and runnable sums. Because they
3228 * represent the very same entity, just at different points in the hierarchy.
3230 * Per the above update_tg_cfs_util() is trivial and simply copies the running
3231 * sum over (but still wrong, because the group entity and group rq do not have
3232 * their PELT windows aligned).
3234 * However, update_tg_cfs_runnable() is more complex. So we have:
3236 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3238 * And since, like util, the runnable part should be directly transferable,
3239 * the following would _appear_ to be the straight forward approach:
3241 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3243 * And per (1) we have:
3245 * ge->avg.runnable_avg == grq->avg.runnable_avg
3249 * ge->load.weight * grq->avg.load_avg
3250 * ge->avg.load_avg = ----------------------------------- (4)
3253 * Except that is wrong!
3255 * Because while for entities historical weight is not important and we
3256 * really only care about our future and therefore can consider a pure
3257 * runnable sum, runqueues can NOT do this.
3259 * We specifically want runqueues to have a load_avg that includes
3260 * historical weights. Those represent the blocked load, the load we expect
3261 * to (shortly) return to us. This only works by keeping the weights as
3262 * integral part of the sum. We therefore cannot decompose as per (3).
3264 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3265 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3266 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3267 * runnable section of these tasks overlap (or not). If they were to perfectly
3268 * align the rq as a whole would be runnable 2/3 of the time. If however we
3269 * always have at least 1 runnable task, the rq as a whole is always runnable.
3271 * So we'll have to approximate.. :/
3273 * Given the constraint:
3275 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3277 * We can construct a rule that adds runnable to a rq by assuming minimal
3280 * On removal, we'll assume each task is equally runnable; which yields:
3282 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3284 * XXX: only do this for the part of runnable > running ?
3289 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3291 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3293 /* Nothing to update */
3298 * The relation between sum and avg is:
3300 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3302 * however, the PELT windows are not aligned between grq and gse.
3305 /* Set new sched_entity's utilization */
3306 se->avg.util_avg = gcfs_rq->avg.util_avg;
3307 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3309 /* Update parent cfs_rq utilization */
3310 add_positive(&cfs_rq->avg.util_avg, delta);
3311 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3315 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3317 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3318 unsigned long runnable_load_avg, load_avg;
3319 u64 runnable_load_sum, load_sum = 0;
3325 gcfs_rq->prop_runnable_sum = 0;
3327 if (runnable_sum >= 0) {
3329 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3330 * the CPU is saturated running == runnable.
3332 runnable_sum += se->avg.load_sum;
3333 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3336 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3337 * assuming all tasks are equally runnable.
3339 if (scale_load_down(gcfs_rq->load.weight)) {
3340 load_sum = div_s64(gcfs_rq->avg.load_sum,
3341 scale_load_down(gcfs_rq->load.weight));
3344 /* But make sure to not inflate se's runnable */
3345 runnable_sum = min(se->avg.load_sum, load_sum);
3349 * runnable_sum can't be lower than running_sum
3350 * Rescale running sum to be in the same range as runnable sum
3351 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3352 * runnable_sum is in [0 : LOAD_AVG_MAX]
3354 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3355 runnable_sum = max(runnable_sum, running_sum);
3357 load_sum = (s64)se_weight(se) * runnable_sum;
3358 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3360 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3361 delta_avg = load_avg - se->avg.load_avg;
3363 se->avg.load_sum = runnable_sum;
3364 se->avg.load_avg = load_avg;
3365 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3366 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3368 runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3369 runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3370 delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3371 delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3373 se->avg.runnable_load_sum = runnable_sum;
3374 se->avg.runnable_load_avg = runnable_load_avg;
3377 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3378 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3382 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3384 cfs_rq->propagate = 1;
3385 cfs_rq->prop_runnable_sum += runnable_sum;
3388 /* Update task and its cfs_rq load average */
3389 static inline int propagate_entity_load_avg(struct sched_entity *se)
3391 struct cfs_rq *cfs_rq, *gcfs_rq;
3393 if (entity_is_task(se))
3396 gcfs_rq = group_cfs_rq(se);
3397 if (!gcfs_rq->propagate)
3400 gcfs_rq->propagate = 0;
3402 cfs_rq = cfs_rq_of(se);
3404 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3406 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3407 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3409 trace_pelt_cfs_tp(cfs_rq);
3410 trace_pelt_se_tp(se);
3416 * Check if we need to update the load and the utilization of a blocked
3419 static inline bool skip_blocked_update(struct sched_entity *se)
3421 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3424 * If sched_entity still have not zero load or utilization, we have to
3427 if (se->avg.load_avg || se->avg.util_avg)
3431 * If there is a pending propagation, we have to update the load and
3432 * the utilization of the sched_entity:
3434 if (gcfs_rq->propagate)
3438 * Otherwise, the load and the utilization of the sched_entity is
3439 * already zero and there is no pending propagation, so it will be a
3440 * waste of time to try to decay it:
3445 #else /* CONFIG_FAIR_GROUP_SCHED */
3447 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3449 static inline int propagate_entity_load_avg(struct sched_entity *se)
3454 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3456 #endif /* CONFIG_FAIR_GROUP_SCHED */
3459 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3460 * @now: current time, as per cfs_rq_clock_pelt()
3461 * @cfs_rq: cfs_rq to update
3463 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3464 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3465 * post_init_entity_util_avg().
3467 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3469 * Returns true if the load decayed or we removed load.
3471 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3472 * call update_tg_load_avg() when this function returns true.
3475 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3477 unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3478 struct sched_avg *sa = &cfs_rq->avg;
3481 if (cfs_rq->removed.nr) {
3483 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3485 raw_spin_lock(&cfs_rq->removed.lock);
3486 swap(cfs_rq->removed.util_avg, removed_util);
3487 swap(cfs_rq->removed.load_avg, removed_load);
3488 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3489 cfs_rq->removed.nr = 0;
3490 raw_spin_unlock(&cfs_rq->removed.lock);
3493 sub_positive(&sa->load_avg, r);
3494 sub_positive(&sa->load_sum, r * divider);
3497 sub_positive(&sa->util_avg, r);
3498 sub_positive(&sa->util_sum, r * divider);
3500 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3505 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
3507 #ifndef CONFIG_64BIT
3509 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3513 cfs_rq_util_change(cfs_rq, 0);
3519 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3520 * @cfs_rq: cfs_rq to attach to
3521 * @se: sched_entity to attach
3522 * @flags: migration hints
3524 * Must call update_cfs_rq_load_avg() before this, since we rely on
3525 * cfs_rq->avg.last_update_time being current.
3527 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3529 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3532 * When we attach the @se to the @cfs_rq, we must align the decay
3533 * window because without that, really weird and wonderful things can
3538 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3539 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3542 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3543 * period_contrib. This isn't strictly correct, but since we're
3544 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3547 se->avg.util_sum = se->avg.util_avg * divider;
3549 se->avg.load_sum = divider;
3550 if (se_weight(se)) {
3552 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3555 se->avg.runnable_load_sum = se->avg.load_sum;
3557 enqueue_load_avg(cfs_rq, se);
3558 cfs_rq->avg.util_avg += se->avg.util_avg;
3559 cfs_rq->avg.util_sum += se->avg.util_sum;
3561 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3563 cfs_rq_util_change(cfs_rq, flags);
3565 trace_pelt_cfs_tp(cfs_rq);
3569 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3570 * @cfs_rq: cfs_rq to detach from
3571 * @se: sched_entity to detach
3573 * Must call update_cfs_rq_load_avg() before this, since we rely on
3574 * cfs_rq->avg.last_update_time being current.
3576 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3578 dequeue_load_avg(cfs_rq, se);
3579 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3580 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3582 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3584 cfs_rq_util_change(cfs_rq, 0);
3586 trace_pelt_cfs_tp(cfs_rq);
3590 * Optional action to be done while updating the load average
3592 #define UPDATE_TG 0x1
3593 #define SKIP_AGE_LOAD 0x2
3594 #define DO_ATTACH 0x4
3596 /* Update task and its cfs_rq load average */
3597 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3599 u64 now = cfs_rq_clock_pelt(cfs_rq);
3603 * Track task load average for carrying it to new CPU after migrated, and
3604 * track group sched_entity load average for task_h_load calc in migration
3606 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3607 __update_load_avg_se(now, cfs_rq, se);
3609 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3610 decayed |= propagate_entity_load_avg(se);
3612 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3615 * DO_ATTACH means we're here from enqueue_entity().
3616 * !last_update_time means we've passed through
3617 * migrate_task_rq_fair() indicating we migrated.
3619 * IOW we're enqueueing a task on a new CPU.
3621 attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
3622 update_tg_load_avg(cfs_rq, 0);
3624 } else if (decayed && (flags & UPDATE_TG))
3625 update_tg_load_avg(cfs_rq, 0);
3628 #ifndef CONFIG_64BIT
3629 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3631 u64 last_update_time_copy;
3632 u64 last_update_time;
3635 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3637 last_update_time = cfs_rq->avg.last_update_time;
3638 } while (last_update_time != last_update_time_copy);
3640 return last_update_time;
3643 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3645 return cfs_rq->avg.last_update_time;
3650 * Synchronize entity load avg of dequeued entity without locking
3653 static void sync_entity_load_avg(struct sched_entity *se)
3655 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3656 u64 last_update_time;
3658 last_update_time = cfs_rq_last_update_time(cfs_rq);
3659 __update_load_avg_blocked_se(last_update_time, se);
3663 * Task first catches up with cfs_rq, and then subtract
3664 * itself from the cfs_rq (task must be off the queue now).
3666 static void remove_entity_load_avg(struct sched_entity *se)
3668 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3669 unsigned long flags;
3672 * tasks cannot exit without having gone through wake_up_new_task() ->
3673 * post_init_entity_util_avg() which will have added things to the
3674 * cfs_rq, so we can remove unconditionally.
3677 sync_entity_load_avg(se);
3679 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3680 ++cfs_rq->removed.nr;
3681 cfs_rq->removed.util_avg += se->avg.util_avg;
3682 cfs_rq->removed.load_avg += se->avg.load_avg;
3683 cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */
3684 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3687 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3689 return cfs_rq->avg.runnable_load_avg;
3692 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3694 return cfs_rq->avg.load_avg;
3697 static inline unsigned long task_util(struct task_struct *p)
3699 return READ_ONCE(p->se.avg.util_avg);
3702 static inline unsigned long _task_util_est(struct task_struct *p)
3704 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3706 return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED);
3709 static inline unsigned long task_util_est(struct task_struct *p)
3711 return max(task_util(p), _task_util_est(p));
3714 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3715 struct task_struct *p)
3717 unsigned int enqueued;
3719 if (!sched_feat(UTIL_EST))
3722 /* Update root cfs_rq's estimated utilization */
3723 enqueued = cfs_rq->avg.util_est.enqueued;
3724 enqueued += _task_util_est(p);
3725 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3729 * Check if a (signed) value is within a specified (unsigned) margin,
3730 * based on the observation that:
3732 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3734 * NOTE: this only works when value + maring < INT_MAX.
3736 static inline bool within_margin(int value, int margin)
3738 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3742 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3744 long last_ewma_diff;
3748 if (!sched_feat(UTIL_EST))
3751 /* Update root cfs_rq's estimated utilization */
3752 ue.enqueued = cfs_rq->avg.util_est.enqueued;
3753 ue.enqueued -= min_t(unsigned int, ue.enqueued, _task_util_est(p));
3754 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3757 * Skip update of task's estimated utilization when the task has not
3758 * yet completed an activation, e.g. being migrated.
3764 * If the PELT values haven't changed since enqueue time,
3765 * skip the util_est update.
3767 ue = p->se.avg.util_est;
3768 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3772 * Skip update of task's estimated utilization when its EWMA is
3773 * already ~1% close to its last activation value.
3775 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3776 last_ewma_diff = ue.enqueued - ue.ewma;
3777 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3781 * To avoid overestimation of actual task utilization, skip updates if
3782 * we cannot grant there is idle time in this CPU.
3784 cpu = cpu_of(rq_of(cfs_rq));
3785 if (task_util(p) > capacity_orig_of(cpu))
3789 * Update Task's estimated utilization
3791 * When *p completes an activation we can consolidate another sample
3792 * of the task size. This is done by storing the current PELT value
3793 * as ue.enqueued and by using this value to update the Exponential
3794 * Weighted Moving Average (EWMA):
3796 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
3797 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
3798 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
3799 * = w * ( last_ewma_diff ) + ewma(t-1)
3800 * = w * (last_ewma_diff + ewma(t-1) / w)
3802 * Where 'w' is the weight of new samples, which is configured to be
3803 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3805 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
3806 ue.ewma += last_ewma_diff;
3807 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
3808 WRITE_ONCE(p->se.avg.util_est, ue);
3811 static inline int task_fits_capacity(struct task_struct *p, long capacity)
3813 return fits_capacity(task_util_est(p), capacity);
3816 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
3818 if (!static_branch_unlikely(&sched_asym_cpucapacity))
3822 rq->misfit_task_load = 0;
3826 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
3827 rq->misfit_task_load = 0;
3831 rq->misfit_task_load = task_h_load(p);
3834 #else /* CONFIG_SMP */
3836 #define UPDATE_TG 0x0
3837 #define SKIP_AGE_LOAD 0x0
3838 #define DO_ATTACH 0x0
3840 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3842 cfs_rq_util_change(cfs_rq, 0);
3845 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3848 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
3850 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3852 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3858 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
3861 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
3863 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
3865 #endif /* CONFIG_SMP */
3867 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3869 #ifdef CONFIG_SCHED_DEBUG
3870 s64 d = se->vruntime - cfs_rq->min_vruntime;
3875 if (d > 3*sysctl_sched_latency)
3876 schedstat_inc(cfs_rq->nr_spread_over);
3881 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3883 u64 vruntime = cfs_rq->min_vruntime;
3886 * The 'current' period is already promised to the current tasks,
3887 * however the extra weight of the new task will slow them down a
3888 * little, place the new task so that it fits in the slot that
3889 * stays open at the end.
3891 if (initial && sched_feat(START_DEBIT))
3892 vruntime += sched_vslice(cfs_rq, se);
3894 /* sleeps up to a single latency don't count. */
3896 unsigned long thresh = sysctl_sched_latency;
3899 * Halve their sleep time's effect, to allow
3900 * for a gentler effect of sleepers:
3902 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3908 /* ensure we never gain time by being placed backwards. */
3909 se->vruntime = max_vruntime(se->vruntime, vruntime);
3912 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3914 static inline void check_schedstat_required(void)
3916 #ifdef CONFIG_SCHEDSTATS
3917 if (schedstat_enabled())
3920 /* Force schedstat enabled if a dependent tracepoint is active */
3921 if (trace_sched_stat_wait_enabled() ||
3922 trace_sched_stat_sleep_enabled() ||
3923 trace_sched_stat_iowait_enabled() ||
3924 trace_sched_stat_blocked_enabled() ||
3925 trace_sched_stat_runtime_enabled()) {
3926 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3927 "stat_blocked and stat_runtime require the "
3928 "kernel parameter schedstats=enable or "
3929 "kernel.sched_schedstats=1\n");
3940 * update_min_vruntime()
3941 * vruntime -= min_vruntime
3945 * update_min_vruntime()
3946 * vruntime += min_vruntime
3948 * this way the vruntime transition between RQs is done when both
3949 * min_vruntime are up-to-date.
3953 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3954 * vruntime -= min_vruntime
3958 * update_min_vruntime()
3959 * vruntime += min_vruntime
3961 * this way we don't have the most up-to-date min_vruntime on the originating
3962 * CPU and an up-to-date min_vruntime on the destination CPU.
3966 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3968 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3969 bool curr = cfs_rq->curr == se;
3972 * If we're the current task, we must renormalise before calling
3976 se->vruntime += cfs_rq->min_vruntime;
3978 update_curr(cfs_rq);
3981 * Otherwise, renormalise after, such that we're placed at the current
3982 * moment in time, instead of some random moment in the past. Being
3983 * placed in the past could significantly boost this task to the
3984 * fairness detriment of existing tasks.
3986 if (renorm && !curr)
3987 se->vruntime += cfs_rq->min_vruntime;
3990 * When enqueuing a sched_entity, we must:
3991 * - Update loads to have both entity and cfs_rq synced with now.
3992 * - Add its load to cfs_rq->runnable_avg
3993 * - For group_entity, update its weight to reflect the new share of
3995 * - Add its new weight to cfs_rq->load.weight
3997 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
3998 update_cfs_group(se);
3999 enqueue_runnable_load_avg(cfs_rq, se);
4000 account_entity_enqueue(cfs_rq, se);
4002 if (flags & ENQUEUE_WAKEUP)
4003 place_entity(cfs_rq, se, 0);
4005 check_schedstat_required();
4006 update_stats_enqueue(cfs_rq, se, flags);
4007 check_spread(cfs_rq, se);
4009 __enqueue_entity(cfs_rq, se);
4012 if (cfs_rq->nr_running == 1) {
4013 list_add_leaf_cfs_rq(cfs_rq);
4014 check_enqueue_throttle(cfs_rq);
4018 static void __clear_buddies_last(struct sched_entity *se)
4020 for_each_sched_entity(se) {
4021 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4022 if (cfs_rq->last != se)
4025 cfs_rq->last = NULL;
4029 static void __clear_buddies_next(struct sched_entity *se)
4031 for_each_sched_entity(se) {
4032 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4033 if (cfs_rq->next != se)
4036 cfs_rq->next = NULL;
4040 static void __clear_buddies_skip(struct sched_entity *se)
4042 for_each_sched_entity(se) {
4043 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4044 if (cfs_rq->skip != se)
4047 cfs_rq->skip = NULL;
4051 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4053 if (cfs_rq->last == se)
4054 __clear_buddies_last(se);
4056 if (cfs_rq->next == se)
4057 __clear_buddies_next(se);
4059 if (cfs_rq->skip == se)
4060 __clear_buddies_skip(se);
4063 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4066 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4069 * Update run-time statistics of the 'current'.
4071 update_curr(cfs_rq);
4074 * When dequeuing a sched_entity, we must:
4075 * - Update loads to have both entity and cfs_rq synced with now.
4076 * - Subtract its load from the cfs_rq->runnable_avg.
4077 * - Subtract its previous weight from cfs_rq->load.weight.
4078 * - For group entity, update its weight to reflect the new share
4079 * of its group cfs_rq.
4081 update_load_avg(cfs_rq, se, UPDATE_TG);
4082 dequeue_runnable_load_avg(cfs_rq, se);
4084 update_stats_dequeue(cfs_rq, se, flags);
4086 clear_buddies(cfs_rq, se);
4088 if (se != cfs_rq->curr)
4089 __dequeue_entity(cfs_rq, se);
4091 account_entity_dequeue(cfs_rq, se);
4094 * Normalize after update_curr(); which will also have moved
4095 * min_vruntime if @se is the one holding it back. But before doing
4096 * update_min_vruntime() again, which will discount @se's position and
4097 * can move min_vruntime forward still more.
4099 if (!(flags & DEQUEUE_SLEEP))
4100 se->vruntime -= cfs_rq->min_vruntime;
4102 /* return excess runtime on last dequeue */
4103 return_cfs_rq_runtime(cfs_rq);
4105 update_cfs_group(se);
4108 * Now advance min_vruntime if @se was the entity holding it back,
4109 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4110 * put back on, and if we advance min_vruntime, we'll be placed back
4111 * further than we started -- ie. we'll be penalized.
4113 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4114 update_min_vruntime(cfs_rq);
4118 * Preempt the current task with a newly woken task if needed:
4121 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4123 unsigned long ideal_runtime, delta_exec;
4124 struct sched_entity *se;
4127 ideal_runtime = sched_slice(cfs_rq, curr);
4128 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4129 if (delta_exec > ideal_runtime) {
4130 resched_curr(rq_of(cfs_rq));
4132 * The current task ran long enough, ensure it doesn't get
4133 * re-elected due to buddy favours.
4135 clear_buddies(cfs_rq, curr);
4140 * Ensure that a task that missed wakeup preemption by a
4141 * narrow margin doesn't have to wait for a full slice.
4142 * This also mitigates buddy induced latencies under load.
4144 if (delta_exec < sysctl_sched_min_granularity)
4147 se = __pick_first_entity(cfs_rq);
4148 delta = curr->vruntime - se->vruntime;
4153 if (delta > ideal_runtime)
4154 resched_curr(rq_of(cfs_rq));
4158 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4160 /* 'current' is not kept within the tree. */
4163 * Any task has to be enqueued before it get to execute on
4164 * a CPU. So account for the time it spent waiting on the
4167 update_stats_wait_end(cfs_rq, se);
4168 __dequeue_entity(cfs_rq, se);
4169 update_load_avg(cfs_rq, se, UPDATE_TG);
4172 update_stats_curr_start(cfs_rq, se);
4176 * Track our maximum slice length, if the CPU's load is at
4177 * least twice that of our own weight (i.e. dont track it
4178 * when there are only lesser-weight tasks around):
4180 if (schedstat_enabled() &&
4181 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
4182 schedstat_set(se->statistics.slice_max,
4183 max((u64)schedstat_val(se->statistics.slice_max),
4184 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4187 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4191 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4194 * Pick the next process, keeping these things in mind, in this order:
4195 * 1) keep things fair between processes/task groups
4196 * 2) pick the "next" process, since someone really wants that to run
4197 * 3) pick the "last" process, for cache locality
4198 * 4) do not run the "skip" process, if something else is available
4200 static struct sched_entity *
4201 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4203 struct sched_entity *left = __pick_first_entity(cfs_rq);
4204 struct sched_entity *se;
4207 * If curr is set we have to see if its left of the leftmost entity
4208 * still in the tree, provided there was anything in the tree at all.
