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 unsigned long cpu_runnable_load(struct rq *rq);
1479 /* Cached statistics for all CPUs within a node */
1483 /* Total compute capacity of CPUs on a node */
1484 unsigned long compute_capacity;
1488 * XXX borrowed from update_sg_lb_stats
1490 static void update_numa_stats(struct numa_stats *ns, int nid)
1494 memset(ns, 0, sizeof(*ns));
1495 for_each_cpu(cpu, cpumask_of_node(nid)) {
1496 struct rq *rq = cpu_rq(cpu);
1498 ns->load += cpu_runnable_load(rq);
1499 ns->compute_capacity += capacity_of(cpu);
1504 struct task_numa_env {
1505 struct task_struct *p;
1507 int src_cpu, src_nid;
1508 int dst_cpu, dst_nid;
1510 struct numa_stats src_stats, dst_stats;
1515 struct task_struct *best_task;
1520 static void task_numa_assign(struct task_numa_env *env,
1521 struct task_struct *p, long imp)
1523 struct rq *rq = cpu_rq(env->dst_cpu);
1525 /* Bail out if run-queue part of active NUMA balance. */
1526 if (xchg(&rq->numa_migrate_on, 1))
1530 * Clear previous best_cpu/rq numa-migrate flag, since task now
1531 * found a better CPU to move/swap.
1533 if (env->best_cpu != -1) {
1534 rq = cpu_rq(env->best_cpu);
1535 WRITE_ONCE(rq->numa_migrate_on, 0);
1539 put_task_struct(env->best_task);
1544 env->best_imp = imp;
1545 env->best_cpu = env->dst_cpu;
1548 static bool load_too_imbalanced(long src_load, long dst_load,
1549 struct task_numa_env *env)
1552 long orig_src_load, orig_dst_load;
1553 long src_capacity, dst_capacity;
1556 * The load is corrected for the CPU capacity available on each node.
1559 * ------------ vs ---------
1560 * src_capacity dst_capacity
1562 src_capacity = env->src_stats.compute_capacity;
1563 dst_capacity = env->dst_stats.compute_capacity;
1565 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1567 orig_src_load = env->src_stats.load;
1568 orig_dst_load = env->dst_stats.load;
1570 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1572 /* Would this change make things worse? */
1573 return (imb > old_imb);
1577 * Maximum NUMA importance can be 1998 (2*999);
1578 * SMALLIMP @ 30 would be close to 1998/64.
1579 * Used to deter task migration.
1584 * This checks if the overall compute and NUMA accesses of the system would
1585 * be improved if the source tasks was migrated to the target dst_cpu taking
1586 * into account that it might be best if task running on the dst_cpu should
1587 * be exchanged with the source task
1589 static void task_numa_compare(struct task_numa_env *env,
1590 long taskimp, long groupimp, bool maymove)
1592 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
1593 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1594 long imp = p_ng ? groupimp : taskimp;
1595 struct task_struct *cur;
1596 long src_load, dst_load;
1597 int dist = env->dist;
1601 if (READ_ONCE(dst_rq->numa_migrate_on))
1605 cur = rcu_dereference(dst_rq->curr);
1606 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1610 * Because we have preemption enabled we can get migrated around and
1611 * end try selecting ourselves (current == env->p) as a swap candidate.
1617 if (maymove && moveimp >= env->best_imp)
1624 * "imp" is the fault differential for the source task between the
1625 * source and destination node. Calculate the total differential for
1626 * the source task and potential destination task. The more negative
1627 * the value is, the more remote accesses that would be expected to
1628 * be incurred if the tasks were swapped.
1630 /* Skip this swap candidate if cannot move to the source cpu */
1631 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
1635 * If dst and source tasks are in the same NUMA group, or not
1636 * in any group then look only at task weights.
1638 cur_ng = rcu_dereference(cur->numa_group);
1639 if (cur_ng == p_ng) {
1640 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1641 task_weight(cur, env->dst_nid, dist);
1643 * Add some hysteresis to prevent swapping the
1644 * tasks within a group over tiny differences.
1650 * Compare the group weights. If a task is all by itself
1651 * (not part of a group), use the task weight instead.
1654 imp += group_weight(cur, env->src_nid, dist) -
1655 group_weight(cur, env->dst_nid, dist);
1657 imp += task_weight(cur, env->src_nid, dist) -
1658 task_weight(cur, env->dst_nid, dist);
1661 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1668 * If the NUMA importance is less than SMALLIMP,
1669 * task migration might only result in ping pong
1670 * of tasks and also hurt performance due to cache
1673 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1677 * In the overloaded case, try and keep the load balanced.
1679 load = task_h_load(env->p) - task_h_load(cur);
1683 dst_load = env->dst_stats.load + load;
1684 src_load = env->src_stats.load - load;
1686 if (load_too_imbalanced(src_load, dst_load, env))
1691 * One idle CPU per node is evaluated for a task numa move.
1692 * Call select_idle_sibling to maybe find a better one.
1696 * select_idle_siblings() uses an per-CPU cpumask that
1697 * can be used from IRQ context.
1699 local_irq_disable();
1700 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1705 task_numa_assign(env, cur, imp);
1710 static void task_numa_find_cpu(struct task_numa_env *env,
1711 long taskimp, long groupimp)
1713 long src_load, dst_load, load;
1714 bool maymove = false;
1717 load = task_h_load(env->p);
1718 dst_load = env->dst_stats.load + load;
1719 src_load = env->src_stats.load - load;
1722 * If the improvement from just moving env->p direction is better
1723 * than swapping tasks around, check if a move is possible.
1725 maymove = !load_too_imbalanced(src_load, dst_load, env);
1727 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1728 /* Skip this CPU if the source task cannot migrate */
1729 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
1733 task_numa_compare(env, taskimp, groupimp, maymove);
1737 static int task_numa_migrate(struct task_struct *p)
1739 struct task_numa_env env = {
1742 .src_cpu = task_cpu(p),
1743 .src_nid = task_node(p),
1745 .imbalance_pct = 112,
1751 unsigned long taskweight, groupweight;
1752 struct sched_domain *sd;
1753 long taskimp, groupimp;
1754 struct numa_group *ng;
1759 * Pick the lowest SD_NUMA domain, as that would have the smallest
1760 * imbalance and would be the first to start moving tasks about.
1762 * And we want to avoid any moving of tasks about, as that would create
1763 * random movement of tasks -- counter the numa conditions we're trying
1767 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1769 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1773 * Cpusets can break the scheduler domain tree into smaller
1774 * balance domains, some of which do not cross NUMA boundaries.
1775 * Tasks that are "trapped" in such domains cannot be migrated
1776 * elsewhere, so there is no point in (re)trying.
1778 if (unlikely(!sd)) {
1779 sched_setnuma(p, task_node(p));
1783 env.dst_nid = p->numa_preferred_nid;
1784 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1785 taskweight = task_weight(p, env.src_nid, dist);
1786 groupweight = group_weight(p, env.src_nid, dist);
1787 update_numa_stats(&env.src_stats, env.src_nid);
1788 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1789 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1790 update_numa_stats(&env.dst_stats, env.dst_nid);
1792 /* Try to find a spot on the preferred nid. */
1793 task_numa_find_cpu(&env, taskimp, groupimp);
1796 * Look at other nodes in these cases:
1797 * - there is no space available on the preferred_nid
1798 * - the task is part of a numa_group that is interleaved across
1799 * multiple NUMA nodes; in order to better consolidate the group,
1800 * we need to check other locations.
1802 ng = deref_curr_numa_group(p);
1803 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
1804 for_each_online_node(nid) {
1805 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1808 dist = node_distance(env.src_nid, env.dst_nid);
1809 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1811 taskweight = task_weight(p, env.src_nid, dist);
1812 groupweight = group_weight(p, env.src_nid, dist);
1815 /* Only consider nodes where both task and groups benefit */
1816 taskimp = task_weight(p, nid, dist) - taskweight;
1817 groupimp = group_weight(p, nid, dist) - groupweight;
1818 if (taskimp < 0 && groupimp < 0)
1823 update_numa_stats(&env.dst_stats, env.dst_nid);
1824 task_numa_find_cpu(&env, taskimp, groupimp);
1829 * If the task is part of a workload that spans multiple NUMA nodes,
1830 * and is migrating into one of the workload's active nodes, remember
1831 * this node as the task's preferred numa node, so the workload can
1833 * A task that migrated to a second choice node will be better off
1834 * trying for a better one later. Do not set the preferred node here.
1837 if (env.best_cpu == -1)
1840 nid = cpu_to_node(env.best_cpu);
1842 if (nid != p->numa_preferred_nid)
1843 sched_setnuma(p, nid);
1846 /* No better CPU than the current one was found. */
1847 if (env.best_cpu == -1)
1850 best_rq = cpu_rq(env.best_cpu);
1851 if (env.best_task == NULL) {
1852 ret = migrate_task_to(p, env.best_cpu);
1853 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1855 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1859 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
1860 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1863 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1864 put_task_struct(env.best_task);
1868 /* Attempt to migrate a task to a CPU on the preferred node. */
1869 static void numa_migrate_preferred(struct task_struct *p)
1871 unsigned long interval = HZ;
1873 /* This task has no NUMA fault statistics yet */
1874 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
1877 /* Periodically retry migrating the task to the preferred node */
1878 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1879 p->numa_migrate_retry = jiffies + interval;
1881 /* Success if task is already running on preferred CPU */
1882 if (task_node(p) == p->numa_preferred_nid)
1885 /* Otherwise, try migrate to a CPU on the preferred node */
1886 task_numa_migrate(p);
1890 * Find out how many nodes on the workload is actively running on. Do this by
1891 * tracking the nodes from which NUMA hinting faults are triggered. This can
1892 * be different from the set of nodes where the workload's memory is currently
1895 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1897 unsigned long faults, max_faults = 0;
1898 int nid, active_nodes = 0;
1900 for_each_online_node(nid) {
1901 faults = group_faults_cpu(numa_group, nid);
1902 if (faults > max_faults)
1903 max_faults = faults;
1906 for_each_online_node(nid) {
1907 faults = group_faults_cpu(numa_group, nid);
1908 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1912 numa_group->max_faults_cpu = max_faults;
1913 numa_group->active_nodes = active_nodes;
1917 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1918 * increments. The more local the fault statistics are, the higher the scan
1919 * period will be for the next scan window. If local/(local+remote) ratio is
1920 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1921 * the scan period will decrease. Aim for 70% local accesses.
1923 #define NUMA_PERIOD_SLOTS 10
1924 #define NUMA_PERIOD_THRESHOLD 7
1927 * Increase the scan period (slow down scanning) if the majority of
1928 * our memory is already on our local node, or if the majority of
1929 * the page accesses are shared with other processes.
1930 * Otherwise, decrease the scan period.
1932 static void update_task_scan_period(struct task_struct *p,
1933 unsigned long shared, unsigned long private)
1935 unsigned int period_slot;
1936 int lr_ratio, ps_ratio;
1939 unsigned long remote = p->numa_faults_locality[0];
1940 unsigned long local = p->numa_faults_locality[1];
1943 * If there were no record hinting faults then either the task is
1944 * completely idle or all activity is areas that are not of interest
1945 * to automatic numa balancing. Related to that, if there were failed
1946 * migration then it implies we are migrating too quickly or the local
1947 * node is overloaded. In either case, scan slower
1949 if (local + shared == 0 || p->numa_faults_locality[2]) {
1950 p->numa_scan_period = min(p->numa_scan_period_max,
1951 p->numa_scan_period << 1);
1953 p->mm->numa_next_scan = jiffies +
1954 msecs_to_jiffies(p->numa_scan_period);
1960 * Prepare to scale scan period relative to the current period.
1961 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1962 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1963 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1965 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1966 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1967 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1969 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1971 * Most memory accesses are local. There is no need to
1972 * do fast NUMA scanning, since memory is already local.
1974 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
1977 diff = slot * period_slot;
1978 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
1980 * Most memory accesses are shared with other tasks.
1981 * There is no point in continuing fast NUMA scanning,
1982 * since other tasks may just move the memory elsewhere.
1984 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
1987 diff = slot * period_slot;
1990 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
1991 * yet they are not on the local NUMA node. Speed up
1992 * NUMA scanning to get the memory moved over.
1994 int ratio = max(lr_ratio, ps_ratio);
1995 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1998 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1999 task_scan_min(p), task_scan_max(p));
2000 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2004 * Get the fraction of time the task has been running since the last
2005 * NUMA placement cycle. The scheduler keeps similar statistics, but
2006 * decays those on a 32ms period, which is orders of magnitude off
2007 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2008 * stats only if the task is so new there are no NUMA statistics yet.
2010 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2012 u64 runtime, delta, now;
2013 /* Use the start of this time slice to avoid calculations. */
2014 now = p->se.exec_start;
2015 runtime = p->se.sum_exec_runtime;
2017 if (p->last_task_numa_placement) {
2018 delta = runtime - p->last_sum_exec_runtime;
2019 *period = now - p->last_task_numa_placement;
2021 /* Avoid time going backwards, prevent potential divide error: */
2022 if (unlikely((s64)*period < 0))
2025 delta = p->se.avg.load_sum;
2026 *period = LOAD_AVG_MAX;
2029 p->last_sum_exec_runtime = runtime;
2030 p->last_task_numa_placement = now;
2036 * Determine the preferred nid for a task in a numa_group. This needs to
2037 * be done in a way that produces consistent results with group_weight,
2038 * otherwise workloads might not converge.
2040 static int preferred_group_nid(struct task_struct *p, int nid)
2045 /* Direct connections between all NUMA nodes. */
2046 if (sched_numa_topology_type == NUMA_DIRECT)
2050 * On a system with glueless mesh NUMA topology, group_weight
2051 * scores nodes according to the number of NUMA hinting faults on
2052 * both the node itself, and on nearby nodes.
2054 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2055 unsigned long score, max_score = 0;
2056 int node, max_node = nid;
2058 dist = sched_max_numa_distance;
2060 for_each_online_node(node) {
2061 score = group_weight(p, node, dist);
2062 if (score > max_score) {
2071 * Finding the preferred nid in a system with NUMA backplane
2072 * interconnect topology is more involved. The goal is to locate
2073 * tasks from numa_groups near each other in the system, and
2074 * untangle workloads from different sides of the system. This requires
2075 * searching down the hierarchy of node groups, recursively searching
2076 * inside the highest scoring group of nodes. The nodemask tricks
2077 * keep the complexity of the search down.
2079 nodes = node_online_map;
2080 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2081 unsigned long max_faults = 0;
2082 nodemask_t max_group = NODE_MASK_NONE;
2085 /* Are there nodes at this distance from each other? */
2086 if (!find_numa_distance(dist))
2089 for_each_node_mask(a, nodes) {
2090 unsigned long faults = 0;
2091 nodemask_t this_group;
2092 nodes_clear(this_group);
2094 /* Sum group's NUMA faults; includes a==b case. */
2095 for_each_node_mask(b, nodes) {
2096 if (node_distance(a, b) < dist) {
2097 faults += group_faults(p, b);
2098 node_set(b, this_group);
2099 node_clear(b, nodes);
2103 /* Remember the top group. */
2104 if (faults > max_faults) {
2105 max_faults = faults;
2106 max_group = this_group;
2108 * subtle: at the smallest distance there is
2109 * just one node left in each "group", the
2110 * winner is the preferred nid.
2115 /* Next round, evaluate the nodes within max_group. */
2123 static void task_numa_placement(struct task_struct *p)
2125 int seq, nid, max_nid = NUMA_NO_NODE;
2126 unsigned long max_faults = 0;
2127 unsigned long fault_types[2] = { 0, 0 };
2128 unsigned long total_faults;
2129 u64 runtime, period;
2130 spinlock_t *group_lock = NULL;
2131 struct numa_group *ng;
2134 * The p->mm->numa_scan_seq field gets updated without
2135 * exclusive access. Use READ_ONCE() here to ensure
2136 * that the field is read in a single access:
2138 seq = READ_ONCE(p->mm->numa_scan_seq);
2139 if (p->numa_scan_seq == seq)
2141 p->numa_scan_seq = seq;
2142 p->numa_scan_period_max = task_scan_max(p);
2144 total_faults = p->numa_faults_locality[0] +
2145 p->numa_faults_locality[1];
2146 runtime = numa_get_avg_runtime(p, &period);
2148 /* If the task is part of a group prevent parallel updates to group stats */
2149 ng = deref_curr_numa_group(p);
2151 group_lock = &ng->lock;
2152 spin_lock_irq(group_lock);
2155 /* Find the node with the highest number of faults */
2156 for_each_online_node(nid) {
2157 /* Keep track of the offsets in numa_faults array */
2158 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2159 unsigned long faults = 0, group_faults = 0;
2162 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2163 long diff, f_diff, f_weight;
2165 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2166 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2167 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2168 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2170 /* Decay existing window, copy faults since last scan */
2171 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2172 fault_types[priv] += p->numa_faults[membuf_idx];
2173 p->numa_faults[membuf_idx] = 0;
2176 * Normalize the faults_from, so all tasks in a group
2177 * count according to CPU use, instead of by the raw
2178 * number of faults. Tasks with little runtime have
2179 * little over-all impact on throughput, and thus their
2180 * faults are less important.
2182 f_weight = div64_u64(runtime << 16, period + 1);
2183 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2185 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2186 p->numa_faults[cpubuf_idx] = 0;
2188 p->numa_faults[mem_idx] += diff;
2189 p->numa_faults[cpu_idx] += f_diff;
2190 faults += p->numa_faults[mem_idx];
2191 p->total_numa_faults += diff;
2194 * safe because we can only change our own group
2196 * mem_idx represents the offset for a given
2197 * nid and priv in a specific region because it
2198 * is at the beginning of the numa_faults array.
2200 ng->faults[mem_idx] += diff;
2201 ng->faults_cpu[mem_idx] += f_diff;
2202 ng->total_faults += diff;
2203 group_faults += ng->faults[mem_idx];
2208 if (faults > max_faults) {
2209 max_faults = faults;
2212 } else if (group_faults > max_faults) {
2213 max_faults = group_faults;
2219 numa_group_count_active_nodes(ng);
2220 spin_unlock_irq(group_lock);
2221 max_nid = preferred_group_nid(p, max_nid);
2225 /* Set the new preferred node */
2226 if (max_nid != p->numa_preferred_nid)
2227 sched_setnuma(p, max_nid);
2230 update_task_scan_period(p, fault_types[0], fault_types[1]);
2233 static inline int get_numa_group(struct numa_group *grp)
2235 return refcount_inc_not_zero(&grp->refcount);
2238 static inline void put_numa_group(struct numa_group *grp)
2240 if (refcount_dec_and_test(&grp->refcount))
2241 kfree_rcu(grp, rcu);
2244 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2247 struct numa_group *grp, *my_grp;
2248 struct task_struct *tsk;
2250 int cpu = cpupid_to_cpu(cpupid);
2253 if (unlikely(!deref_curr_numa_group(p))) {
2254 unsigned int size = sizeof(struct numa_group) +
2255 4*nr_node_ids*sizeof(unsigned long);
2257 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2261 refcount_set(&grp->refcount, 1);
2262 grp->active_nodes = 1;
2263 grp->max_faults_cpu = 0;
2264 spin_lock_init(&grp->lock);
2266 /* Second half of the array tracks nids where faults happen */
2267 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2270 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2271 grp->faults[i] = p->numa_faults[i];
2273 grp->total_faults = p->total_numa_faults;
2276 rcu_assign_pointer(p->numa_group, grp);
2280 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2282 if (!cpupid_match_pid(tsk, cpupid))
2285 grp = rcu_dereference(tsk->numa_group);
2289 my_grp = deref_curr_numa_group(p);
2294 * Only join the other group if its bigger; if we're the bigger group,
2295 * the other task will join us.
2297 if (my_grp->nr_tasks > grp->nr_tasks)
2301 * Tie-break on the grp address.
2303 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2306 /* Always join threads in the same process. */
2307 if (tsk->mm == current->mm)
2310 /* Simple filter to avoid false positives due to PID collisions */
2311 if (flags & TNF_SHARED)
2314 /* Update priv based on whether false sharing was detected */
2317 if (join && !get_numa_group(grp))
2325 BUG_ON(irqs_disabled());
2326 double_lock_irq(&my_grp->lock, &grp->lock);
2328 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2329 my_grp->faults[i] -= p->numa_faults[i];
2330 grp->faults[i] += p->numa_faults[i];
2332 my_grp->total_faults -= p->total_numa_faults;
2333 grp->total_faults += p->total_numa_faults;
2338 spin_unlock(&my_grp->lock);
2339 spin_unlock_irq(&grp->lock);
2341 rcu_assign_pointer(p->numa_group, grp);
2343 put_numa_group(my_grp);
2352 * Get rid of NUMA staticstics associated with a task (either current or dead).
2353 * If @final is set, the task is dead and has reached refcount zero, so we can
2354 * safely free all relevant data structures. Otherwise, there might be
2355 * concurrent reads from places like load balancing and procfs, and we should
2356 * reset the data back to default state without freeing ->numa_faults.
2358 void task_numa_free(struct task_struct *p, bool final)
2360 /* safe: p either is current or is being freed by current */
2361 struct numa_group *grp = rcu_dereference_raw(p->numa_group);
2362 unsigned long *numa_faults = p->numa_faults;
2363 unsigned long flags;
2370 spin_lock_irqsave(&grp->lock, flags);
2371 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2372 grp->faults[i] -= p->numa_faults[i];
2373 grp->total_faults -= p->total_numa_faults;
2376 spin_unlock_irqrestore(&grp->lock, flags);
2377 RCU_INIT_POINTER(p->numa_group, NULL);
2378 put_numa_group(grp);
2382 p->numa_faults = NULL;
2385 p->total_numa_faults = 0;
2386 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2392 * Got a PROT_NONE fault for a page on @node.
2394 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2396 struct task_struct *p = current;
2397 bool migrated = flags & TNF_MIGRATED;
2398 int cpu_node = task_node(current);
2399 int local = !!(flags & TNF_FAULT_LOCAL);
2400 struct numa_group *ng;
2403 if (!static_branch_likely(&sched_numa_balancing))
2406 /* for example, ksmd faulting in a user's mm */
2410 /* Allocate buffer to track faults on a per-node basis */
2411 if (unlikely(!p->numa_faults)) {
2412 int size = sizeof(*p->numa_faults) *
2413 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2415 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2416 if (!p->numa_faults)
2419 p->total_numa_faults = 0;
2420 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2424 * First accesses are treated as private, otherwise consider accesses
2425 * to be private if the accessing pid has not changed
2427 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2430 priv = cpupid_match_pid(p, last_cpupid);
2431 if (!priv && !(flags & TNF_NO_GROUP))
2432 task_numa_group(p, last_cpupid, flags, &priv);
2436 * If a workload spans multiple NUMA nodes, a shared fault that
2437 * occurs wholly within the set of nodes that the workload is
2438 * actively using should be counted as local. This allows the
2439 * scan rate to slow down when a workload has settled down.
2441 ng = deref_curr_numa_group(p);
2442 if (!priv && !local && ng && ng->active_nodes > 1 &&
2443 numa_is_active_node(cpu_node, ng) &&
2444 numa_is_active_node(mem_node, ng))
2448 * Retry to migrate task to preferred node periodically, in case it
2449 * previously failed, or the scheduler moved us.
2451 if (time_after(jiffies, p->numa_migrate_retry)) {
2452 task_numa_placement(p);
2453 numa_migrate_preferred(p);
2457 p->numa_pages_migrated += pages;
2458 if (flags & TNF_MIGRATE_FAIL)
2459 p->numa_faults_locality[2] += pages;
2461 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2462 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2463 p->numa_faults_locality[local] += pages;
2466 static void reset_ptenuma_scan(struct task_struct *p)
2469 * We only did a read acquisition of the mmap sem, so
2470 * p->mm->numa_scan_seq is written to without exclusive access
2471 * and the update is not guaranteed to be atomic. That's not
2472 * much of an issue though, since this is just used for
2473 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2474 * expensive, to avoid any form of compiler optimizations:
2476 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2477 p->mm->numa_scan_offset = 0;
2481 * The expensive part of numa migration is done from task_work context.
2482 * Triggered from task_tick_numa().
2484 static void task_numa_work(struct callback_head *work)
2486 unsigned long migrate, next_scan, now = jiffies;
2487 struct task_struct *p = current;
2488 struct mm_struct *mm = p->mm;
2489 u64 runtime = p->se.sum_exec_runtime;
2490 struct vm_area_struct *vma;
2491 unsigned long start, end;
2492 unsigned long nr_pte_updates = 0;
2493 long pages, virtpages;
2495 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2499 * Who cares about NUMA placement when they're dying.
2501 * NOTE: make sure not to dereference p->mm before this check,
2502 * exit_task_work() happens _after_ exit_mm() so we could be called
2503 * without p->mm even though we still had it when we enqueued this
2506 if (p->flags & PF_EXITING)
2509 if (!mm->numa_next_scan) {
2510 mm->numa_next_scan = now +
2511 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2515 * Enforce maximal scan/migration frequency..
2517 migrate = mm->numa_next_scan;
2518 if (time_before(now, migrate))
2521 if (p->numa_scan_period == 0) {
2522 p->numa_scan_period_max = task_scan_max(p);
2523 p->numa_scan_period = task_scan_start(p);
2526 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2527 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2531 * Delay this task enough that another task of this mm will likely win
2532 * the next time around.
2534 p->node_stamp += 2 * TICK_NSEC;
2536 start = mm->numa_scan_offset;
2537 pages = sysctl_numa_balancing_scan_size;
2538 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2539 virtpages = pages * 8; /* Scan up to this much virtual space */
2544 if (!down_read_trylock(&mm->mmap_sem))
2546 vma = find_vma(mm, start);
2548 reset_ptenuma_scan(p);
2552 for (; vma; vma = vma->vm_next) {
2553 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2554 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2559 * Shared library pages mapped by multiple processes are not
2560 * migrated as it is expected they are cache replicated. Avoid
2561 * hinting faults in read-only file-backed mappings or the vdso
2562 * as migrating the pages will be of marginal benefit.
2565 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2569 * Skip inaccessible VMAs to avoid any confusion between
2570 * PROT_NONE and NUMA hinting ptes
2572 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2576 start = max(start, vma->vm_start);
2577 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2578 end = min(end, vma->vm_end);
2579 nr_pte_updates = change_prot_numa(vma, start, end);
2582 * Try to scan sysctl_numa_balancing_size worth of
2583 * hpages that have at least one present PTE that
2584 * is not already pte-numa. If the VMA contains
2585 * areas that are unused or already full of prot_numa
2586 * PTEs, scan up to virtpages, to skip through those
2590 pages -= (end - start) >> PAGE_SHIFT;
2591 virtpages -= (end - start) >> PAGE_SHIFT;
2594 if (pages <= 0 || virtpages <= 0)
2598 } while (end != vma->vm_end);
2603 * It is possible to reach the end of the VMA list but the last few
2604 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2605 * would find the !migratable VMA on the next scan but not reset the
2606 * scanner to the start so check it now.
2609 mm->numa_scan_offset = start;
2611 reset_ptenuma_scan(p);
2612 up_read(&mm->mmap_sem);
2615 * Make sure tasks use at least 32x as much time to run other code
2616 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2617 * Usually update_task_scan_period slows down scanning enough; on an
2618 * overloaded system we need to limit overhead on a per task basis.
2620 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2621 u64 diff = p->se.sum_exec_runtime - runtime;
2622 p->node_stamp += 32 * diff;
2626 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
2629 struct mm_struct *mm = p->mm;
2632 mm_users = atomic_read(&mm->mm_users);
2633 if (mm_users == 1) {
2634 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2635 mm->numa_scan_seq = 0;
2639 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
2640 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2641 /* Protect against double add, see task_tick_numa and task_numa_work */
2642 p->numa_work.next = &p->numa_work;
2643 p->numa_faults = NULL;
2644 RCU_INIT_POINTER(p->numa_group, NULL);
2645 p->last_task_numa_placement = 0;
2646 p->last_sum_exec_runtime = 0;
2648 init_task_work(&p->numa_work, task_numa_work);
2650 /* New address space, reset the preferred nid */
2651 if (!(clone_flags & CLONE_VM)) {
2652 p->numa_preferred_nid = NUMA_NO_NODE;
2657 * New thread, keep existing numa_preferred_nid which should be copied
2658 * already by arch_dup_task_struct but stagger when scans start.
2663 delay = min_t(unsigned int, task_scan_max(current),
2664 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
2665 delay += 2 * TICK_NSEC;
2666 p->node_stamp = delay;
2671 * Drive the periodic memory faults..
2673 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2675 struct callback_head *work = &curr->numa_work;
2679 * We don't care about NUMA placement if we don't have memory.
2681 if ((curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
2685 * Using runtime rather than walltime has the dual advantage that
2686 * we (mostly) drive the selection from busy threads and that the
2687 * task needs to have done some actual work before we bother with
2690 now = curr->se.sum_exec_runtime;
2691 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2693 if (now > curr->node_stamp + period) {
2694 if (!curr->node_stamp)
2695 curr->numa_scan_period = task_scan_start(curr);
2696 curr->node_stamp += period;
2698 if (!time_before(jiffies, curr->mm->numa_next_scan))
2699 task_work_add(curr, work, true);
2703 static void update_scan_period(struct task_struct *p, int new_cpu)
2705 int src_nid = cpu_to_node(task_cpu(p));
2706 int dst_nid = cpu_to_node(new_cpu);
2708 if (!static_branch_likely(&sched_numa_balancing))
2711 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2714 if (src_nid == dst_nid)
2718 * Allow resets if faults have been trapped before one scan
2719 * has completed. This is most likely due to a new task that
2720 * is pulled cross-node due to wakeups or load balancing.
2722 if (p->numa_scan_seq) {
2724 * Avoid scan adjustments if moving to the preferred
2725 * node or if the task was not previously running on
2726 * the preferred node.
2728 if (dst_nid == p->numa_preferred_nid ||
2729 (p->numa_preferred_nid != NUMA_NO_NODE &&
2730 src_nid != p->numa_preferred_nid))
2734 p->numa_scan_period = task_scan_start(p);
2738 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2742 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2746 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2750 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2754 #endif /* CONFIG_NUMA_BALANCING */
2757 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2759 update_load_add(&cfs_rq->load, se->load.weight);
2761 if (entity_is_task(se)) {
2762 struct rq *rq = rq_of(cfs_rq);
2764 account_numa_enqueue(rq, task_of(se));
2765 list_add(&se->group_node, &rq->cfs_tasks);
2768 cfs_rq->nr_running++;
2772 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2774 update_load_sub(&cfs_rq->load, se->load.weight);
2776 if (entity_is_task(se)) {
2777 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2778 list_del_init(&se->group_node);
2781 cfs_rq->nr_running--;
2785 * Signed add and clamp on underflow.
2787 * Explicitly do a load-store to ensure the intermediate value never hits
2788 * memory. This allows lockless observations without ever seeing the negative
2791 #define add_positive(_ptr, _val) do { \
2792 typeof(_ptr) ptr = (_ptr); \
2793 typeof(_val) val = (_val); \
2794 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2798 if (val < 0 && res > var) \
2801 WRITE_ONCE(*ptr, res); \
2805 * Unsigned subtract and clamp on underflow.
2807 * Explicitly do a load-store to ensure the intermediate value never hits
2808 * memory. This allows lockless observations without ever seeing the negative
2811 #define sub_positive(_ptr, _val) do { \
2812 typeof(_ptr) ptr = (_ptr); \
2813 typeof(*ptr) val = (_val); \
2814 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2818 WRITE_ONCE(*ptr, res); \
2822 * Remove and clamp on negative, from a local variable.
2824 * A variant of sub_positive(), which does not use explicit load-store
2825 * and is thus optimized for local variable updates.
2827 #define lsub_positive(_ptr, _val) do { \
2828 typeof(_ptr) ptr = (_ptr); \
2829 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
2834 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2836 cfs_rq->runnable_weight += se->runnable_weight;
2838 cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2839 cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2843 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2845 cfs_rq->runnable_weight -= se->runnable_weight;
2847 sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2848 sub_positive(&cfs_rq->avg.runnable_load_sum,
2849 se_runnable(se) * se->avg.runnable_load_sum);
2853 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2855 cfs_rq->avg.load_avg += se->avg.load_avg;
2856 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2860 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2862 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2863 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2867 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2869 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2871 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2873 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2876 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2877 unsigned long weight, unsigned long runnable)
2880 /* commit outstanding execution time */
2881 if (cfs_rq->curr == se)
2882 update_curr(cfs_rq);
2883 account_entity_dequeue(cfs_rq, se);
2884 dequeue_runnable_load_avg(cfs_rq, se);
2886 dequeue_load_avg(cfs_rq, se);
2888 se->runnable_weight = runnable;
2889 update_load_set(&se->load, weight);
2893 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2895 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2896 se->avg.runnable_load_avg =
2897 div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2901 enqueue_load_avg(cfs_rq, se);
2903 account_entity_enqueue(cfs_rq, se);
2904 enqueue_runnable_load_avg(cfs_rq, se);
2908 void reweight_task(struct task_struct *p, int prio)
2910 struct sched_entity *se = &p->se;
2911 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2912 struct load_weight *load = &se->load;
2913 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2915 reweight_entity(cfs_rq, se, weight, weight);
2916 load->inv_weight = sched_prio_to_wmult[prio];
2919 #ifdef CONFIG_FAIR_GROUP_SCHED
2922 * All this does is approximate the hierarchical proportion which includes that
2923 * global sum we all love to hate.
2925 * That is, the weight of a group entity, is the proportional share of the
2926 * group weight based on the group runqueue weights. That is:
2928 * tg->weight * grq->load.weight
2929 * ge->load.weight = ----------------------------- (1)
2930 * \Sum grq->load.weight
2932 * Now, because computing that sum is prohibitively expensive to compute (been
2933 * there, done that) we approximate it with this average stuff. The average
2934 * moves slower and therefore the approximation is cheaper and more stable.
2936 * So instead of the above, we substitute:
2938 * grq->load.weight -> grq->avg.load_avg (2)
2940 * which yields the following:
2942 * tg->weight * grq->avg.load_avg
2943 * ge->load.weight = ------------------------------ (3)
2946 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2948 * That is shares_avg, and it is right (given the approximation (2)).
2950 * The problem with it is that because the average is slow -- it was designed
2951 * to be exactly that of course -- this leads to transients in boundary
2952 * conditions. In specific, the case where the group was idle and we start the
2953 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2954 * yielding bad latency etc..
2956 * Now, in that special case (1) reduces to:
2958 * tg->weight * grq->load.weight
2959 * ge->load.weight = ----------------------------- = tg->weight (4)
2962 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2964 * So what we do is modify our approximation (3) to approach (4) in the (near)
2969 * tg->weight * grq->load.weight
2970 * --------------------------------------------------- (5)
2971 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2973 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2974 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2977 * tg->weight * grq->load.weight
2978 * ge->load.weight = ----------------------------- (6)
2983 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2984 * max(grq->load.weight, grq->avg.load_avg)
2986 * And that is shares_weight and is icky. In the (near) UP case it approaches
2987 * (4) while in the normal case it approaches (3). It consistently
2988 * overestimates the ge->load.weight and therefore:
2990 * \Sum ge->load.weight >= tg->weight
2994 static long calc_group_shares(struct cfs_rq *cfs_rq)
2996 long tg_weight, tg_shares, load, shares;
2997 struct task_group *tg = cfs_rq->tg;
2999 tg_shares = READ_ONCE(tg->shares);
3001 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3003 tg_weight = atomic_long_read(&tg->load_avg);
3005 /* Ensure tg_weight >= load */
3006 tg_weight -= cfs_rq->tg_load_avg_contrib;
3009 shares = (tg_shares * load);
3011 shares /= tg_weight;
3014 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3015 * of a group with small tg->shares value. It is a floor value which is
3016 * assigned as a minimum load.weight to the sched_entity representing
3017 * the group on a CPU.
3019 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3020 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3021 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3022 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3025 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3029 * This calculates the effective runnable weight for a group entity based on
3030 * the group entity weight calculated above.
3032 * Because of the above approximation (2), our group entity weight is
3033 * an load_avg based ratio (3). This means that it includes blocked load and
3034 * does not represent the runnable weight.
3036 * Approximate the group entity's runnable weight per ratio from the group
3039 * grq->avg.runnable_load_avg
3040 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
3043 * However, analogous to above, since the avg numbers are slow, this leads to
3044 * transients in the from-idle case. Instead we use:
3046 * ge->runnable_weight = ge->load.weight *
3048 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
3049 * ----------------------------------------------------- (8)
3050 * max(grq->avg.load_avg, grq->load.weight)
3052 * Where these max() serve both to use the 'instant' values to fix the slow
3053 * from-idle and avoid the /0 on to-idle, similar to (6).
3055 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
3057 long runnable, load_avg;
3059 load_avg = max(cfs_rq->avg.load_avg,
3060 scale_load_down(cfs_rq->load.weight));
3062 runnable = max(cfs_rq->avg.runnable_load_avg,
3063 scale_load_down(cfs_rq->runnable_weight));
3067 runnable /= load_avg;
3069 return clamp_t(long, runnable, MIN_SHARES, shares);
3071 #endif /* CONFIG_SMP */
3073 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3076 * Recomputes the group entity based on the current state of its group
3079 static void update_cfs_group(struct sched_entity *se)
3081 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3082 long shares, runnable;
3087 if (throttled_hierarchy(gcfs_rq))
3091 runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
3093 if (likely(se->load.weight == shares))
3096 shares = calc_group_shares(gcfs_rq);
3097 runnable = calc_group_runnable(gcfs_rq, shares);
3100 reweight_entity(cfs_rq_of(se), se, shares, runnable);
3103 #else /* CONFIG_FAIR_GROUP_SCHED */
3104 static inline void update_cfs_group(struct sched_entity *se)
3107 #endif /* CONFIG_FAIR_GROUP_SCHED */
3109 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3111 struct rq *rq = rq_of(cfs_rq);
3113 if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) {
3115 * There are a few boundary cases this might miss but it should
3116 * get called often enough that that should (hopefully) not be
3119 * It will not get called when we go idle, because the idle
3120 * thread is a different class (!fair), nor will the utilization
3121 * number include things like RT tasks.
3123 * As is, the util number is not freq-invariant (we'd have to
3124 * implement arch_scale_freq_capacity() for that).
3128 cpufreq_update_util(rq, flags);
3133 #ifdef CONFIG_FAIR_GROUP_SCHED
3135 * update_tg_load_avg - update the tg's load avg
3136 * @cfs_rq: the cfs_rq whose avg changed
3137 * @force: update regardless of how small the difference
3139 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3140 * However, because tg->load_avg is a global value there are performance
3143 * In order to avoid having to look at the other cfs_rq's, we use a
3144 * differential update where we store the last value we propagated. This in
3145 * turn allows skipping updates if the differential is 'small'.
3147 * Updating tg's load_avg is necessary before update_cfs_share().
3149 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3151 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3154 * No need to update load_avg for root_task_group as it is not used.
3156 if (cfs_rq->tg == &root_task_group)
3159 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3160 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3161 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3166 * Called within set_task_rq() right before setting a task's CPU. The
3167 * caller only guarantees p->pi_lock is held; no other assumptions,
3168 * including the state of rq->lock, should be made.
3170 void set_task_rq_fair(struct sched_entity *se,
3171 struct cfs_rq *prev, struct cfs_rq *next)
3173 u64 p_last_update_time;
3174 u64 n_last_update_time;
3176 if (!sched_feat(ATTACH_AGE_LOAD))
3180 * We are supposed to update the task to "current" time, then its up to
3181 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3182 * getting what current time is, so simply throw away the out-of-date
3183 * time. This will result in the wakee task is less decayed, but giving
3184 * the wakee more load sounds not bad.
3186 if (!(se->avg.last_update_time && prev))
3189 #ifndef CONFIG_64BIT
3191 u64 p_last_update_time_copy;
3192 u64 n_last_update_time_copy;
3195 p_last_update_time_copy = prev->load_last_update_time_copy;
3196 n_last_update_time_copy = next->load_last_update_time_copy;
3200 p_last_update_time = prev->avg.last_update_time;
3201 n_last_update_time = next->avg.last_update_time;
3203 } while (p_last_update_time != p_last_update_time_copy ||
3204 n_last_update_time != n_last_update_time_copy);
3207 p_last_update_time = prev->avg.last_update_time;
3208 n_last_update_time = next->avg.last_update_time;
3210 __update_load_avg_blocked_se(p_last_update_time, se);
3211 se->avg.last_update_time = n_last_update_time;
3216 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3217 * propagate its contribution. The key to this propagation is the invariant
3218 * that for each group:
3220 * ge->avg == grq->avg (1)
3222 * _IFF_ we look at the pure running and runnable sums. Because they
3223 * represent the very same entity, just at different points in the hierarchy.
3225 * Per the above update_tg_cfs_util() is trivial and simply copies the running
3226 * sum over (but still wrong, because the group entity and group rq do not have
3227 * their PELT windows aligned).
3229 * However, update_tg_cfs_runnable() is more complex. So we have:
3231 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3233 * And since, like util, the runnable part should be directly transferable,
3234 * the following would _appear_ to be the straight forward approach:
3236 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3238 * And per (1) we have:
3240 * ge->avg.runnable_avg == grq->avg.runnable_avg
3244 * ge->load.weight * grq->avg.load_avg
3245 * ge->avg.load_avg = ----------------------------------- (4)
3248 * Except that is wrong!
3250 * Because while for entities historical weight is not important and we
3251 * really only care about our future and therefore can consider a pure
3252 * runnable sum, runqueues can NOT do this.
3254 * We specifically want runqueues to have a load_avg that includes
3255 * historical weights. Those represent the blocked load, the load we expect
3256 * to (shortly) return to us. This only works by keeping the weights as
3257 * integral part of the sum. We therefore cannot decompose as per (3).
3259 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3260 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3261 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3262 * runnable section of these tasks overlap (or not). If they were to perfectly
3263 * align the rq as a whole would be runnable 2/3 of the time. If however we
3264 * always have at least 1 runnable task, the rq as a whole is always runnable.
3266 * So we'll have to approximate.. :/
3268 * Given the constraint:
3270 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3272 * We can construct a rule that adds runnable to a rq by assuming minimal
3275 * On removal, we'll assume each task is equally runnable; which yields:
3277 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3279 * XXX: only do this for the part of runnable > running ?
3284 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3286 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3288 /* Nothing to update */
3293 * The relation between sum and avg is:
3295 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3297 * however, the PELT windows are not aligned between grq and gse.
3300 /* Set new sched_entity's utilization */
3301 se->avg.util_avg = gcfs_rq->avg.util_avg;
3302 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3304 /* Update parent cfs_rq utilization */
3305 add_positive(&cfs_rq->avg.util_avg, delta);
3306 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3310 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3312 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3313 unsigned long runnable_load_avg, load_avg;
3314 u64 runnable_load_sum, load_sum = 0;
3320 gcfs_rq->prop_runnable_sum = 0;
3322 if (runnable_sum >= 0) {
3324 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3325 * the CPU is saturated running == runnable.
3327 runnable_sum += se->avg.load_sum;
3328 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3331 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3332 * assuming all tasks are equally runnable.