4210 if (!left || (curr && entity_before(curr, left)))
4213 se = left; /* ideally we run the leftmost entity */
4216 * Avoid running the skip buddy, if running something else can
4217 * be done without getting too unfair.
4219 if (cfs_rq->skip == se) {
4220 struct sched_entity *second;
4223 second = __pick_first_entity(cfs_rq);
4225 second = __pick_next_entity(se);
4226 if (!second || (curr && entity_before(curr, second)))
4230 if (second && wakeup_preempt_entity(second, left) < 1)
4235 * Prefer last buddy, try to return the CPU to a preempted task.
4237 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4241 * Someone really wants this to run. If it's not unfair, run it.
4243 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4246 clear_buddies(cfs_rq, se);
4251 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4253 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4256 * If still on the runqueue then deactivate_task()
4257 * was not called and update_curr() has to be done:
4260 update_curr(cfs_rq);
4262 /* throttle cfs_rqs exceeding runtime */
4263 check_cfs_rq_runtime(cfs_rq);
4265 check_spread(cfs_rq, prev);
4268 update_stats_wait_start(cfs_rq, prev);
4269 /* Put 'current' back into the tree. */
4270 __enqueue_entity(cfs_rq, prev);
4271 /* in !on_rq case, update occurred at dequeue */
4272 update_load_avg(cfs_rq, prev, 0);
4274 cfs_rq->curr = NULL;
4278 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4281 * Update run-time statistics of the 'current'.
4283 update_curr(cfs_rq);
4286 * Ensure that runnable average is periodically updated.
4288 update_load_avg(cfs_rq, curr, UPDATE_TG);
4289 update_cfs_group(curr);
4291 #ifdef CONFIG_SCHED_HRTICK
4293 * queued ticks are scheduled to match the slice, so don't bother
4294 * validating it and just reschedule.
4297 resched_curr(rq_of(cfs_rq));
4301 * don't let the period tick interfere with the hrtick preemption
4303 if (!sched_feat(DOUBLE_TICK) &&
4304 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4308 if (cfs_rq->nr_running > 1)
4309 check_preempt_tick(cfs_rq, curr);
4313 /**************************************************
4314 * CFS bandwidth control machinery
4317 #ifdef CONFIG_CFS_BANDWIDTH
4319 #ifdef CONFIG_JUMP_LABEL
4320 static struct static_key __cfs_bandwidth_used;
4322 static inline bool cfs_bandwidth_used(void)
4324 return static_key_false(&__cfs_bandwidth_used);
4327 void cfs_bandwidth_usage_inc(void)
4329 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4332 void cfs_bandwidth_usage_dec(void)
4334 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4336 #else /* CONFIG_JUMP_LABEL */
4337 static bool cfs_bandwidth_used(void)
4342 void cfs_bandwidth_usage_inc(void) {}
4343 void cfs_bandwidth_usage_dec(void) {}
4344 #endif /* CONFIG_JUMP_LABEL */
4347 * default period for cfs group bandwidth.
4348 * default: 0.1s, units: nanoseconds
4350 static inline u64 default_cfs_period(void)
4352 return 100000000ULL;
4355 static inline u64 sched_cfs_bandwidth_slice(void)
4357 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4361 * Replenish runtime according to assigned quota. We use sched_clock_cpu
4362 * directly instead of rq->clock to avoid adding additional synchronization
4365 * requires cfs_b->lock
4367 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4369 if (cfs_b->quota != RUNTIME_INF)
4370 cfs_b->runtime = cfs_b->quota;
4373 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4375 return &tg->cfs_bandwidth;
4378 /* returns 0 on failure to allocate runtime */
4379 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4381 struct task_group *tg = cfs_rq->tg;
4382 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4383 u64 amount = 0, min_amount;
4385 /* note: this is a positive sum as runtime_remaining <= 0 */
4386 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4388 raw_spin_lock(&cfs_b->lock);
4389 if (cfs_b->quota == RUNTIME_INF)
4390 amount = min_amount;
4392 start_cfs_bandwidth(cfs_b);
4394 if (cfs_b->runtime > 0) {
4395 amount = min(cfs_b->runtime, min_amount);
4396 cfs_b->runtime -= amount;
4400 raw_spin_unlock(&cfs_b->lock);
4402 cfs_rq->runtime_remaining += amount;
4404 return cfs_rq->runtime_remaining > 0;
4407 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4409 /* dock delta_exec before expiring quota (as it could span periods) */
4410 cfs_rq->runtime_remaining -= delta_exec;
4412 if (likely(cfs_rq->runtime_remaining > 0))
4415 if (cfs_rq->throttled)
4418 * if we're unable to extend our runtime we resched so that the active
4419 * hierarchy can be throttled
4421 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4422 resched_curr(rq_of(cfs_rq));
4425 static __always_inline
4426 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4428 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4431 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4434 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4436 return cfs_bandwidth_used() && cfs_rq->throttled;
4439 /* check whether cfs_rq, or any parent, is throttled */
4440 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4442 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4446 * Ensure that neither of the group entities corresponding to src_cpu or
4447 * dest_cpu are members of a throttled hierarchy when performing group
4448 * load-balance operations.
4450 static inline int throttled_lb_pair(struct task_group *tg,
4451 int src_cpu, int dest_cpu)
4453 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4455 src_cfs_rq = tg->cfs_rq[src_cpu];
4456 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4458 return throttled_hierarchy(src_cfs_rq) ||
4459 throttled_hierarchy(dest_cfs_rq);
4462 static int tg_unthrottle_up(struct task_group *tg, void *data)
4464 struct rq *rq = data;
4465 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4467 cfs_rq->throttle_count--;
4468 if (!cfs_rq->throttle_count) {
4469 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4470 cfs_rq->throttled_clock_task;
4472 /* Add cfs_rq with already running entity in the list */
4473 if (cfs_rq->nr_running >= 1)
4474 list_add_leaf_cfs_rq(cfs_rq);
4480 static int tg_throttle_down(struct task_group *tg, void *data)
4482 struct rq *rq = data;
4483 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4485 /* group is entering throttled state, stop time */
4486 if (!cfs_rq->throttle_count) {
4487 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4488 list_del_leaf_cfs_rq(cfs_rq);
4490 cfs_rq->throttle_count++;
4495 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4497 struct rq *rq = rq_of(cfs_rq);
4498 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4499 struct sched_entity *se;
4500 long task_delta, idle_task_delta, dequeue = 1;
4503 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4505 /* freeze hierarchy runnable averages while throttled */
4507 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4510 task_delta = cfs_rq->h_nr_running;
4511 idle_task_delta = cfs_rq->idle_h_nr_running;
4512 for_each_sched_entity(se) {
4513 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4514 /* throttled entity or throttle-on-deactivate */
4519 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4520 qcfs_rq->h_nr_running -= task_delta;
4521 qcfs_rq->idle_h_nr_running -= idle_task_delta;
4523 if (qcfs_rq->load.weight)
4528 sub_nr_running(rq, task_delta);
4530 cfs_rq->throttled = 1;
4531 cfs_rq->throttled_clock = rq_clock(rq);
4532 raw_spin_lock(&cfs_b->lock);
4533 empty = list_empty(&cfs_b->throttled_cfs_rq);
4536 * Add to the _head_ of the list, so that an already-started
4537 * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
4538 * not running add to the tail so that later runqueues don't get starved.
4540 if (cfs_b->distribute_running)
4541 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4543 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4546 * If we're the first throttled task, make sure the bandwidth
4550 start_cfs_bandwidth(cfs_b);
4552 raw_spin_unlock(&cfs_b->lock);
4555 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4557 struct rq *rq = rq_of(cfs_rq);
4558 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4559 struct sched_entity *se;
4561 long task_delta, idle_task_delta;
4563 se = cfs_rq->tg->se[cpu_of(rq)];
4565 cfs_rq->throttled = 0;
4567 update_rq_clock(rq);
4569 raw_spin_lock(&cfs_b->lock);
4570 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4571 list_del_rcu(&cfs_rq->throttled_list);
4572 raw_spin_unlock(&cfs_b->lock);
4574 /* update hierarchical throttle state */
4575 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4577 if (!cfs_rq->load.weight)
4580 task_delta = cfs_rq->h_nr_running;
4581 idle_task_delta = cfs_rq->idle_h_nr_running;
4582 for_each_sched_entity(se) {
4586 cfs_rq = cfs_rq_of(se);
4588 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4589 cfs_rq->h_nr_running += task_delta;
4590 cfs_rq->idle_h_nr_running += idle_task_delta;
4592 if (cfs_rq_throttled(cfs_rq))
4596 assert_list_leaf_cfs_rq(rq);
4599 add_nr_running(rq, task_delta);
4601 /* Determine whether we need to wake up potentially idle CPU: */
4602 if (rq->curr == rq->idle && rq->cfs.nr_running)
4606 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, u64 remaining)
4608 struct cfs_rq *cfs_rq;
4610 u64 starting_runtime = remaining;
4613 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4615 struct rq *rq = rq_of(cfs_rq);
4618 rq_lock_irqsave(rq, &rf);
4619 if (!cfs_rq_throttled(cfs_rq))
4622 /* By the above check, this should never be true */
4623 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
4625 runtime = -cfs_rq->runtime_remaining + 1;
4626 if (runtime > remaining)
4627 runtime = remaining;
4628 remaining -= runtime;
4630 cfs_rq->runtime_remaining += runtime;
4632 /* we check whether we're throttled above */
4633 if (cfs_rq->runtime_remaining > 0)
4634 unthrottle_cfs_rq(cfs_rq);
4637 rq_unlock_irqrestore(rq, &rf);
4644 return starting_runtime - remaining;
4648 * Responsible for refilling a task_group's bandwidth and unthrottling its
4649 * cfs_rqs as appropriate. If there has been no activity within the last
4650 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4651 * used to track this state.
4653 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
4658 /* no need to continue the timer with no bandwidth constraint */
4659 if (cfs_b->quota == RUNTIME_INF)
4660 goto out_deactivate;
4662 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4663 cfs_b->nr_periods += overrun;
4666 * idle depends on !throttled (for the case of a large deficit), and if
4667 * we're going inactive then everything else can be deferred
4669 if (cfs_b->idle && !throttled)
4670 goto out_deactivate;
4672 __refill_cfs_bandwidth_runtime(cfs_b);
4675 /* mark as potentially idle for the upcoming period */
4680 /* account preceding periods in which throttling occurred */
4681 cfs_b->nr_throttled += overrun;
4684 * This check is repeated as we are holding onto the new bandwidth while
4685 * we unthrottle. This can potentially race with an unthrottled group
4686 * trying to acquire new bandwidth from the global pool. This can result
4687 * in us over-using our runtime if it is all used during this loop, but
4688 * only by limited amounts in that extreme case.
4690 while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
4691 runtime = cfs_b->runtime;
4692 cfs_b->distribute_running = 1;
4693 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4694 /* we can't nest cfs_b->lock while distributing bandwidth */
4695 runtime = distribute_cfs_runtime(cfs_b, runtime);
4696 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4698 cfs_b->distribute_running = 0;
4699 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4701 lsub_positive(&cfs_b->runtime, runtime);
4705 * While we are ensured activity in the period following an
4706 * unthrottle, this also covers the case in which the new bandwidth is
4707 * insufficient to cover the existing bandwidth deficit. (Forcing the
4708 * timer to remain active while there are any throttled entities.)
4718 /* a cfs_rq won't donate quota below this amount */
4719 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4720 /* minimum remaining period time to redistribute slack quota */
4721 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4722 /* how long we wait to gather additional slack before distributing */
4723 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4726 * Are we near the end of the current quota period?
4728 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4729 * hrtimer base being cleared by hrtimer_start. In the case of
4730 * migrate_hrtimers, base is never cleared, so we are fine.
4732 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4734 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4737 /* if the call-back is running a quota refresh is already occurring */
4738 if (hrtimer_callback_running(refresh_timer))
4741 /* is a quota refresh about to occur? */
4742 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4743 if (remaining < min_expire)
4749 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4751 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4753 /* if there's a quota refresh soon don't bother with slack */
4754 if (runtime_refresh_within(cfs_b, min_left))
4757 /* don't push forwards an existing deferred unthrottle */
4758 if (cfs_b->slack_started)
4760 cfs_b->slack_started = true;
4762 hrtimer_start(&cfs_b->slack_timer,
4763 ns_to_ktime(cfs_bandwidth_slack_period),
4767 /* we know any runtime found here is valid as update_curr() precedes return */
4768 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4770 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4771 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4773 if (slack_runtime <= 0)
4776 raw_spin_lock(&cfs_b->lock);
4777 if (cfs_b->quota != RUNTIME_INF) {
4778 cfs_b->runtime += slack_runtime;
4780 /* we are under rq->lock, defer unthrottling using a timer */
4781 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4782 !list_empty(&cfs_b->throttled_cfs_rq))
4783 start_cfs_slack_bandwidth(cfs_b);
4785 raw_spin_unlock(&cfs_b->lock);
4787 /* even if it's not valid for return we don't want to try again */
4788 cfs_rq->runtime_remaining -= slack_runtime;
4791 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4793 if (!cfs_bandwidth_used())
4796 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4799 __return_cfs_rq_runtime(cfs_rq);
4803 * This is done with a timer (instead of inline with bandwidth return) since
4804 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4806 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4808 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4809 unsigned long flags;
4811 /* confirm we're still not at a refresh boundary */
4812 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4813 cfs_b->slack_started = false;
4814 if (cfs_b->distribute_running) {
4815 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4819 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4820 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4824 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4825 runtime = cfs_b->runtime;
4828 cfs_b->distribute_running = 1;
4830 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4835 runtime = distribute_cfs_runtime(cfs_b, runtime);
4837 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4838 lsub_positive(&cfs_b->runtime, runtime);
4839 cfs_b->distribute_running = 0;
4840 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4844 * When a group wakes up we want to make sure that its quota is not already
4845 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4846 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4848 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4850 if (!cfs_bandwidth_used())
4853 /* an active group must be handled by the update_curr()->put() path */
4854 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4857 /* ensure the group is not already throttled */
4858 if (cfs_rq_throttled(cfs_rq))
4861 /* update runtime allocation */
4862 account_cfs_rq_runtime(cfs_rq, 0);
4863 if (cfs_rq->runtime_remaining <= 0)
4864 throttle_cfs_rq(cfs_rq);
4867 static void sync_throttle(struct task_group *tg, int cpu)
4869 struct cfs_rq *pcfs_rq, *cfs_rq;
4871 if (!cfs_bandwidth_used())
4877 cfs_rq = tg->cfs_rq[cpu];
4878 pcfs_rq = tg->parent->cfs_rq[cpu];
4880 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4881 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4884 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4885 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4887 if (!cfs_bandwidth_used())
4890 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4894 * it's possible for a throttled entity to be forced into a running
4895 * state (e.g. set_curr_task), in this case we're finished.
4897 if (cfs_rq_throttled(cfs_rq))
4900 throttle_cfs_rq(cfs_rq);
4904 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4906 struct cfs_bandwidth *cfs_b =
4907 container_of(timer, struct cfs_bandwidth, slack_timer);
4909 do_sched_cfs_slack_timer(cfs_b);
4911 return HRTIMER_NORESTART;
4914 extern const u64 max_cfs_quota_period;
4916 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4918 struct cfs_bandwidth *cfs_b =
4919 container_of(timer, struct cfs_bandwidth, period_timer);
4920 unsigned long flags;
4925 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4927 overrun = hrtimer_forward_now(timer, cfs_b->period);
4932 u64 new, old = ktime_to_ns(cfs_b->period);
4935 * Grow period by a factor of 2 to avoid losing precision.
4936 * Precision loss in the quota/period ratio can cause __cfs_schedulable
4940 if (new < max_cfs_quota_period) {
4941 cfs_b->period = ns_to_ktime(new);
4944 pr_warn_ratelimited(
4945 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
4947 div_u64(new, NSEC_PER_USEC),
4948 div_u64(cfs_b->quota, NSEC_PER_USEC));
4950 pr_warn_ratelimited(
4951 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
4953 div_u64(old, NSEC_PER_USEC),
4954 div_u64(cfs_b->quota, NSEC_PER_USEC));
4957 /* reset count so we don't come right back in here */
4961 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
4964 cfs_b->period_active = 0;
4965 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4967 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4970 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4972 raw_spin_lock_init(&cfs_b->lock);
4974 cfs_b->quota = RUNTIME_INF;
4975 cfs_b->period = ns_to_ktime(default_cfs_period());
4977 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4978 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4979 cfs_b->period_timer.function = sched_cfs_period_timer;
4980 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4981 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4982 cfs_b->distribute_running = 0;
4983 cfs_b->slack_started = false;
4986 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4988 cfs_rq->runtime_enabled = 0;
4989 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4992 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4994 lockdep_assert_held(&cfs_b->lock);
4996 if (cfs_b->period_active)
4999 cfs_b->period_active = 1;
5000 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5001 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5004 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5006 /* init_cfs_bandwidth() was not called */
5007 if (!cfs_b->throttled_cfs_rq.next)
5010 hrtimer_cancel(&cfs_b->period_timer);
5011 hrtimer_cancel(&cfs_b->slack_timer);
5015 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5017 * The race is harmless, since modifying bandwidth settings of unhooked group
5018 * bits doesn't do much.
5021 /* cpu online calback */
5022 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5024 struct task_group *tg;
5026 lockdep_assert_held(&rq->lock);
5029 list_for_each_entry_rcu(tg, &task_groups, list) {
5030 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5031 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5033 raw_spin_lock(&cfs_b->lock);
5034 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5035 raw_spin_unlock(&cfs_b->lock);
5040 /* cpu offline callback */
5041 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5043 struct task_group *tg;
5045 lockdep_assert_held(&rq->lock);
5048 list_for_each_entry_rcu(tg, &task_groups, list) {
5049 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5051 if (!cfs_rq->runtime_enabled)
5055 * clock_task is not advancing so we just need to make sure
5056 * there's some valid quota amount
5058 cfs_rq->runtime_remaining = 1;
5060 * Offline rq is schedulable till CPU is completely disabled
5061 * in take_cpu_down(), so we prevent new cfs throttling here.
5063 cfs_rq->runtime_enabled = 0;
5065 if (cfs_rq_throttled(cfs_rq))
5066 unthrottle_cfs_rq(cfs_rq);
5071 #else /* CONFIG_CFS_BANDWIDTH */
5073 static inline bool cfs_bandwidth_used(void)
5078 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5079 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5080 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5081 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5082 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5084 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5089 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5094 static inline int throttled_lb_pair(struct task_group *tg,
5095 int src_cpu, int dest_cpu)
5100 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5102 #ifdef CONFIG_FAIR_GROUP_SCHED
5103 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5106 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5110 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5111 static inline void update_runtime_enabled(struct rq *rq) {}
5112 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5114 #endif /* CONFIG_CFS_BANDWIDTH */
5116 /**************************************************
5117 * CFS operations on tasks:
5120 #ifdef CONFIG_SCHED_HRTICK
5121 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5123 struct sched_entity *se = &p->se;
5124 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5126 SCHED_WARN_ON(task_rq(p) != rq);
5128 if (rq->cfs.h_nr_running > 1) {
5129 u64 slice = sched_slice(cfs_rq, se);
5130 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5131 s64 delta = slice - ran;
5138 hrtick_start(rq, delta);
5143 * called from enqueue/dequeue and updates the hrtick when the
5144 * current task is from our class and nr_running is low enough
5147 static void hrtick_update(struct rq *rq)
5149 struct task_struct *curr = rq->curr;
5151 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5154 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5155 hrtick_start_fair(rq, curr);
5157 #else /* !CONFIG_SCHED_HRTICK */
5159 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5163 static inline void hrtick_update(struct rq *rq)
5169 static inline unsigned long cpu_util(int cpu);
5171 static inline bool cpu_overutilized(int cpu)
5173 return !fits_capacity(cpu_util(cpu), capacity_of(cpu));
5176 static inline void update_overutilized_status(struct rq *rq)
5178 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
5179 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5180 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
5184 static inline void update_overutilized_status(struct rq *rq) { }
5188 * The enqueue_task method is called before nr_running is
5189 * increased. Here we update the fair scheduling stats and
5190 * then put the task into the rbtree:
5193 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5195 struct cfs_rq *cfs_rq;
5196 struct sched_entity *se = &p->se;
5197 int idle_h_nr_running = task_has_idle_policy(p);
5200 * The code below (indirectly) updates schedutil which looks at
5201 * the cfs_rq utilization to select a frequency.
5202 * Let's add the task's estimated utilization to the cfs_rq's
5203 * estimated utilization, before we update schedutil.
5205 util_est_enqueue(&rq->cfs, p);
5208 * If in_iowait is set, the code below may not trigger any cpufreq
5209 * utilization updates, so do it here explicitly with the IOWAIT flag
5213 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5215 for_each_sched_entity(se) {
5218 cfs_rq = cfs_rq_of(se);
5219 enqueue_entity(cfs_rq, se, flags);
5222 * end evaluation on encountering a throttled cfs_rq
5224 * note: in the case of encountering a throttled cfs_rq we will
5225 * post the final h_nr_running increment below.
5227 if (cfs_rq_throttled(cfs_rq))
5229 cfs_rq->h_nr_running++;
5230 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5232 flags = ENQUEUE_WAKEUP;
5235 for_each_sched_entity(se) {
5236 cfs_rq = cfs_rq_of(se);
5237 cfs_rq->h_nr_running++;
5238 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5240 if (cfs_rq_throttled(cfs_rq))
5243 update_load_avg(cfs_rq, se, UPDATE_TG);
5244 update_cfs_group(se);
5248 add_nr_running(rq, 1);
5250 * Since new tasks are assigned an initial util_avg equal to
5251 * half of the spare capacity of their CPU, tiny tasks have the
5252 * ability to cross the overutilized threshold, which will
5253 * result in the load balancer ruining all the task placement
5254 * done by EAS. As a way to mitigate that effect, do not account
5255 * for the first enqueue operation of new tasks during the
5256 * overutilized flag detection.
5258 * A better way of solving this problem would be to wait for
5259 * the PELT signals of tasks to converge before taking them
5260 * into account, but that is not straightforward to implement,
5261 * and the following generally works well enough in practice.
5263 if (flags & ENQUEUE_WAKEUP)
5264 update_overutilized_status(rq);
5268 if (cfs_bandwidth_used()) {
5270 * When bandwidth control is enabled; the cfs_rq_throttled()
5271 * breaks in the above iteration can result in incomplete
5272 * leaf list maintenance, resulting in triggering the assertion
5275 for_each_sched_entity(se) {
5276 cfs_rq = cfs_rq_of(se);
5278 if (list_add_leaf_cfs_rq(cfs_rq))
5283 assert_list_leaf_cfs_rq(rq);
5288 static void set_next_buddy(struct sched_entity *se);
5291 * The dequeue_task method is called before nr_running is
5292 * decreased. We remove the task from the rbtree and
5293 * update the fair scheduling stats:
5295 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5297 struct cfs_rq *cfs_rq;
5298 struct sched_entity *se = &p->se;
5299 int task_sleep = flags & DEQUEUE_SLEEP;
5300 int idle_h_nr_running = task_has_idle_policy(p);
5302 for_each_sched_entity(se) {
5303 cfs_rq = cfs_rq_of(se);
5304 dequeue_entity(cfs_rq, se, flags);
5307 * end evaluation on encountering a throttled cfs_rq
5309 * note: in the case of encountering a throttled cfs_rq we will
5310 * post the final h_nr_running decrement below.
5312 if (cfs_rq_throttled(cfs_rq))
5314 cfs_rq->h_nr_running--;
5315 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5317 /* Don't dequeue parent if it has other entities besides us */
5318 if (cfs_rq->load.weight) {
5319 /* Avoid re-evaluating load for this entity: */
5320 se = parent_entity(se);
5322 * Bias pick_next to pick a task from this cfs_rq, as
5323 * p is sleeping when it is within its sched_slice.
5325 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5329 flags |= DEQUEUE_SLEEP;
5332 for_each_sched_entity(se) {
5333 cfs_rq = cfs_rq_of(se);
5334 cfs_rq->h_nr_running--;
5335 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5337 if (cfs_rq_throttled(cfs_rq))
5340 update_load_avg(cfs_rq, se, UPDATE_TG);
5341 update_cfs_group(se);
5345 sub_nr_running(rq, 1);
5347 util_est_dequeue(&rq->cfs, p, task_sleep);
5353 /* Working cpumask for: load_balance, load_balance_newidle. */
5354 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5355 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5357 #ifdef CONFIG_NO_HZ_COMMON
5360 cpumask_var_t idle_cpus_mask;
5362 int has_blocked; /* Idle CPUS has blocked load */
5363 unsigned long next_balance; /* in jiffy units */
5364 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5365 } nohz ____cacheline_aligned;
5367 #endif /* CONFIG_NO_HZ_COMMON */
5369 /* CPU only has SCHED_IDLE tasks enqueued */
5370 static int sched_idle_cpu(int cpu)
5372 struct rq *rq = cpu_rq(cpu);
5374 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
5378 static unsigned long cpu_load(struct rq *rq)
5380 return cfs_rq_load_avg(&rq->cfs);
5383 static unsigned long capacity_of(int cpu)
5385 return cpu_rq(cpu)->cpu_capacity;
5388 static void record_wakee(struct task_struct *p)
5391 * Only decay a single time; tasks that have less then 1 wakeup per
5392 * jiffy will not have built up many flips.
5394 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5395 current->wakee_flips >>= 1;
5396 current->wakee_flip_decay_ts = jiffies;
5399 if (current->last_wakee != p) {
5400 current->last_wakee = p;
5401 current->wakee_flips++;
5406 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5408 * A waker of many should wake a different task than the one last awakened
5409 * at a frequency roughly N times higher than one of its wakees.
5411 * In order to determine whether we should let the load spread vs consolidating
5412 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5413 * partner, and a factor of lls_size higher frequency in the other.
5415 * With both conditions met, we can be relatively sure that the relationship is
5416 * non-monogamous, with partner count exceeding socket size.
5418 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5419 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5422 static int wake_wide(struct task_struct *p)
5424 unsigned int master = current->wakee_flips;
5425 unsigned int slave = p->wakee_flips;
5426 int factor = this_cpu_read(sd_llc_size);
5429 swap(master, slave);
5430 if (slave < factor || master < slave * factor)
5436 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5437 * soonest. For the purpose of speed we only consider the waking and previous
5440 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5441 * cache-affine and is (or will be) idle.
5443 * wake_affine_weight() - considers the weight to reflect the average
5444 * scheduling latency of the CPUs. This seems to work
5445 * for the overloaded case.
5448 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5451 * If this_cpu is idle, it implies the wakeup is from interrupt
5452 * context. Only allow the move if cache is shared. Otherwise an
5453 * interrupt intensive workload could force all tasks onto one
5454 * node depending on the IO topology or IRQ affinity settings.
5456 * If the prev_cpu is idle and cache affine then avoid a migration.
5457 * There is no guarantee that the cache hot data from an interrupt
5458 * is more important than cache hot data on the prev_cpu and from
5459 * a cpufreq perspective, it's better to have higher utilisation
5462 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5463 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5465 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5468 return nr_cpumask_bits;
5472 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5473 int this_cpu, int prev_cpu, int sync)
5475 s64 this_eff_load, prev_eff_load;
5476 unsigned long task_load;
5478 this_eff_load = cpu_load(cpu_rq(this_cpu));
5481 unsigned long current_load = task_h_load(current);
5483 if (current_load > this_eff_load)
5486 this_eff_load -= current_load;
5489 task_load = task_h_load(p);
5491 this_eff_load += task_load;
5492 if (sched_feat(WA_BIAS))
5493 this_eff_load *= 100;
5494 this_eff_load *= capacity_of(prev_cpu);
5496 prev_eff_load = cpu_load(cpu_rq(prev_cpu));
5497 prev_eff_load -= task_load;
5498 if (sched_feat(WA_BIAS))
5499 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5500 prev_eff_load *= capacity_of(this_cpu);
5503 * If sync, adjust the weight of prev_eff_load such that if
5504 * prev_eff == this_eff that select_idle_sibling() will consider
5505 * stacking the wakee on top of the waker if no other CPU is
5511 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5514 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5515 int this_cpu, int prev_cpu, int sync)
5517 int target = nr_cpumask_bits;
5519 if (sched_feat(WA_IDLE))
5520 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5522 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5523 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5525 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5526 if (target == nr_cpumask_bits)
5529 schedstat_inc(sd->ttwu_move_affine);
5530 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5534 static struct sched_group *
5535 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5536 int this_cpu, int sd_flag);
5539 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5542 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5544 unsigned long load, min_load = ULONG_MAX;
5545 unsigned int min_exit_latency = UINT_MAX;
5546 u64 latest_idle_timestamp = 0;
5547 int least_loaded_cpu = this_cpu;
5548 int shallowest_idle_cpu = -1, si_cpu = -1;
5551 /* Check if we have any choice: */
5552 if (group->group_weight == 1)
5553 return cpumask_first(sched_group_span(group));
5555 /* Traverse only the allowed CPUs */
5556 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
5557 if (available_idle_cpu(i)) {
5558 struct rq *rq = cpu_rq(i);
5559 struct cpuidle_state *idle = idle_get_state(rq);
5560 if (idle && idle->exit_latency < min_exit_latency) {
5562 * We give priority to a CPU whose idle state
5563 * has the smallest exit latency irrespective
5564 * of any idle timestamp.
5566 min_exit_latency = idle->exit_latency;
5567 latest_idle_timestamp = rq->idle_stamp;
5568 shallowest_idle_cpu = i;
5569 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5570 rq->idle_stamp > latest_idle_timestamp) {
5572 * If equal or no active idle state, then
5573 * the most recently idled CPU might have
5576 latest_idle_timestamp = rq->idle_stamp;
5577 shallowest_idle_cpu = i;
5579 } else if (shallowest_idle_cpu == -1 && si_cpu == -1) {
5580 if (sched_idle_cpu(i)) {
5585 load = cpu_load(cpu_rq(i));
5586 if (load < min_load) {
5588 least_loaded_cpu = i;
5593 if (shallowest_idle_cpu != -1)
5594 return shallowest_idle_cpu;
5597 return least_loaded_cpu;
5600 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5601 int cpu, int prev_cpu, int sd_flag)
5605 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
5609 * We need task's util for cpu_util_without, sync it up to
5610 * prev_cpu's last_update_time.
5612 if (!(sd_flag & SD_BALANCE_FORK))
5613 sync_entity_load_avg(&p->se);
5616 struct sched_group *group;
5617 struct sched_domain *tmp;
5620 if (!(sd->flags & sd_flag)) {
5625 group = find_idlest_group(sd, p, cpu, sd_flag);
5631 new_cpu = find_idlest_group_cpu(group, p, cpu);
5632 if (new_cpu == cpu) {
5633 /* Now try balancing at a lower domain level of 'cpu': */
5638 /* Now try balancing at a lower domain level of 'new_cpu': */
5640 weight = sd->span_weight;
5642 for_each_domain(cpu, tmp) {
5643 if (weight <= tmp->span_weight)
5645 if (tmp->flags & sd_flag)
5653 #ifdef CONFIG_SCHED_SMT
5654 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5655 EXPORT_SYMBOL_GPL(sched_smt_present);
5657 static inline void set_idle_cores(int cpu, int val)
5659 struct sched_domain_shared *sds;
5661 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5663 WRITE_ONCE(sds->has_idle_cores, val);
5666 static inline bool test_idle_cores(int cpu, bool def)
5668 struct sched_domain_shared *sds;
5670 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5672 return READ_ONCE(sds->has_idle_cores);
5678 * Scans the local SMT mask to see if the entire core is idle, and records this
5679 * information in sd_llc_shared->has_idle_cores.
5681 * Since SMT siblings share all cache levels, inspecting this limited remote
5682 * state should be fairly cheap.
5684 void __update_idle_core(struct rq *rq)
5686 int core = cpu_of(rq);
5690 if (test_idle_cores(core, true))
5693 for_each_cpu(cpu, cpu_smt_mask(core)) {
5697 if (!available_idle_cpu(cpu))
5701 set_idle_cores(core, 1);
5707 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5708 * there are no idle cores left in the system; tracked through
5709 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5711 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5713 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5716 if (!static_branch_likely(&sched_smt_present))
5719 if (!test_idle_cores(target, false))
5722 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
5724 for_each_cpu_wrap(core, cpus, target) {
5727 for_each_cpu(cpu, cpu_smt_mask(core)) {
5728 __cpumask_clear_cpu(cpu, cpus);
5729 if (!available_idle_cpu(cpu))
5738 * Failed to find an idle core; stop looking for one.
5740 set_idle_cores(target, 0);
5746 * Scan the local SMT mask for idle CPUs.
5748 static int select_idle_smt(struct task_struct *p, int target)
5750 int cpu, si_cpu = -1;
5752 if (!static_branch_likely(&sched_smt_present))
5755 for_each_cpu(cpu, cpu_smt_mask(target)) {
5756 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
5758 if (available_idle_cpu(cpu))
5760 if (si_cpu == -1 && sched_idle_cpu(cpu))
5767 #else /* CONFIG_SCHED_SMT */
5769 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5774 static inline int select_idle_smt(struct task_struct *p, int target)
5779 #endif /* CONFIG_SCHED_SMT */
5782 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5783 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5784 * average idle time for this rq (as found in rq->avg_idle).
5786 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
5788 struct sched_domain *this_sd;
5789 u64 avg_cost, avg_idle;
5792 int this = smp_processor_id();
5793 int cpu, nr = INT_MAX, si_cpu = -1;
5795 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
5800 * Due to large variance we need a large fuzz factor; hackbench in
5801 * particularly is sensitive here.
5803 avg_idle = this_rq()->avg_idle / 512;
5804 avg_cost = this_sd->avg_scan_cost + 1;
5806 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
5809 if (sched_feat(SIS_PROP)) {
5810 u64 span_avg = sd->span_weight * avg_idle;
5811 if (span_avg > 4*avg_cost)
5812 nr = div_u64(span_avg, avg_cost);
5817 time = cpu_clock(this);
5819 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
5822 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
5824 if (available_idle_cpu(cpu))
5826 if (si_cpu == -1 && sched_idle_cpu(cpu))
5830 time = cpu_clock(this) - time;
5831 cost = this_sd->avg_scan_cost;
5832 delta = (s64)(time - cost) / 8;
5833 this_sd->avg_scan_cost += delta;
5839 * Try and locate an idle core/thread in the LLC cache domain.
5841 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5843 struct sched_domain *sd;
5844 int i, recent_used_cpu;
5846 if (available_idle_cpu(target) || sched_idle_cpu(target))
5850 * If the previous CPU is cache affine and idle, don't be stupid:
5852 if (prev != target && cpus_share_cache(prev, target) &&
5853 (available_idle_cpu(prev) || sched_idle_cpu(prev)))
5856 /* Check a recently used CPU as a potential idle candidate: */
5857 recent_used_cpu = p->recent_used_cpu;
5858 if (recent_used_cpu != prev &&
5859 recent_used_cpu != target &&
5860 cpus_share_cache(recent_used_cpu, target) &&
5861 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
5862 cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr)) {
5864 * Replace recent_used_cpu with prev as it is a potential
5865 * candidate for the next wake:
5867 p->recent_used_cpu = prev;
5868 return recent_used_cpu;
5871 sd = rcu_dereference(per_cpu(sd_llc, target));
5875 i = select_idle_core(p, sd, target);
5876 if ((unsigned)i < nr_cpumask_bits)
5879 i = select_idle_cpu(p, sd, target);
5880 if ((unsigned)i < nr_cpumask_bits)
5883 i = select_idle_smt(p, target);
5884 if ((unsigned)i < nr_cpumask_bits)
5891 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
5892 * @cpu: the CPU to get the utilization of
5894 * The unit of the return value must be the one of capacity so we can compare
5895 * the utilization with the capacity of the CPU that is available for CFS task
5896 * (ie cpu_capacity).
5898 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5899 * recent utilization of currently non-runnable tasks on a CPU. It represents
5900 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5901 * capacity_orig is the cpu_capacity available at the highest frequency
5902 * (arch_scale_freq_capacity()).
5903 * The utilization of a CPU converges towards a sum equal to or less than the
5904 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5905 * the running time on this CPU scaled by capacity_curr.
5907 * The estimated utilization of a CPU is defined to be the maximum between its
5908 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
5909 * currently RUNNABLE on that CPU.
5910 * This allows to properly represent the expected utilization of a CPU which
5911 * has just got a big task running since a long sleep period. At the same time
5912 * however it preserves the benefits of the "blocked utilization" in
5913 * describing the potential for other tasks waking up on the same CPU.
5915 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5916 * higher than capacity_orig because of unfortunate rounding in
5917 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5918 * the average stabilizes with the new running time. We need to check that the
5919 * utilization stays within the range of [0..capacity_orig] and cap it if
5920 * necessary. Without utilization capping, a group could be seen as overloaded
5921 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5922 * available capacity. We allow utilization to overshoot capacity_curr (but not
5923 * capacity_orig) as it useful for predicting the capacity required after task
5924 * migrations (scheduler-driven DVFS).
5926 * Return: the (estimated) utilization for the specified CPU
5928 static inline unsigned long cpu_util(int cpu)
5930 struct cfs_rq *cfs_rq;
5933 cfs_rq = &cpu_rq(cpu)->cfs;
5934 util = READ_ONCE(cfs_rq->avg.util_avg);
5936 if (sched_feat(UTIL_EST))
5937 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
5939 return min_t(unsigned long, util, capacity_orig_of(cpu));
5943 * cpu_util_without: compute cpu utilization without any contributions from *p
5944 * @cpu: the CPU which utilization is requested
5945 * @p: the task which utilization should be discounted
5947 * The utilization of a CPU is defined by the utilization of tasks currently
5948 * enqueued on that CPU as well as tasks which are currently sleeping after an
5949 * execution on that CPU.
5951 * This method returns the utilization of the specified CPU by discounting the
5952 * utilization of the specified task, whenever the task is currently
5953 * contributing to the CPU utilization.
5955 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
5957 struct cfs_rq *cfs_rq;
5960 /* Task has no contribution or is new */
5961 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
5962 return cpu_util(cpu);
5964 cfs_rq = &cpu_rq(cpu)->cfs;
5965 util = READ_ONCE(cfs_rq->avg.util_avg);
5967 /* Discount task's util from CPU's util */
5968 lsub_positive(&util, task_util(p));
5973 * a) if *p is the only task sleeping on this CPU, then:
5974 * cpu_util (== task_util) > util_est (== 0)
5975 * and thus we return:
5976 * cpu_util_without = (cpu_util - task_util) = 0
5978 * b) if other tasks are SLEEPING on this CPU, which is now exiting
5980 * cpu_util >= task_util
5981 * cpu_util > util_est (== 0)
5982 * and thus we discount *p's blocked utilization to return:
5983 * cpu_util_without = (cpu_util - task_util) >= 0
5985 * c) if other tasks are RUNNABLE on that CPU and
5986 * util_est > cpu_util
5987 * then we use util_est since it returns a more restrictive
5988 * estimation of the spare capacity on that CPU, by just
5989 * considering the expected utilization of tasks already
5990 * runnable on that CPU.
5992 * Cases a) and b) are covered by the above code, while case c) is
5993 * covered by the following code when estimated utilization is
5996 if (sched_feat(UTIL_EST)) {
5997 unsigned int estimated =
5998 READ_ONCE(cfs_rq->avg.util_est.enqueued);
6001 * Despite the following checks we still have a small window
6002 * for a possible race, when an execl's select_task_rq_fair()
6003 * races with LB's detach_task():
6006 * p->on_rq = TASK_ON_RQ_MIGRATING;
6007 * ---------------------------------- A
6008 * deactivate_task() \
6009 * dequeue_task() + RaceTime
6010 * util_est_dequeue() /
6011 * ---------------------------------- B
6013 * The additional check on "current == p" it's required to
6014 * properly fix the execl regression and it helps in further
6015 * reducing the chances for the above race.
6017 if (unlikely(task_on_rq_queued(p) || current == p))
6018 lsub_positive(&estimated, _task_util_est(p));
6020 util = max(util, estimated);
6024 * Utilization (estimated) can exceed the CPU capacity, thus let's
6025 * clamp to the maximum CPU capacity to ensure consistency with
6026 * the cpu_util call.
6028 return min_t(unsigned long, util, capacity_orig_of(cpu));
6032 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6033 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6035 * In that case WAKE_AFFINE doesn't make sense and we'll let
6036 * BALANCE_WAKE sort things out.
6038 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6040 long min_cap, max_cap;
6042 if (!static_branch_unlikely(&sched_asym_cpucapacity))
6045 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6046 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6048 /* Minimum capacity is close to max, no need to abort wake_affine */
6049 if (max_cap - min_cap < max_cap >> 3)
6052 /* Bring task utilization in sync with prev_cpu */
6053 sync_entity_load_avg(&p->se);
6055 return !task_fits_capacity(p, min_cap);
6059 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6062 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6064 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6065 unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
6068 * If @p migrates from @cpu to another, remove its contribution. Or,
6069 * if @p migrates from another CPU to @cpu, add its contribution. In
6070 * the other cases, @cpu is not impacted by the migration, so the
6071 * util_avg should already be correct.
6073 if (task_cpu(p) == cpu && dst_cpu != cpu)
6074 sub_positive(&util, task_util(p));
6075 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6076 util += task_util(p);
6078 if (sched_feat(UTIL_EST)) {
6079 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6082 * During wake-up, the task isn't enqueued yet and doesn't
6083 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6084 * so just add it (if needed) to "simulate" what will be
6085 * cpu_util() after the task has been enqueued.
6088 util_est += _task_util_est(p);
6090 util = max(util, util_est);
6093 return min(util, capacity_orig_of(cpu));
6097 * compute_energy(): Estimates the energy that @pd would consume if @p was
6098 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6099 * landscape of @pd's CPUs after the task migration, and uses the Energy Model
6100 * to compute what would be the energy if we decided to actually migrate that
6104 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
6106 struct cpumask *pd_mask = perf_domain_span(pd);
6107 unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask));
6108 unsigned long max_util = 0, sum_util = 0;
6112 * The capacity state of CPUs of the current rd can be driven by CPUs
6113 * of another rd if they belong to the same pd. So, account for the
6114 * utilization of these CPUs too by masking pd with cpu_online_mask
6115 * instead of the rd span.