3334 if (scale_load_down(gcfs_rq->load.weight)) {
3335 load_sum = div_s64(gcfs_rq->avg.load_sum,
3336 scale_load_down(gcfs_rq->load.weight));
3339 /* But make sure to not inflate se's runnable */
3340 runnable_sum = min(se->avg.load_sum, load_sum);
3344 * runnable_sum can't be lower than running_sum
3345 * Rescale running sum to be in the same range as runnable sum
3346 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3347 * runnable_sum is in [0 : LOAD_AVG_MAX]
3349 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3350 runnable_sum = max(runnable_sum, running_sum);
3352 load_sum = (s64)se_weight(se) * runnable_sum;
3353 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3355 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3356 delta_avg = load_avg - se->avg.load_avg;
3358 se->avg.load_sum = runnable_sum;
3359 se->avg.load_avg = load_avg;
3360 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3361 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3363 runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3364 runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3365 delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3366 delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3368 se->avg.runnable_load_sum = runnable_sum;
3369 se->avg.runnable_load_avg = runnable_load_avg;
3372 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3373 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3377 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3379 cfs_rq->propagate = 1;
3380 cfs_rq->prop_runnable_sum += runnable_sum;
3383 /* Update task and its cfs_rq load average */
3384 static inline int propagate_entity_load_avg(struct sched_entity *se)
3386 struct cfs_rq *cfs_rq, *gcfs_rq;
3388 if (entity_is_task(se))
3391 gcfs_rq = group_cfs_rq(se);
3392 if (!gcfs_rq->propagate)
3395 gcfs_rq->propagate = 0;
3397 cfs_rq = cfs_rq_of(se);
3399 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3401 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3402 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3404 trace_pelt_cfs_tp(cfs_rq);
3405 trace_pelt_se_tp(se);
3411 * Check if we need to update the load and the utilization of a blocked
3414 static inline bool skip_blocked_update(struct sched_entity *se)
3416 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3419 * If sched_entity still have not zero load or utilization, we have to
3422 if (se->avg.load_avg || se->avg.util_avg)
3426 * If there is a pending propagation, we have to update the load and
3427 * the utilization of the sched_entity:
3429 if (gcfs_rq->propagate)
3433 * Otherwise, the load and the utilization of the sched_entity is
3434 * already zero and there is no pending propagation, so it will be a
3435 * waste of time to try to decay it:
3440 #else /* CONFIG_FAIR_GROUP_SCHED */
3442 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3444 static inline int propagate_entity_load_avg(struct sched_entity *se)
3449 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3451 #endif /* CONFIG_FAIR_GROUP_SCHED */
3454 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3455 * @now: current time, as per cfs_rq_clock_pelt()
3456 * @cfs_rq: cfs_rq to update
3458 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3459 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3460 * post_init_entity_util_avg().
3462 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3464 * Returns true if the load decayed or we removed load.
3466 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3467 * call update_tg_load_avg() when this function returns true.
3470 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3472 unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3473 struct sched_avg *sa = &cfs_rq->avg;
3476 if (cfs_rq->removed.nr) {
3478 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3480 raw_spin_lock(&cfs_rq->removed.lock);
3481 swap(cfs_rq->removed.util_avg, removed_util);
3482 swap(cfs_rq->removed.load_avg, removed_load);
3483 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3484 cfs_rq->removed.nr = 0;
3485 raw_spin_unlock(&cfs_rq->removed.lock);
3488 sub_positive(&sa->load_avg, r);
3489 sub_positive(&sa->load_sum, r * divider);
3492 sub_positive(&sa->util_avg, r);
3493 sub_positive(&sa->util_sum, r * divider);
3495 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3500 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
3502 #ifndef CONFIG_64BIT
3504 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3511 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3512 * @cfs_rq: cfs_rq to attach to
3513 * @se: sched_entity to attach
3514 * @flags: migration hints
3516 * Must call update_cfs_rq_load_avg() before this, since we rely on
3517 * cfs_rq->avg.last_update_time being current.
3519 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3521 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3524 * When we attach the @se to the @cfs_rq, we must align the decay
3525 * window because without that, really weird and wonderful things can
3530 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3531 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3534 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3535 * period_contrib. This isn't strictly correct, but since we're
3536 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3539 se->avg.util_sum = se->avg.util_avg * divider;
3541 se->avg.load_sum = divider;
3542 if (se_weight(se)) {
3544 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3547 se->avg.runnable_load_sum = se->avg.load_sum;
3549 enqueue_load_avg(cfs_rq, se);
3550 cfs_rq->avg.util_avg += se->avg.util_avg;
3551 cfs_rq->avg.util_sum += se->avg.util_sum;
3553 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3555 cfs_rq_util_change(cfs_rq, flags);
3557 trace_pelt_cfs_tp(cfs_rq);
3561 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3562 * @cfs_rq: cfs_rq to detach from
3563 * @se: sched_entity to detach
3565 * Must call update_cfs_rq_load_avg() before this, since we rely on
3566 * cfs_rq->avg.last_update_time being current.
3568 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3570 dequeue_load_avg(cfs_rq, se);
3571 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3572 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3574 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3576 cfs_rq_util_change(cfs_rq, 0);
3578 trace_pelt_cfs_tp(cfs_rq);
3582 * Optional action to be done while updating the load average
3584 #define UPDATE_TG 0x1
3585 #define SKIP_AGE_LOAD 0x2
3586 #define DO_ATTACH 0x4
3588 /* Update task and its cfs_rq load average */
3589 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3591 u64 now = cfs_rq_clock_pelt(cfs_rq);
3595 * Track task load average for carrying it to new CPU after migrated, and
3596 * track group sched_entity load average for task_h_load calc in migration
3598 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3599 __update_load_avg_se(now, cfs_rq, se);
3601 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3602 decayed |= propagate_entity_load_avg(se);
3604 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3607 * DO_ATTACH means we're here from enqueue_entity().
3608 * !last_update_time means we've passed through
3609 * migrate_task_rq_fair() indicating we migrated.
3611 * IOW we're enqueueing a task on a new CPU.
3613 attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
3614 update_tg_load_avg(cfs_rq, 0);
3616 } else if (decayed) {
3617 cfs_rq_util_change(cfs_rq, 0);
3619 if (flags & UPDATE_TG)
3620 update_tg_load_avg(cfs_rq, 0);
3624 #ifndef CONFIG_64BIT
3625 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3627 u64 last_update_time_copy;
3628 u64 last_update_time;
3631 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3633 last_update_time = cfs_rq->avg.last_update_time;
3634 } while (last_update_time != last_update_time_copy);
3636 return last_update_time;
3639 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3641 return cfs_rq->avg.last_update_time;
3646 * Synchronize entity load avg of dequeued entity without locking
3649 static void sync_entity_load_avg(struct sched_entity *se)
3651 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3652 u64 last_update_time;
3654 last_update_time = cfs_rq_last_update_time(cfs_rq);
3655 __update_load_avg_blocked_se(last_update_time, se);
3659 * Task first catches up with cfs_rq, and then subtract
3660 * itself from the cfs_rq (task must be off the queue now).
3662 static void remove_entity_load_avg(struct sched_entity *se)
3664 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3665 unsigned long flags;
3668 * tasks cannot exit without having gone through wake_up_new_task() ->
3669 * post_init_entity_util_avg() which will have added things to the
3670 * cfs_rq, so we can remove unconditionally.
3673 sync_entity_load_avg(se);
3675 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3676 ++cfs_rq->removed.nr;
3677 cfs_rq->removed.util_avg += se->avg.util_avg;
3678 cfs_rq->removed.load_avg += se->avg.load_avg;
3679 cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */
3680 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3683 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3685 return cfs_rq->avg.runnable_load_avg;
3688 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3690 return cfs_rq->avg.load_avg;
3693 static inline unsigned long task_util(struct task_struct *p)
3695 return READ_ONCE(p->se.avg.util_avg);
3698 static inline unsigned long _task_util_est(struct task_struct *p)
3700 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3702 return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED);
3705 static inline unsigned long task_util_est(struct task_struct *p)
3707 return max(task_util(p), _task_util_est(p));
3710 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3711 struct task_struct *p)
3713 unsigned int enqueued;
3715 if (!sched_feat(UTIL_EST))
3718 /* Update root cfs_rq's estimated utilization */
3719 enqueued = cfs_rq->avg.util_est.enqueued;
3720 enqueued += _task_util_est(p);
3721 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3725 * Check if a (signed) value is within a specified (unsigned) margin,
3726 * based on the observation that:
3728 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3730 * NOTE: this only works when value + maring < INT_MAX.
3732 static inline bool within_margin(int value, int margin)
3734 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3738 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3740 long last_ewma_diff;
3744 if (!sched_feat(UTIL_EST))
3747 /* Update root cfs_rq's estimated utilization */
3748 ue.enqueued = cfs_rq->avg.util_est.enqueued;
3749 ue.enqueued -= min_t(unsigned int, ue.enqueued, _task_util_est(p));
3750 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3753 * Skip update of task's estimated utilization when the task has not
3754 * yet completed an activation, e.g. being migrated.
3760 * If the PELT values haven't changed since enqueue time,
3761 * skip the util_est update.
3763 ue = p->se.avg.util_est;
3764 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3768 * Skip update of task's estimated utilization when its EWMA is
3769 * already ~1% close to its last activation value.
3771 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3772 last_ewma_diff = ue.enqueued - ue.ewma;
3773 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3777 * To avoid overestimation of actual task utilization, skip updates if
3778 * we cannot grant there is idle time in this CPU.
3780 cpu = cpu_of(rq_of(cfs_rq));
3781 if (task_util(p) > capacity_orig_of(cpu))
3785 * Update Task's estimated utilization
3787 * When *p completes an activation we can consolidate another sample
3788 * of the task size. This is done by storing the current PELT value
3789 * as ue.enqueued and by using this value to update the Exponential
3790 * Weighted Moving Average (EWMA):
3792 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
3793 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
3794 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
3795 * = w * ( last_ewma_diff ) + ewma(t-1)
3796 * = w * (last_ewma_diff + ewma(t-1) / w)
3798 * Where 'w' is the weight of new samples, which is configured to be
3799 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3801 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
3802 ue.ewma += last_ewma_diff;
3803 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
3804 WRITE_ONCE(p->se.avg.util_est, ue);
3807 static inline int task_fits_capacity(struct task_struct *p, long capacity)
3809 return fits_capacity(task_util_est(p), capacity);
3812 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
3814 if (!static_branch_unlikely(&sched_asym_cpucapacity))
3817 if (!p || p->nr_cpus_allowed == 1) {
3818 rq->misfit_task_load = 0;
3822 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
3823 rq->misfit_task_load = 0;
3828 * Make sure that misfit_task_load will not be null even if
3829 * task_h_load() returns 0.
3831 rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1);
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");
3934 static inline bool cfs_bandwidth_used(void);
3941 * update_min_vruntime()
3942 * vruntime -= min_vruntime
3946 * update_min_vruntime()
3947 * vruntime += min_vruntime
3949 * this way the vruntime transition between RQs is done when both
3950 * min_vruntime are up-to-date.
3954 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3955 * vruntime -= min_vruntime
3959 * update_min_vruntime()
3960 * vruntime += min_vruntime
3962 * this way we don't have the most up-to-date min_vruntime on the originating
3963 * CPU and an up-to-date min_vruntime on the destination CPU.
3967 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3969 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3970 bool curr = cfs_rq->curr == se;
3973 * If we're the current task, we must renormalise before calling
3977 se->vruntime += cfs_rq->min_vruntime;
3979 update_curr(cfs_rq);
3982 * Otherwise, renormalise after, such that we're placed at the current
3983 * moment in time, instead of some random moment in the past. Being
3984 * placed in the past could significantly boost this task to the
3985 * fairness detriment of existing tasks.
3987 if (renorm && !curr)
3988 se->vruntime += cfs_rq->min_vruntime;
3991 * When enqueuing a sched_entity, we must:
3992 * - Update loads to have both entity and cfs_rq synced with now.
3993 * - Add its load to cfs_rq->runnable_avg
3994 * - For group_entity, update its weight to reflect the new share of
3996 * - Add its new weight to cfs_rq->load.weight
3998 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
3999 update_cfs_group(se);
4000 enqueue_runnable_load_avg(cfs_rq, se);
4001 account_entity_enqueue(cfs_rq, se);
4003 if (flags & ENQUEUE_WAKEUP)
4004 place_entity(cfs_rq, se, 0);
4006 check_schedstat_required();
4007 update_stats_enqueue(cfs_rq, se, flags);
4008 check_spread(cfs_rq, se);
4010 __enqueue_entity(cfs_rq, se);
4014 * When bandwidth control is enabled, cfs might have been removed
4015 * because of a parent been throttled but cfs->nr_running > 1. Try to
4016 * add it unconditionnally.
4018 if (cfs_rq->nr_running == 1 || cfs_bandwidth_used())
4019 list_add_leaf_cfs_rq(cfs_rq);
4021 if (cfs_rq->nr_running == 1)
4022 check_enqueue_throttle(cfs_rq);
4025 static void __clear_buddies_last(struct sched_entity *se)
4027 for_each_sched_entity(se) {
4028 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4029 if (cfs_rq->last != se)
4032 cfs_rq->last = NULL;
4036 static void __clear_buddies_next(struct sched_entity *se)
4038 for_each_sched_entity(se) {
4039 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4040 if (cfs_rq->next != se)
4043 cfs_rq->next = NULL;
4047 static void __clear_buddies_skip(struct sched_entity *se)
4049 for_each_sched_entity(se) {
4050 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4051 if (cfs_rq->skip != se)
4054 cfs_rq->skip = NULL;
4058 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4060 if (cfs_rq->last == se)
4061 __clear_buddies_last(se);
4063 if (cfs_rq->next == se)
4064 __clear_buddies_next(se);
4066 if (cfs_rq->skip == se)
4067 __clear_buddies_skip(se);
4070 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4073 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4076 * Update run-time statistics of the 'current'.
4078 update_curr(cfs_rq);
4081 * When dequeuing a sched_entity, we must:
4082 * - Update loads to have both entity and cfs_rq synced with now.
4083 * - Subtract its load from the cfs_rq->runnable_avg.
4084 * - Subtract its previous weight from cfs_rq->load.weight.
4085 * - For group entity, update its weight to reflect the new share
4086 * of its group cfs_rq.
4088 update_load_avg(cfs_rq, se, UPDATE_TG);
4089 dequeue_runnable_load_avg(cfs_rq, se);
4091 update_stats_dequeue(cfs_rq, se, flags);
4093 clear_buddies(cfs_rq, se);
4095 if (se != cfs_rq->curr)
4096 __dequeue_entity(cfs_rq, se);
4098 account_entity_dequeue(cfs_rq, se);
4101 * Normalize after update_curr(); which will also have moved
4102 * min_vruntime if @se is the one holding it back. But before doing
4103 * update_min_vruntime() again, which will discount @se's position and
4104 * can move min_vruntime forward still more.
4106 if (!(flags & DEQUEUE_SLEEP))
4107 se->vruntime -= cfs_rq->min_vruntime;
4109 /* return excess runtime on last dequeue */
4110 return_cfs_rq_runtime(cfs_rq);
4112 update_cfs_group(se);
4115 * Now advance min_vruntime if @se was the entity holding it back,
4116 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4117 * put back on, and if we advance min_vruntime, we'll be placed back
4118 * further than we started -- ie. we'll be penalized.
4120 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4121 update_min_vruntime(cfs_rq);
4125 * Preempt the current task with a newly woken task if needed:
4128 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4130 unsigned long ideal_runtime, delta_exec;
4131 struct sched_entity *se;
4134 ideal_runtime = sched_slice(cfs_rq, curr);
4135 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4136 if (delta_exec > ideal_runtime) {
4137 resched_curr(rq_of(cfs_rq));
4139 * The current task ran long enough, ensure it doesn't get
4140 * re-elected due to buddy favours.
4142 clear_buddies(cfs_rq, curr);
4147 * Ensure that a task that missed wakeup preemption by a
4148 * narrow margin doesn't have to wait for a full slice.
4149 * This also mitigates buddy induced latencies under load.
4151 if (delta_exec < sysctl_sched_min_granularity)
4154 se = __pick_first_entity(cfs_rq);
4155 delta = curr->vruntime - se->vruntime;
4160 if (delta > ideal_runtime)
4161 resched_curr(rq_of(cfs_rq));
4165 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4167 /* 'current' is not kept within the tree. */
4170 * Any task has to be enqueued before it get to execute on
4171 * a CPU. So account for the time it spent waiting on the
4174 update_stats_wait_end(cfs_rq, se);
4175 __dequeue_entity(cfs_rq, se);
4176 update_load_avg(cfs_rq, se, UPDATE_TG);
4179 update_stats_curr_start(cfs_rq, se);
4183 * Track our maximum slice length, if the CPU's load is at
4184 * least twice that of our own weight (i.e. dont track it
4185 * when there are only lesser-weight tasks around):
4187 if (schedstat_enabled() &&
4188 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
4189 schedstat_set(se->statistics.slice_max,
4190 max((u64)schedstat_val(se->statistics.slice_max),
4191 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4194 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4198 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4201 * Pick the next process, keeping these things in mind, in this order:
4202 * 1) keep things fair between processes/task groups
4203 * 2) pick the "next" process, since someone really wants that to run
4204 * 3) pick the "last" process, for cache locality
4205 * 4) do not run the "skip" process, if something else is available
4207 static struct sched_entity *
4208 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4210 struct sched_entity *left = __pick_first_entity(cfs_rq);
4211 struct sched_entity *se;
4214 * If curr is set we have to see if its left of the leftmost entity
4215 * still in the tree, provided there was anything in the tree at all.
4217 if (!left || (curr && entity_before(curr, left)))
4220 se = left; /* ideally we run the leftmost entity */
4223 * Avoid running the skip buddy, if running something else can
4224 * be done without getting too unfair.
4226 if (cfs_rq->skip == se) {
4227 struct sched_entity *second;
4230 second = __pick_first_entity(cfs_rq);
4232 second = __pick_next_entity(se);
4233 if (!second || (curr && entity_before(curr, second)))
4237 if (second && wakeup_preempt_entity(second, left) < 1)
4242 * Prefer last buddy, try to return the CPU to a preempted task.
4244 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4248 * Someone really wants this to run. If it's not unfair, run it.
4250 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4253 clear_buddies(cfs_rq, se);
4258 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4260 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4263 * If still on the runqueue then deactivate_task()
4264 * was not called and update_curr() has to be done:
4267 update_curr(cfs_rq);
4269 /* throttle cfs_rqs exceeding runtime */
4270 check_cfs_rq_runtime(cfs_rq);
4272 check_spread(cfs_rq, prev);
4275 update_stats_wait_start(cfs_rq, prev);
4276 /* Put 'current' back into the tree. */
4277 __enqueue_entity(cfs_rq, prev);
4278 /* in !on_rq case, update occurred at dequeue */
4279 update_load_avg(cfs_rq, prev, 0);
4281 cfs_rq->curr = NULL;
4285 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4288 * Update run-time statistics of the 'current'.
4290 update_curr(cfs_rq);
4293 * Ensure that runnable average is periodically updated.
4295 update_load_avg(cfs_rq, curr, UPDATE_TG);
4296 update_cfs_group(curr);
4298 #ifdef CONFIG_SCHED_HRTICK
4300 * queued ticks are scheduled to match the slice, so don't bother
4301 * validating it and just reschedule.
4304 resched_curr(rq_of(cfs_rq));
4308 * don't let the period tick interfere with the hrtick preemption
4310 if (!sched_feat(DOUBLE_TICK) &&
4311 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4315 if (cfs_rq->nr_running > 1)
4316 check_preempt_tick(cfs_rq, curr);
4320 /**************************************************
4321 * CFS bandwidth control machinery
4324 #ifdef CONFIG_CFS_BANDWIDTH
4326 #ifdef CONFIG_JUMP_LABEL
4327 static struct static_key __cfs_bandwidth_used;
4329 static inline bool cfs_bandwidth_used(void)
4331 return static_key_false(&__cfs_bandwidth_used);
4334 void cfs_bandwidth_usage_inc(void)
4336 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4339 void cfs_bandwidth_usage_dec(void)
4341 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4343 #else /* CONFIG_JUMP_LABEL */
4344 static bool cfs_bandwidth_used(void)
4349 void cfs_bandwidth_usage_inc(void) {}
4350 void cfs_bandwidth_usage_dec(void) {}
4351 #endif /* CONFIG_JUMP_LABEL */
4354 * default period for cfs group bandwidth.
4355 * default: 0.1s, units: nanoseconds
4357 static inline u64 default_cfs_period(void)
4359 return 100000000ULL;
4362 static inline u64 sched_cfs_bandwidth_slice(void)
4364 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4368 * Replenish runtime according to assigned quota. We use sched_clock_cpu
4369 * directly instead of rq->clock to avoid adding additional synchronization
4372 * requires cfs_b->lock
4374 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4376 if (cfs_b->quota != RUNTIME_INF)
4377 cfs_b->runtime = cfs_b->quota;
4380 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4382 return &tg->cfs_bandwidth;
4385 /* returns 0 on failure to allocate runtime */
4386 static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b,
4387 struct cfs_rq *cfs_rq, u64 target_runtime)
4389 u64 min_amount, amount = 0;
4391 lockdep_assert_held(&cfs_b->lock);
4393 /* note: this is a positive sum as runtime_remaining <= 0 */
4394 min_amount = target_runtime - cfs_rq->runtime_remaining;
4396 if (cfs_b->quota == RUNTIME_INF)
4397 amount = min_amount;
4399 start_cfs_bandwidth(cfs_b);
4401 if (cfs_b->runtime > 0) {
4402 amount = min(cfs_b->runtime, min_amount);
4403 cfs_b->runtime -= amount;
4408 cfs_rq->runtime_remaining += amount;
4410 return cfs_rq->runtime_remaining > 0;
4413 /* returns 0 on failure to allocate runtime */
4414 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4416 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4419 raw_spin_lock(&cfs_b->lock);
4420 ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice());
4421 raw_spin_unlock(&cfs_b->lock);
4426 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4428 /* dock delta_exec before expiring quota (as it could span periods) */
4429 cfs_rq->runtime_remaining -= delta_exec;
4431 if (likely(cfs_rq->runtime_remaining > 0))
4434 if (cfs_rq->throttled)
4437 * if we're unable to extend our runtime we resched so that the active
4438 * hierarchy can be throttled
4440 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4441 resched_curr(rq_of(cfs_rq));
4444 static __always_inline
4445 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4447 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4450 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4453 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4455 return cfs_bandwidth_used() && cfs_rq->throttled;
4458 /* check whether cfs_rq, or any parent, is throttled */
4459 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4461 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4465 * Ensure that neither of the group entities corresponding to src_cpu or
4466 * dest_cpu are members of a throttled hierarchy when performing group
4467 * load-balance operations.
4469 static inline int throttled_lb_pair(struct task_group *tg,
4470 int src_cpu, int dest_cpu)
4472 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4474 src_cfs_rq = tg->cfs_rq[src_cpu];
4475 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4477 return throttled_hierarchy(src_cfs_rq) ||
4478 throttled_hierarchy(dest_cfs_rq);
4481 static int tg_unthrottle_up(struct task_group *tg, void *data)
4483 struct rq *rq = data;
4484 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4486 cfs_rq->throttle_count--;
4487 if (!cfs_rq->throttle_count) {
4488 cfs_rq->throttled_clock_pelt_time += rq_clock_pelt(rq) -
4489 cfs_rq->throttled_clock_pelt;
4491 /* Add cfs_rq with already running entity in the list */
4492 if (cfs_rq->nr_running >= 1)
4493 list_add_leaf_cfs_rq(cfs_rq);
4499 static int tg_throttle_down(struct task_group *tg, void *data)
4501 struct rq *rq = data;
4502 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4504 /* group is entering throttled state, stop time */
4505 if (!cfs_rq->throttle_count) {
4506 cfs_rq->throttled_clock_pelt = rq_clock_pelt(rq);
4507 list_del_leaf_cfs_rq(cfs_rq);
4509 cfs_rq->throttle_count++;
4514 static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
4516 struct rq *rq = rq_of(cfs_rq);
4517 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4518 struct sched_entity *se;
4519 long task_delta, idle_task_delta, dequeue = 1;
4521 raw_spin_lock(&cfs_b->lock);
4522 /* This will start the period timer if necessary */
4523 if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) {
4525 * We have raced with bandwidth becoming available, and if we
4526 * actually throttled the timer might not unthrottle us for an
4527 * entire period. We additionally needed to make sure that any
4528 * subsequent check_cfs_rq_runtime calls agree not to throttle
4529 * us, as we may commit to do cfs put_prev+pick_next, so we ask
4530 * for 1ns of runtime rather than just check cfs_b.
4534 list_add_tail_rcu(&cfs_rq->throttled_list,
4535 &cfs_b->throttled_cfs_rq);
4537 raw_spin_unlock(&cfs_b->lock);
4540 return false; /* Throttle no longer required. */
4542 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4544 /* freeze hierarchy runnable averages while throttled */
4546 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4549 task_delta = cfs_rq->h_nr_running;
4550 idle_task_delta = cfs_rq->idle_h_nr_running;
4551 for_each_sched_entity(se) {
4552 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4553 /* throttled entity or throttle-on-deactivate */
4558 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4559 qcfs_rq->h_nr_running -= task_delta;
4560 qcfs_rq->idle_h_nr_running -= idle_task_delta;
4562 if (qcfs_rq->load.weight)
4567 sub_nr_running(rq, task_delta);
4570 * Note: distribution will already see us throttled via the
4571 * throttled-list. rq->lock protects completion.
4573 cfs_rq->throttled = 1;
4574 cfs_rq->throttled_clock = rq_clock(rq);
4578 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4580 struct rq *rq = rq_of(cfs_rq);
4581 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4582 struct sched_entity *se;
4583 long task_delta, idle_task_delta;
4585 se = cfs_rq->tg->se[cpu_of(rq)];
4587 cfs_rq->throttled = 0;
4589 update_rq_clock(rq);
4591 raw_spin_lock(&cfs_b->lock);
4592 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4593 list_del_rcu(&cfs_rq->throttled_list);
4594 raw_spin_unlock(&cfs_b->lock);
4596 /* update hierarchical throttle state */
4597 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4599 if (!cfs_rq->load.weight)
4602 task_delta = cfs_rq->h_nr_running;
4603 idle_task_delta = cfs_rq->idle_h_nr_running;
4604 for_each_sched_entity(se) {
4607 cfs_rq = cfs_rq_of(se);
4608 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4610 cfs_rq->h_nr_running += task_delta;
4611 cfs_rq->idle_h_nr_running += idle_task_delta;
4613 /* end evaluation on encountering a throttled cfs_rq */
4614 if (cfs_rq_throttled(cfs_rq))
4615 goto unthrottle_throttle;
4618 for_each_sched_entity(se) {
4619 cfs_rq = cfs_rq_of(se);
4621 cfs_rq->h_nr_running += task_delta;
4622 cfs_rq->idle_h_nr_running += idle_task_delta;
4625 /* end evaluation on encountering a throttled cfs_rq */
4626 if (cfs_rq_throttled(cfs_rq))
4627 goto unthrottle_throttle;
4630 * One parent has been throttled and cfs_rq removed from the
4631 * list. Add it back to not break the leaf list.
4633 if (throttled_hierarchy(cfs_rq))
4634 list_add_leaf_cfs_rq(cfs_rq);
4637 /* At this point se is NULL and we are at root level*/
4638 add_nr_running(rq, task_delta);
4640 unthrottle_throttle:
4642 * The cfs_rq_throttled() breaks in the above iteration can result in
4643 * incomplete leaf list maintenance, resulting in triggering the
4646 for_each_sched_entity(se) {
4647 cfs_rq = cfs_rq_of(se);
4649 if (list_add_leaf_cfs_rq(cfs_rq))
4653 assert_list_leaf_cfs_rq(rq);
4655 /* Determine whether we need to wake up potentially idle CPU: */
4656 if (rq->curr == rq->idle && rq->cfs.nr_running)
4660 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, u64 remaining)
4662 struct cfs_rq *cfs_rq;
4664 u64 starting_runtime = remaining;
4667 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4669 struct rq *rq = rq_of(cfs_rq);
4672 rq_lock_irqsave(rq, &rf);
4673 if (!cfs_rq_throttled(cfs_rq))
4676 /* By the above check, this should never be true */
4677 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
4679 runtime = -cfs_rq->runtime_remaining + 1;
4680 if (runtime > remaining)
4681 runtime = remaining;
4682 remaining -= runtime;
4684 cfs_rq->runtime_remaining += runtime;
4686 /* we check whether we're throttled above */
4687 if (cfs_rq->runtime_remaining > 0)
4688 unthrottle_cfs_rq(cfs_rq);
4691 rq_unlock_irqrestore(rq, &rf);
4698 return starting_runtime - remaining;
4702 * Responsible for refilling a task_group's bandwidth and unthrottling its
4703 * cfs_rqs as appropriate. If there has been no activity within the last
4704 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4705 * used to track this state.
4707 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
4712 /* no need to continue the timer with no bandwidth constraint */
4713 if (cfs_b->quota == RUNTIME_INF)
4714 goto out_deactivate;
4716 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4717 cfs_b->nr_periods += overrun;
4720 * idle depends on !throttled (for the case of a large deficit), and if
4721 * we're going inactive then everything else can be deferred
4723 if (cfs_b->idle && !throttled)
4724 goto out_deactivate;
4726 __refill_cfs_bandwidth_runtime(cfs_b);
4729 /* mark as potentially idle for the upcoming period */
4734 /* account preceding periods in which throttling occurred */
4735 cfs_b->nr_throttled += overrun;
4738 * This check is repeated as we are holding onto the new bandwidth while
4739 * we unthrottle. This can potentially race with an unthrottled group
4740 * trying to acquire new bandwidth from the global pool. This can result
4741 * in us over-using our runtime if it is all used during this loop, but
4742 * only by limited amounts in that extreme case.
4744 while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
4745 runtime = cfs_b->runtime;
4746 cfs_b->distribute_running = 1;
4747 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4748 /* we can't nest cfs_b->lock while distributing bandwidth */
4749 runtime = distribute_cfs_runtime(cfs_b, runtime);
4750 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4752 cfs_b->distribute_running = 0;
4753 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4755 lsub_positive(&cfs_b->runtime, runtime);
4759 * While we are ensured activity in the period following an
4760 * unthrottle, this also covers the case in which the new bandwidth is
4761 * insufficient to cover the existing bandwidth deficit. (Forcing the
4762 * timer to remain active while there are any throttled entities.)
4772 /* a cfs_rq won't donate quota below this amount */
4773 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4774 /* minimum remaining period time to redistribute slack quota */
4775 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4776 /* how long we wait to gather additional slack before distributing */
4777 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4780 * Are we near the end of the current quota period?
4782 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4783 * hrtimer base being cleared by hrtimer_start. In the case of
4784 * migrate_hrtimers, base is never cleared, so we are fine.
4786 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4788 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4791 /* if the call-back is running a quota refresh is already occurring */
4792 if (hrtimer_callback_running(refresh_timer))
4795 /* is a quota refresh about to occur? */
4796 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4797 if (remaining < (s64)min_expire)
4803 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4805 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4807 /* if there's a quota refresh soon don't bother with slack */
4808 if (runtime_refresh_within(cfs_b, min_left))
4811 /* don't push forwards an existing deferred unthrottle */
4812 if (cfs_b->slack_started)
4814 cfs_b->slack_started = true;
4816 hrtimer_start(&cfs_b->slack_timer,
4817 ns_to_ktime(cfs_bandwidth_slack_period),
4821 /* we know any runtime found here is valid as update_curr() precedes return */
4822 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4824 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4825 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4827 if (slack_runtime <= 0)
4830 raw_spin_lock(&cfs_b->lock);
4831 if (cfs_b->quota != RUNTIME_INF) {
4832 cfs_b->runtime += slack_runtime;
4834 /* we are under rq->lock, defer unthrottling using a timer */
4835 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4836 !list_empty(&cfs_b->throttled_cfs_rq))
4837 start_cfs_slack_bandwidth(cfs_b);
4839 raw_spin_unlock(&cfs_b->lock);
4841 /* even if it's not valid for return we don't want to try again */
4842 cfs_rq->runtime_remaining -= slack_runtime;
4845 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4847 if (!cfs_bandwidth_used())
4850 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4853 __return_cfs_rq_runtime(cfs_rq);
4857 * This is done with a timer (instead of inline with bandwidth return) since
4858 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4860 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4862 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4863 unsigned long flags;
4865 /* confirm we're still not at a refresh boundary */
4866 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4867 cfs_b->slack_started = false;
4868 if (cfs_b->distribute_running) {
4869 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4873 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4874 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4878 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4879 runtime = cfs_b->runtime;
4882 cfs_b->distribute_running = 1;
4884 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4889 runtime = distribute_cfs_runtime(cfs_b, runtime);
4891 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4892 lsub_positive(&cfs_b->runtime, runtime);
4893 cfs_b->distribute_running = 0;
4894 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4898 * When a group wakes up we want to make sure that its quota is not already
4899 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4900 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4902 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4904 if (!cfs_bandwidth_used())
4907 /* an active group must be handled by the update_curr()->put() path */
4908 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4911 /* ensure the group is not already throttled */
4912 if (cfs_rq_throttled(cfs_rq))
4915 /* update runtime allocation */
4916 account_cfs_rq_runtime(cfs_rq, 0);
4917 if (cfs_rq->runtime_remaining <= 0)
4918 throttle_cfs_rq(cfs_rq);
4921 static void sync_throttle(struct task_group *tg, int cpu)
4923 struct cfs_rq *pcfs_rq, *cfs_rq;
4925 if (!cfs_bandwidth_used())
4931 cfs_rq = tg->cfs_rq[cpu];
4932 pcfs_rq = tg->parent->cfs_rq[cpu];
4934 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4935 cfs_rq->throttled_clock_pelt = rq_clock_pelt(cpu_rq(cpu));
4938 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4939 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4941 if (!cfs_bandwidth_used())
4944 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4948 * it's possible for a throttled entity to be forced into a running
4949 * state (e.g. set_curr_task), in this case we're finished.
4951 if (cfs_rq_throttled(cfs_rq))
4954 return throttle_cfs_rq(cfs_rq);
4957 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4959 struct cfs_bandwidth *cfs_b =
4960 container_of(timer, struct cfs_bandwidth, slack_timer);
4962 do_sched_cfs_slack_timer(cfs_b);
4964 return HRTIMER_NORESTART;
4967 extern const u64 max_cfs_quota_period;
4969 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4971 struct cfs_bandwidth *cfs_b =
4972 container_of(timer, struct cfs_bandwidth, period_timer);
4973 unsigned long flags;
4978 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4980 overrun = hrtimer_forward_now(timer, cfs_b->period);
4984 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
4987 u64 new, old = ktime_to_ns(cfs_b->period);
4990 * Grow period by a factor of 2 to avoid losing precision.
4991 * Precision loss in the quota/period ratio can cause __cfs_schedulable
4995 if (new < max_cfs_quota_period) {
4996 cfs_b->period = ns_to_ktime(new);
4999 pr_warn_ratelimited(
5000 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5002 div_u64(new, NSEC_PER_USEC),
5003 div_u64(cfs_b->quota, NSEC_PER_USEC));
5005 pr_warn_ratelimited(
5006 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5008 div_u64(old, NSEC_PER_USEC),
5009 div_u64(cfs_b->quota, NSEC_PER_USEC));
5012 /* reset count so we don't come right back in here */
5017 cfs_b->period_active = 0;
5018 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5020 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
5023 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5025 raw_spin_lock_init(&cfs_b->lock);
5027 cfs_b->quota = RUNTIME_INF;
5028 cfs_b->period = ns_to_ktime(default_cfs_period());
5030 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
5031 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
5032 cfs_b->period_timer.function = sched_cfs_period_timer;
5033 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5034 cfs_b->slack_timer.function = sched_cfs_slack_timer;
5035 cfs_b->distribute_running = 0;
5036 cfs_b->slack_started = false;
5039 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5041 cfs_rq->runtime_enabled = 0;
5042 INIT_LIST_HEAD(&cfs_rq->throttled_list);
5045 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5047 lockdep_assert_held(&cfs_b->lock);
5049 if (cfs_b->period_active)
5052 cfs_b->period_active = 1;
5053 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5054 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5057 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5059 /* init_cfs_bandwidth() was not called */
5060 if (!cfs_b->throttled_cfs_rq.next)
5063 hrtimer_cancel(&cfs_b->period_timer);
5064 hrtimer_cancel(&cfs_b->slack_timer);
5068 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5070 * The race is harmless, since modifying bandwidth settings of unhooked group
5071 * bits doesn't do much.
5074 /* cpu online calback */
5075 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5077 struct task_group *tg;
5079 lockdep_assert_held(&rq->lock);
5082 list_for_each_entry_rcu(tg, &task_groups, list) {
5083 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5084 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5086 raw_spin_lock(&cfs_b->lock);
5087 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5088 raw_spin_unlock(&cfs_b->lock);
5093 /* cpu offline callback */
5094 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5096 struct task_group *tg;
5098 lockdep_assert_held(&rq->lock);
5101 list_for_each_entry_rcu(tg, &task_groups, list) {
5102 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5104 if (!cfs_rq->runtime_enabled)
5108 * clock_task is not advancing so we just need to make sure
5109 * there's some valid quota amount
5111 cfs_rq->runtime_remaining = 1;
5113 * Offline rq is schedulable till CPU is completely disabled
5114 * in take_cpu_down(), so we prevent new cfs throttling here.
5116 cfs_rq->runtime_enabled = 0;
5118 if (cfs_rq_throttled(cfs_rq))
5119 unthrottle_cfs_rq(cfs_rq);
5124 #else /* CONFIG_CFS_BANDWIDTH */
5126 static inline bool cfs_bandwidth_used(void)
5131 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5132 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5133 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5134 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5135 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5137 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5142 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5147 static inline int throttled_lb_pair(struct task_group *tg,
5148 int src_cpu, int dest_cpu)
5153 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5155 #ifdef CONFIG_FAIR_GROUP_SCHED
5156 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5159 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5163 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5164 static inline void update_runtime_enabled(struct rq *rq) {}
5165 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5167 #endif /* CONFIG_CFS_BANDWIDTH */
5169 /**************************************************
5170 * CFS operations on tasks:
5173 #ifdef CONFIG_SCHED_HRTICK
5174 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5176 struct sched_entity *se = &p->se;
5177 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5179 SCHED_WARN_ON(task_rq(p) != rq);
5181 if (rq->cfs.h_nr_running > 1) {
5182 u64 slice = sched_slice(cfs_rq, se);
5183 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5184 s64 delta = slice - ran;
5191 hrtick_start(rq, delta);
5196 * called from enqueue/dequeue and updates the hrtick when the
5197 * current task is from our class and nr_running is low enough
5200 static void hrtick_update(struct rq *rq)
5202 struct task_struct *curr = rq->curr;
5204 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5207 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5208 hrtick_start_fair(rq, curr);
5210 #else /* !CONFIG_SCHED_HRTICK */
5212 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5216 static inline void hrtick_update(struct rq *rq)
5222 static inline unsigned long cpu_util(int cpu);
5224 static inline bool cpu_overutilized(int cpu)
5226 return !fits_capacity(cpu_util(cpu), capacity_of(cpu));
5229 static inline void update_overutilized_status(struct rq *rq)
5231 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
5232 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5233 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
5237 static inline void update_overutilized_status(struct rq *rq) { }
5241 * The enqueue_task method is called before nr_running is
5242 * increased. Here we update the fair scheduling stats and
5243 * then put the task into the rbtree:
5246 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5248 struct cfs_rq *cfs_rq;
5249 struct sched_entity *se = &p->se;
5250 int idle_h_nr_running = task_has_idle_policy(p);
5251 int task_new = !(flags & ENQUEUE_WAKEUP);
5254 * The code below (indirectly) updates schedutil which looks at
5255 * the cfs_rq utilization to select a frequency.
5256 * Let's add the task's estimated utilization to the cfs_rq's
5257 * estimated utilization, before we update schedutil.
5259 util_est_enqueue(&rq->cfs, p);
5262 * If in_iowait is set, the code below may not trigger any cpufreq
5263 * utilization updates, so do it here explicitly with the IOWAIT flag
5267 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5269 for_each_sched_entity(se) {
5272 cfs_rq = cfs_rq_of(se);
5273 enqueue_entity(cfs_rq, se, flags);
5275 cfs_rq->h_nr_running++;
5276 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5278 /* end evaluation on encountering a throttled cfs_rq */
5279 if (cfs_rq_throttled(cfs_rq))
5280 goto enqueue_throttle;
5282 flags = ENQUEUE_WAKEUP;
5285 for_each_sched_entity(se) {
5286 cfs_rq = cfs_rq_of(se);
5288 update_load_avg(cfs_rq, se, UPDATE_TG);
5289 update_cfs_group(se);
5291 cfs_rq->h_nr_running++;
5292 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5294 /* end evaluation on encountering a throttled cfs_rq */
5295 if (cfs_rq_throttled(cfs_rq))
5296 goto enqueue_throttle;
5299 * One parent has been throttled and cfs_rq removed from the
5300 * list. Add it back to not break the leaf list.
5302 if (throttled_hierarchy(cfs_rq))
5303 list_add_leaf_cfs_rq(cfs_rq);
5308 add_nr_running(rq, 1);
5310 * Since new tasks are assigned an initial util_avg equal to
5311 * half of the spare capacity of their CPU, tiny tasks have the
5312 * ability to cross the overutilized threshold, which will
5313 * result in the load balancer ruining all the task placement
5314 * done by EAS. As a way to mitigate that effect, do not account
5315 * for the first enqueue operation of new tasks during the
5316 * overutilized flag detection.
5318 * A better way of solving this problem would be to wait for
5319 * the PELT signals of tasks to converge before taking them
5320 * into account, but that is not straightforward to implement,
5321 * and the following generally works well enough in practice.
5324 update_overutilized_status(rq);
5328 if (cfs_bandwidth_used()) {
5330 * When bandwidth control is enabled; the cfs_rq_throttled()
5331 * breaks in the above iteration can result in incomplete
5332 * leaf list maintenance, resulting in triggering the assertion
5335 for_each_sched_entity(se) {
5336 cfs_rq = cfs_rq_of(se);
5338 if (list_add_leaf_cfs_rq(cfs_rq))
5343 assert_list_leaf_cfs_rq(rq);
5348 static void set_next_buddy(struct sched_entity *se);
5351 * The dequeue_task method is called before nr_running is
5352 * decreased. We remove the task from the rbtree and
5353 * update the fair scheduling stats:
5355 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5357 struct cfs_rq *cfs_rq;
5358 struct sched_entity *se = &p->se;
5359 int task_sleep = flags & DEQUEUE_SLEEP;
5360 int idle_h_nr_running = task_has_idle_policy(p);
5362 for_each_sched_entity(se) {
5363 cfs_rq = cfs_rq_of(se);
5364 dequeue_entity(cfs_rq, se, flags);
5366 cfs_rq->h_nr_running--;
5367 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5369 /* end evaluation on encountering a throttled cfs_rq */
5370 if (cfs_rq_throttled(cfs_rq))
5371 goto dequeue_throttle;
5373 /* Don't dequeue parent if it has other entities besides us */
5374 if (cfs_rq->load.weight) {
5375 /* Avoid re-evaluating load for this entity: */
5376 se = parent_entity(se);
5378 * Bias pick_next to pick a task from this cfs_rq, as
5379 * p is sleeping when it is within its sched_slice.