6117 * If an entire pd is outside of the current rd, it will not appear in
6118 * its pd list and will not be accounted by compute_energy().
6120 for_each_cpu_and(cpu, pd_mask, cpu_online_mask) {
6121 unsigned long cpu_util, util_cfs = cpu_util_next(cpu, p, dst_cpu);
6122 struct task_struct *tsk = cpu == dst_cpu ? p : NULL;
6125 * Busy time computation: utilization clamping is not
6126 * required since the ratio (sum_util / cpu_capacity)
6127 * is already enough to scale the EM reported power
6128 * consumption at the (eventually clamped) cpu_capacity.
6130 sum_util += schedutil_cpu_util(cpu, util_cfs, cpu_cap,
6134 * Performance domain frequency: utilization clamping
6135 * must be considered since it affects the selection
6136 * of the performance domain frequency.
6137 * NOTE: in case RT tasks are running, by default the
6138 * FREQUENCY_UTIL's utilization can be max OPP.
6140 cpu_util = schedutil_cpu_util(cpu, util_cfs, cpu_cap,
6141 FREQUENCY_UTIL, tsk);
6142 max_util = max(max_util, cpu_util);
6145 return em_pd_energy(pd->em_pd, max_util, sum_util);
6149 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6150 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6151 * spare capacity in each performance domain and uses it as a potential
6152 * candidate to execute the task. Then, it uses the Energy Model to figure
6153 * out which of the CPU candidates is the most energy-efficient.
6155 * The rationale for this heuristic is as follows. In a performance domain,
6156 * all the most energy efficient CPU candidates (according to the Energy
6157 * Model) are those for which we'll request a low frequency. When there are
6158 * several CPUs for which the frequency request will be the same, we don't
6159 * have enough data to break the tie between them, because the Energy Model
6160 * only includes active power costs. With this model, if we assume that
6161 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6162 * the maximum spare capacity in a performance domain is guaranteed to be among
6163 * the best candidates of the performance domain.
6165 * In practice, it could be preferable from an energy standpoint to pack
6166 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6167 * but that could also hurt our chances to go cluster idle, and we have no
6168 * ways to tell with the current Energy Model if this is actually a good
6169 * idea or not. So, find_energy_efficient_cpu() basically favors
6170 * cluster-packing, and spreading inside a cluster. That should at least be
6171 * a good thing for latency, and this is consistent with the idea that most
6172 * of the energy savings of EAS come from the asymmetry of the system, and
6173 * not so much from breaking the tie between identical CPUs. That's also the
6174 * reason why EAS is enabled in the topology code only for systems where
6175 * SD_ASYM_CPUCAPACITY is set.
6177 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6178 * they don't have any useful utilization data yet and it's not possible to
6179 * forecast their impact on energy consumption. Consequently, they will be
6180 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6181 * to be energy-inefficient in some use-cases. The alternative would be to
6182 * bias new tasks towards specific types of CPUs first, or to try to infer
6183 * their util_avg from the parent task, but those heuristics could hurt
6184 * other use-cases too. So, until someone finds a better way to solve this,
6185 * let's keep things simple by re-using the existing slow path.
6187 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6189 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
6190 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6191 unsigned long cpu_cap, util, base_energy = 0;
6192 int cpu, best_energy_cpu = prev_cpu;
6193 struct sched_domain *sd;
6194 struct perf_domain *pd;
6197 pd = rcu_dereference(rd->pd);
6198 if (!pd || READ_ONCE(rd->overutilized))
6202 * Energy-aware wake-up happens on the lowest sched_domain starting
6203 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6205 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6206 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6211 sync_entity_load_avg(&p->se);
6212 if (!task_util_est(p))
6215 for (; pd; pd = pd->next) {
6216 unsigned long cur_delta, spare_cap, max_spare_cap = 0;
6217 unsigned long base_energy_pd;
6218 int max_spare_cap_cpu = -1;
6220 /* Compute the 'base' energy of the pd, without @p */
6221 base_energy_pd = compute_energy(p, -1, pd);
6222 base_energy += base_energy_pd;
6224 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
6225 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6228 /* Skip CPUs that will be overutilized. */
6229 util = cpu_util_next(cpu, p, cpu);
6230 cpu_cap = capacity_of(cpu);
6231 if (!fits_capacity(util, cpu_cap))
6234 /* Always use prev_cpu as a candidate. */
6235 if (cpu == prev_cpu) {
6236 prev_delta = compute_energy(p, prev_cpu, pd);
6237 prev_delta -= base_energy_pd;
6238 best_delta = min(best_delta, prev_delta);
6242 * Find the CPU with the maximum spare capacity in
6243 * the performance domain
6245 spare_cap = cpu_cap - util;
6246 if (spare_cap > max_spare_cap) {
6247 max_spare_cap = spare_cap;
6248 max_spare_cap_cpu = cpu;
6252 /* Evaluate the energy impact of using this CPU. */
6253 if (max_spare_cap_cpu >= 0 && max_spare_cap_cpu != prev_cpu) {
6254 cur_delta = compute_energy(p, max_spare_cap_cpu, pd);
6255 cur_delta -= base_energy_pd;
6256 if (cur_delta < best_delta) {
6257 best_delta = cur_delta;
6258 best_energy_cpu = max_spare_cap_cpu;
6266 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6267 * least 6% of the energy used by prev_cpu.
6269 if (prev_delta == ULONG_MAX)
6270 return best_energy_cpu;
6272 if ((prev_delta - best_delta) > ((prev_delta + base_energy) >> 4))
6273 return best_energy_cpu;
6284 * select_task_rq_fair: Select target runqueue for the waking task in domains
6285 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6286 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6288 * Balances load by selecting the idlest CPU in the idlest group, or under
6289 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6291 * Returns the target CPU number.
6293 * preempt must be disabled.
6296 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6298 struct sched_domain *tmp, *sd = NULL;
6299 int cpu = smp_processor_id();
6300 int new_cpu = prev_cpu;
6301 int want_affine = 0;
6302 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6304 if (sd_flag & SD_BALANCE_WAKE) {
6307 if (sched_energy_enabled()) {
6308 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
6314 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu) &&
6315 cpumask_test_cpu(cpu, p->cpus_ptr);
6319 for_each_domain(cpu, tmp) {
6320 if (!(tmp->flags & SD_LOAD_BALANCE))
6324 * If both 'cpu' and 'prev_cpu' are part of this domain,
6325 * cpu is a valid SD_WAKE_AFFINE target.
6327 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6328 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6329 if (cpu != prev_cpu)
6330 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6332 sd = NULL; /* Prefer wake_affine over balance flags */
6336 if (tmp->flags & sd_flag)
6338 else if (!want_affine)
6344 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6345 } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6348 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6351 current->recent_used_cpu = cpu;
6358 static void detach_entity_cfs_rq(struct sched_entity *se);
6361 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6362 * cfs_rq_of(p) references at time of call are still valid and identify the
6363 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6365 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6368 * As blocked tasks retain absolute vruntime the migration needs to
6369 * deal with this by subtracting the old and adding the new
6370 * min_vruntime -- the latter is done by enqueue_entity() when placing
6371 * the task on the new runqueue.
6373 if (p->state == TASK_WAKING) {
6374 struct sched_entity *se = &p->se;
6375 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6378 #ifndef CONFIG_64BIT
6379 u64 min_vruntime_copy;
6382 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6384 min_vruntime = cfs_rq->min_vruntime;
6385 } while (min_vruntime != min_vruntime_copy);
6387 min_vruntime = cfs_rq->min_vruntime;
6390 se->vruntime -= min_vruntime;
6393 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6395 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6396 * rq->lock and can modify state directly.
6398 lockdep_assert_held(&task_rq(p)->lock);
6399 detach_entity_cfs_rq(&p->se);
6403 * We are supposed to update the task to "current" time, then
6404 * its up to date and ready to go to new CPU/cfs_rq. But we
6405 * have difficulty in getting what current time is, so simply
6406 * throw away the out-of-date time. This will result in the
6407 * wakee task is less decayed, but giving the wakee more load
6410 remove_entity_load_avg(&p->se);
6413 /* Tell new CPU we are migrated */
6414 p->se.avg.last_update_time = 0;
6416 /* We have migrated, no longer consider this task hot */
6417 p->se.exec_start = 0;
6419 update_scan_period(p, new_cpu);
6422 static void task_dead_fair(struct task_struct *p)
6424 remove_entity_load_avg(&p->se);
6426 #endif /* CONFIG_SMP */
6428 static unsigned long wakeup_gran(struct sched_entity *se)
6430 unsigned long gran = sysctl_sched_wakeup_granularity;
6433 * Since its curr running now, convert the gran from real-time
6434 * to virtual-time in his units.
6436 * By using 'se' instead of 'curr' we penalize light tasks, so
6437 * they get preempted easier. That is, if 'se' < 'curr' then
6438 * the resulting gran will be larger, therefore penalizing the
6439 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6440 * be smaller, again penalizing the lighter task.
6442 * This is especially important for buddies when the leftmost
6443 * task is higher priority than the buddy.
6445 return calc_delta_fair(gran, se);
6449 * Should 'se' preempt 'curr'.
6463 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6465 s64 gran, vdiff = curr->vruntime - se->vruntime;
6470 gran = wakeup_gran(se);
6477 static void set_last_buddy(struct sched_entity *se)
6479 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6482 for_each_sched_entity(se) {
6483 if (SCHED_WARN_ON(!se->on_rq))
6485 cfs_rq_of(se)->last = se;
6489 static void set_next_buddy(struct sched_entity *se)
6491 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6494 for_each_sched_entity(se) {
6495 if (SCHED_WARN_ON(!se->on_rq))
6497 cfs_rq_of(se)->next = se;
6501 static void set_skip_buddy(struct sched_entity *se)
6503 for_each_sched_entity(se)
6504 cfs_rq_of(se)->skip = se;
6508 * Preempt the current task with a newly woken task if needed:
6510 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6512 struct task_struct *curr = rq->curr;
6513 struct sched_entity *se = &curr->se, *pse = &p->se;
6514 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6515 int scale = cfs_rq->nr_running >= sched_nr_latency;
6516 int next_buddy_marked = 0;
6518 if (unlikely(se == pse))
6522 * This is possible from callers such as attach_tasks(), in which we
6523 * unconditionally check_prempt_curr() after an enqueue (which may have
6524 * lead to a throttle). This both saves work and prevents false
6525 * next-buddy nomination below.
6527 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6530 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6531 set_next_buddy(pse);
6532 next_buddy_marked = 1;
6536 * We can come here with TIF_NEED_RESCHED already set from new task
6539 * Note: this also catches the edge-case of curr being in a throttled
6540 * group (e.g. via set_curr_task), since update_curr() (in the
6541 * enqueue of curr) will have resulted in resched being set. This
6542 * prevents us from potentially nominating it as a false LAST_BUDDY
6545 if (test_tsk_need_resched(curr))
6548 /* Idle tasks are by definition preempted by non-idle tasks. */
6549 if (unlikely(task_has_idle_policy(curr)) &&
6550 likely(!task_has_idle_policy(p)))
6554 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6555 * is driven by the tick):
6557 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6560 find_matching_se(&se, &pse);
6561 update_curr(cfs_rq_of(se));
6563 if (wakeup_preempt_entity(se, pse) == 1) {
6565 * Bias pick_next to pick the sched entity that is
6566 * triggering this preemption.
6568 if (!next_buddy_marked)
6569 set_next_buddy(pse);
6578 * Only set the backward buddy when the current task is still
6579 * on the rq. This can happen when a wakeup gets interleaved
6580 * with schedule on the ->pre_schedule() or idle_balance()
6581 * point, either of which can * drop the rq lock.
6583 * Also, during early boot the idle thread is in the fair class,
6584 * for obvious reasons its a bad idea to schedule back to it.
6586 if (unlikely(!se->on_rq || curr == rq->idle))
6589 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6593 static struct task_struct *
6594 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6596 struct cfs_rq *cfs_rq = &rq->cfs;
6597 struct sched_entity *se;
6598 struct task_struct *p;
6602 if (!cfs_rq->nr_running)
6605 #ifdef CONFIG_FAIR_GROUP_SCHED
6606 if (!prev || prev->sched_class != &fair_sched_class)
6610 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6611 * likely that a next task is from the same cgroup as the current.
6613 * Therefore attempt to avoid putting and setting the entire cgroup
6614 * hierarchy, only change the part that actually changes.
6618 struct sched_entity *curr = cfs_rq->curr;
6621 * Since we got here without doing put_prev_entity() we also
6622 * have to consider cfs_rq->curr. If it is still a runnable
6623 * entity, update_curr() will update its vruntime, otherwise
6624 * forget we've ever seen it.
6628 update_curr(cfs_rq);
6633 * This call to check_cfs_rq_runtime() will do the
6634 * throttle and dequeue its entity in the parent(s).
6635 * Therefore the nr_running test will indeed
6638 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6641 if (!cfs_rq->nr_running)
6648 se = pick_next_entity(cfs_rq, curr);
6649 cfs_rq = group_cfs_rq(se);
6655 * Since we haven't yet done put_prev_entity and if the selected task
6656 * is a different task than we started out with, try and touch the
6657 * least amount of cfs_rqs.
6660 struct sched_entity *pse = &prev->se;
6662 while (!(cfs_rq = is_same_group(se, pse))) {
6663 int se_depth = se->depth;
6664 int pse_depth = pse->depth;
6666 if (se_depth <= pse_depth) {
6667 put_prev_entity(cfs_rq_of(pse), pse);
6668 pse = parent_entity(pse);
6670 if (se_depth >= pse_depth) {
6671 set_next_entity(cfs_rq_of(se), se);
6672 se = parent_entity(se);
6676 put_prev_entity(cfs_rq, pse);
6677 set_next_entity(cfs_rq, se);
6684 put_prev_task(rq, prev);
6687 se = pick_next_entity(cfs_rq, NULL);
6688 set_next_entity(cfs_rq, se);
6689 cfs_rq = group_cfs_rq(se);
6694 done: __maybe_unused;
6697 * Move the next running task to the front of
6698 * the list, so our cfs_tasks list becomes MRU
6701 list_move(&p->se.group_node, &rq->cfs_tasks);
6704 if (hrtick_enabled(rq))
6705 hrtick_start_fair(rq, p);
6707 update_misfit_status(p, rq);
6715 new_tasks = newidle_balance(rq, rf);
6718 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
6719 * possible for any higher priority task to appear. In that case we
6720 * must re-start the pick_next_entity() loop.
6729 * rq is about to be idle, check if we need to update the
6730 * lost_idle_time of clock_pelt
6732 update_idle_rq_clock_pelt(rq);
6738 * Account for a descheduled task:
6740 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6742 struct sched_entity *se = &prev->se;
6743 struct cfs_rq *cfs_rq;
6745 for_each_sched_entity(se) {
6746 cfs_rq = cfs_rq_of(se);
6747 put_prev_entity(cfs_rq, se);
6752 * sched_yield() is very simple
6754 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6756 static void yield_task_fair(struct rq *rq)
6758 struct task_struct *curr = rq->curr;
6759 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6760 struct sched_entity *se = &curr->se;
6763 * Are we the only task in the tree?
6765 if (unlikely(rq->nr_running == 1))
6768 clear_buddies(cfs_rq, se);
6770 if (curr->policy != SCHED_BATCH) {
6771 update_rq_clock(rq);
6773 * Update run-time statistics of the 'current'.
6775 update_curr(cfs_rq);
6777 * Tell update_rq_clock() that we've just updated,
6778 * so we don't do microscopic update in schedule()
6779 * and double the fastpath cost.
6781 rq_clock_skip_update(rq);
6787 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6789 struct sched_entity *se = &p->se;
6791 /* throttled hierarchies are not runnable */
6792 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6795 /* Tell the scheduler that we'd really like pse to run next. */
6798 yield_task_fair(rq);
6804 /**************************************************
6805 * Fair scheduling class load-balancing methods.
6809 * The purpose of load-balancing is to achieve the same basic fairness the
6810 * per-CPU scheduler provides, namely provide a proportional amount of compute
6811 * time to each task. This is expressed in the following equation:
6813 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6815 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
6816 * W_i,0 is defined as:
6818 * W_i,0 = \Sum_j w_i,j (2)
6820 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
6821 * is derived from the nice value as per sched_prio_to_weight[].
6823 * The weight average is an exponential decay average of the instantaneous
6826 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6828 * C_i is the compute capacity of CPU i, typically it is the
6829 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6830 * can also include other factors [XXX].
6832 * To achieve this balance we define a measure of imbalance which follows
6833 * directly from (1):
6835 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6837 * We them move tasks around to minimize the imbalance. In the continuous
6838 * function space it is obvious this converges, in the discrete case we get
6839 * a few fun cases generally called infeasible weight scenarios.
6842 * - infeasible weights;
6843 * - local vs global optima in the discrete case. ]
6848 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6849 * for all i,j solution, we create a tree of CPUs that follows the hardware
6850 * topology where each level pairs two lower groups (or better). This results
6851 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
6852 * tree to only the first of the previous level and we decrease the frequency
6853 * of load-balance at each level inv. proportional to the number of CPUs in
6859 * \Sum { --- * --- * 2^i } = O(n) (5)
6861 * `- size of each group
6862 * | | `- number of CPUs doing load-balance
6864 * `- sum over all levels
6866 * Coupled with a limit on how many tasks we can migrate every balance pass,
6867 * this makes (5) the runtime complexity of the balancer.
6869 * An important property here is that each CPU is still (indirectly) connected
6870 * to every other CPU in at most O(log n) steps:
6872 * The adjacency matrix of the resulting graph is given by:
6875 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6878 * And you'll find that:
6880 * A^(log_2 n)_i,j != 0 for all i,j (7)
6882 * Showing there's indeed a path between every CPU in at most O(log n) steps.
6883 * The task movement gives a factor of O(m), giving a convergence complexity
6886 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6891 * In order to avoid CPUs going idle while there's still work to do, new idle
6892 * balancing is more aggressive and has the newly idle CPU iterate up the domain
6893 * tree itself instead of relying on other CPUs to bring it work.
6895 * This adds some complexity to both (5) and (8) but it reduces the total idle
6903 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6906 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6911 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6913 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
6915 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6918 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6919 * rewrite all of this once again.]
6922 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6924 enum fbq_type { regular, remote, all };
6927 * group_type describes the group of CPUs at the moment of the load balance.
6928 * The enum is ordered by pulling priority, with the group with lowest priority
6929 * first so the groupe_type can be simply compared when selecting the busiest
6930 * group. see update_sd_pick_busiest().
6933 group_has_spare = 0,
6941 enum migration_type {
6948 #define LBF_ALL_PINNED 0x01
6949 #define LBF_NEED_BREAK 0x02
6950 #define LBF_DST_PINNED 0x04
6951 #define LBF_SOME_PINNED 0x08
6952 #define LBF_NOHZ_STATS 0x10
6953 #define LBF_NOHZ_AGAIN 0x20
6956 struct sched_domain *sd;
6964 struct cpumask *dst_grpmask;
6966 enum cpu_idle_type idle;
6968 /* The set of CPUs under consideration for load-balancing */
6969 struct cpumask *cpus;
6974 unsigned int loop_break;
6975 unsigned int loop_max;
6977 enum fbq_type fbq_type;
6978 enum migration_type migration_type;
6979 struct list_head tasks;
6983 * Is this task likely cache-hot:
6985 static int task_hot(struct task_struct *p, struct lb_env *env)
6989 lockdep_assert_held(&env->src_rq->lock);
6991 if (p->sched_class != &fair_sched_class)
6994 if (unlikely(task_has_idle_policy(p)))
6998 * Buddy candidates are cache hot:
7000 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7001 (&p->se == cfs_rq_of(&p->se)->next ||
7002 &p->se == cfs_rq_of(&p->se)->last))
7005 if (sysctl_sched_migration_cost == -1)
7007 if (sysctl_sched_migration_cost == 0)
7010 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7012 return delta < (s64)sysctl_sched_migration_cost;
7015 #ifdef CONFIG_NUMA_BALANCING
7017 * Returns 1, if task migration degrades locality
7018 * Returns 0, if task migration improves locality i.e migration preferred.
7019 * Returns -1, if task migration is not affected by locality.
7021 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7023 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7024 unsigned long src_weight, dst_weight;
7025 int src_nid, dst_nid, dist;
7027 if (!static_branch_likely(&sched_numa_balancing))
7030 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7033 src_nid = cpu_to_node(env->src_cpu);
7034 dst_nid = cpu_to_node(env->dst_cpu);
7036 if (src_nid == dst_nid)
7039 /* Migrating away from the preferred node is always bad. */
7040 if (src_nid == p->numa_preferred_nid) {
7041 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7047 /* Encourage migration to the preferred node. */
7048 if (dst_nid == p->numa_preferred_nid)
7051 /* Leaving a core idle is often worse than degrading locality. */
7052 if (env->idle == CPU_IDLE)
7055 dist = node_distance(src_nid, dst_nid);
7057 src_weight = group_weight(p, src_nid, dist);
7058 dst_weight = group_weight(p, dst_nid, dist);
7060 src_weight = task_weight(p, src_nid, dist);
7061 dst_weight = task_weight(p, dst_nid, dist);
7064 return dst_weight < src_weight;
7068 static inline int migrate_degrades_locality(struct task_struct *p,
7076 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7079 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7083 lockdep_assert_held(&env->src_rq->lock);
7086 * We do not migrate tasks that are:
7087 * 1) throttled_lb_pair, or
7088 * 2) cannot be migrated to this CPU due to cpus_ptr, or
7089 * 3) running (obviously), or
7090 * 4) are cache-hot on their current CPU.