5381 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5385 flags |= DEQUEUE_SLEEP;
5388 for_each_sched_entity(se) {
5389 cfs_rq = cfs_rq_of(se);
5391 update_load_avg(cfs_rq, se, UPDATE_TG);
5392 update_cfs_group(se);
5394 cfs_rq->h_nr_running--;
5395 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5397 /* end evaluation on encountering a throttled cfs_rq */
5398 if (cfs_rq_throttled(cfs_rq))
5399 goto dequeue_throttle;
5405 sub_nr_running(rq, 1);
5407 util_est_dequeue(&rq->cfs, p, task_sleep);
5413 /* Working cpumask for: load_balance, load_balance_newidle. */
5414 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5415 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5417 #ifdef CONFIG_NO_HZ_COMMON
5420 cpumask_var_t idle_cpus_mask;
5422 int has_blocked; /* Idle CPUS has blocked load */
5423 unsigned long next_balance; /* in jiffy units */
5424 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5425 } nohz ____cacheline_aligned;
5427 #endif /* CONFIG_NO_HZ_COMMON */
5429 /* CPU only has SCHED_IDLE tasks enqueued */
5430 static int sched_idle_cpu(int cpu)
5432 struct rq *rq = cpu_rq(cpu);
5434 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
5438 static unsigned long cpu_runnable_load(struct rq *rq)
5440 return cfs_rq_runnable_load_avg(&rq->cfs);
5443 static unsigned long capacity_of(int cpu)
5445 return cpu_rq(cpu)->cpu_capacity;
5448 static unsigned long cpu_avg_load_per_task(int cpu)
5450 struct rq *rq = cpu_rq(cpu);
5451 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5452 unsigned long load_avg = cpu_runnable_load(rq);
5455 return load_avg / nr_running;
5460 static void record_wakee(struct task_struct *p)
5463 * Only decay a single time; tasks that have less then 1 wakeup per
5464 * jiffy will not have built up many flips.
5466 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5467 current->wakee_flips >>= 1;
5468 current->wakee_flip_decay_ts = jiffies;
5471 if (current->last_wakee != p) {
5472 current->last_wakee = p;
5473 current->wakee_flips++;
5478 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5480 * A waker of many should wake a different task than the one last awakened
5481 * at a frequency roughly N times higher than one of its wakees.
5483 * In order to determine whether we should let the load spread vs consolidating
5484 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5485 * partner, and a factor of lls_size higher frequency in the other.
5487 * With both conditions met, we can be relatively sure that the relationship is
5488 * non-monogamous, with partner count exceeding socket size.
5490 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5491 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5494 static int wake_wide(struct task_struct *p)
5496 unsigned int master = current->wakee_flips;
5497 unsigned int slave = p->wakee_flips;
5498 int factor = this_cpu_read(sd_llc_size);
5501 swap(master, slave);
5502 if (slave < factor || master < slave * factor)
5508 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5509 * soonest. For the purpose of speed we only consider the waking and previous
5512 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5513 * cache-affine and is (or will be) idle.
5515 * wake_affine_weight() - considers the weight to reflect the average
5516 * scheduling latency of the CPUs. This seems to work
5517 * for the overloaded case.
5520 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5523 * If this_cpu is idle, it implies the wakeup is from interrupt
5524 * context. Only allow the move if cache is shared. Otherwise an
5525 * interrupt intensive workload could force all tasks onto one
5526 * node depending on the IO topology or IRQ affinity settings.
5528 * If the prev_cpu is idle and cache affine then avoid a migration.
5529 * There is no guarantee that the cache hot data from an interrupt
5530 * is more important than cache hot data on the prev_cpu and from
5531 * a cpufreq perspective, it's better to have higher utilisation
5534 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5535 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5537 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5540 return nr_cpumask_bits;
5544 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5545 int this_cpu, int prev_cpu, int sync)
5547 s64 this_eff_load, prev_eff_load;
5548 unsigned long task_load;
5550 this_eff_load = cpu_runnable_load(cpu_rq(this_cpu));
5553 unsigned long current_load = task_h_load(current);
5555 if (current_load > this_eff_load)
5558 this_eff_load -= current_load;
5561 task_load = task_h_load(p);
5563 this_eff_load += task_load;
5564 if (sched_feat(WA_BIAS))
5565 this_eff_load *= 100;
5566 this_eff_load *= capacity_of(prev_cpu);
5568 prev_eff_load = cpu_runnable_load(cpu_rq(prev_cpu));
5569 prev_eff_load -= task_load;
5570 if (sched_feat(WA_BIAS))
5571 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5572 prev_eff_load *= capacity_of(this_cpu);
5575 * If sync, adjust the weight of prev_eff_load such that if
5576 * prev_eff == this_eff that select_idle_sibling() will consider
5577 * stacking the wakee on top of the waker if no other CPU is
5583 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5586 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5587 int this_cpu, int prev_cpu, int sync)
5589 int target = nr_cpumask_bits;
5591 if (sched_feat(WA_IDLE))
5592 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5594 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5595 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5597 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5598 if (target == nr_cpumask_bits)
5601 schedstat_inc(sd->ttwu_move_affine);
5602 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5606 static unsigned long cpu_util_without(int cpu, struct task_struct *p);
5608 static unsigned long capacity_spare_without(int cpu, struct task_struct *p)
5610 return max_t(long, capacity_of(cpu) - cpu_util_without(cpu, p), 0);
5614 * find_idlest_group finds and returns the least busy CPU group within the
5617 * Assumes p is allowed on at least one CPU in sd.
5619 static struct sched_group *
5620 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5621 int this_cpu, int sd_flag)
5623 struct sched_group *idlest = NULL, *group = sd->groups;
5624 struct sched_group *most_spare_sg = NULL;
5625 unsigned long min_runnable_load = ULONG_MAX;
5626 unsigned long this_runnable_load = ULONG_MAX;
5627 unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5628 unsigned long most_spare = 0, this_spare = 0;
5629 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5630 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5631 (sd->imbalance_pct-100) / 100;
5634 unsigned long load, avg_load, runnable_load;
5635 unsigned long spare_cap, max_spare_cap;
5639 /* Skip over this group if it has no CPUs allowed */
5640 if (!cpumask_intersects(sched_group_span(group),
5644 local_group = cpumask_test_cpu(this_cpu,
5645 sched_group_span(group));
5648 * Tally up the load of all CPUs in the group and find
5649 * the group containing the CPU with most spare capacity.
5655 for_each_cpu(i, sched_group_span(group)) {
5656 load = cpu_runnable_load(cpu_rq(i));
5657 runnable_load += load;
5659 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5661 spare_cap = capacity_spare_without(i, p);
5663 if (spare_cap > max_spare_cap)
5664 max_spare_cap = spare_cap;
5667 /* Adjust by relative CPU capacity of the group */
5668 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5669 group->sgc->capacity;
5670 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5671 group->sgc->capacity;
5674 this_runnable_load = runnable_load;
5675 this_avg_load = avg_load;
5676 this_spare = max_spare_cap;
5678 if (min_runnable_load > (runnable_load + imbalance)) {
5680 * The runnable load is significantly smaller
5681 * so we can pick this new CPU:
5683 min_runnable_load = runnable_load;
5684 min_avg_load = avg_load;
5686 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5687 (100*min_avg_load > imbalance_scale*avg_load)) {
5689 * The runnable loads are close so take the
5690 * blocked load into account through avg_load:
5692 min_avg_load = avg_load;
5696 if (most_spare < max_spare_cap) {
5697 most_spare = max_spare_cap;
5698 most_spare_sg = group;
5701 } while (group = group->next, group != sd->groups);
5704 * The cross-over point between using spare capacity or least load
5705 * is too conservative for high utilization tasks on partially
5706 * utilized systems if we require spare_capacity > task_util(p),
5707 * so we allow for some task stuffing by using
5708 * spare_capacity > task_util(p)/2.
5710 * Spare capacity can't be used for fork because the utilization has
5711 * not been set yet, we must first select a rq to compute the initial
5714 if (sd_flag & SD_BALANCE_FORK)
5717 if (this_spare > task_util(p) / 2 &&
5718 imbalance_scale*this_spare > 100*most_spare)
5721 if (most_spare > task_util(p) / 2)
5722 return most_spare_sg;
5729 * When comparing groups across NUMA domains, it's possible for the
5730 * local domain to be very lightly loaded relative to the remote
5731 * domains but "imbalance" skews the comparison making remote CPUs
5732 * look much more favourable. When considering cross-domain, add
5733 * imbalance to the runnable load on the remote node and consider
5736 if ((sd->flags & SD_NUMA) &&
5737 min_runnable_load + imbalance >= this_runnable_load)
5740 if (min_runnable_load > (this_runnable_load + imbalance))
5743 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5744 (100*this_avg_load < imbalance_scale*min_avg_load))
5751 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5754 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5756 unsigned long load, min_load = ULONG_MAX;
5757 unsigned int min_exit_latency = UINT_MAX;
5758 u64 latest_idle_timestamp = 0;
5759 int least_loaded_cpu = this_cpu;
5760 int shallowest_idle_cpu = -1, si_cpu = -1;
5763 /* Check if we have any choice: */
5764 if (group->group_weight == 1)
5765 return cpumask_first(sched_group_span(group));
5767 /* Traverse only the allowed CPUs */
5768 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
5769 if (available_idle_cpu(i)) {
5770 struct rq *rq = cpu_rq(i);
5771 struct cpuidle_state *idle = idle_get_state(rq);
5772 if (idle && idle->exit_latency < min_exit_latency) {
5774 * We give priority to a CPU whose idle state
5775 * has the smallest exit latency irrespective
5776 * of any idle timestamp.
5778 min_exit_latency = idle->exit_latency;
5779 latest_idle_timestamp = rq->idle_stamp;
5780 shallowest_idle_cpu = i;
5781 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5782 rq->idle_stamp > latest_idle_timestamp) {
5784 * If equal or no active idle state, then
5785 * the most recently idled CPU might have
5788 latest_idle_timestamp = rq->idle_stamp;
5789 shallowest_idle_cpu = i;
5791 } else if (shallowest_idle_cpu == -1 && si_cpu == -1) {
5792 if (sched_idle_cpu(i)) {
5797 load = cpu_runnable_load(cpu_rq(i));
5798 if (load < min_load) {
5800 least_loaded_cpu = i;
5805 if (shallowest_idle_cpu != -1)
5806 return shallowest_idle_cpu;
5809 return least_loaded_cpu;
5812 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5813 int cpu, int prev_cpu, int sd_flag)
5817 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
5821 * We need task's util for capacity_spare_without, sync it up to
5822 * prev_cpu's last_update_time.
5824 if (!(sd_flag & SD_BALANCE_FORK))
5825 sync_entity_load_avg(&p->se);
5828 struct sched_group *group;
5829 struct sched_domain *tmp;
5832 if (!(sd->flags & sd_flag)) {
5837 group = find_idlest_group(sd, p, cpu, sd_flag);
5843 new_cpu = find_idlest_group_cpu(group, p, cpu);
5844 if (new_cpu == cpu) {
5845 /* Now try balancing at a lower domain level of 'cpu': */
5850 /* Now try balancing at a lower domain level of 'new_cpu': */
5852 weight = sd->span_weight;
5854 for_each_domain(cpu, tmp) {
5855 if (weight <= tmp->span_weight)
5857 if (tmp->flags & sd_flag)
5865 #ifdef CONFIG_SCHED_SMT
5866 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5867 EXPORT_SYMBOL_GPL(sched_smt_present);
5869 static inline void set_idle_cores(int cpu, int val)
5871 struct sched_domain_shared *sds;
5873 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5875 WRITE_ONCE(sds->has_idle_cores, val);
5878 static inline bool test_idle_cores(int cpu, bool def)
5880 struct sched_domain_shared *sds;
5882 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5884 return READ_ONCE(sds->has_idle_cores);
5890 * Scans the local SMT mask to see if the entire core is idle, and records this
5891 * information in sd_llc_shared->has_idle_cores.
5893 * Since SMT siblings share all cache levels, inspecting this limited remote
5894 * state should be fairly cheap.
5896 void __update_idle_core(struct rq *rq)
5898 int core = cpu_of(rq);
5902 if (test_idle_cores(core, true))
5905 for_each_cpu(cpu, cpu_smt_mask(core)) {
5909 if (!available_idle_cpu(cpu))
5913 set_idle_cores(core, 1);
5919 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5920 * there are no idle cores left in the system; tracked through
5921 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5923 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5925 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5928 if (!static_branch_likely(&sched_smt_present))
5931 if (!test_idle_cores(target, false))
5934 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
5936 for_each_cpu_wrap(core, cpus, target) {
5939 for_each_cpu(cpu, cpu_smt_mask(core)) {
5940 __cpumask_clear_cpu(cpu, cpus);
5941 if (!available_idle_cpu(cpu))
5950 * Failed to find an idle core; stop looking for one.
5952 set_idle_cores(target, 0);
5958 * Scan the local SMT mask for idle CPUs.
5960 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5962 int cpu, si_cpu = -1;
5964 if (!static_branch_likely(&sched_smt_present))
5967 for_each_cpu(cpu, cpu_smt_mask(target)) {
5968 if (!cpumask_test_cpu(cpu, p->cpus_ptr) ||
5969 !cpumask_test_cpu(cpu, sched_domain_span(sd)))
5971 if (available_idle_cpu(cpu))
5973 if (si_cpu == -1 && sched_idle_cpu(cpu))
5980 #else /* CONFIG_SCHED_SMT */
5982 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5987 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5992 #endif /* CONFIG_SCHED_SMT */
5995 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5996 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5997 * average idle time for this rq (as found in rq->avg_idle).
5999 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
6001 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6002 struct sched_domain *this_sd;
6003 u64 avg_cost, avg_idle;
6006 int this = smp_processor_id();
6007 int cpu, nr = INT_MAX, si_cpu = -1;
6009 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6014 * Due to large variance we need a large fuzz factor; hackbench in
6015 * particularly is sensitive here.
6017 avg_idle = this_rq()->avg_idle / 512;
6018 avg_cost = this_sd->avg_scan_cost + 1;
6020 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
6023 if (sched_feat(SIS_PROP)) {
6024 u64 span_avg = sd->span_weight * avg_idle;
6025 if (span_avg > 4*avg_cost)
6026 nr = div_u64(span_avg, avg_cost);
6031 time = cpu_clock(this);
6033 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6035 for_each_cpu_wrap(cpu, cpus, target) {
6038 if (available_idle_cpu(cpu))
6040 if (si_cpu == -1 && sched_idle_cpu(cpu))
6044 time = cpu_clock(this) - time;
6045 cost = this_sd->avg_scan_cost;
6046 delta = (s64)(time - cost) / 8;
6047 this_sd->avg_scan_cost += delta;
6053 * Try and locate an idle core/thread in the LLC cache domain.
6055 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6057 struct sched_domain *sd;
6058 int i, recent_used_cpu;
6060 if (available_idle_cpu(target) || sched_idle_cpu(target))
6064 * If the previous CPU is cache affine and idle, don't be stupid:
6066 if (prev != target && cpus_share_cache(prev, target) &&
6067 (available_idle_cpu(prev) || sched_idle_cpu(prev)))
6070 /* Check a recently used CPU as a potential idle candidate: */
6071 recent_used_cpu = p->recent_used_cpu;
6072 if (recent_used_cpu != prev &&
6073 recent_used_cpu != target &&
6074 cpus_share_cache(recent_used_cpu, target) &&
6075 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
6076 cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr)) {
6078 * Replace recent_used_cpu with prev as it is a potential
6079 * candidate for the next wake:
6081 p->recent_used_cpu = prev;
6082 return recent_used_cpu;
6085 sd = rcu_dereference(per_cpu(sd_llc, target));
6089 i = select_idle_core(p, sd, target);
6090 if ((unsigned)i < nr_cpumask_bits)
6093 i = select_idle_cpu(p, sd, target);
6094 if ((unsigned)i < nr_cpumask_bits)
6097 i = select_idle_smt(p, sd, target);
6098 if ((unsigned)i < nr_cpumask_bits)
6105 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6106 * @cpu: the CPU to get the utilization of
6108 * The unit of the return value must be the one of capacity so we can compare
6109 * the utilization with the capacity of the CPU that is available for CFS task
6110 * (ie cpu_capacity).
6112 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6113 * recent utilization of currently non-runnable tasks on a CPU. It represents
6114 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6115 * capacity_orig is the cpu_capacity available at the highest frequency
6116 * (arch_scale_freq_capacity()).
6117 * The utilization of a CPU converges towards a sum equal to or less than the
6118 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6119 * the running time on this CPU scaled by capacity_curr.
6121 * The estimated utilization of a CPU is defined to be the maximum between its
6122 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6123 * currently RUNNABLE on that CPU.
6124 * This allows to properly represent the expected utilization of a CPU which
6125 * has just got a big task running since a long sleep period. At the same time
6126 * however it preserves the benefits of the "blocked utilization" in
6127 * describing the potential for other tasks waking up on the same CPU.
6129 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6130 * higher than capacity_orig because of unfortunate rounding in
6131 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6132 * the average stabilizes with the new running time. We need to check that the
6133 * utilization stays within the range of [0..capacity_orig] and cap it if
6134 * necessary. Without utilization capping, a group could be seen as overloaded
6135 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6136 * available capacity. We allow utilization to overshoot capacity_curr (but not
6137 * capacity_orig) as it useful for predicting the capacity required after task
6138 * migrations (scheduler-driven DVFS).
6140 * Return: the (estimated) utilization for the specified CPU
6142 static inline unsigned long cpu_util(int cpu)
6144 struct cfs_rq *cfs_rq;
6147 cfs_rq = &cpu_rq(cpu)->cfs;
6148 util = READ_ONCE(cfs_rq->avg.util_avg);
6150 if (sched_feat(UTIL_EST))
6151 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6153 return min_t(unsigned long, util, capacity_orig_of(cpu));
6157 * cpu_util_without: compute cpu utilization without any contributions from *p
6158 * @cpu: the CPU which utilization is requested
6159 * @p: the task which utilization should be discounted
6161 * The utilization of a CPU is defined by the utilization of tasks currently
6162 * enqueued on that CPU as well as tasks which are currently sleeping after an
6163 * execution on that CPU.
6165 * This method returns the utilization of the specified CPU by discounting the
6166 * utilization of the specified task, whenever the task is currently
6167 * contributing to the CPU utilization.
6169 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6171 struct cfs_rq *cfs_rq;
6174 /* Task has no contribution or is new */
6175 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6176 return cpu_util(cpu);
6178 cfs_rq = &cpu_rq(cpu)->cfs;
6179 util = READ_ONCE(cfs_rq->avg.util_avg);
6181 /* Discount task's util from CPU's util */
6182 lsub_positive(&util, task_util(p));
6187 * a) if *p is the only task sleeping on this CPU, then:
6188 * cpu_util (== task_util) > util_est (== 0)
6189 * and thus we return:
6190 * cpu_util_without = (cpu_util - task_util) = 0
6192 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6194 * cpu_util >= task_util
6195 * cpu_util > util_est (== 0)
6196 * and thus we discount *p's blocked utilization to return:
6197 * cpu_util_without = (cpu_util - task_util) >= 0
6199 * c) if other tasks are RUNNABLE on that CPU and
6200 * util_est > cpu_util
6201 * then we use util_est since it returns a more restrictive
6202 * estimation of the spare capacity on that CPU, by just
6203 * considering the expected utilization of tasks already
6204 * runnable on that CPU.
6206 * Cases a) and b) are covered by the above code, while case c) is
6207 * covered by the following code when estimated utilization is
6210 if (sched_feat(UTIL_EST)) {
6211 unsigned int estimated =
6212 READ_ONCE(cfs_rq->avg.util_est.enqueued);
6215 * Despite the following checks we still have a small window
6216 * for a possible race, when an execl's select_task_rq_fair()
6217 * races with LB's detach_task():
6220 * p->on_rq = TASK_ON_RQ_MIGRATING;
6221 * ---------------------------------- A
6222 * deactivate_task() \
6223 * dequeue_task() + RaceTime
6224 * util_est_dequeue() /
6225 * ---------------------------------- B
6227 * The additional check on "current == p" it's required to
6228 * properly fix the execl regression and it helps in further
6229 * reducing the chances for the above race.
6231 if (unlikely(task_on_rq_queued(p) || current == p))
6232 lsub_positive(&estimated, _task_util_est(p));
6234 util = max(util, estimated);
6238 * Utilization (estimated) can exceed the CPU capacity, thus let's
6239 * clamp to the maximum CPU capacity to ensure consistency with
6240 * the cpu_util call.
6242 return min_t(unsigned long, util, capacity_orig_of(cpu));
6246 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6247 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6249 * In that case WAKE_AFFINE doesn't make sense and we'll let
6250 * BALANCE_WAKE sort things out.
6252 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6254 long min_cap, max_cap;
6256 if (!static_branch_unlikely(&sched_asym_cpucapacity))
6259 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6260 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6262 /* Minimum capacity is close to max, no need to abort wake_affine */
6263 if (max_cap - min_cap < max_cap >> 3)
6266 /* Bring task utilization in sync with prev_cpu */
6267 sync_entity_load_avg(&p->se);
6269 return !task_fits_capacity(p, min_cap);
6273 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6276 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6278 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6279 unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
6282 * If @p migrates from @cpu to another, remove its contribution. Or,
6283 * if @p migrates from another CPU to @cpu, add its contribution. In
6284 * the other cases, @cpu is not impacted by the migration, so the
6285 * util_avg should already be correct.
6287 if (task_cpu(p) == cpu && dst_cpu != cpu)
6288 sub_positive(&util, task_util(p));
6289 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6290 util += task_util(p);
6292 if (sched_feat(UTIL_EST)) {
6293 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6296 * During wake-up, the task isn't enqueued yet and doesn't
6297 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6298 * so just add it (if needed) to "simulate" what will be
6299 * cpu_util() after the task has been enqueued.
6302 util_est += _task_util_est(p);
6304 util = max(util, util_est);
6307 return min(util, capacity_orig_of(cpu));
6311 * compute_energy(): Estimates the energy that @pd would consume if @p was
6312 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6313 * landscape of @pd's CPUs after the task migration, and uses the Energy Model
6314 * to compute what would be the energy if we decided to actually migrate that
6318 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
6320 struct cpumask *pd_mask = perf_domain_span(pd);
6321 unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask));
6322 unsigned long max_util = 0, sum_util = 0;
6326 * The capacity state of CPUs of the current rd can be driven by CPUs
6327 * of another rd if they belong to the same pd. So, account for the
6328 * utilization of these CPUs too by masking pd with cpu_online_mask
6329 * instead of the rd span.
6331 * If an entire pd is outside of the current rd, it will not appear in
6332 * its pd list and will not be accounted by compute_energy().
6334 for_each_cpu_and(cpu, pd_mask, cpu_online_mask) {
6335 unsigned long cpu_util, util_cfs = cpu_util_next(cpu, p, dst_cpu);
6336 struct task_struct *tsk = cpu == dst_cpu ? p : NULL;
6339 * Busy time computation: utilization clamping is not
6340 * required since the ratio (sum_util / cpu_capacity)
6341 * is already enough to scale the EM reported power
6342 * consumption at the (eventually clamped) cpu_capacity.
6344 sum_util += schedutil_cpu_util(cpu, util_cfs, cpu_cap,
6348 * Performance domain frequency: utilization clamping
6349 * must be considered since it affects the selection
6350 * of the performance domain frequency.
6351 * NOTE: in case RT tasks are running, by default the
6352 * FREQUENCY_UTIL's utilization can be max OPP.
6354 cpu_util = schedutil_cpu_util(cpu, util_cfs, cpu_cap,
6355 FREQUENCY_UTIL, tsk);
6356 max_util = max(max_util, cpu_util);
6359 return em_pd_energy(pd->em_pd, max_util, sum_util);
6363 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6364 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6365 * spare capacity in each performance domain and uses it as a potential
6366 * candidate to execute the task. Then, it uses the Energy Model to figure
6367 * out which of the CPU candidates is the most energy-efficient.
6369 * The rationale for this heuristic is as follows. In a performance domain,
6370 * all the most energy efficient CPU candidates (according to the Energy
6371 * Model) are those for which we'll request a low frequency. When there are
6372 * several CPUs for which the frequency request will be the same, we don't
6373 * have enough data to break the tie between them, because the Energy Model
6374 * only includes active power costs. With this model, if we assume that
6375 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6376 * the maximum spare capacity in a performance domain is guaranteed to be among
6377 * the best candidates of the performance domain.
6379 * In practice, it could be preferable from an energy standpoint to pack
6380 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6381 * but that could also hurt our chances to go cluster idle, and we have no
6382 * ways to tell with the current Energy Model if this is actually a good
6383 * idea or not. So, find_energy_efficient_cpu() basically favors
6384 * cluster-packing, and spreading inside a cluster. That should at least be
6385 * a good thing for latency, and this is consistent with the idea that most
6386 * of the energy savings of EAS come from the asymmetry of the system, and
6387 * not so much from breaking the tie between identical CPUs. That's also the
6388 * reason why EAS is enabled in the topology code only for systems where
6389 * SD_ASYM_CPUCAPACITY is set.
6391 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6392 * they don't have any useful utilization data yet and it's not possible to
6393 * forecast their impact on energy consumption. Consequently, they will be
6394 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6395 * to be energy-inefficient in some use-cases. The alternative would be to
6396 * bias new tasks towards specific types of CPUs first, or to try to infer
6397 * their util_avg from the parent task, but those heuristics could hurt
6398 * other use-cases too. So, until someone finds a better way to solve this,
6399 * let's keep things simple by re-using the existing slow path.
6401 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6403 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
6404 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6405 unsigned long cpu_cap, util, base_energy = 0;
6406 int cpu, best_energy_cpu = prev_cpu;
6407 struct sched_domain *sd;
6408 struct perf_domain *pd;
6411 pd = rcu_dereference(rd->pd);
6412 if (!pd || READ_ONCE(rd->overutilized))
6416 * Energy-aware wake-up happens on the lowest sched_domain starting
6417 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6419 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6420 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6425 sync_entity_load_avg(&p->se);
6426 if (!task_util_est(p))
6429 for (; pd; pd = pd->next) {
6430 unsigned long cur_delta, spare_cap, max_spare_cap = 0;
6431 unsigned long base_energy_pd;
6432 int max_spare_cap_cpu = -1;
6434 /* Compute the 'base' energy of the pd, without @p */
6435 base_energy_pd = compute_energy(p, -1, pd);
6436 base_energy += base_energy_pd;
6438 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
6439 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6442 /* Skip CPUs that will be overutilized. */
6443 util = cpu_util_next(cpu, p, cpu);
6444 cpu_cap = capacity_of(cpu);
6445 if (!fits_capacity(util, cpu_cap))
6448 /* Always use prev_cpu as a candidate. */
6449 if (cpu == prev_cpu) {
6450 prev_delta = compute_energy(p, prev_cpu, pd);
6451 prev_delta -= base_energy_pd;
6452 best_delta = min(best_delta, prev_delta);
6456 * Find the CPU with the maximum spare capacity in
6457 * the performance domain
6459 spare_cap = cpu_cap - util;
6460 if (spare_cap > max_spare_cap) {
6461 max_spare_cap = spare_cap;
6462 max_spare_cap_cpu = cpu;
6466 /* Evaluate the energy impact of using this CPU. */
6467 if (max_spare_cap_cpu >= 0 && max_spare_cap_cpu != prev_cpu) {
6468 cur_delta = compute_energy(p, max_spare_cap_cpu, pd);
6469 cur_delta -= base_energy_pd;
6470 if (cur_delta < best_delta) {
6471 best_delta = cur_delta;
6472 best_energy_cpu = max_spare_cap_cpu;
6480 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6481 * least 6% of the energy used by prev_cpu.
6483 if (prev_delta == ULONG_MAX)
6484 return best_energy_cpu;
6486 if ((prev_delta - best_delta) > ((prev_delta + base_energy) >> 4))
6487 return best_energy_cpu;
6498 * select_task_rq_fair: Select target runqueue for the waking task in domains
6499 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6500 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6502 * Balances load by selecting the idlest CPU in the idlest group, or under
6503 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6505 * Returns the target CPU number.
6507 * preempt must be disabled.
6510 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6512 struct sched_domain *tmp, *sd = NULL;
6513 int cpu = smp_processor_id();
6514 int new_cpu = prev_cpu;
6515 int want_affine = 0;
6516 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6518 if (sd_flag & SD_BALANCE_WAKE) {
6521 if (sched_energy_enabled()) {
6522 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
6528 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu) &&
6529 cpumask_test_cpu(cpu, p->cpus_ptr);
6533 for_each_domain(cpu, tmp) {
6534 if (!(tmp->flags & SD_LOAD_BALANCE))
6538 * If both 'cpu' and 'prev_cpu' are part of this domain,
6539 * cpu is a valid SD_WAKE_AFFINE target.
6541 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6542 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6543 if (cpu != prev_cpu)
6544 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6546 sd = NULL; /* Prefer wake_affine over balance flags */
6550 if (tmp->flags & sd_flag)
6552 else if (!want_affine)
6558 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6559 } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6562 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6565 current->recent_used_cpu = cpu;
6572 static void detach_entity_cfs_rq(struct sched_entity *se);
6575 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6576 * cfs_rq_of(p) references at time of call are still valid and identify the
6577 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6579 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6582 * As blocked tasks retain absolute vruntime the migration needs to
6583 * deal with this by subtracting the old and adding the new
6584 * min_vruntime -- the latter is done by enqueue_entity() when placing
6585 * the task on the new runqueue.
6587 if (p->state == TASK_WAKING) {
6588 struct sched_entity *se = &p->se;
6589 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6592 #ifndef CONFIG_64BIT
6593 u64 min_vruntime_copy;
6596 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6598 min_vruntime = cfs_rq->min_vruntime;
6599 } while (min_vruntime != min_vruntime_copy);
6601 min_vruntime = cfs_rq->min_vruntime;
6604 se->vruntime -= min_vruntime;
6607 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6609 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6610 * rq->lock and can modify state directly.
6612 lockdep_assert_held(&task_rq(p)->lock);
6613 detach_entity_cfs_rq(&p->se);
6617 * We are supposed to update the task to "current" time, then
6618 * its up to date and ready to go to new CPU/cfs_rq. But we
6619 * have difficulty in getting what current time is, so simply
6620 * throw away the out-of-date time. This will result in the
6621 * wakee task is less decayed, but giving the wakee more load
6624 remove_entity_load_avg(&p->se);
6627 /* Tell new CPU we are migrated */
6628 p->se.avg.last_update_time = 0;
6630 /* We have migrated, no longer consider this task hot */
6631 p->se.exec_start = 0;
6633 update_scan_period(p, new_cpu);
6636 static void task_dead_fair(struct task_struct *p)
6638 remove_entity_load_avg(&p->se);
6642 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6647 return newidle_balance(rq, rf) != 0;
6649 #endif /* CONFIG_SMP */
6651 static unsigned long wakeup_gran(struct sched_entity *se)
6653 unsigned long gran = sysctl_sched_wakeup_granularity;
6656 * Since its curr running now, convert the gran from real-time
6657 * to virtual-time in his units.
6659 * By using 'se' instead of 'curr' we penalize light tasks, so
6660 * they get preempted easier. That is, if 'se' < 'curr' then
6661 * the resulting gran will be larger, therefore penalizing the
6662 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6663 * be smaller, again penalizing the lighter task.
6665 * This is especially important for buddies when the leftmost
6666 * task is higher priority than the buddy.
6668 return calc_delta_fair(gran, se);
6672 * Should 'se' preempt 'curr'.
6686 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6688 s64 gran, vdiff = curr->vruntime - se->vruntime;
6693 gran = wakeup_gran(se);
6700 static void set_last_buddy(struct sched_entity *se)
6702 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6705 for_each_sched_entity(se) {
6706 if (SCHED_WARN_ON(!se->on_rq))
6708 cfs_rq_of(se)->last = se;
6712 static void set_next_buddy(struct sched_entity *se)
6714 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6717 for_each_sched_entity(se) {
6718 if (SCHED_WARN_ON(!se->on_rq))
6720 cfs_rq_of(se)->next = se;
6724 static void set_skip_buddy(struct sched_entity *se)
6726 for_each_sched_entity(se)
6727 cfs_rq_of(se)->skip = se;
6731 * Preempt the current task with a newly woken task if needed:
6733 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6735 struct task_struct *curr = rq->curr;
6736 struct sched_entity *se = &curr->se, *pse = &p->se;
6737 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6738 int scale = cfs_rq->nr_running >= sched_nr_latency;
6739 int next_buddy_marked = 0;
6741 if (unlikely(se == pse))
6745 * This is possible from callers such as attach_tasks(), in which we
6746 * unconditionally check_prempt_curr() after an enqueue (which may have
6747 * lead to a throttle). This both saves work and prevents false
6748 * next-buddy nomination below.
6750 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6753 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6754 set_next_buddy(pse);
6755 next_buddy_marked = 1;
6759 * We can come here with TIF_NEED_RESCHED already set from new task
6762 * Note: this also catches the edge-case of curr being in a throttled
6763 * group (e.g. via set_curr_task), since update_curr() (in the
6764 * enqueue of curr) will have resulted in resched being set. This
6765 * prevents us from potentially nominating it as a false LAST_BUDDY
6768 if (test_tsk_need_resched(curr))
6771 /* Idle tasks are by definition preempted by non-idle tasks. */
6772 if (unlikely(task_has_idle_policy(curr)) &&
6773 likely(!task_has_idle_policy(p)))
6777 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6778 * is driven by the tick):
6780 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6783 find_matching_se(&se, &pse);
6784 update_curr(cfs_rq_of(se));
6786 if (wakeup_preempt_entity(se, pse) == 1) {
6788 * Bias pick_next to pick the sched entity that is
6789 * triggering this preemption.
6791 if (!next_buddy_marked)
6792 set_next_buddy(pse);
6801 * Only set the backward buddy when the current task is still
6802 * on the rq. This can happen when a wakeup gets interleaved
6803 * with schedule on the ->pre_schedule() or idle_balance()
6804 * point, either of which can * drop the rq lock.
6806 * Also, during early boot the idle thread is in the fair class,
6807 * for obvious reasons its a bad idea to schedule back to it.
6809 if (unlikely(!se->on_rq || curr == rq->idle))
6812 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6816 static struct task_struct *
6817 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6819 struct cfs_rq *cfs_rq = &rq->cfs;
6820 struct sched_entity *se;
6821 struct task_struct *p;
6825 if (!sched_fair_runnable(rq))
6828 #ifdef CONFIG_FAIR_GROUP_SCHED
6829 if (!prev || prev->sched_class != &fair_sched_class)
6833 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6834 * likely that a next task is from the same cgroup as the current.
6836 * Therefore attempt to avoid putting and setting the entire cgroup
6837 * hierarchy, only change the part that actually changes.
6841 struct sched_entity *curr = cfs_rq->curr;
6844 * Since we got here without doing put_prev_entity() we also
6845 * have to consider cfs_rq->curr. If it is still a runnable
6846 * entity, update_curr() will update its vruntime, otherwise
6847 * forget we've ever seen it.
6851 update_curr(cfs_rq);
6856 * This call to check_cfs_rq_runtime() will do the
6857 * throttle and dequeue its entity in the parent(s).
6858 * Therefore the nr_running test will indeed
6861 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6864 if (!cfs_rq->nr_running)
6871 se = pick_next_entity(cfs_rq, curr);
6872 cfs_rq = group_cfs_rq(se);
6878 * Since we haven't yet done put_prev_entity and if the selected task
6879 * is a different task than we started out with, try and touch the
6880 * least amount of cfs_rqs.
6883 struct sched_entity *pse = &prev->se;
6885 while (!(cfs_rq = is_same_group(se, pse))) {
6886 int se_depth = se->depth;
6887 int pse_depth = pse->depth;
6889 if (se_depth <= pse_depth) {
6890 put_prev_entity(cfs_rq_of(pse), pse);
6891 pse = parent_entity(pse);
6893 if (se_depth >= pse_depth) {
6894 set_next_entity(cfs_rq_of(se), se);
6895 se = parent_entity(se);
6899 put_prev_entity(cfs_rq, pse);
6900 set_next_entity(cfs_rq, se);
6907 put_prev_task(rq, prev);
6910 se = pick_next_entity(cfs_rq, NULL);
6911 set_next_entity(cfs_rq, se);
6912 cfs_rq = group_cfs_rq(se);
6917 done: __maybe_unused;
6920 * Move the next running task to the front of
6921 * the list, so our cfs_tasks list becomes MRU
6924 list_move(&p->se.group_node, &rq->cfs_tasks);
6927 if (hrtick_enabled(rq))
6928 hrtick_start_fair(rq, p);
6930 update_misfit_status(p, rq);
6938 new_tasks = newidle_balance(rq, rf);
6941 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
6942 * possible for any higher priority task to appear. In that case we
6943 * must re-start the pick_next_entity() loop.
6952 * rq is about to be idle, check if we need to update the
6953 * lost_idle_time of clock_pelt
6955 update_idle_rq_clock_pelt(rq);
6961 * Account for a descheduled task:
6963 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6965 struct sched_entity *se = &prev->se;
6966 struct cfs_rq *cfs_rq;
6968 for_each_sched_entity(se) {
6969 cfs_rq = cfs_rq_of(se);
6970 put_prev_entity(cfs_rq, se);
6975 * sched_yield() is very simple
6977 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6979 static void yield_task_fair(struct rq *rq)
6981 struct task_struct *curr = rq->curr;
6982 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6983 struct sched_entity *se = &curr->se;
6986 * Are we the only task in the tree?
6988 if (unlikely(rq->nr_running == 1))
6991 clear_buddies(cfs_rq, se);
6993 if (curr->policy != SCHED_BATCH) {
6994 update_rq_clock(rq);
6996 * Update run-time statistics of the 'current'.
6998 update_curr(cfs_rq);
7000 * Tell update_rq_clock() that we've just updated,
7001 * so we don't do microscopic update in schedule()
7002 * and double the fastpath cost.
7004 rq_clock_skip_update(rq);
7010 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
7012 struct sched_entity *se = &p->se;
7014 /* throttled hierarchies are not runnable */
7015 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
7018 /* Tell the scheduler that we'd really like pse to run next. */
7021 yield_task_fair(rq);
7027 /**************************************************
7028 * Fair scheduling class load-balancing methods.
7032 * The purpose of load-balancing is to achieve the same basic fairness the
7033 * per-CPU scheduler provides, namely provide a proportional amount of compute
7034 * time to each task. This is expressed in the following equation:
7036 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7038 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7039 * W_i,0 is defined as:
7041 * W_i,0 = \Sum_j w_i,j (2)
7043 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7044 * is derived from the nice value as per sched_prio_to_weight[].
7046 * The weight average is an exponential decay average of the instantaneous
7049 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7051 * C_i is the compute capacity of CPU i, typically it is the
7052 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7053 * can also include other factors [XXX].
7055 * To achieve this balance we define a measure of imbalance which follows
7056 * directly from (1):
7058 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7060 * We them move tasks around to minimize the imbalance. In the continuous
7061 * function space it is obvious this converges, in the discrete case we get
7062 * a few fun cases generally called infeasible weight scenarios.
7065 * - infeasible weights;
7066 * - local vs global optima in the discrete case. ]
7071 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7072 * for all i,j solution, we create a tree of CPUs that follows the hardware
7073 * topology where each level pairs two lower groups (or better). This results
7074 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7075 * tree to only the first of the previous level and we decrease the frequency
7076 * of load-balance at each level inv. proportional to the number of CPUs in
7082 * \Sum { --- * --- * 2^i } = O(n) (5)
7084 * `- size of each group
7085 * | | `- number of CPUs doing load-balance
7087 * `- sum over all levels
7089 * Coupled with a limit on how many tasks we can migrate every balance pass,
7090 * this makes (5) the runtime complexity of the balancer.
7092 * An important property here is that each CPU is still (indirectly) connected
7093 * to every other CPU in at most O(log n) steps:
7095 * The adjacency matrix of the resulting graph is given by:
7098 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7101 * And you'll find that:
7103 * A^(log_2 n)_i,j != 0 for all i,j (7)
7105 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7106 * The task movement gives a factor of O(m), giving a convergence complexity
7109 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7114 * In order to avoid CPUs going idle while there's still work to do, new idle
7115 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7116 * tree itself instead of relying on other CPUs to bring it work.
7118 * This adds some complexity to both (5) and (8) but it reduces the total idle
7126 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7129 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7134 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7136 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7138 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7141 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7142 * rewrite all of this once again.]
7145 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7147 enum fbq_type { regular, remote, all };
7156 #define LBF_ALL_PINNED 0x01
7157 #define LBF_NEED_BREAK 0x02
7158 #define LBF_DST_PINNED 0x04
7159 #define LBF_SOME_PINNED 0x08
7160 #define LBF_NOHZ_STATS 0x10
7161 #define LBF_NOHZ_AGAIN 0x20
7164 struct sched_domain *sd;
7172 struct cpumask *dst_grpmask;
7174 enum cpu_idle_type idle;
7176 /* The set of CPUs under consideration for load-balancing */
7177 struct cpumask *cpus;
7182 unsigned int loop_break;
7183 unsigned int loop_max;
7185 enum fbq_type fbq_type;
7186 enum group_type src_grp_type;
7187 struct list_head tasks;
7191 * Is this task likely cache-hot:
7193 static int task_hot(struct task_struct *p, struct lb_env *env)
7197 lockdep_assert_held(&env->src_rq->lock);
7199 if (p->sched_class != &fair_sched_class)
7202 if (unlikely(task_has_idle_policy(p)))
7206 * Buddy candidates are cache hot:
7208 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7209 (&p->se == cfs_rq_of(&p->se)->next ||
7210 &p->se == cfs_rq_of(&p->se)->last))
7213 if (sysctl_sched_migration_cost == -1)
7215 if (sysctl_sched_migration_cost == 0)
7218 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7220 return delta < (s64)sysctl_sched_migration_cost;
7223 #ifdef CONFIG_NUMA_BALANCING
7225 * Returns 1, if task migration degrades locality
7226 * Returns 0, if task migration improves locality i.e migration preferred.
7227 * Returns -1, if task migration is not affected by locality.
7229 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7231 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7232 unsigned long src_weight, dst_weight;
7233 int src_nid, dst_nid, dist;
7235 if (!static_branch_likely(&sched_numa_balancing))
7238 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7241 src_nid = cpu_to_node(env->src_cpu);
7242 dst_nid = cpu_to_node(env->dst_cpu);
7244 if (src_nid == dst_nid)
7247 /* Migrating away from the preferred node is always bad. */
7248 if (src_nid == p->numa_preferred_nid) {
7249 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7255 /* Encourage migration to the preferred node. */
7256 if (dst_nid == p->numa_preferred_nid)
7259 /* Leaving a core idle is often worse than degrading locality. */
7260 if (env->idle == CPU_IDLE)
7263 dist = node_distance(src_nid, dst_nid);
7265 src_weight = group_weight(p, src_nid, dist);
7266 dst_weight = group_weight(p, dst_nid, dist);
7268 src_weight = task_weight(p, src_nid, dist);
7269 dst_weight = task_weight(p, dst_nid, dist);
7272 return dst_weight < src_weight;
7276 static inline int migrate_degrades_locality(struct task_struct *p,
7284 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7287 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7291 lockdep_assert_held(&env->src_rq->lock);
7294 * We do not migrate tasks that are:
7295 * 1) throttled_lb_pair, or
7296 * 2) cannot be migrated to this CPU due to cpus_ptr, or
7297 * 3) running (obviously), or
7298 * 4) are cache-hot on their current CPU.