7092 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7095 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
7098 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7100 env->flags |= LBF_SOME_PINNED;
7103 * Remember if this task can be migrated to any other CPU in
7104 * our sched_group. We may want to revisit it if we couldn't
7105 * meet load balance goals by pulling other tasks on src_cpu.
7107 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7108 * already computed one in current iteration.
7110 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7113 /* Prevent to re-select dst_cpu via env's CPUs: */
7114 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7115 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
7116 env->flags |= LBF_DST_PINNED;
7117 env->new_dst_cpu = cpu;
7125 /* Record that we found atleast one task that could run on dst_cpu */
7126 env->flags &= ~LBF_ALL_PINNED;
7128 if (task_running(env->src_rq, p)) {
7129 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7134 * Aggressive migration if:
7135 * 1) destination numa is preferred
7136 * 2) task is cache cold, or
7137 * 3) too many balance attempts have failed.
7139 tsk_cache_hot = migrate_degrades_locality(p, env);
7140 if (tsk_cache_hot == -1)
7141 tsk_cache_hot = task_hot(p, env);
7143 if (tsk_cache_hot <= 0 ||
7144 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7145 if (tsk_cache_hot == 1) {
7146 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7147 schedstat_inc(p->se.statistics.nr_forced_migrations);
7152 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7157 * detach_task() -- detach the task for the migration specified in env
7159 static void detach_task(struct task_struct *p, struct lb_env *env)
7161 lockdep_assert_held(&env->src_rq->lock);
7163 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7164 set_task_cpu(p, env->dst_cpu);
7168 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7169 * part of active balancing operations within "domain".
7171 * Returns a task if successful and NULL otherwise.
7173 static struct task_struct *detach_one_task(struct lb_env *env)
7175 struct task_struct *p;
7177 lockdep_assert_held(&env->src_rq->lock);
7179 list_for_each_entry_reverse(p,
7180 &env->src_rq->cfs_tasks, se.group_node) {
7181 if (!can_migrate_task(p, env))
7184 detach_task(p, env);
7187 * Right now, this is only the second place where
7188 * lb_gained[env->idle] is updated (other is detach_tasks)
7189 * so we can safely collect stats here rather than
7190 * inside detach_tasks().
7192 schedstat_inc(env->sd->lb_gained[env->idle]);
7198 static const unsigned int sched_nr_migrate_break = 32;
7201 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
7202 * busiest_rq, as part of a balancing operation within domain "sd".
7204 * Returns number of detached tasks if successful and 0 otherwise.
7206 static int detach_tasks(struct lb_env *env)
7208 struct list_head *tasks = &env->src_rq->cfs_tasks;
7209 unsigned long util, load;
7210 struct task_struct *p;
7213 lockdep_assert_held(&env->src_rq->lock);
7215 if (env->imbalance <= 0)
7218 while (!list_empty(tasks)) {
7220 * We don't want to steal all, otherwise we may be treated likewise,
7221 * which could at worst lead to a livelock crash.
7223 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7226 p = list_last_entry(tasks, struct task_struct, se.group_node);
7229 /* We've more or less seen every task there is, call it quits */
7230 if (env->loop > env->loop_max)
7233 /* take a breather every nr_migrate tasks */
7234 if (env->loop > env->loop_break) {
7235 env->loop_break += sched_nr_migrate_break;
7236 env->flags |= LBF_NEED_BREAK;
7240 if (!can_migrate_task(p, env))
7243 switch (env->migration_type) {
7245 load = task_h_load(p);
7247 if (sched_feat(LB_MIN) &&
7248 load < 16 && !env->sd->nr_balance_failed)
7251 if (load/2 > env->imbalance)
7254 env->imbalance -= load;
7258 util = task_util_est(p);
7260 if (util > env->imbalance)
7263 env->imbalance -= util;
7270 case migrate_misfit:
7271 /* This is not a misfit task */
7272 if (task_fits_capacity(p, capacity_of(env->src_cpu)))
7279 detach_task(p, env);
7280 list_add(&p->se.group_node, &env->tasks);
7284 #ifdef CONFIG_PREEMPTION
7286 * NEWIDLE balancing is a source of latency, so preemptible
7287 * kernels will stop after the first task is detached to minimize
7288 * the critical section.
7290 if (env->idle == CPU_NEWLY_IDLE)
7295 * We only want to steal up to the prescribed amount of
7298 if (env->imbalance <= 0)
7303 list_move(&p->se.group_node, tasks);
7307 * Right now, this is one of only two places we collect this stat
7308 * so we can safely collect detach_one_task() stats here rather
7309 * than inside detach_one_task().
7311 schedstat_add(env->sd->lb_gained[env->idle], detached);
7317 * attach_task() -- attach the task detached by detach_task() to its new rq.
7319 static void attach_task(struct rq *rq, struct task_struct *p)
7321 lockdep_assert_held(&rq->lock);
7323 BUG_ON(task_rq(p) != rq);
7324 activate_task(rq, p, ENQUEUE_NOCLOCK);
7325 check_preempt_curr(rq, p, 0);
7329 * attach_one_task() -- attaches the task returned from detach_one_task() to
7332 static void attach_one_task(struct rq *rq, struct task_struct *p)
7337 update_rq_clock(rq);
7343 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7346 static void attach_tasks(struct lb_env *env)
7348 struct list_head *tasks = &env->tasks;
7349 struct task_struct *p;
7352 rq_lock(env->dst_rq, &rf);
7353 update_rq_clock(env->dst_rq);
7355 while (!list_empty(tasks)) {
7356 p = list_first_entry(tasks, struct task_struct, se.group_node);
7357 list_del_init(&p->se.group_node);
7359 attach_task(env->dst_rq, p);
7362 rq_unlock(env->dst_rq, &rf);
7365 #ifdef CONFIG_NO_HZ_COMMON
7366 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7368 if (cfs_rq->avg.load_avg)
7371 if (cfs_rq->avg.util_avg)
7377 static inline bool others_have_blocked(struct rq *rq)
7379 if (READ_ONCE(rq->avg_rt.util_avg))
7382 if (READ_ONCE(rq->avg_dl.util_avg))
7385 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7386 if (READ_ONCE(rq->avg_irq.util_avg))
7393 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
7395 rq->last_blocked_load_update_tick = jiffies;
7398 rq->has_blocked_load = 0;
7401 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
7402 static inline bool others_have_blocked(struct rq *rq) { return false; }
7403 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
7406 #ifdef CONFIG_FAIR_GROUP_SCHED
7408 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7410 if (cfs_rq->load.weight)
7413 if (cfs_rq->avg.load_sum)
7416 if (cfs_rq->avg.util_sum)
7419 if (cfs_rq->avg.runnable_load_sum)
7425 static void update_blocked_averages(int cpu)
7427 struct rq *rq = cpu_rq(cpu);
7428 struct cfs_rq *cfs_rq, *pos;
7429 const struct sched_class *curr_class;
7433 rq_lock_irqsave(rq, &rf);
7434 update_rq_clock(rq);
7437 * Iterates the task_group tree in a bottom up fashion, see
7438 * list_add_leaf_cfs_rq() for details.
7440 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7441 struct sched_entity *se;
7443 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq))
7444 update_tg_load_avg(cfs_rq, 0);
7446 /* Propagate pending load changes to the parent, if any: */
7447 se = cfs_rq->tg->se[cpu];
7448 if (se && !skip_blocked_update(se))
7449 update_load_avg(cfs_rq_of(se), se, 0);
7452 * There can be a lot of idle CPU cgroups. Don't let fully
7453 * decayed cfs_rqs linger on the list.
7455 if (cfs_rq_is_decayed(cfs_rq))
7456 list_del_leaf_cfs_rq(cfs_rq);
7458 /* Don't need periodic decay once load/util_avg are null */
7459 if (cfs_rq_has_blocked(cfs_rq))
7463 curr_class = rq->curr->sched_class;
7464 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class);
7465 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class);
7466 update_irq_load_avg(rq, 0);
7467 /* Don't need periodic decay once load/util_avg are null */
7468 if (others_have_blocked(rq))
7471 update_blocked_load_status(rq, !done);
7472 rq_unlock_irqrestore(rq, &rf);
7476 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7477 * This needs to be done in a top-down fashion because the load of a child
7478 * group is a fraction of its parents load.
7480 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7482 struct rq *rq = rq_of(cfs_rq);
7483 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7484 unsigned long now = jiffies;
7487 if (cfs_rq->last_h_load_update == now)
7490 WRITE_ONCE(cfs_rq->h_load_next, NULL);
7491 for_each_sched_entity(se) {
7492 cfs_rq = cfs_rq_of(se);
7493 WRITE_ONCE(cfs_rq->h_load_next, se);
7494 if (cfs_rq->last_h_load_update == now)
7499 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7500 cfs_rq->last_h_load_update = now;
7503 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
7504 load = cfs_rq->h_load;
7505 load = div64_ul(load * se->avg.load_avg,
7506 cfs_rq_load_avg(cfs_rq) + 1);
7507 cfs_rq = group_cfs_rq(se);
7508 cfs_rq->h_load = load;
7509 cfs_rq->last_h_load_update = now;
7513 static unsigned long task_h_load(struct task_struct *p)
7515 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7517 update_cfs_rq_h_load(cfs_rq);
7518 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7519 cfs_rq_load_avg(cfs_rq) + 1);
7522 static inline void update_blocked_averages(int cpu)
7524 struct rq *rq = cpu_rq(cpu);
7525 struct cfs_rq *cfs_rq = &rq->cfs;
7526 const struct sched_class *curr_class;
7529 rq_lock_irqsave(rq, &rf);
7530 update_rq_clock(rq);
7531 update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
7533 curr_class = rq->curr->sched_class;
7534 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class);
7535 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class);
7536 update_irq_load_avg(rq, 0);
7537 update_blocked_load_status(rq, cfs_rq_has_blocked(cfs_rq) || others_have_blocked(rq));
7538 rq_unlock_irqrestore(rq, &rf);
7541 static unsigned long task_h_load(struct task_struct *p)
7543 return p->se.avg.load_avg;
7547 /********** Helpers for find_busiest_group ************************/
7550 * sg_lb_stats - stats of a sched_group required for load_balancing
7552 struct sg_lb_stats {
7553 unsigned long avg_load; /*Avg load across the CPUs of the group */
7554 unsigned long group_load; /* Total load over the CPUs of the group */
7555 unsigned long group_capacity;
7556 unsigned long group_util; /* Total utilization of the group */
7557 unsigned int sum_nr_running; /* Nr of tasks running in the group */
7558 unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
7559 unsigned int idle_cpus;
7560 unsigned int group_weight;
7561 enum group_type group_type;
7562 unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
7563 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
7564 #ifdef CONFIG_NUMA_BALANCING
7565 unsigned int nr_numa_running;
7566 unsigned int nr_preferred_running;
7571 * sd_lb_stats - Structure to store the statistics of a sched_domain
7572 * during load balancing.
7574 struct sd_lb_stats {
7575 struct sched_group *busiest; /* Busiest group in this sd */
7576 struct sched_group *local; /* Local group in this sd */
7577 unsigned long total_load; /* Total load of all groups in sd */
7578 unsigned long total_capacity; /* Total capacity of all groups in sd */
7579 unsigned long avg_load; /* Average load across all groups in sd */
7580 unsigned int prefer_sibling; /* tasks should go to sibling first */
7582 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7583 struct sg_lb_stats local_stat; /* Statistics of the local group */
7586 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7589 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7590 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7591 * We must however set busiest_stat::group_type and
7592 * busiest_stat::idle_cpus to the worst busiest group because
7593 * update_sd_pick_busiest() reads these before assignment.
7595 *sds = (struct sd_lb_stats){
7599 .total_capacity = 0UL,
7601 .idle_cpus = UINT_MAX,
7602 .group_type = group_has_spare,
7607 static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu)
7609 struct rq *rq = cpu_rq(cpu);
7610 unsigned long max = arch_scale_cpu_capacity(cpu);
7611 unsigned long used, free;
7614 irq = cpu_util_irq(rq);
7616 if (unlikely(irq >= max))
7619 used = READ_ONCE(rq->avg_rt.util_avg);
7620 used += READ_ONCE(rq->avg_dl.util_avg);
7622 if (unlikely(used >= max))
7627 return scale_irq_capacity(free, irq, max);
7630 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7632 unsigned long capacity = scale_rt_capacity(sd, cpu);
7633 struct sched_group *sdg = sd->groups;
7635 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
7640 cpu_rq(cpu)->cpu_capacity = capacity;
7641 sdg->sgc->capacity = capacity;
7642 sdg->sgc->min_capacity = capacity;
7643 sdg->sgc->max_capacity = capacity;
7646 void update_group_capacity(struct sched_domain *sd, int cpu)
7648 struct sched_domain *child = sd->child;
7649 struct sched_group *group, *sdg = sd->groups;
7650 unsigned long capacity, min_capacity, max_capacity;
7651 unsigned long interval;
7653 interval = msecs_to_jiffies(sd->balance_interval);
7654 interval = clamp(interval, 1UL, max_load_balance_interval);
7655 sdg->sgc->next_update = jiffies + interval;
7658 update_cpu_capacity(sd, cpu);
7663 min_capacity = ULONG_MAX;
7666 if (child->flags & SD_OVERLAP) {
7668 * SD_OVERLAP domains cannot assume that child groups
7669 * span the current group.
7672 for_each_cpu(cpu, sched_group_span(sdg)) {
7673 struct sched_group_capacity *sgc;
7674 struct rq *rq = cpu_rq(cpu);
7677 * build_sched_domains() -> init_sched_groups_capacity()
7678 * gets here before we've attached the domains to the
7681 * Use capacity_of(), which is set irrespective of domains
7682 * in update_cpu_capacity().
7684 * This avoids capacity from being 0 and
7685 * causing divide-by-zero issues on boot.
7687 if (unlikely(!rq->sd)) {
7688 capacity += capacity_of(cpu);
7690 sgc = rq->sd->groups->sgc;
7691 capacity += sgc->capacity;
7694 min_capacity = min(capacity, min_capacity);
7695 max_capacity = max(capacity, max_capacity);
7699 * !SD_OVERLAP domains can assume that child groups
7700 * span the current group.
7703 group = child->groups;
7705 struct sched_group_capacity *sgc = group->sgc;
7707 capacity += sgc->capacity;
7708 min_capacity = min(sgc->min_capacity, min_capacity);
7709 max_capacity = max(sgc->max_capacity, max_capacity);
7710 group = group->next;
7711 } while (group != child->groups);
7714 sdg->sgc->capacity = capacity;
7715 sdg->sgc->min_capacity = min_capacity;
7716 sdg->sgc->max_capacity = max_capacity;
7720 * Check whether the capacity of the rq has been noticeably reduced by side
7721 * activity. The imbalance_pct is used for the threshold.
7722 * Return true is the capacity is reduced
7725 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7727 return ((rq->cpu_capacity * sd->imbalance_pct) <
7728 (rq->cpu_capacity_orig * 100));
7732 * Check whether a rq has a misfit task and if it looks like we can actually
7733 * help that task: we can migrate the task to a CPU of higher capacity, or
7734 * the task's current CPU is heavily pressured.
7736 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
7738 return rq->misfit_task_load &&
7739 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
7740 check_cpu_capacity(rq, sd));
7744 * Group imbalance indicates (and tries to solve) the problem where balancing
7745 * groups is inadequate due to ->cpus_ptr constraints.
7747 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
7748 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
7751 * { 0 1 2 3 } { 4 5 6 7 }
7754 * If we were to balance group-wise we'd place two tasks in the first group and
7755 * two tasks in the second group. Clearly this is undesired as it will overload
7756 * cpu 3 and leave one of the CPUs in the second group unused.
7758 * The current solution to this issue is detecting the skew in the first group
7759 * by noticing the lower domain failed to reach balance and had difficulty
7760 * moving tasks due to affinity constraints.
7762 * When this is so detected; this group becomes a candidate for busiest; see
7763 * update_sd_pick_busiest(). And calculate_imbalance() and
7764 * find_busiest_group() avoid some of the usual balance conditions to allow it
7765 * to create an effective group imbalance.
7767 * This is a somewhat tricky proposition since the next run might not find the
7768 * group imbalance and decide the groups need to be balanced again. A most
7769 * subtle and fragile situation.
7772 static inline int sg_imbalanced(struct sched_group *group)
7774 return group->sgc->imbalance;
7778 * group_has_capacity returns true if the group has spare capacity that could
7779 * be used by some tasks.
7780 * We consider that a group has spare capacity if the * number of task is
7781 * smaller than the number of CPUs or if the utilization is lower than the
7782 * available capacity for CFS tasks.
7783 * For the latter, we use a threshold to stabilize the state, to take into
7784 * account the variance of the tasks' load and to return true if the available
7785 * capacity in meaningful for the load balancer.
7786 * As an example, an available capacity of 1% can appear but it doesn't make
7787 * any benefit for the load balance.
7790 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
7792 if (sgs->sum_nr_running < sgs->group_weight)
7795 if ((sgs->group_capacity * 100) >
7796 (sgs->group_util * imbalance_pct))
7803 * group_is_overloaded returns true if the group has more tasks than it can
7805 * group_is_overloaded is not equals to !group_has_capacity because a group
7806 * with the exact right number of tasks, has no more spare capacity but is not
7807 * overloaded so both group_has_capacity and group_is_overloaded return
7811 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
7813 if (sgs->sum_nr_running <= sgs->group_weight)
7816 if ((sgs->group_capacity * 100) <
7817 (sgs->group_util * imbalance_pct))
7824 * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller
7825 * per-CPU capacity than sched_group ref.
7828 group_smaller_min_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7830 return fits_capacity(sg->sgc->min_capacity, ref->sgc->min_capacity);
7834 * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller
7835 * per-CPU capacity_orig than sched_group ref.
7838 group_smaller_max_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7840 return fits_capacity(sg->sgc->max_capacity, ref->sgc->max_capacity);
7844 group_type group_classify(unsigned int imbalance_pct,
7845 struct sched_group *group,
7846 struct sg_lb_stats *sgs)
7848 if (group_is_overloaded(imbalance_pct, sgs))
7849 return group_overloaded;
7851 if (sg_imbalanced(group))
7852 return group_imbalanced;
7854 if (sgs->group_asym_packing)
7855 return group_asym_packing;
7857 if (sgs->group_misfit_task_load)
7858 return group_misfit_task;
7860 if (!group_has_capacity(imbalance_pct, sgs))
7861 return group_fully_busy;
7863 return group_has_spare;
7866 static bool update_nohz_stats(struct rq *rq, bool force)
7868 #ifdef CONFIG_NO_HZ_COMMON
7869 unsigned int cpu = rq->cpu;
7871 if (!rq->has_blocked_load)
7874 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
7877 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
7880 update_blocked_averages(cpu);
7882 return rq->has_blocked_load;
7889 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7890 * @env: The load balancing environment.
7891 * @group: sched_group whose statistics are to be updated.
7892 * @sgs: variable to hold the statistics for this group.
7893 * @sg_status: Holds flag indicating the status of the sched_group
7895 static inline void update_sg_lb_stats(struct lb_env *env,
7896 struct sched_group *group,
7897 struct sg_lb_stats *sgs,
7900 int i, nr_running, local_group;
7902 memset(sgs, 0, sizeof(*sgs));
7904 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(group));
7906 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
7907 struct rq *rq = cpu_rq(i);
7909 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
7910 env->flags |= LBF_NOHZ_AGAIN;
7912 sgs->group_load += cpu_load(rq);
7913 sgs->group_util += cpu_util(i);
7914 sgs->sum_h_nr_running += rq->cfs.h_nr_running;
7916 nr_running = rq->nr_running;
7917 sgs->sum_nr_running += nr_running;
7920 *sg_status |= SG_OVERLOAD;
7922 if (cpu_overutilized(i))
7923 *sg_status |= SG_OVERUTILIZED;
7925 #ifdef CONFIG_NUMA_BALANCING
7926 sgs->nr_numa_running += rq->nr_numa_running;
7927 sgs->nr_preferred_running += rq->nr_preferred_running;
7930 * No need to call idle_cpu() if nr_running is not 0
7932 if (!nr_running && idle_cpu(i)) {
7934 /* Idle cpu can't have misfit task */
7941 /* Check for a misfit task on the cpu */
7942 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
7943 sgs->group_misfit_task_load < rq->misfit_task_load) {
7944 sgs->group_misfit_task_load = rq->misfit_task_load;
7945 *sg_status |= SG_OVERLOAD;
7949 /* Check if dst CPU is idle and preferred to this group */
7950 if (env->sd->flags & SD_ASYM_PACKING &&
7951 env->idle != CPU_NOT_IDLE &&
7952 sgs->sum_h_nr_running &&
7953 sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu)) {
7954 sgs->group_asym_packing = 1;
7957 sgs->group_capacity = group->sgc->capacity;
7959 sgs->group_weight = group->group_weight;
7961 sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
7963 /* Computing avg_load makes sense only when group is overloaded */
7964 if (sgs->group_type == group_overloaded)
7965 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
7966 sgs->group_capacity;
7970 * update_sd_pick_busiest - return 1 on busiest group
7971 * @env: The load balancing environment.