7300 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7303 /* Disregard pcpu kthreads; they are where they need to be. */
7304 if (kthread_is_per_cpu(p))
7307 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
7310 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7312 env->flags |= LBF_SOME_PINNED;
7315 * Remember if this task can be migrated to any other CPU in
7316 * our sched_group. We may want to revisit it if we couldn't
7317 * meet load balance goals by pulling other tasks on src_cpu.
7319 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7320 * already computed one in current iteration.
7322 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7325 /* Prevent to re-select dst_cpu via env's CPUs: */
7326 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7327 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
7328 env->flags |= LBF_DST_PINNED;
7329 env->new_dst_cpu = cpu;
7337 /* Record that we found atleast one task that could run on dst_cpu */
7338 env->flags &= ~LBF_ALL_PINNED;
7340 if (task_running(env->src_rq, p)) {
7341 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7346 * Aggressive migration if:
7347 * 1) destination numa is preferred
7348 * 2) task is cache cold, or
7349 * 3) too many balance attempts have failed.
7351 tsk_cache_hot = migrate_degrades_locality(p, env);
7352 if (tsk_cache_hot == -1)
7353 tsk_cache_hot = task_hot(p, env);
7355 if (tsk_cache_hot <= 0 ||
7356 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7357 if (tsk_cache_hot == 1) {
7358 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7359 schedstat_inc(p->se.statistics.nr_forced_migrations);
7364 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7369 * detach_task() -- detach the task for the migration specified in env
7371 static void detach_task(struct task_struct *p, struct lb_env *env)
7373 lockdep_assert_held(&env->src_rq->lock);
7375 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7376 set_task_cpu(p, env->dst_cpu);
7380 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7381 * part of active balancing operations within "domain".
7383 * Returns a task if successful and NULL otherwise.
7385 static struct task_struct *detach_one_task(struct lb_env *env)
7387 struct task_struct *p;
7389 lockdep_assert_held(&env->src_rq->lock);
7391 list_for_each_entry_reverse(p,
7392 &env->src_rq->cfs_tasks, se.group_node) {
7393 if (!can_migrate_task(p, env))
7396 detach_task(p, env);
7399 * Right now, this is only the second place where
7400 * lb_gained[env->idle] is updated (other is detach_tasks)
7401 * so we can safely collect stats here rather than
7402 * inside detach_tasks().
7404 schedstat_inc(env->sd->lb_gained[env->idle]);
7410 static const unsigned int sched_nr_migrate_break = 32;
7413 * detach_tasks() -- tries to detach up to imbalance runnable load from
7414 * busiest_rq, as part of a balancing operation within domain "sd".
7416 * Returns number of detached tasks if successful and 0 otherwise.
7418 static int detach_tasks(struct lb_env *env)
7420 struct list_head *tasks = &env->src_rq->cfs_tasks;
7421 struct task_struct *p;
7425 lockdep_assert_held(&env->src_rq->lock);
7427 if (env->imbalance <= 0)
7430 while (!list_empty(tasks)) {
7432 * We don't want to steal all, otherwise we may be treated likewise,
7433 * which could at worst lead to a livelock crash.
7435 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7438 p = list_last_entry(tasks, struct task_struct, se.group_node);
7441 /* We've more or less seen every task there is, call it quits */
7442 if (env->loop > env->loop_max)
7445 /* take a breather every nr_migrate tasks */
7446 if (env->loop > env->loop_break) {
7447 env->loop_break += sched_nr_migrate_break;
7448 env->flags |= LBF_NEED_BREAK;
7452 if (!can_migrate_task(p, env))
7456 * Depending of the number of CPUs and tasks and the
7457 * cgroup hierarchy, task_h_load() can return a null
7458 * value. Make sure that env->imbalance decreases
7459 * otherwise detach_tasks() will stop only after
7460 * detaching up to loop_max tasks.
7462 load = max_t(unsigned long, task_h_load(p), 1);
7465 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7468 if ((load / 2) > env->imbalance)
7471 detach_task(p, env);
7472 list_add(&p->se.group_node, &env->tasks);
7475 env->imbalance -= load;
7477 #ifdef CONFIG_PREEMPTION
7479 * NEWIDLE balancing is a source of latency, so preemptible
7480 * kernels will stop after the first task is detached to minimize
7481 * the critical section.
7483 if (env->idle == CPU_NEWLY_IDLE)
7488 * We only want to steal up to the prescribed amount of
7491 if (env->imbalance <= 0)
7496 list_move(&p->se.group_node, tasks);
7500 * Right now, this is one of only two places we collect this stat
7501 * so we can safely collect detach_one_task() stats here rather
7502 * than inside detach_one_task().
7504 schedstat_add(env->sd->lb_gained[env->idle], detached);
7510 * attach_task() -- attach the task detached by detach_task() to its new rq.
7512 static void attach_task(struct rq *rq, struct task_struct *p)
7514 lockdep_assert_held(&rq->lock);
7516 BUG_ON(task_rq(p) != rq);
7517 activate_task(rq, p, ENQUEUE_NOCLOCK);
7518 check_preempt_curr(rq, p, 0);
7522 * attach_one_task() -- attaches the task returned from detach_one_task() to
7525 static void attach_one_task(struct rq *rq, struct task_struct *p)
7530 update_rq_clock(rq);
7536 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7539 static void attach_tasks(struct lb_env *env)
7541 struct list_head *tasks = &env->tasks;
7542 struct task_struct *p;
7545 rq_lock(env->dst_rq, &rf);
7546 update_rq_clock(env->dst_rq);
7548 while (!list_empty(tasks)) {
7549 p = list_first_entry(tasks, struct task_struct, se.group_node);
7550 list_del_init(&p->se.group_node);
7552 attach_task(env->dst_rq, p);
7555 rq_unlock(env->dst_rq, &rf);
7558 #ifdef CONFIG_NO_HZ_COMMON
7559 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7561 if (cfs_rq->avg.load_avg)
7564 if (cfs_rq->avg.util_avg)
7570 static inline bool others_have_blocked(struct rq *rq)
7572 if (READ_ONCE(rq->avg_rt.util_avg))
7575 if (READ_ONCE(rq->avg_dl.util_avg))
7578 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7579 if (READ_ONCE(rq->avg_irq.util_avg))
7586 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
7588 rq->last_blocked_load_update_tick = jiffies;
7591 rq->has_blocked_load = 0;
7594 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
7595 static inline bool others_have_blocked(struct rq *rq) { return false; }
7596 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
7599 static bool __update_blocked_others(struct rq *rq, bool *done)
7601 const struct sched_class *curr_class;
7602 u64 now = rq_clock_pelt(rq);
7606 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
7607 * DL and IRQ signals have been updated before updating CFS.
7609 curr_class = rq->curr->sched_class;
7611 decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
7612 update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
7613 update_irq_load_avg(rq, 0);
7615 if (others_have_blocked(rq))
7621 #ifdef CONFIG_FAIR_GROUP_SCHED
7623 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7625 if (cfs_rq->load.weight)
7628 if (cfs_rq->avg.load_sum)
7631 if (cfs_rq->avg.util_sum)
7634 if (cfs_rq->avg.runnable_load_sum)
7640 static bool __update_blocked_fair(struct rq *rq, bool *done)
7642 struct cfs_rq *cfs_rq, *pos;
7643 bool decayed = false;
7644 int cpu = cpu_of(rq);
7647 * Iterates the task_group tree in a bottom up fashion, see
7648 * list_add_leaf_cfs_rq() for details.
7650 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7651 struct sched_entity *se;
7653 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
7654 update_tg_load_avg(cfs_rq, 0);
7656 if (cfs_rq == &rq->cfs)
7660 /* Propagate pending load changes to the parent, if any: */
7661 se = cfs_rq->tg->se[cpu];
7662 if (se && !skip_blocked_update(se))
7663 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
7666 * There can be a lot of idle CPU cgroups. Don't let fully
7667 * decayed cfs_rqs linger on the list.
7669 if (cfs_rq_is_decayed(cfs_rq))
7670 list_del_leaf_cfs_rq(cfs_rq);
7672 /* Don't need periodic decay once load/util_avg are null */
7673 if (cfs_rq_has_blocked(cfs_rq))
7681 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7682 * This needs to be done in a top-down fashion because the load of a child
7683 * group is a fraction of its parents load.
7685 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7687 struct rq *rq = rq_of(cfs_rq);
7688 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7689 unsigned long now = jiffies;
7692 if (cfs_rq->last_h_load_update == now)
7695 WRITE_ONCE(cfs_rq->h_load_next, NULL);
7696 for_each_sched_entity(se) {
7697 cfs_rq = cfs_rq_of(se);
7698 WRITE_ONCE(cfs_rq->h_load_next, se);
7699 if (cfs_rq->last_h_load_update == now)
7704 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7705 cfs_rq->last_h_load_update = now;
7708 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
7709 load = cfs_rq->h_load;
7710 load = div64_ul(load * se->avg.load_avg,
7711 cfs_rq_load_avg(cfs_rq) + 1);
7712 cfs_rq = group_cfs_rq(se);
7713 cfs_rq->h_load = load;
7714 cfs_rq->last_h_load_update = now;
7718 static unsigned long task_h_load(struct task_struct *p)
7720 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7722 update_cfs_rq_h_load(cfs_rq);
7723 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7724 cfs_rq_load_avg(cfs_rq) + 1);
7727 static bool __update_blocked_fair(struct rq *rq, bool *done)
7729 struct cfs_rq *cfs_rq = &rq->cfs;
7732 decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
7733 if (cfs_rq_has_blocked(cfs_rq))
7739 static unsigned long task_h_load(struct task_struct *p)
7741 return p->se.avg.load_avg;
7745 static void update_blocked_averages(int cpu)
7747 bool decayed = false, done = true;
7748 struct rq *rq = cpu_rq(cpu);
7751 rq_lock_irqsave(rq, &rf);
7752 update_rq_clock(rq);
7754 decayed |= __update_blocked_others(rq, &done);
7755 decayed |= __update_blocked_fair(rq, &done);
7757 update_blocked_load_status(rq, !done);
7759 cpufreq_update_util(rq, 0);
7760 rq_unlock_irqrestore(rq, &rf);
7763 /********** Helpers for find_busiest_group ************************/
7766 * sg_lb_stats - stats of a sched_group required for load_balancing
7768 struct sg_lb_stats {
7769 unsigned long avg_load; /*Avg load across the CPUs of the group */
7770 unsigned long group_load; /* Total load over the CPUs of the group */
7771 unsigned long load_per_task;
7772 unsigned long group_capacity;
7773 unsigned long group_util; /* Total utilization of the group */
7774 unsigned int sum_nr_running; /* Nr tasks running in the group */
7775 unsigned int idle_cpus;
7776 unsigned int group_weight;
7777 enum group_type group_type;
7778 int group_no_capacity;
7779 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
7780 #ifdef CONFIG_NUMA_BALANCING
7781 unsigned int nr_numa_running;
7782 unsigned int nr_preferred_running;
7787 * sd_lb_stats - Structure to store the statistics of a sched_domain
7788 * during load balancing.
7790 struct sd_lb_stats {
7791 struct sched_group *busiest; /* Busiest group in this sd */
7792 struct sched_group *local; /* Local group in this sd */
7793 unsigned long total_running;
7794 unsigned long total_load; /* Total load of all groups in sd */
7795 unsigned long total_capacity; /* Total capacity of all groups in sd */
7796 unsigned long avg_load; /* Average load across all groups in sd */
7798 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7799 struct sg_lb_stats local_stat; /* Statistics of the local group */
7802 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7805 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7806 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7807 * We must however clear busiest_stat::avg_load because
7808 * update_sd_pick_busiest() reads this before assignment.
7810 *sds = (struct sd_lb_stats){
7813 .total_running = 0UL,
7815 .total_capacity = 0UL,
7818 .sum_nr_running = 0,
7819 .group_type = group_other,
7824 static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu)
7826 struct rq *rq = cpu_rq(cpu);
7827 unsigned long max = arch_scale_cpu_capacity(cpu);
7828 unsigned long used, free;
7831 irq = cpu_util_irq(rq);
7833 if (unlikely(irq >= max))
7836 used = READ_ONCE(rq->avg_rt.util_avg);
7837 used += READ_ONCE(rq->avg_dl.util_avg);
7839 if (unlikely(used >= max))
7844 return scale_irq_capacity(free, irq, max);
7847 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7849 unsigned long capacity = scale_rt_capacity(sd, cpu);
7850 struct sched_group *sdg = sd->groups;
7852 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
7857 cpu_rq(cpu)->cpu_capacity = capacity;
7858 sdg->sgc->capacity = capacity;
7859 sdg->sgc->min_capacity = capacity;
7860 sdg->sgc->max_capacity = capacity;
7863 void update_group_capacity(struct sched_domain *sd, int cpu)
7865 struct sched_domain *child = sd->child;
7866 struct sched_group *group, *sdg = sd->groups;
7867 unsigned long capacity, min_capacity, max_capacity;
7868 unsigned long interval;
7870 interval = msecs_to_jiffies(sd->balance_interval);
7871 interval = clamp(interval, 1UL, max_load_balance_interval);
7872 sdg->sgc->next_update = jiffies + interval;
7875 update_cpu_capacity(sd, cpu);
7880 min_capacity = ULONG_MAX;
7883 if (child->flags & SD_OVERLAP) {
7885 * SD_OVERLAP domains cannot assume that child groups
7886 * span the current group.
7889 for_each_cpu(cpu, sched_group_span(sdg)) {
7890 struct sched_group_capacity *sgc;
7891 struct rq *rq = cpu_rq(cpu);
7894 * build_sched_domains() -> init_sched_groups_capacity()
7895 * gets here before we've attached the domains to the
7898 * Use capacity_of(), which is set irrespective of domains
7899 * in update_cpu_capacity().
7901 * This avoids capacity from being 0 and
7902 * causing divide-by-zero issues on boot.
7904 if (unlikely(!rq->sd)) {
7905 capacity += capacity_of(cpu);
7907 sgc = rq->sd->groups->sgc;
7908 capacity += sgc->capacity;
7911 min_capacity = min(capacity, min_capacity);
7912 max_capacity = max(capacity, max_capacity);
7916 * !SD_OVERLAP domains can assume that child groups
7917 * span the current group.
7920 group = child->groups;
7922 struct sched_group_capacity *sgc = group->sgc;
7924 capacity += sgc->capacity;
7925 min_capacity = min(sgc->min_capacity, min_capacity);
7926 max_capacity = max(sgc->max_capacity, max_capacity);
7927 group = group->next;
7928 } while (group != child->groups);
7931 sdg->sgc->capacity = capacity;
7932 sdg->sgc->min_capacity = min_capacity;
7933 sdg->sgc->max_capacity = max_capacity;
7937 * Check whether the capacity of the rq has been noticeably reduced by side
7938 * activity. The imbalance_pct is used for the threshold.
7939 * Return true is the capacity is reduced
7942 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7944 return ((rq->cpu_capacity * sd->imbalance_pct) <
7945 (rq->cpu_capacity_orig * 100));
7949 * Check whether a rq has a misfit task and if it looks like we can actually
7950 * help that task: we can migrate the task to a CPU of higher capacity, or
7951 * the task's current CPU is heavily pressured.
7953 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
7955 return rq->misfit_task_load &&
7956 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
7957 check_cpu_capacity(rq, sd));
7961 * Group imbalance indicates (and tries to solve) the problem where balancing
7962 * groups is inadequate due to ->cpus_ptr constraints.
7964 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
7965 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
7968 * { 0 1 2 3 } { 4 5 6 7 }
7971 * If we were to balance group-wise we'd place two tasks in the first group and
7972 * two tasks in the second group. Clearly this is undesired as it will overload
7973 * cpu 3 and leave one of the CPUs in the second group unused.
7975 * The current solution to this issue is detecting the skew in the first group
7976 * by noticing the lower domain failed to reach balance and had difficulty
7977 * moving tasks due to affinity constraints.
7979 * When this is so detected; this group becomes a candidate for busiest; see
7980 * update_sd_pick_busiest(). And calculate_imbalance() and
7981 * find_busiest_group() avoid some of the usual balance conditions to allow it
7982 * to create an effective group imbalance.
7984 * This is a somewhat tricky proposition since the next run might not find the
7985 * group imbalance and decide the groups need to be balanced again. A most
7986 * subtle and fragile situation.
7989 static inline int sg_imbalanced(struct sched_group *group)
7991 return group->sgc->imbalance;
7995 * group_has_capacity returns true if the group has spare capacity that could
7996 * be used by some tasks.
7997 * We consider that a group has spare capacity if the * number of task is
7998 * smaller than the number of CPUs or if the utilization is lower than the
7999 * available capacity for CFS tasks.
8000 * For the latter, we use a threshold to stabilize the state, to take into
8001 * account the variance of the tasks' load and to return true if the available
8002 * capacity in meaningful for the load balancer.
8003 * As an example, an available capacity of 1% can appear but it doesn't make
8004 * any benefit for the load balance.
8007 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
8009 if (sgs->sum_nr_running < sgs->group_weight)
8012 if ((sgs->group_capacity * 100) >
8013 (sgs->group_util * env->sd->imbalance_pct))
8020 * group_is_overloaded returns true if the group has more tasks than it can
8022 * group_is_overloaded is not equals to !group_has_capacity because a group
8023 * with the exact right number of tasks, has no more spare capacity but is not
8024 * overloaded so both group_has_capacity and group_is_overloaded return
8028 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
8030 if (sgs->sum_nr_running <= sgs->group_weight)
8033 if ((sgs->group_capacity * 100) <
8034 (sgs->group_util * env->sd->imbalance_pct))
8041 * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller
8042 * per-CPU capacity than sched_group ref.
8045 group_smaller_min_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
8047 return fits_capacity(sg->sgc->min_capacity, ref->sgc->min_capacity);
8051 * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller
8052 * per-CPU capacity_orig than sched_group ref.
8055 group_smaller_max_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
8057 return fits_capacity(sg->sgc->max_capacity, ref->sgc->max_capacity);
8061 group_type group_classify(struct sched_group *group,
8062 struct sg_lb_stats *sgs)
8064 if (sgs->group_no_capacity)
8065 return group_overloaded;
8067 if (sg_imbalanced(group))
8068 return group_imbalanced;
8070 if (sgs->group_misfit_task_load)
8071 return group_misfit_task;
8076 static bool update_nohz_stats(struct rq *rq, bool force)
8078 #ifdef CONFIG_NO_HZ_COMMON
8079 unsigned int cpu = rq->cpu;
8081 if (!rq->has_blocked_load)
8084 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
8087 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
8090 update_blocked_averages(cpu);
8092 return rq->has_blocked_load;
8099 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8100 * @env: The load balancing environment.
8101 * @group: sched_group whose statistics are to be updated.
8102 * @sgs: variable to hold the statistics for this group.
8103 * @sg_status: Holds flag indicating the status of the sched_group
8105 static inline void update_sg_lb_stats(struct lb_env *env,
8106 struct sched_group *group,
8107 struct sg_lb_stats *sgs,
8112 memset(sgs, 0, sizeof(*sgs));
8114 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8115 struct rq *rq = cpu_rq(i);
8117 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
8118 env->flags |= LBF_NOHZ_AGAIN;
8120 sgs->group_load += cpu_runnable_load(rq);
8121 sgs->group_util += cpu_util(i);
8122 sgs->sum_nr_running += rq->cfs.h_nr_running;
8124 nr_running = rq->nr_running;
8126 *sg_status |= SG_OVERLOAD;
8128 if (cpu_overutilized(i))
8129 *sg_status |= SG_OVERUTILIZED;
8131 #ifdef CONFIG_NUMA_BALANCING
8132 sgs->nr_numa_running += rq->nr_numa_running;
8133 sgs->nr_preferred_running += rq->nr_preferred_running;
8136 * No need to call idle_cpu() if nr_running is not 0
8138 if (!nr_running && idle_cpu(i))
8141 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8142 sgs->group_misfit_task_load < rq->misfit_task_load) {
8143 sgs->group_misfit_task_load = rq->misfit_task_load;
8144 *sg_status |= SG_OVERLOAD;
8148 /* Adjust by relative CPU capacity of the group */
8149 sgs->group_capacity = group->sgc->capacity;
8150 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
8152 if (sgs->sum_nr_running)
8153 sgs->load_per_task = sgs->group_load / sgs->sum_nr_running;
8155 sgs->group_weight = group->group_weight;
8157 sgs->group_no_capacity = group_is_overloaded(env, sgs);
8158 sgs->group_type = group_classify(group, sgs);
8162 * update_sd_pick_busiest - return 1 on busiest group
8163 * @env: The load balancing environment.
8164 * @sds: sched_domain statistics
8165 * @sg: sched_group candidate to be checked for being the busiest
8166 * @sgs: sched_group statistics
8168 * Determine if @sg is a busier group than the previously selected
8171 * Return: %true if @sg is a busier group than the previously selected
8172 * busiest group. %false otherwise.
8174 static bool update_sd_pick_busiest(struct lb_env *env,
8175 struct sd_lb_stats *sds,
8176 struct sched_group *sg,
8177 struct sg_lb_stats *sgs)
8179 struct sg_lb_stats *busiest = &sds->busiest_stat;
8182 * Don't try to pull misfit tasks we can't help.
8183 * We can use max_capacity here as reduction in capacity on some
8184 * CPUs in the group should either be possible to resolve
8185 * internally or be covered by avg_load imbalance (eventually).
8187 if (sgs->group_type == group_misfit_task &&
8188 (!group_smaller_max_cpu_capacity(sg, sds->local) ||
8189 !group_has_capacity(env, &sds->local_stat)))
8192 if (sgs->group_type > busiest->group_type)
8195 if (sgs->group_type < busiest->group_type)
8198 if (sgs->avg_load <= busiest->avg_load)
8201 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
8205 * Candidate sg has no more than one task per CPU and
8206 * has higher per-CPU capacity. Migrating tasks to less
8207 * capable CPUs may harm throughput. Maximize throughput,
8208 * power/energy consequences are not considered.