7972 * @sds: sched_domain statistics
7973 * @sg: sched_group candidate to be checked for being the busiest
7974 * @sgs: sched_group statistics
7976 * Determine if @sg is a busier group than the previously selected
7979 * Return: %true if @sg is a busier group than the previously selected
7980 * busiest group. %false otherwise.
7982 static bool update_sd_pick_busiest(struct lb_env *env,
7983 struct sd_lb_stats *sds,
7984 struct sched_group *sg,
7985 struct sg_lb_stats *sgs)
7987 struct sg_lb_stats *busiest = &sds->busiest_stat;
7989 /* Make sure that there is at least one task to pull */
7990 if (!sgs->sum_h_nr_running)
7994 * Don't try to pull misfit tasks we can't help.
7995 * We can use max_capacity here as reduction in capacity on some
7996 * CPUs in the group should either be possible to resolve
7997 * internally or be covered by avg_load imbalance (eventually).
7999 if (sgs->group_type == group_misfit_task &&
8000 (!group_smaller_max_cpu_capacity(sg, sds->local) ||
8001 sds->local_stat.group_type != group_has_spare))
8004 if (sgs->group_type > busiest->group_type)
8007 if (sgs->group_type < busiest->group_type)
8011 * The candidate and the current busiest group are the same type of
8012 * group. Let check which one is the busiest according to the type.
8015 switch (sgs->group_type) {
8016 case group_overloaded:
8017 /* Select the overloaded group with highest avg_load. */
8018 if (sgs->avg_load <= busiest->avg_load)
8022 case group_imbalanced:
8024 * Select the 1st imbalanced group as we don't have any way to
8025 * choose one more than another.
8029 case group_asym_packing:
8030 /* Prefer to move from lowest priority CPU's work */
8031 if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu))
8035 case group_misfit_task:
8037 * If we have more than one misfit sg go with the biggest
8040 if (sgs->group_misfit_task_load < busiest->group_misfit_task_load)
8044 case group_fully_busy:
8046 * Select the fully busy group with highest avg_load. In
8047 * theory, there is no need to pull task from such kind of
8048 * group because tasks have all compute capacity that they need
8049 * but we can still improve the overall throughput by reducing
8050 * contention when accessing shared HW resources.
8052 * XXX for now avg_load is not computed and always 0 so we
8053 * select the 1st one.
8055 if (sgs->avg_load <= busiest->avg_load)
8059 case group_has_spare:
8061 * Select not overloaded group with lowest number of
8062 * idle cpus. We could also compare the spare capacity
8063 * which is more stable but it can end up that the
8064 * group has less spare capacity but finally more idle
8065 * CPUs which means less opportunity to pull tasks.
8067 if (sgs->idle_cpus >= busiest->idle_cpus)
8073 * Candidate sg has no more than one task per CPU and has higher
8074 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
8075 * throughput. Maximize throughput, power/energy consequences are not
8078 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
8079 (sgs->group_type <= group_fully_busy) &&
8080 (group_smaller_min_cpu_capacity(sds->local, sg)))
8086 #ifdef CONFIG_NUMA_BALANCING
8087 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8089 if (sgs->sum_h_nr_running > sgs->nr_numa_running)
8091 if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
8096 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8098 if (rq->nr_running > rq->nr_numa_running)
8100 if (rq->nr_running > rq->nr_preferred_running)
8105 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8110 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8114 #endif /* CONFIG_NUMA_BALANCING */
8120 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
8121 * @denv: The ched_domain level to look for idlest group.
8122 * @group: sched_group whose statistics are to be updated.
8123 * @sgs: variable to hold the statistics for this group.
8125 static inline void update_sg_wakeup_stats(struct sched_domain *sd,
8126 struct sched_group *group,
8127 struct sg_lb_stats *sgs,
8128 struct task_struct *p)
8132 memset(sgs, 0, sizeof(*sgs));
8134 for_each_cpu(i, sched_group_span(group)) {
8135 struct rq *rq = cpu_rq(i);
8137 sgs->group_load += cpu_load(rq);
8138 sgs->group_util += cpu_util_without(i, p);
8139 sgs->sum_h_nr_running += rq->cfs.h_nr_running;
8141 nr_running = rq->nr_running;
8142 sgs->sum_nr_running += nr_running;
8145 * No need to call idle_cpu() if nr_running is not 0
8147 if (!nr_running && idle_cpu(i))
8153 /* Check if task fits in the group */
8154 if (sd->flags & SD_ASYM_CPUCAPACITY &&
8155 !task_fits_capacity(p, group->sgc->max_capacity)) {
8156 sgs->group_misfit_task_load = 1;
8159 sgs->group_capacity = group->sgc->capacity;
8161 sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
8164 * Computing avg_load makes sense only when group is fully busy or
8167 if (sgs->group_type < group_fully_busy)
8168 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
8169 sgs->group_capacity;
8172 static bool update_pick_idlest(struct sched_group *idlest,
8173 struct sg_lb_stats *idlest_sgs,
8174 struct sched_group *group,
8175 struct sg_lb_stats *sgs)
8177 if (sgs->group_type < idlest_sgs->group_type)
8180 if (sgs->group_type > idlest_sgs->group_type)
8184 * The candidate and the current idlest group are the same type of
8185 * group. Let check which one is the idlest according to the type.
8188 switch (sgs->group_type) {
8189 case group_overloaded:
8190 case group_fully_busy:
8191 /* Select the group with lowest avg_load. */
8192 if (idlest_sgs->avg_load <= sgs->avg_load)
8196 case group_imbalanced:
8197 case group_asym_packing:
8198 /* Those types are not used in the slow wakeup path */
8201 case group_misfit_task:
8202 /* Select group with the highest max capacity */
8203 if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
8207 case group_has_spare:
8208 /* Select group with most idle CPUs */
8209 if (idlest_sgs->idle_cpus >= sgs->idle_cpus)
8218 * find_idlest_group() finds and returns the least busy CPU group within the
8221 * Assumes p is allowed on at least one CPU in sd.
8223 static struct sched_group *
8224 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
8225 int this_cpu, int sd_flag)
8227 struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
8228 struct sg_lb_stats local_sgs, tmp_sgs;
8229 struct sg_lb_stats *sgs;
8230 unsigned long imbalance;
8231 struct sg_lb_stats idlest_sgs = {
8232 .avg_load = UINT_MAX,
8233 .group_type = group_overloaded,
8236 imbalance = scale_load_down(NICE_0_LOAD) *
8237 (sd->imbalance_pct-100) / 100;
8242 /* Skip over this group if it has no CPUs allowed */
8243 if (!cpumask_intersects(sched_group_span(group),
8247 local_group = cpumask_test_cpu(this_cpu,
8248 sched_group_span(group));
8257 update_sg_wakeup_stats(sd, group, sgs, p);
8259 if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
8264 } while (group = group->next, group != sd->groups);
8267 /* There is no idlest group to push tasks to */
8272 * If the local group is idler than the selected idlest group
8273 * don't try and push the task.
8275 if (local_sgs.group_type < idlest_sgs.group_type)
8279 * If the local group is busier than the selected idlest group
8280 * try and push the task.
8282 if (local_sgs.group_type > idlest_sgs.group_type)
8285 switch (local_sgs.group_type) {
8286 case group_overloaded:
8287 case group_fully_busy:
8289 * When comparing groups across NUMA domains, it's possible for
8290 * the local domain to be very lightly loaded relative to the
8291 * remote domains but "imbalance" skews the comparison making
8292 * remote CPUs look much more favourable. When considering
8293 * cross-domain, add imbalance to the load on the remote node
8294 * and consider staying local.
8297 if ((sd->flags & SD_NUMA) &&
8298 ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
8302 * If the local group is less loaded than the selected
8303 * idlest group don't try and push any tasks.
8305 if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
8308 if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
8312 case group_imbalanced:
8313 case group_asym_packing:
8314 /* Those type are not used in the slow wakeup path */
8317 case group_misfit_task:
8318 /* Select group with the highest max capacity */
8319 if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
8323 case group_has_spare:
8324 if (sd->flags & SD_NUMA) {
8325 #ifdef CONFIG_NUMA_BALANCING
8328 * If there is spare capacity at NUMA, try to select
8329 * the preferred node
8331 if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
8334 idlest_cpu = cpumask_first(sched_group_span(idlest));
8335 if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
8339 * Otherwise, keep the task on this node to stay close
8340 * its wakeup source and improve locality. If there is
8341 * a real need of migration, periodic load balance will
8344 if (local_sgs.idle_cpus)
8349 * Select group with highest number of idle CPUs. We could also
8350 * compare the utilization which is more stable but it can end
8351 * up that the group has less spare capacity but finally more
8352 * idle CPUs which means more opportunity to run task.
8354 if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
8363 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8364 * @env: The load balancing environment.
8365 * @sds: variable to hold the statistics for this sched_domain.
8368 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8370 struct sched_domain *child = env->sd->child;
8371 struct sched_group *sg = env->sd->groups;
8372 struct sg_lb_stats *local = &sds->local_stat;
8373 struct sg_lb_stats tmp_sgs;
8376 #ifdef CONFIG_NO_HZ_COMMON
8377 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
8378 env->flags |= LBF_NOHZ_STATS;
8382 struct sg_lb_stats *sgs = &tmp_sgs;
8385 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
8390 if (env->idle != CPU_NEWLY_IDLE ||
8391 time_after_eq(jiffies, sg->sgc->next_update))
8392 update_group_capacity(env->sd, env->dst_cpu);
8395 update_sg_lb_stats(env, sg, sgs, &sg_status);
8401 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8403 sds->busiest_stat = *sgs;
8407 /* Now, start updating sd_lb_stats */
8408 sds->total_load += sgs->group_load;
8409 sds->total_capacity += sgs->group_capacity;
8412 } while (sg != env->sd->groups);
8414 /* Tag domain that child domain prefers tasks go to siblings first */
8415 sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
8417 #ifdef CONFIG_NO_HZ_COMMON
8418 if ((env->flags & LBF_NOHZ_AGAIN) &&
8419 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8421 WRITE_ONCE(nohz.next_blocked,
8422 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8426 if (env->sd->flags & SD_NUMA)
8427 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8429 if (!env->sd->parent) {
8430 struct root_domain *rd = env->dst_rq->rd;
8432 /* update overload indicator if we are at root domain */
8433 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
8435 /* Update over-utilization (tipping point, U >= 0) indicator */
8436 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
8437 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
8438 } else if (sg_status & SG_OVERUTILIZED) {
8439 struct root_domain *rd = env->dst_rq->rd;
8441 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
8442 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
8447 * calculate_imbalance - Calculate the amount of imbalance present within the
8448 * groups of a given sched_domain during load balance.
8449 * @env: load balance environment
8450 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8452 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8454 struct sg_lb_stats *local, *busiest;
8456 local = &sds->local_stat;
8457 busiest = &sds->busiest_stat;
8459 if (busiest->group_type == group_misfit_task) {
8460 /* Set imbalance to allow misfit tasks to be balanced. */
8461 env->migration_type = migrate_misfit;
8466 if (busiest->group_type == group_asym_packing) {
8468 * In case of asym capacity, we will try to migrate all load to
8469 * the preferred CPU.
8471 env->migration_type = migrate_task;
8472 env->imbalance = busiest->sum_h_nr_running;
8476 if (busiest->group_type == group_imbalanced) {
8478 * In the group_imb case we cannot rely on group-wide averages
8479 * to ensure CPU-load equilibrium, try to move any task to fix
8480 * the imbalance. The next load balance will take care of
8481 * balancing back the system.
8483 env->migration_type = migrate_task;
8489 * Try to use spare capacity of local group without overloading it or
8492 if (local->group_type == group_has_spare) {
8493 if (busiest->group_type > group_fully_busy) {
8495 * If busiest is overloaded, try to fill spare
8496 * capacity. This might end up creating spare capacity
8497 * in busiest or busiest still being overloaded but
8498 * there is no simple way to directly compute the
8499 * amount of load to migrate in order to balance the
8502 env->migration_type = migrate_util;
8503 env->imbalance = max(local->group_capacity, local->group_util) -
8507 * In some cases, the group's utilization is max or even
8508 * higher than capacity because of migrations but the
8509 * local CPU is (newly) idle. There is at least one
8510 * waiting task in this overloaded busiest group. Let's
8513 if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) {
8514 env->migration_type = migrate_task;
8521 if (busiest->group_weight == 1 || sds->prefer_sibling) {
8522 unsigned int nr_diff = busiest->sum_nr_running;
8524 * When prefer sibling, evenly spread running tasks on
8527 env->migration_type = migrate_task;
8528 lsub_positive(&nr_diff, local->sum_nr_running);
8529 env->imbalance = nr_diff >> 1;
8534 * If there is no overload, we just want to even the number of
8537 env->migration_type = migrate_task;
8538 env->imbalance = max_t(long, 0, (local->idle_cpus -
8539 busiest->idle_cpus) >> 1);
8544 * Local is fully busy but has to take more load to relieve the
8547 if (local->group_type < group_overloaded) {
8549 * Local will become overloaded so the avg_load metrics are
8553 local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
8554 local->group_capacity;
8556 sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
8557 sds->total_capacity;
8561 * Both group are or will become overloaded and we're trying to get all
8562 * the CPUs to the average_load, so we don't want to push ourselves
8563 * above the average load, nor do we wish to reduce the max loaded CPU
8564 * below the average load. At the same time, we also don't want to
8565 * reduce the group load below the group capacity. Thus we look for
8566 * the minimum possible imbalance.
8568 env->migration_type = migrate_load;
8569 env->imbalance = min(
8570 (busiest->avg_load - sds->avg_load) * busiest->group_capacity,
8571 (sds->avg_load - local->avg_load) * local->group_capacity
8572 ) / SCHED_CAPACITY_SCALE;
8575 /******* find_busiest_group() helpers end here *********************/
8578 * Decision matrix according to the local and busiest group type:
8580 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
8581 * has_spare nr_idle balanced N/A N/A balanced balanced
8582 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
8583 * misfit_task force N/A N/A N/A force force
8584 * asym_packing force force N/A N/A force force
8585 * imbalanced force force N/A N/A force force
8586 * overloaded force force N/A N/A force avg_load
8588 * N/A : Not Applicable because already filtered while updating
8590 * balanced : The system is balanced for these 2 groups.
8591 * force : Calculate the imbalance as load migration is probably needed.
8592 * avg_load : Only if imbalance is significant enough.
8593 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
8594 * different in groups.
8598 * find_busiest_group - Returns the busiest group within the sched_domain
8599 * if there is an imbalance.
8601 * Also calculates the amount of runnable load which should be moved
8602 * to restore balance.
8604 * @env: The load balancing environment.
8606 * Return: - The busiest group if imbalance exists.
8608 static struct sched_group *find_busiest_group(struct lb_env *env)
8610 struct sg_lb_stats *local, *busiest;
8611 struct sd_lb_stats sds;
8613 init_sd_lb_stats(&sds);
8616 * Compute the various statistics relevant for load balancing at
8619 update_sd_lb_stats(env, &sds);
8621 if (sched_energy_enabled()) {
8622 struct root_domain *rd = env->dst_rq->rd;
8624 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
8628 local = &sds.local_stat;
8629 busiest = &sds.busiest_stat;
8631 /* There is no busy sibling group to pull tasks from */
8635 /* Misfit tasks should be dealt with regardless of the avg load */
8636 if (busiest->group_type == group_misfit_task)
8639 /* ASYM feature bypasses nice load balance check */
8640 if (busiest->group_type == group_asym_packing)
8644 * If the busiest group is imbalanced the below checks don't
8645 * work because they assume all things are equal, which typically
8646 * isn't true due to cpus_ptr constraints and the like.
8648 if (busiest->group_type == group_imbalanced)
8652 * If the local group is busier than the selected busiest group
8653 * don't try and pull any tasks.
8655 if (local->group_type > busiest->group_type)
8659 * When groups are overloaded, use the avg_load to ensure fairness
8662 if (local->group_type == group_overloaded) {
8664 * If the local group is more loaded than the selected
8665 * busiest group don't try to pull any tasks.
8667 if (local->avg_load >= busiest->avg_load)
8670 /* XXX broken for overlapping NUMA groups */
8671 sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
8675 * Don't pull any tasks if this group is already above the
8676 * domain average load.
8678 if (local->avg_load >= sds.avg_load)
8682 * If the busiest group is more loaded, use imbalance_pct to be
8685 if (100 * busiest->avg_load <=
8686 env->sd->imbalance_pct * local->avg_load)
8690 /* Try to move all excess tasks to child's sibling domain */
8691 if (sds.prefer_sibling && local->group_type == group_has_spare &&
8692 busiest->sum_nr_running > local->sum_nr_running + 1)
8695 if (busiest->group_type != group_overloaded) {
8696 if (env->idle == CPU_NOT_IDLE)
8698 * If the busiest group is not overloaded (and as a
8699 * result the local one too) but this CPU is already
8700 * busy, let another idle CPU try to pull task.
8704 if (busiest->group_weight > 1 &&
8705 local->idle_cpus <= (busiest->idle_cpus + 1))
8707 * If the busiest group is not overloaded
8708 * and there is no imbalance between this and busiest
8709 * group wrt idle CPUs, it is balanced. The imbalance
8710 * becomes significant if the diff is greater than 1
8711 * otherwise we might end up to just move the imbalance
8712 * on another group. Of course this applies only if
8713 * there is more than 1 CPU per group.
8717 if (busiest->sum_h_nr_running == 1)
8719 * busiest doesn't have any tasks waiting to run
8725 /* Looks like there is an imbalance. Compute it */
8726 calculate_imbalance(env, &sds);
8727 return env->imbalance ? sds.busiest : NULL;
8735 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
8737 static struct rq *find_busiest_queue(struct lb_env *env,
8738 struct sched_group *group)
8740 struct rq *busiest = NULL, *rq;
8741 unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
8742 unsigned int busiest_nr = 0;
8745 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8746 unsigned long capacity, load, util;
8747 unsigned int nr_running;
8751 rt = fbq_classify_rq(rq);
8754 * We classify groups/runqueues into three groups:
8755 * - regular: there are !numa tasks
8756 * - remote: there are numa tasks that run on the 'wrong' node
8757 * - all: there is no distinction
8759 * In order to avoid migrating ideally placed numa tasks,
8760 * ignore those when there's better options.
8762 * If we ignore the actual busiest queue to migrate another
8763 * task, the next balance pass can still reduce the busiest
8764 * queue by moving tasks around inside the node.
8766 * If we cannot move enough load due to this classification
8767 * the next pass will adjust the group classification and
8768 * allow migration of more tasks.
8770 * Both cases only affect the total convergence complexity.
8772 if (rt > env->fbq_type)
8775 capacity = capacity_of(i);
8776 nr_running = rq->cfs.h_nr_running;
8779 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
8780 * eventually lead to active_balancing high->low capacity.
8781 * Higher per-CPU capacity is considered better than balancing
8784 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8785 capacity_of(env->dst_cpu) < capacity &&
8789 switch (env->migration_type) {
8792 * When comparing with load imbalance, use cpu_load()
8793 * which is not scaled with the CPU capacity.
8795 load = cpu_load(rq);
8797 if (nr_running == 1 && load > env->imbalance &&
8798 !check_cpu_capacity(rq, env->sd))
8802 * For the load comparisons with the other CPUs,
8803 * consider the cpu_load() scaled with the CPU
8804 * capacity, so that the load can be moved away
8805 * from the CPU that is potentially running at a
8808 * Thus we're looking for max(load_i / capacity_i),
8809 * crosswise multiplication to rid ourselves of the
8810 * division works out to:
8811 * load_i * capacity_j > load_j * capacity_i;
8812 * where j is our previous maximum.
8814 if (load * busiest_capacity > busiest_load * capacity) {
8815 busiest_load = load;
8816 busiest_capacity = capacity;
8822 util = cpu_util(cpu_of(rq));
8824 if (busiest_util < util) {
8825 busiest_util = util;
8831 if (busiest_nr < nr_running) {
8832 busiest_nr = nr_running;
8837 case migrate_misfit:
8839 * For ASYM_CPUCAPACITY domains with misfit tasks we
8840 * simply seek the "biggest" misfit task.
8842 if (rq->misfit_task_load > busiest_load) {
8843 busiest_load = rq->misfit_task_load;
8856 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8857 * so long as it is large enough.
8859 #define MAX_PINNED_INTERVAL 512
8862 asym_active_balance(struct lb_env *env)
8865 * ASYM_PACKING needs to force migrate tasks from busy but
8866 * lower priority CPUs in order to pack all tasks in the
8867 * highest priority CPUs.
8869 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
8870 sched_asym_prefer(env->dst_cpu, env->src_cpu);
8874 voluntary_active_balance(struct lb_env *env)
8876 struct sched_domain *sd = env->sd;
8878 if (asym_active_balance(env))
8882 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8883 * It's worth migrating the task if the src_cpu's capacity is reduced
8884 * because of other sched_class or IRQs if more capacity stays
8885 * available on dst_cpu.
8887 if ((env->idle != CPU_NOT_IDLE) &&
8888 (env->src_rq->cfs.h_nr_running == 1)) {
8889 if ((check_cpu_capacity(env->src_rq, sd)) &&
8890 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8894 if (env->migration_type == migrate_misfit)
8900 static int need_active_balance(struct lb_env *env)
8902 struct sched_domain *sd = env->sd;
8904 if (voluntary_active_balance(env))
8907 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8910 static int active_load_balance_cpu_stop(void *data);
8912 static int should_we_balance(struct lb_env *env)
8914 struct sched_group *sg = env->sd->groups;
8915 int cpu, balance_cpu = -1;
8918 * Ensure the balancing environment is consistent; can happen
8919 * when the softirq triggers 'during' hotplug.
8921 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8925 * In the newly idle case, we will allow all the CPUs
8926 * to do the newly idle load balance.
8928 if (env->idle == CPU_NEWLY_IDLE)
8931 /* Try to find first idle CPU */
8932 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8940 if (balance_cpu == -1)
8941 balance_cpu = group_balance_cpu(sg);
8944 * First idle CPU or the first CPU(busiest) in this sched group
8945 * is eligible for doing load balancing at this and above domains.
8947 return balance_cpu == env->dst_cpu;
8951 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8952 * tasks if there is an imbalance.
8954 static int load_balance(int this_cpu, struct rq *this_rq,
8955 struct sched_domain *sd, enum cpu_idle_type idle,
8956 int *continue_balancing)
8958 int ld_moved, cur_ld_moved, active_balance = 0;
8959 struct sched_domain *sd_parent = sd->parent;
8960 struct sched_group *group;
8963 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8965 struct lb_env env = {
8967 .dst_cpu = this_cpu,
8969 .dst_grpmask = sched_group_span(sd->groups),
8971 .loop_break = sched_nr_migrate_break,
8974 .tasks = LIST_HEAD_INIT(env.tasks),
8977 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8979 schedstat_inc(sd->lb_count[idle]);
8982 if (!should_we_balance(&env)) {
8983 *continue_balancing = 0;
8987 group = find_busiest_group(&env);
8989 schedstat_inc(sd->lb_nobusyg[idle]);
8993 busiest = find_busiest_queue(&env, group);
8995 schedstat_inc(sd->lb_nobusyq[idle]);
8999 BUG_ON(busiest == env.dst_rq);
9001 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
9003 env.src_cpu = busiest->cpu;
9004 env.src_rq = busiest;
9007 if (busiest->nr_running > 1) {
9009 * Attempt to move tasks. If find_busiest_group has found
9010 * an imbalance but busiest->nr_running <= 1, the group is
9011 * still unbalanced. ld_moved simply stays zero, so it is
9012 * correctly treated as an imbalance.
9014 env.flags |= LBF_ALL_PINNED;
9015 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
9018 rq_lock_irqsave(busiest, &rf);
9019 update_rq_clock(busiest);
9022 * cur_ld_moved - load moved in current iteration
9023 * ld_moved - cumulative load moved across iterations
9025 cur_ld_moved = detach_tasks(&env);
9028 * We've detached some tasks from busiest_rq. Every
9029 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
9030 * unlock busiest->lock, and we are able to be sure
9031 * that nobody can manipulate the tasks in parallel.
9032 * See task_rq_lock() family for the details.
9035 rq_unlock(busiest, &rf);
9039 ld_moved += cur_ld_moved;
9042 local_irq_restore(rf.flags);
9044 if (env.flags & LBF_NEED_BREAK) {
9045 env.flags &= ~LBF_NEED_BREAK;
9050 * Revisit (affine) tasks on src_cpu that couldn't be moved to
9051 * us and move them to an alternate dst_cpu in our sched_group
9052 * where they can run. The upper limit on how many times we
9053 * iterate on same src_cpu is dependent on number of CPUs in our
9056 * This changes load balance semantics a bit on who can move
9057 * load to a given_cpu. In addition to the given_cpu itself
9058 * (or a ilb_cpu acting on its behalf where given_cpu is
9059 * nohz-idle), we now have balance_cpu in a position to move
9060 * load to given_cpu. In rare situations, this may cause
9061 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
9062 * _independently_ and at _same_ time to move some load to
9063 * given_cpu) causing exceess load to be moved to given_cpu.
9064 * This however should not happen so much in practice and
9065 * moreover subsequent load balance cycles should correct the
9066 * excess load moved.
9068 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
9070 /* Prevent to re-select dst_cpu via env's CPUs */
9071 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
9073 env.dst_rq = cpu_rq(env.new_dst_cpu);
9074 env.dst_cpu = env.new_dst_cpu;
9075 env.flags &= ~LBF_DST_PINNED;
9077 env.loop_break = sched_nr_migrate_break;
9080 * Go back to "more_balance" rather than "redo" since we
9081 * need to continue with same src_cpu.
9087 * We failed to reach balance because of affinity.
9090 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9092 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
9093 *group_imbalance = 1;
9096 /* All tasks on this runqueue were pinned by CPU affinity */
9097 if (unlikely(env.flags & LBF_ALL_PINNED)) {
9098 __cpumask_clear_cpu(cpu_of(busiest), cpus);
9100 * Attempting to continue load balancing at the current
9101 * sched_domain level only makes sense if there are
9102 * active CPUs remaining as possible busiest CPUs to
9103 * pull load from which are not contained within the
9104 * destination group that is receiving any migrated
9107 if (!cpumask_subset(cpus, env.dst_grpmask)) {
9109 env.loop_break = sched_nr_migrate_break;
9112 goto out_all_pinned;
9117 schedstat_inc(sd->lb_failed[idle]);
9119 * Increment the failure counter only on periodic balance.
9120 * We do not want newidle balance, which can be very
9121 * frequent, pollute the failure counter causing
9122 * excessive cache_hot migrations and active balances.
9124 if (idle != CPU_NEWLY_IDLE)
9125 sd->nr_balance_failed++;
9127 if (need_active_balance(&env)) {
9128 unsigned long flags;
9130 raw_spin_lock_irqsave(&busiest->lock, flags);
9133 * Don't kick the active_load_balance_cpu_stop,
9134 * if the curr task on busiest CPU can't be
9135 * moved to this_cpu:
9137 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
9138 raw_spin_unlock_irqrestore(&busiest->lock,
9140 env.flags |= LBF_ALL_PINNED;
9141 goto out_one_pinned;
9145 * ->active_balance synchronizes accesses to
9146 * ->active_balance_work. Once set, it's cleared
9147 * only after active load balance is finished.
9149 if (!busiest->active_balance) {
9150 busiest->active_balance = 1;
9151 busiest->push_cpu = this_cpu;
9154 raw_spin_unlock_irqrestore(&busiest->lock, flags);
9156 if (active_balance) {
9157 stop_one_cpu_nowait(cpu_of(busiest),
9158 active_load_balance_cpu_stop, busiest,
9159 &busiest->active_balance_work);
9162 /* We've kicked active balancing, force task migration. */
9163 sd->nr_balance_failed = sd->cache_nice_tries+1;
9166 sd->nr_balance_failed = 0;
9168 if (likely(!active_balance) || voluntary_active_balance(&env)) {
9169 /* We were unbalanced, so reset the balancing interval */
9170 sd->balance_interval = sd->min_interval;
9173 * If we've begun active balancing, start to back off. This
9174 * case may not be covered by the all_pinned logic if there
9175 * is only 1 task on the busy runqueue (because we don't call
9178 if (sd->balance_interval < sd->max_interval)
9179 sd->balance_interval *= 2;
9186 * We reach balance although we may have faced some affinity
9187 * constraints. Clear the imbalance flag only if other tasks got
9188 * a chance to move and fix the imbalance.
9190 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
9191 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9193 if (*group_imbalance)
9194 *group_imbalance = 0;
9199 * We reach balance because all tasks are pinned at this level so
9200 * we can't migrate them. Let the imbalance flag set so parent level
9201 * can try to migrate them.
9203 schedstat_inc(sd->lb_balanced[idle]);
9205 sd->nr_balance_failed = 0;
9211 * newidle_balance() disregards balance intervals, so we could
9212 * repeatedly reach this code, which would lead to balance_interval
9213 * skyrocketting in a short amount of time. Skip the balance_interval
9214 * increase logic to avoid that.
9216 if (env.idle == CPU_NEWLY_IDLE)
9219 /* tune up the balancing interval */
9220 if ((env.flags & LBF_ALL_PINNED &&
9221 sd->balance_interval < MAX_PINNED_INTERVAL) ||
9222 sd->balance_interval < sd->max_interval)
9223 sd->balance_interval *= 2;
9228 static inline unsigned long
9229 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
9231 unsigned long interval = sd->balance_interval;
9234 interval *= sd->busy_factor;
9236 /* scale ms to jiffies */
9237 interval = msecs_to_jiffies(interval);
9238 interval = clamp(interval, 1UL, max_load_balance_interval);
9244 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
9246 unsigned long interval, next;
9248 /* used by idle balance, so cpu_busy = 0 */
9249 interval = get_sd_balance_interval(sd, 0);
9250 next = sd->last_balance + interval;
9252 if (time_after(*next_balance, next))
9253 *next_balance = next;
9257 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9258 * running tasks off the busiest CPU onto idle CPUs. It requires at
9259 * least 1 task to be running on each physical CPU where possible, and
9260 * avoids physical / logical imbalances.
9262 static int active_load_balance_cpu_stop(void *data)
9264 struct rq *busiest_rq = data;
9265 int busiest_cpu = cpu_of(busiest_rq);
9266 int target_cpu = busiest_rq->push_cpu;
9267 struct rq *target_rq = cpu_rq(target_cpu);
9268 struct sched_domain *sd;
9269 struct task_struct *p = NULL;
9272 rq_lock_irq(busiest_rq, &rf);
9274 * Between queueing the stop-work and running it is a hole in which
9275 * CPUs can become inactive. We should not move tasks from or to
9278 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
9281 /* Make sure the requested CPU hasn't gone down in the meantime: */
9282 if (unlikely(busiest_cpu != smp_processor_id() ||
9283 !busiest_rq->active_balance))
9286 /* Is there any task to move? */
9287 if (busiest_rq->nr_running <= 1)
9291 * This condition is "impossible", if it occurs
9292 * we need to fix it. Originally reported by
9293 * Bjorn Helgaas on a 128-CPU setup.
9295 BUG_ON(busiest_rq == target_rq);
9297 /* Search for an sd spanning us and the target CPU. */
9299 for_each_domain(target_cpu, sd) {
9300 if ((sd->flags & SD_LOAD_BALANCE) &&
9301 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9306 struct lb_env env = {
9308 .dst_cpu = target_cpu,
9309 .dst_rq = target_rq,
9310 .src_cpu = busiest_rq->cpu,
9311 .src_rq = busiest_rq,
9314 * can_migrate_task() doesn't need to compute new_dst_cpu
9315 * for active balancing. Since we have CPU_IDLE, but no
9316 * @dst_grpmask we need to make that test go away with lying
9319 .flags = LBF_DST_PINNED,
9322 schedstat_inc(sd->alb_count);
9323 update_rq_clock(busiest_rq);
9325 p = detach_one_task(&env);
9327 schedstat_inc(sd->alb_pushed);
9328 /* Active balancing done, reset the failure counter. */
9329 sd->nr_balance_failed = 0;
9331 schedstat_inc(sd->alb_failed);
9336 busiest_rq->active_balance = 0;
9337 rq_unlock(busiest_rq, &rf);
9340 attach_one_task(target_rq, p);
9347 static DEFINE_SPINLOCK(balancing);
9350 * Scale the max load_balance interval with the number of CPUs in the system.
9351 * This trades load-balance latency on larger machines for less cross talk.
9353 void update_max_interval(void)
9355 max_load_balance_interval = HZ*num_online_cpus()/10;
9359 * It checks each scheduling domain to see if it is due to be balanced,
9360 * and initiates a balancing operation if so.
9362 * Balancing parameters are set up in init_sched_domains.
9364 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9366 int continue_balancing = 1;
9368 unsigned long interval;
9369 struct sched_domain *sd;
9370 /* Earliest time when we have to do rebalance again */
9371 unsigned long next_balance = jiffies + 60*HZ;
9372 int update_next_balance = 0;
9373 int need_serialize, need_decay = 0;
9377 for_each_domain(cpu, sd) {
9379 * Decay the newidle max times here because this is a regular
9380 * visit to all the domains. Decay ~1% per second.
9382 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9383 sd->max_newidle_lb_cost =
9384 (sd->max_newidle_lb_cost * 253) / 256;
9385 sd->next_decay_max_lb_cost = jiffies + HZ;
9388 max_cost += sd->max_newidle_lb_cost;
9390 if (!(sd->flags & SD_LOAD_BALANCE))
9394 * Stop the load balance at this level. There is another
9395 * CPU in our sched group which is doing load balancing more
9398 if (!continue_balancing) {
9404 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9406 need_serialize = sd->flags & SD_SERIALIZE;
9407 if (need_serialize) {
9408 if (!spin_trylock(&balancing))
9412 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9413 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9415 * The LBF_DST_PINNED logic could have changed
9416 * env->dst_cpu, so we can't know our idle
9417 * state even if we migrated tasks. Update it.
9419 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9421 sd->last_balance = jiffies;
9422 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9425 spin_unlock(&balancing);
9427 if (time_after(next_balance, sd->last_balance + interval)) {
9428 next_balance = sd->last_balance + interval;
9429 update_next_balance = 1;
9434 * Ensure the rq-wide value also decays but keep it at a
9435 * reasonable floor to avoid funnies with rq->avg_idle.
9437 rq->max_idle_balance_cost =
9438 max((u64)sysctl_sched_migration_cost, max_cost);
9443 * next_balance will be updated only when there is a need.
9444 * When the cpu is attached to null domain for ex, it will not be
9447 if (likely(update_next_balance)) {
9448 rq->next_balance = next_balance;
9450 #ifdef CONFIG_NO_HZ_COMMON
9452 * If this CPU has been elected to perform the nohz idle
9453 * balance. Other idle CPUs have already rebalanced with
9454 * nohz_idle_balance() and nohz.next_balance has been
9455 * updated accordingly. This CPU is now running the idle load
9456 * balance for itself and we need to update the
9457 * nohz.next_balance accordingly.
9459 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9460 nohz.next_balance = rq->next_balance;
9465 static inline int on_null_domain(struct rq *rq)
9467 return unlikely(!rcu_dereference_sched(rq->sd));
9470 #ifdef CONFIG_NO_HZ_COMMON
9472 * idle load balancing details
9473 * - When one of the busy CPUs notice that there may be an idle rebalancing
9474 * needed, they will kick the idle load balancer, which then does idle
9475 * load balancing for all the idle CPUs.
9476 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
9480 static inline int find_new_ilb(void)
9484 for_each_cpu_and(ilb, nohz.idle_cpus_mask,
9485 housekeeping_cpumask(HK_FLAG_MISC)) {
9494 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
9495 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
9497 static void kick_ilb(unsigned int flags)
9501 nohz.next_balance++;
9503 ilb_cpu = find_new_ilb();
9505 if (ilb_cpu >= nr_cpu_ids)
9508 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
9509 if (flags & NOHZ_KICK_MASK)
9513 * Use smp_send_reschedule() instead of resched_cpu().
9514 * This way we generate a sched IPI on the target CPU which
9515 * is idle. And the softirq performing nohz idle load balance
9516 * will be run before returning from the IPI.
9518 smp_send_reschedule(ilb_cpu);
9522 * Current decision point for kicking the idle load balancer in the presence
9523 * of idle CPUs in the system.
9525 static void nohz_balancer_kick(struct rq *rq)
9527 unsigned long now = jiffies;
9528 struct sched_domain_shared *sds;
9529 struct sched_domain *sd;
9530 int nr_busy, i, cpu = rq->cpu;
9531 unsigned int flags = 0;
9533 if (unlikely(rq->idle_balance))
9537 * We may be recently in ticked or tickless idle mode. At the first
9538 * busy tick after returning from idle, we will update the busy stats.
9540 nohz_balance_exit_idle(rq);
9543 * None are in tickless mode and hence no need for NOHZ idle load
9546 if (likely(!atomic_read(&nohz.nr_cpus)))
9549 if (READ_ONCE(nohz.has_blocked) &&
9550 time_after(now, READ_ONCE(nohz.next_blocked)))
9551 flags = NOHZ_STATS_KICK;
9553 if (time_before(now, nohz.next_balance))
9556 if (rq->nr_running >= 2) {
9557 flags = NOHZ_KICK_MASK;
9563 sd = rcu_dereference(rq->sd);
9566 * If there's a CFS task and the current CPU has reduced
9567 * capacity; kick the ILB to see if there's a better CPU to run
9570 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
9571 flags = NOHZ_KICK_MASK;
9576 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
9579 * When ASYM_PACKING; see if there's a more preferred CPU
9580 * currently idle; in which case, kick the ILB to move tasks
9583 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
9584 if (sched_asym_prefer(i, cpu)) {
9585 flags = NOHZ_KICK_MASK;
9591 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
9594 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
9595 * to run the misfit task on.
9597 if (check_misfit_status(rq, sd)) {
9598 flags = NOHZ_KICK_MASK;
9603 * For asymmetric systems, we do not want to nicely balance
9604 * cache use, instead we want to embrace asymmetry and only
9605 * ensure tasks have enough CPU capacity.
9607 * Skip the LLC logic because it's not relevant in that case.
9612 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9615 * If there is an imbalance between LLC domains (IOW we could
9616 * increase the overall cache use), we need some less-loaded LLC
9617 * domain to pull some load. Likewise, we may need to spread
9618 * load within the current LLC domain (e.g. packed SMT cores but
9619 * other CPUs are idle). We can't really know from here how busy
9620 * the others are - so just get a nohz balance going if it looks
9621 * like this LLC domain has tasks we could move.
9623 nr_busy = atomic_read(&sds->nr_busy_cpus);
9625 flags = NOHZ_KICK_MASK;
9636 static void set_cpu_sd_state_busy(int cpu)
9638 struct sched_domain *sd;
9641 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9643 if (!sd || !sd->nohz_idle)
9647 atomic_inc(&sd->shared->nr_busy_cpus);
9652 void nohz_balance_exit_idle(struct rq *rq)
9654 SCHED_WARN_ON(rq != this_rq());
9656 if (likely(!rq->nohz_tick_stopped))
9659 rq->nohz_tick_stopped = 0;
9660 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
9661 atomic_dec(&nohz.nr_cpus);
9663 set_cpu_sd_state_busy(rq->cpu);
9666 static void set_cpu_sd_state_idle(int cpu)
9668 struct sched_domain *sd;
9671 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9673 if (!sd || sd->nohz_idle)
9677 atomic_dec(&sd->shared->nr_busy_cpus);
9683 * This routine will record that the CPU is going idle with tick stopped.
9684 * This info will be used in performing idle load balancing in the future.
9686 void nohz_balance_enter_idle(int cpu)
9688 struct rq *rq = cpu_rq(cpu);
9690 SCHED_WARN_ON(cpu != smp_processor_id());
9692 /* If this CPU is going down, then nothing needs to be done: */
9693 if (!cpu_active(cpu))
9696 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9697 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9701 * Can be set safely without rq->lock held
9702 * If a clear happens, it will have evaluated last additions because
9703 * rq->lock is held during the check and the clear
9705 rq->has_blocked_load = 1;
9708 * The tick is still stopped but load could have been added in the
9709 * meantime. We set the nohz.has_blocked flag to trig a check of the
9710 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
9711 * of nohz.has_blocked can only happen after checking the new load
9713 if (rq->nohz_tick_stopped)
9716 /* If we're a completely isolated CPU, we don't play: */
9717 if (on_null_domain(rq))
9720 rq->nohz_tick_stopped = 1;
9722 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9723 atomic_inc(&nohz.nr_cpus);
9726 * Ensures that if nohz_idle_balance() fails to observe our
9727 * @idle_cpus_mask store, it must observe the @has_blocked
9730 smp_mb__after_atomic();
9732 set_cpu_sd_state_idle(cpu);
9736 * Each time a cpu enter idle, we assume that it has blocked load and
9737 * enable the periodic update of the load of idle cpus
9739 WRITE_ONCE(nohz.has_blocked, 1);
9743 * Internal function that runs load balance for all idle cpus. The load balance
9744 * can be a simple update of blocked load or a complete load balance with
9745 * tasks movement depending of flags.
9746 * The function returns false if the loop has stopped before running
9747 * through all idle CPUs.
9749 static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
9750 enum cpu_idle_type idle)
9752 /* Earliest time when we have to do rebalance again */
9753 unsigned long now = jiffies;
9754 unsigned long next_balance = now + 60*HZ;
9755 bool has_blocked_load = false;
9756 int update_next_balance = 0;
9757 int this_cpu = this_rq->cpu;
9762 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
9765 * We assume there will be no idle load after this update and clear
9766 * the has_blocked flag. If a cpu enters idle in the mean time, it will
9767 * set the has_blocked flag and trig another update of idle load.
9768 * Because a cpu that becomes idle, is added to idle_cpus_mask before
9769 * setting the flag, we are sure to not clear the state and not
9770 * check the load of an idle cpu.
9772 WRITE_ONCE(nohz.has_blocked, 0);
9775 * Ensures that if we miss the CPU, we must see the has_blocked
9776 * store from nohz_balance_enter_idle().
9780 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9781 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9785 * If this CPU gets work to do, stop the load balancing
9786 * work being done for other CPUs. Next load
9787 * balancing owner will pick it up.
9789 if (need_resched()) {
9790 has_blocked_load = true;
9794 rq = cpu_rq(balance_cpu);
9796 has_blocked_load |= update_nohz_stats(rq, true);
9799 * If time for next balance is due,
9802 if (time_after_eq(jiffies, rq->next_balance)) {
9805 rq_lock_irqsave(rq, &rf);
9806 update_rq_clock(rq);
9807 rq_unlock_irqrestore(rq, &rf);
9809 if (flags & NOHZ_BALANCE_KICK)
9810 rebalance_domains(rq, CPU_IDLE);
9813 if (time_after(next_balance, rq->next_balance)) {
9814 next_balance = rq->next_balance;
9815 update_next_balance = 1;
9819 /* Newly idle CPU doesn't need an update */
9820 if (idle != CPU_NEWLY_IDLE) {
9821 update_blocked_averages(this_cpu);
9822 has_blocked_load |= this_rq->has_blocked_load;
9825 if (flags & NOHZ_BALANCE_KICK)
9826 rebalance_domains(this_rq, CPU_IDLE);
9828 WRITE_ONCE(nohz.next_blocked,
9829 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
9831 /* The full idle balance loop has been done */
9835 /* There is still blocked load, enable periodic update */
9836 if (has_blocked_load)
9837 WRITE_ONCE(nohz.has_blocked, 1);
9840 * next_balance will be updated only when there is a need.
9841 * When the CPU is attached to null domain for ex, it will not be
9844 if (likely(update_next_balance))
9845 nohz.next_balance = next_balance;
9851 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9852 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9854 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9856 int this_cpu = this_rq->cpu;
9859 if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
9862 if (idle != CPU_IDLE) {
9863 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9867 /* could be _relaxed() */
9868 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9869 if (!(flags & NOHZ_KICK_MASK))
9872 _nohz_idle_balance(this_rq, flags, idle);
9877 static void nohz_newidle_balance(struct rq *this_rq)
9879 int this_cpu = this_rq->cpu;
9882 * This CPU doesn't want to be disturbed by scheduler
9885 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
9888 /* Will wake up very soon. No time for doing anything else*/
9889 if (this_rq->avg_idle < sysctl_sched_migration_cost)
9892 /* Don't need to update blocked load of idle CPUs*/
9893 if (!READ_ONCE(nohz.has_blocked) ||
9894 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
9897 raw_spin_unlock(&this_rq->lock);
9899 * This CPU is going to be idle and blocked load of idle CPUs
9900 * need to be updated. Run the ilb locally as it is a good
9901 * candidate for ilb instead of waking up another idle CPU.
9902 * Kick an normal ilb if we failed to do the update.
9904 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
9905 kick_ilb(NOHZ_STATS_KICK);
9906 raw_spin_lock(&this_rq->lock);
9909 #else /* !CONFIG_NO_HZ_COMMON */
9910 static inline void nohz_balancer_kick(struct rq *rq) { }
9912 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9917 static inline void nohz_newidle_balance(struct rq *this_rq) { }
9918 #endif /* CONFIG_NO_HZ_COMMON */
9921 * idle_balance is called by schedule() if this_cpu is about to become
9922 * idle. Attempts to pull tasks from other CPUs.
9924 int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
9926 unsigned long next_balance = jiffies + HZ;
9927 int this_cpu = this_rq->cpu;
9928 struct sched_domain *sd;
9929 int pulled_task = 0;
9932 update_misfit_status(NULL, this_rq);
9934 * We must set idle_stamp _before_ calling idle_balance(), such that we
9935 * measure the duration of idle_balance() as idle time.
9937 this_rq->idle_stamp = rq_clock(this_rq);
9940 * Do not pull tasks towards !active CPUs...
9942 if (!cpu_active(this_cpu))
9946 * This is OK, because current is on_cpu, which avoids it being picked
9947 * for load-balance and preemption/IRQs are still disabled avoiding
9948 * further scheduler activity on it and we're being very careful to
9949 * re-start the picking loop.
9951 rq_unpin_lock(this_rq, rf);
9953 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
9954 !READ_ONCE(this_rq->rd->overload)) {
9957 sd = rcu_dereference_check_sched_domain(this_rq->sd);
9959 update_next_balance(sd, &next_balance);
9962 nohz_newidle_balance(this_rq);
9967 raw_spin_unlock(&this_rq->lock);
9969 update_blocked_averages(this_cpu);
9971 for_each_domain(this_cpu, sd) {
9972 int continue_balancing = 1;
9973 u64 t0, domain_cost;
9975 if (!(sd->flags & SD_LOAD_BALANCE))
9978 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
9979 update_next_balance(sd, &next_balance);
9983 if (sd->flags & SD_BALANCE_NEWIDLE) {
9984 t0 = sched_clock_cpu(this_cpu);
9986 pulled_task = load_balance(this_cpu, this_rq,
9988 &continue_balancing);
9990 domain_cost = sched_clock_cpu(this_cpu) - t0;
9991 if (domain_cost > sd->max_newidle_lb_cost)
9992 sd->max_newidle_lb_cost = domain_cost;
9994 curr_cost += domain_cost;
9997 update_next_balance(sd, &next_balance);
10000 * Stop searching for tasks to pull if there are
10001 * now runnable tasks on this rq.
10003 if (pulled_task || this_rq->nr_running > 0)
10008 raw_spin_lock(&this_rq->lock);
10010 if (curr_cost > this_rq->max_idle_balance_cost)
10011 this_rq->max_idle_balance_cost = curr_cost;
10015 * While browsing the domains, we released the rq lock, a task could
10016 * have been enqueued in the meantime. Since we're not going idle,
10017 * pretend we pulled a task.
10019 if (this_rq->cfs.h_nr_running && !pulled_task)
10022 /* Move the next balance forward */
10023 if (time_after(this_rq->next_balance, next_balance))
10024 this_rq->next_balance = next_balance;
10026 /* Is there a task of a high priority class? */
10027 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
10031 this_rq->idle_stamp = 0;
10033 rq_repin_lock(this_rq, rf);
10035 return pulled_task;
10039 * run_rebalance_domains is triggered when needed from the scheduler tick.
10040 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
10042 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
10044 struct rq *this_rq = this_rq();
10045 enum cpu_idle_type idle = this_rq->idle_balance ?
10046 CPU_IDLE : CPU_NOT_IDLE;
10049 * If this CPU has a pending nohz_balance_kick, then do the
10050 * balancing on behalf of the other idle CPUs whose ticks are
10051 * stopped. Do nohz_idle_balance *before* rebalance_domains to
10052 * give the idle CPUs a chance to load balance. Else we may
10053 * load balance only within the local sched_domain hierarchy
10054 * and abort nohz_idle_balance altogether if we pull some load.
10056 if (nohz_idle_balance(this_rq, idle))
10059 /* normal load balance */
10060 update_blocked_averages(this_rq->cpu);
10061 rebalance_domains(this_rq, idle);
10065 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
10067 void trigger_load_balance(struct rq *rq)
10069 /* Don't need to rebalance while attached to NULL domain */
10070 if (unlikely(on_null_domain(rq)))
10073 if (time_after_eq(jiffies, rq->next_balance))
10074 raise_softirq(SCHED_SOFTIRQ);
10076 nohz_balancer_kick(rq);
10079 static void rq_online_fair(struct rq *rq)
10083 update_runtime_enabled(rq);
10086 static void rq_offline_fair(struct rq *rq)
10090 /* Ensure any throttled groups are reachable by pick_next_task */
10091 unthrottle_offline_cfs_rqs(rq);
10094 #endif /* CONFIG_SMP */
10097 * scheduler tick hitting a task of our scheduling class.
10099 * NOTE: This function can be called remotely by the tick offload that
10100 * goes along full dynticks. Therefore no local assumption can be made
10101 * and everything must be accessed through the @rq and @curr passed in
10104 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
10106 struct cfs_rq *cfs_rq;
10107 struct sched_entity *se = &curr->se;
10109 for_each_sched_entity(se) {
10110 cfs_rq = cfs_rq_of(se);
10111 entity_tick(cfs_rq, se, queued);
10114 if (static_branch_unlikely(&sched_numa_balancing))
10115 task_tick_numa(rq, curr);
10117 update_misfit_status(curr, rq);
10118 update_overutilized_status(task_rq(curr));
10122 * called on fork with the child task as argument from the parent's context
10123 * - child not yet on the tasklist
10124 * - preemption disabled
10126 static void task_fork_fair(struct task_struct *p)
10128 struct cfs_rq *cfs_rq;
10129 struct sched_entity *se = &p->se, *curr;
10130 struct rq *rq = this_rq();
10131 struct rq_flags rf;
10134 update_rq_clock(rq);
10136 cfs_rq = task_cfs_rq(current);
10137 curr = cfs_rq->curr;
10139 update_curr(cfs_rq);
10140 se->vruntime = curr->vruntime;
10142 place_entity(cfs_rq, se, 1);
10144 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
10146 * Upon rescheduling, sched_class::put_prev_task() will place
10147 * 'current' within the tree based on its new key value.
10149 swap(curr->vruntime, se->vruntime);
10153 se->vruntime -= cfs_rq->min_vruntime;
10154 rq_unlock(rq, &rf);
10158 * Priority of the task has changed. Check to see if we preempt
10159 * the current task.
10162 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
10164 if (!task_on_rq_queued(p))
10168 * Reschedule if we are currently running on this runqueue and
10169 * our priority decreased, or if we are not currently running on
10170 * this runqueue and our priority is higher than the current's
10172 if (rq->curr == p) {
10173 if (p->prio > oldprio)
10176 check_preempt_curr(rq, p, 0);
10179 static inline bool vruntime_normalized(struct task_struct *p)
10181 struct sched_entity *se = &p->se;
10184 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
10185 * the dequeue_entity(.flags=0) will already have normalized the
10192 * When !on_rq, vruntime of the task has usually NOT been normalized.
10193 * But there are some cases where it has already been normalized:
10195 * - A forked child which is waiting for being woken up by
10196 * wake_up_new_task().
10197 * - A task which has been woken up by try_to_wake_up() and
10198 * waiting for actually being woken up by sched_ttwu_pending().
10200 if (!se->sum_exec_runtime ||
10201 (p->state == TASK_WAKING && p->sched_remote_wakeup))
10207 #ifdef CONFIG_FAIR_GROUP_SCHED
10209 * Propagate the changes of the sched_entity across the tg tree to make it
10210 * visible to the root
10212 static void propagate_entity_cfs_rq(struct sched_entity *se)
10214 struct cfs_rq *cfs_rq;
10216 /* Start to propagate at parent */
10219 for_each_sched_entity(se) {
10220 cfs_rq = cfs_rq_of(se);
10222 if (cfs_rq_throttled(cfs_rq))
10225 update_load_avg(cfs_rq, se, UPDATE_TG);
10229 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
10232 static void detach_entity_cfs_rq(struct sched_entity *se)
10234 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10236 /* Catch up with the cfs_rq and remove our load when we leave */
10237 update_load_avg(cfs_rq, se, 0);
10238 detach_entity_load_avg(cfs_rq, se);
10239 update_tg_load_avg(cfs_rq, false);
10240 propagate_entity_cfs_rq(se);
10243 static void attach_entity_cfs_rq(struct sched_entity *se)
10245 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10247 #ifdef CONFIG_FAIR_GROUP_SCHED
10249 * Since the real-depth could have been changed (only FAIR
10250 * class maintain depth value), reset depth properly.
10252 se->depth = se->parent ? se->parent->depth + 1 : 0;
10255 /* Synchronize entity with its cfs_rq */
10256 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
10257 attach_entity_load_avg(cfs_rq, se, 0);
10258 update_tg_load_avg(cfs_rq, false);
10259 propagate_entity_cfs_rq(se);
10262 static void detach_task_cfs_rq(struct task_struct *p)
10264 struct sched_entity *se = &p->se;
10265 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10267 if (!vruntime_normalized(p)) {
10269 * Fix up our vruntime so that the current sleep doesn't
10270 * cause 'unlimited' sleep bonus.
10272 place_entity(cfs_rq, se, 0);
10273 se->vruntime -= cfs_rq->min_vruntime;
10276 detach_entity_cfs_rq(se);
10279 static void attach_task_cfs_rq(struct task_struct *p)
10281 struct sched_entity *se = &p->se;
10282 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10284 attach_entity_cfs_rq(se);
10286 if (!vruntime_normalized(p))
10287 se->vruntime += cfs_rq->min_vruntime;
10290 static void switched_from_fair(struct rq *rq, struct task_struct *p)
10292 detach_task_cfs_rq(p);
10295 static void switched_to_fair(struct rq *rq, struct task_struct *p)
10297 attach_task_cfs_rq(p);
10299 if (task_on_rq_queued(p)) {
10301 * We were most likely switched from sched_rt, so
10302 * kick off the schedule if running, otherwise just see
10303 * if we can still preempt the current task.
10308 check_preempt_curr(rq, p, 0);
10312 /* Account for a task changing its policy or group.
10314 * This routine is mostly called to set cfs_rq->curr field when a task
10315 * migrates between groups/classes.
10317 static void set_next_task_fair(struct rq *rq, struct task_struct *p)
10319 struct sched_entity *se = &p->se;
10322 if (task_on_rq_queued(p)) {
10324 * Move the next running task to the front of the list, so our
10325 * cfs_tasks list becomes MRU one.
10327 list_move(&se->group_node, &rq->cfs_tasks);
10331 for_each_sched_entity(se) {
10332 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10334 set_next_entity(cfs_rq, se);
10335 /* ensure bandwidth has been allocated on our new cfs_rq */
10336 account_cfs_rq_runtime(cfs_rq, 0);
10340 void init_cfs_rq(struct cfs_rq *cfs_rq)
10342 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
10343 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
10344 #ifndef CONFIG_64BIT
10345 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
10348 raw_spin_lock_init(&cfs_rq->removed.lock);
10352 #ifdef CONFIG_FAIR_GROUP_SCHED
10353 static void task_set_group_fair(struct task_struct *p)
10355 struct sched_entity *se = &p->se;
10357 set_task_rq(p, task_cpu(p));
10358 se->depth = se->parent ? se->parent->depth + 1 : 0;
10361 static void task_move_group_fair(struct task_struct *p)
10363 detach_task_cfs_rq(p);
10364 set_task_rq(p, task_cpu(p));
10367 /* Tell se's cfs_rq has been changed -- migrated */
10368 p->se.avg.last_update_time = 0;
10370 attach_task_cfs_rq(p);
10373 static void task_change_group_fair(struct task_struct *p, int type)
10376 case TASK_SET_GROUP:
10377 task_set_group_fair(p);
10380 case TASK_MOVE_GROUP:
10381 task_move_group_fair(p);
10386 void free_fair_sched_group(struct task_group *tg)
10390 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
10392 for_each_possible_cpu(i) {
10394 kfree(tg->cfs_rq[i]);
10403 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10405 struct sched_entity *se;
10406 struct cfs_rq *cfs_rq;
10409 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
10412 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
10416 tg->shares = NICE_0_LOAD;
10418 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
10420 for_each_possible_cpu(i) {
10421 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
10422 GFP_KERNEL, cpu_to_node(i));
10426 se = kzalloc_node(sizeof(struct sched_entity),
10427 GFP_KERNEL, cpu_to_node(i));
10431 init_cfs_rq(cfs_rq);
10432 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
10433 init_entity_runnable_average(se);
10444 void online_fair_sched_group(struct task_group *tg)
10446 struct sched_entity *se;
10447 struct rq_flags rf;
10451 for_each_possible_cpu(i) {
10454 rq_lock_irq(rq, &rf);
10455 update_rq_clock(rq);
10456 attach_entity_cfs_rq(se);
10457 sync_throttle(tg, i);
10458 rq_unlock_irq(rq, &rf);
10462 void unregister_fair_sched_group(struct task_group *tg)
10464 unsigned long flags;
10468 for_each_possible_cpu(cpu) {
10470 remove_entity_load_avg(tg->se[cpu]);
10473 * Only empty task groups can be destroyed; so we can speculatively
10474 * check on_list without danger of it being re-added.
10476 if (!tg->cfs_rq[cpu]->on_list)
10481 raw_spin_lock_irqsave(&rq->lock, flags);
10482 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
10483 raw_spin_unlock_irqrestore(&rq->lock, flags);
10487 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
10488 struct sched_entity *se, int cpu,
10489 struct sched_entity *parent)
10491 struct rq *rq = cpu_rq(cpu);
10495 init_cfs_rq_runtime(cfs_rq);
10497 tg->cfs_rq[cpu] = cfs_rq;
10500 /* se could be NULL for root_task_group */
10505 se->cfs_rq = &rq->cfs;
10508 se->cfs_rq = parent->my_q;
10509 se->depth = parent->depth + 1;
10513 /* guarantee group entities always have weight */
10514 update_load_set(&se->load, NICE_0_LOAD);
10515 se->parent = parent;
10518 static DEFINE_MUTEX(shares_mutex);
10520 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10525 * We can't change the weight of the root cgroup.
10530 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10532 mutex_lock(&shares_mutex);
10533 if (tg->shares == shares)
10536 tg->shares = shares;
10537 for_each_possible_cpu(i) {
10538 struct rq *rq = cpu_rq(i);
10539 struct sched_entity *se = tg->se[i];
10540 struct rq_flags rf;
10542 /* Propagate contribution to hierarchy */
10543 rq_lock_irqsave(rq, &rf);
10544 update_rq_clock(rq);
10545 for_each_sched_entity(se) {
10546 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
10547 update_cfs_group(se);
10549 rq_unlock_irqrestore(rq, &rf);
10553 mutex_unlock(&shares_mutex);
10556 #else /* CONFIG_FAIR_GROUP_SCHED */
10558 void free_fair_sched_group(struct task_group *tg) { }
10560 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10565 void online_fair_sched_group(struct task_group *tg) { }
10567 void unregister_fair_sched_group(struct task_group *tg) { }
10569 #endif /* CONFIG_FAIR_GROUP_SCHED */
10572 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10574 struct sched_entity *se = &task->se;
10575 unsigned int rr_interval = 0;
10578 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10581 if (rq->cfs.load.weight)
10582 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10584 return rr_interval;
10588 * All the scheduling class methods:
10590 const struct sched_class fair_sched_class = {
10591 .next = &idle_sched_class,
10592 .enqueue_task = enqueue_task_fair,
10593 .dequeue_task = dequeue_task_fair,
10594 .yield_task = yield_task_fair,
10595 .yield_to_task = yield_to_task_fair,
10597 .check_preempt_curr = check_preempt_wakeup,
10599 .pick_next_task = pick_next_task_fair,
10601 .put_prev_task = put_prev_task_fair,
10602 .set_next_task = set_next_task_fair,
10605 .select_task_rq = select_task_rq_fair,
10606 .migrate_task_rq = migrate_task_rq_fair,
10608 .rq_online = rq_online_fair,
10609 .rq_offline = rq_offline_fair,
10611 .task_dead = task_dead_fair,
10612 .set_cpus_allowed = set_cpus_allowed_common,
10615 .task_tick = task_tick_fair,
10616 .task_fork = task_fork_fair,
10618 .prio_changed = prio_changed_fair,
10619 .switched_from = switched_from_fair,
10620 .switched_to = switched_to_fair,
10622 .get_rr_interval = get_rr_interval_fair,
10624 .update_curr = update_curr_fair,
10626 #ifdef CONFIG_FAIR_GROUP_SCHED
10627 .task_change_group = task_change_group_fair,
10630 #ifdef CONFIG_UCLAMP_TASK
10631 .uclamp_enabled = 1,
10635 #ifdef CONFIG_SCHED_DEBUG
10636 void print_cfs_stats(struct seq_file *m, int cpu)
10638 struct cfs_rq *cfs_rq, *pos;
10641 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
10642 print_cfs_rq(m, cpu, cfs_rq);
10646 #ifdef CONFIG_NUMA_BALANCING
10647 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10650 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10651 struct numa_group *ng;
10654 ng = rcu_dereference(p->numa_group);
10655 for_each_online_node(node) {
10656 if (p->numa_faults) {
10657 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10658 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10661 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
10662 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
10664 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10668 #endif /* CONFIG_NUMA_BALANCING */
10669 #endif /* CONFIG_SCHED_DEBUG */
10671 __init void init_sched_fair_class(void)
10674 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10676 #ifdef CONFIG_NO_HZ_COMMON
10677 nohz.next_balance = jiffies;
10678 nohz.next_blocked = jiffies;
10679 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
10686 * Helper functions to facilitate extracting info from tracepoints.
10689 const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq)
10692 return cfs_rq ? &cfs_rq->avg : NULL;
10697 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg);
10699 char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len)
10703 strlcpy(str, "(null)", len);
10708 cfs_rq_tg_path(cfs_rq, str, len);
10711 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path);
10713 int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq)
10715 return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1;
10717 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu);
10719 const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq)
10722 return rq ? &rq->avg_rt : NULL;
10727 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt);
10729 const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq)
10732 return rq ? &rq->avg_dl : NULL;
10737 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl);
10739 const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq)
10741 #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ)
10742 return rq ? &rq->avg_irq : NULL;
10747 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq);
10749 int sched_trace_rq_cpu(struct rq *rq)
10751 return rq ? cpu_of(rq) : -1;
10753 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu);
10755 const struct cpumask *sched_trace_rd_span(struct root_domain *rd)
10758 return rd ? rd->span : NULL;
10763 EXPORT_SYMBOL_GPL(sched_trace_rd_span);