8210 if (sgs->sum_nr_running <= sgs->group_weight &&
8211 group_smaller_min_cpu_capacity(sds->local, sg))
8215 * If we have more than one misfit sg go with the biggest misfit.
8217 if (sgs->group_type == group_misfit_task &&
8218 sgs->group_misfit_task_load < busiest->group_misfit_task_load)
8222 /* This is the busiest node in its class. */
8223 if (!(env->sd->flags & SD_ASYM_PACKING))
8226 /* No ASYM_PACKING if target CPU is already busy */
8227 if (env->idle == CPU_NOT_IDLE)
8230 * ASYM_PACKING needs to move all the work to the highest
8231 * prority CPUs in the group, therefore mark all groups
8232 * of lower priority than ourself as busy.
8234 if (sgs->sum_nr_running &&
8235 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
8239 /* Prefer to move from lowest priority CPU's work */
8240 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
8241 sg->asym_prefer_cpu))
8248 #ifdef CONFIG_NUMA_BALANCING
8249 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8251 if (sgs->sum_nr_running > sgs->nr_numa_running)
8253 if (sgs->sum_nr_running > sgs->nr_preferred_running)
8258 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8260 if (rq->nr_running > rq->nr_numa_running)
8262 if (rq->nr_running > rq->nr_preferred_running)
8267 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8272 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8276 #endif /* CONFIG_NUMA_BALANCING */
8279 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8280 * @env: The load balancing environment.
8281 * @sds: variable to hold the statistics for this sched_domain.
8283 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8285 struct sched_domain *child = env->sd->child;
8286 struct sched_group *sg = env->sd->groups;
8287 struct sg_lb_stats *local = &sds->local_stat;
8288 struct sg_lb_stats tmp_sgs;
8289 bool prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
8292 #ifdef CONFIG_NO_HZ_COMMON
8293 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
8294 env->flags |= LBF_NOHZ_STATS;
8298 struct sg_lb_stats *sgs = &tmp_sgs;
8301 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
8306 if (env->idle != CPU_NEWLY_IDLE ||
8307 time_after_eq(jiffies, sg->sgc->next_update))
8308 update_group_capacity(env->sd, env->dst_cpu);
8311 update_sg_lb_stats(env, sg, sgs, &sg_status);
8317 * In case the child domain prefers tasks go to siblings
8318 * first, lower the sg capacity so that we'll try
8319 * and move all the excess tasks away. We lower the capacity
8320 * of a group only if the local group has the capacity to fit
8321 * these excess tasks. The extra check prevents the case where
8322 * you always pull from the heaviest group when it is already
8323 * under-utilized (possible with a large weight task outweighs
8324 * the tasks on the system).
8326 if (prefer_sibling && sds->local &&
8327 group_has_capacity(env, local) &&
8328 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
8329 sgs->group_no_capacity = 1;
8330 sgs->group_type = group_classify(sg, sgs);
8333 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8335 sds->busiest_stat = *sgs;
8339 /* Now, start updating sd_lb_stats */
8340 sds->total_running += sgs->sum_nr_running;
8341 sds->total_load += sgs->group_load;
8342 sds->total_capacity += sgs->group_capacity;
8345 } while (sg != env->sd->groups);
8347 #ifdef CONFIG_NO_HZ_COMMON
8348 if ((env->flags & LBF_NOHZ_AGAIN) &&
8349 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8351 WRITE_ONCE(nohz.next_blocked,
8352 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8356 if (env->sd->flags & SD_NUMA)
8357 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8359 if (!env->sd->parent) {
8360 struct root_domain *rd = env->dst_rq->rd;
8362 /* update overload indicator if we are at root domain */
8363 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
8365 /* Update over-utilization (tipping point, U >= 0) indicator */
8366 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
8367 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
8368 } else if (sg_status & SG_OVERUTILIZED) {
8369 struct root_domain *rd = env->dst_rq->rd;
8371 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
8372 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
8377 * check_asym_packing - Check to see if the group is packed into the
8380 * This is primarily intended to used at the sibling level. Some
8381 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8382 * case of POWER7, it can move to lower SMT modes only when higher
8383 * threads are idle. When in lower SMT modes, the threads will
8384 * perform better since they share less core resources. Hence when we
8385 * have idle threads, we want them to be the higher ones.
8387 * This packing function is run on idle threads. It checks to see if
8388 * the busiest CPU in this domain (core in the P7 case) has a higher
8389 * CPU number than the packing function is being run on. Here we are
8390 * assuming lower CPU number will be equivalent to lower a SMT thread
8393 * Return: 1 when packing is required and a task should be moved to
8394 * this CPU. The amount of the imbalance is returned in env->imbalance.
8396 * @env: The load balancing environment.
8397 * @sds: Statistics of the sched_domain which is to be packed
8399 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8403 if (!(env->sd->flags & SD_ASYM_PACKING))
8406 if (env->idle == CPU_NOT_IDLE)
8412 busiest_cpu = sds->busiest->asym_prefer_cpu;
8413 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
8416 env->imbalance = sds->busiest_stat.group_load;
8422 * fix_small_imbalance - Calculate the minor imbalance that exists
8423 * amongst the groups of a sched_domain, during
8425 * @env: The load balancing environment.
8426 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8429 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8431 unsigned long tmp, capa_now = 0, capa_move = 0;
8432 unsigned int imbn = 2;
8433 unsigned long scaled_busy_load_per_task;
8434 struct sg_lb_stats *local, *busiest;
8436 local = &sds->local_stat;
8437 busiest = &sds->busiest_stat;
8439 if (!local->sum_nr_running)
8440 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8441 else if (busiest->load_per_task > local->load_per_task)
8444 scaled_busy_load_per_task =
8445 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8446 busiest->group_capacity;
8448 if (busiest->avg_load + scaled_busy_load_per_task >=
8449 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8450 env->imbalance = busiest->load_per_task;
8455 * OK, we don't have enough imbalance to justify moving tasks,
8456 * however we may be able to increase total CPU capacity used by
8460 capa_now += busiest->group_capacity *
8461 min(busiest->load_per_task, busiest->avg_load);
8462 capa_now += local->group_capacity *
8463 min(local->load_per_task, local->avg_load);
8464 capa_now /= SCHED_CAPACITY_SCALE;
8466 /* Amount of load we'd subtract */
8467 if (busiest->avg_load > scaled_busy_load_per_task) {
8468 capa_move += busiest->group_capacity *
8469 min(busiest->load_per_task,
8470 busiest->avg_load - scaled_busy_load_per_task);
8473 /* Amount of load we'd add */
8474 if (busiest->avg_load * busiest->group_capacity <
8475 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8476 tmp = (busiest->avg_load * busiest->group_capacity) /
8477 local->group_capacity;
8479 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8480 local->group_capacity;
8482 capa_move += local->group_capacity *
8483 min(local->load_per_task, local->avg_load + tmp);
8484 capa_move /= SCHED_CAPACITY_SCALE;
8486 /* Move if we gain throughput */
8487 if (capa_move > capa_now)
8488 env->imbalance = busiest->load_per_task;
8492 * calculate_imbalance - Calculate the amount of imbalance present within the
8493 * groups of a given sched_domain during load balance.
8494 * @env: load balance environment
8495 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8497 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8499 unsigned long max_pull, load_above_capacity = ~0UL;
8500 struct sg_lb_stats *local, *busiest;
8502 local = &sds->local_stat;
8503 busiest = &sds->busiest_stat;
8505 if (busiest->group_type == group_imbalanced) {
8507 * In the group_imb case we cannot rely on group-wide averages
8508 * to ensure CPU-load equilibrium, look at wider averages. XXX
8510 busiest->load_per_task =
8511 min(busiest->load_per_task, sds->avg_load);
8515 * Avg load of busiest sg can be less and avg load of local sg can
8516 * be greater than avg load across all sgs of sd because avg load
8517 * factors in sg capacity and sgs with smaller group_type are
8518 * skipped when updating the busiest sg:
8520 if (busiest->group_type != group_misfit_task &&
8521 (busiest->avg_load <= sds->avg_load ||
8522 local->avg_load >= sds->avg_load)) {
8524 return fix_small_imbalance(env, sds);
8528 * If there aren't any idle CPUs, avoid creating some.
8530 if (busiest->group_type == group_overloaded &&
8531 local->group_type == group_overloaded) {
8532 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
8533 if (load_above_capacity > busiest->group_capacity) {
8534 load_above_capacity -= busiest->group_capacity;
8535 load_above_capacity *= scale_load_down(NICE_0_LOAD);
8536 load_above_capacity /= busiest->group_capacity;
8538 load_above_capacity = ~0UL;
8542 * We're trying to get all the CPUs to the average_load, so we don't
8543 * want to push ourselves above the average load, nor do we wish to
8544 * reduce the max loaded CPU below the average load. At the same time,
8545 * we also don't want to reduce the group load below the group
8546 * capacity. Thus we look for the minimum possible imbalance.
8548 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8550 /* How much load to actually move to equalise the imbalance */
8551 env->imbalance = min(
8552 max_pull * busiest->group_capacity,
8553 (sds->avg_load - local->avg_load) * local->group_capacity
8554 ) / SCHED_CAPACITY_SCALE;
8556 /* Boost imbalance to allow misfit task to be balanced. */
8557 if (busiest->group_type == group_misfit_task) {
8558 env->imbalance = max_t(long, env->imbalance,
8559 busiest->group_misfit_task_load);
8563 * if *imbalance is less than the average load per runnable task
8564 * there is no guarantee that any tasks will be moved so we'll have
8565 * a think about bumping its value to force at least one task to be
8568 if (env->imbalance < busiest->load_per_task)
8569 return fix_small_imbalance(env, sds);
8572 /******* find_busiest_group() helpers end here *********************/
8575 * find_busiest_group - Returns the busiest group within the sched_domain
8576 * if there is an imbalance.
8578 * Also calculates the amount of runnable load which should be moved
8579 * to restore balance.
8581 * @env: The load balancing environment.
8583 * Return: - The busiest group if imbalance exists.
8585 static struct sched_group *find_busiest_group(struct lb_env *env)
8587 struct sg_lb_stats *local, *busiest;
8588 struct sd_lb_stats sds;
8590 init_sd_lb_stats(&sds);
8593 * Compute the various statistics relavent for load balancing at
8596 update_sd_lb_stats(env, &sds);
8598 if (sched_energy_enabled()) {
8599 struct root_domain *rd = env->dst_rq->rd;
8601 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
8605 local = &sds.local_stat;
8606 busiest = &sds.busiest_stat;
8608 /* ASYM feature bypasses nice load balance check */
8609 if (check_asym_packing(env, &sds))
8612 /* There is no busy sibling group to pull tasks from */
8613 if (!sds.busiest || busiest->sum_nr_running == 0)
8616 /* XXX broken for overlapping NUMA groups */
8617 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8618 / sds.total_capacity;
8621 * If the busiest group is imbalanced the below checks don't
8622 * work because they assume all things are equal, which typically
8623 * isn't true due to cpus_ptr constraints and the like.
8625 if (busiest->group_type == group_imbalanced)
8629 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8630 * capacities from resulting in underutilization due to avg_load.
8632 if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8633 busiest->group_no_capacity)
8636 /* Misfit tasks should be dealt with regardless of the avg load */
8637 if (busiest->group_type == group_misfit_task)
8641 * If the local group is busier than the selected busiest group
8642 * don't try and pull any tasks.
8644 if (local->avg_load >= busiest->avg_load)
8648 * Don't pull any tasks if this group is already above the domain
8651 if (local->avg_load >= sds.avg_load)
8654 if (env->idle == CPU_IDLE) {
8656 * This CPU is idle. If the busiest group is not overloaded
8657 * and there is no imbalance between this and busiest group
8658 * wrt idle CPUs, it is balanced. The imbalance becomes
8659 * significant if the diff is greater than 1 otherwise we
8660 * might end up to just move the imbalance on another group
8662 if ((busiest->group_type != group_overloaded) &&
8663 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8667 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8668 * imbalance_pct to be conservative.
8670 if (100 * busiest->avg_load <=
8671 env->sd->imbalance_pct * local->avg_load)
8676 /* Looks like there is an imbalance. Compute it */
8677 env->src_grp_type = busiest->group_type;
8678 calculate_imbalance(env, &sds);
8679 return env->imbalance ? sds.busiest : NULL;
8687 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
8689 static struct rq *find_busiest_queue(struct lb_env *env,
8690 struct sched_group *group)
8692 struct rq *busiest = NULL, *rq;
8693 unsigned long busiest_load = 0, busiest_capacity = 1;
8696 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8697 unsigned long capacity, load;
8701 rt = fbq_classify_rq(rq);
8704 * We classify groups/runqueues into three groups:
8705 * - regular: there are !numa tasks
8706 * - remote: there are numa tasks that run on the 'wrong' node
8707 * - all: there is no distinction
8709 * In order to avoid migrating ideally placed numa tasks,
8710 * ignore those when there's better options.
8712 * If we ignore the actual busiest queue to migrate another
8713 * task, the next balance pass can still reduce the busiest
8714 * queue by moving tasks around inside the node.
8716 * If we cannot move enough load due to this classification
8717 * the next pass will adjust the group classification and
8718 * allow migration of more tasks.
8720 * Both cases only affect the total convergence complexity.
8722 if (rt > env->fbq_type)
8726 * For ASYM_CPUCAPACITY domains with misfit tasks we simply
8727 * seek the "biggest" misfit task.
8729 if (env->src_grp_type == group_misfit_task) {
8730 if (rq->misfit_task_load > busiest_load) {
8731 busiest_load = rq->misfit_task_load;
8738 capacity = capacity_of(i);
8741 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
8742 * eventually lead to active_balancing high->low capacity.
8743 * Higher per-CPU capacity is considered better than balancing
8746 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8747 capacity_of(env->dst_cpu) < capacity &&
8748 rq->nr_running == 1)
8751 load = cpu_runnable_load(rq);
8754 * When comparing with imbalance, use cpu_runnable_load()
8755 * which is not scaled with the CPU capacity.
8758 if (rq->nr_running == 1 && load > env->imbalance &&
8759 !check_cpu_capacity(rq, env->sd))
8763 * For the load comparisons with the other CPU's, consider
8764 * the cpu_runnable_load() scaled with the CPU capacity, so
8765 * that the load can be moved away from the CPU that is
8766 * potentially running at a lower capacity.
8768 * Thus we're looking for max(load_i / capacity_i), crosswise
8769 * multiplication to rid ourselves of the division works out
8770 * to: load_i * capacity_j > load_j * capacity_i; where j is
8771 * our previous maximum.
8773 if (load * busiest_capacity > busiest_load * capacity) {
8774 busiest_load = load;
8775 busiest_capacity = capacity;
8784 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8785 * so long as it is large enough.
8787 #define MAX_PINNED_INTERVAL 512
8790 asym_active_balance(struct lb_env *env)
8793 * ASYM_PACKING needs to force migrate tasks from busy but
8794 * lower priority CPUs in order to pack all tasks in the
8795 * highest priority CPUs.
8797 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
8798 sched_asym_prefer(env->dst_cpu, env->src_cpu);
8802 voluntary_active_balance(struct lb_env *env)
8804 struct sched_domain *sd = env->sd;
8806 if (asym_active_balance(env))
8810 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8811 * It's worth migrating the task if the src_cpu's capacity is reduced
8812 * because of other sched_class or IRQs if more capacity stays
8813 * available on dst_cpu.
8815 if ((env->idle != CPU_NOT_IDLE) &&
8816 (env->src_rq->cfs.h_nr_running == 1)) {
8817 if ((check_cpu_capacity(env->src_rq, sd)) &&
8818 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8822 if (env->src_grp_type == group_misfit_task)
8828 static int need_active_balance(struct lb_env *env)
8830 struct sched_domain *sd = env->sd;
8832 if (voluntary_active_balance(env))
8835 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8838 static int active_load_balance_cpu_stop(void *data);
8840 static int should_we_balance(struct lb_env *env)
8842 struct sched_group *sg = env->sd->groups;
8843 int cpu, balance_cpu = -1;
8846 * Ensure the balancing environment is consistent; can happen
8847 * when the softirq triggers 'during' hotplug.
8849 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8853 * In the newly idle case, we will allow all the CPUs
8854 * to do the newly idle load balance.
8856 if (env->idle == CPU_NEWLY_IDLE)
8859 /* Try to find first idle CPU */
8860 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8868 if (balance_cpu == -1)
8869 balance_cpu = group_balance_cpu(sg);
8872 * First idle CPU or the first CPU(busiest) in this sched group
8873 * is eligible for doing load balancing at this and above domains.
8875 return balance_cpu == env->dst_cpu;
8879 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8880 * tasks if there is an imbalance.
8882 static int load_balance(int this_cpu, struct rq *this_rq,
8883 struct sched_domain *sd, enum cpu_idle_type idle,
8884 int *continue_balancing)
8886 int ld_moved, cur_ld_moved, active_balance = 0;
8887 struct sched_domain *sd_parent = sd->parent;
8888 struct sched_group *group;
8891 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8893 struct lb_env env = {
8895 .dst_cpu = this_cpu,
8897 .dst_grpmask = sched_group_span(sd->groups),
8899 .loop_break = sched_nr_migrate_break,
8902 .tasks = LIST_HEAD_INIT(env.tasks),
8905 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8907 schedstat_inc(sd->lb_count[idle]);
8910 if (!should_we_balance(&env)) {
8911 *continue_balancing = 0;
8915 group = find_busiest_group(&env);
8917 schedstat_inc(sd->lb_nobusyg[idle]);
8921 busiest = find_busiest_queue(&env, group);
8923 schedstat_inc(sd->lb_nobusyq[idle]);
8927 BUG_ON(busiest == env.dst_rq);
8929 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8931 env.src_cpu = busiest->cpu;
8932 env.src_rq = busiest;
8935 if (busiest->nr_running > 1) {
8937 * Attempt to move tasks. If find_busiest_group has found
8938 * an imbalance but busiest->nr_running <= 1, the group is
8939 * still unbalanced. ld_moved simply stays zero, so it is
8940 * correctly treated as an imbalance.
8942 env.flags |= LBF_ALL_PINNED;
8943 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8946 rq_lock_irqsave(busiest, &rf);
8947 update_rq_clock(busiest);
8950 * cur_ld_moved - load moved in current iteration
8951 * ld_moved - cumulative load moved across iterations
8953 cur_ld_moved = detach_tasks(&env);
8956 * We've detached some tasks from busiest_rq. Every
8957 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8958 * unlock busiest->lock, and we are able to be sure
8959 * that nobody can manipulate the tasks in parallel.
8960 * See task_rq_lock() family for the details.
8963 rq_unlock(busiest, &rf);
8967 ld_moved += cur_ld_moved;
8970 local_irq_restore(rf.flags);
8972 if (env.flags & LBF_NEED_BREAK) {
8973 env.flags &= ~LBF_NEED_BREAK;
8978 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8979 * us and move them to an alternate dst_cpu in our sched_group
8980 * where they can run. The upper limit on how many times we
8981 * iterate on same src_cpu is dependent on number of CPUs in our
8984 * This changes load balance semantics a bit on who can move
8985 * load to a given_cpu. In addition to the given_cpu itself
8986 * (or a ilb_cpu acting on its behalf where given_cpu is
8987 * nohz-idle), we now have balance_cpu in a position to move
8988 * load to given_cpu. In rare situations, this may cause
8989 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8990 * _independently_ and at _same_ time to move some load to
8991 * given_cpu) causing exceess load to be moved to given_cpu.
8992 * This however should not happen so much in practice and
8993 * moreover subsequent load balance cycles should correct the
8994 * excess load moved.
8996 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8998 /* Prevent to re-select dst_cpu via env's CPUs */
8999 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
9001 env.dst_rq = cpu_rq(env.new_dst_cpu);
9002 env.dst_cpu = env.new_dst_cpu;
9003 env.flags &= ~LBF_DST_PINNED;
9005 env.loop_break = sched_nr_migrate_break;
9008 * Go back to "more_balance" rather than "redo" since we
9009 * need to continue with same src_cpu.
9015 * We failed to reach balance because of affinity.
9018 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9020 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
9021 *group_imbalance = 1;
9024 /* All tasks on this runqueue were pinned by CPU affinity */
9025 if (unlikely(env.flags & LBF_ALL_PINNED)) {
9026 __cpumask_clear_cpu(cpu_of(busiest), cpus);
9028 * Attempting to continue load balancing at the current
9029 * sched_domain level only makes sense if there are
9030 * active CPUs remaining as possible busiest CPUs to
9031 * pull load from which are not contained within the
9032 * destination group that is receiving any migrated
9035 if (!cpumask_subset(cpus, env.dst_grpmask)) {
9037 env.loop_break = sched_nr_migrate_break;
9040 goto out_all_pinned;
9045 schedstat_inc(sd->lb_failed[idle]);
9047 * Increment the failure counter only on periodic balance.
9048 * We do not want newidle balance, which can be very
9049 * frequent, pollute the failure counter causing
9050 * excessive cache_hot migrations and active balances.
9052 if (idle != CPU_NEWLY_IDLE)
9053 sd->nr_balance_failed++;
9055 if (need_active_balance(&env)) {
9056 unsigned long flags;
9058 raw_spin_lock_irqsave(&busiest->lock, flags);
9061 * Don't kick the active_load_balance_cpu_stop,
9062 * if the curr task on busiest CPU can't be
9063 * moved to this_cpu:
9065 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
9066 raw_spin_unlock_irqrestore(&busiest->lock,
9068 env.flags |= LBF_ALL_PINNED;
9069 goto out_one_pinned;
9073 * ->active_balance synchronizes accesses to
9074 * ->active_balance_work. Once set, it's cleared
9075 * only after active load balance is finished.
9077 if (!busiest->active_balance) {
9078 busiest->active_balance = 1;
9079 busiest->push_cpu = this_cpu;
9082 raw_spin_unlock_irqrestore(&busiest->lock, flags);
9084 if (active_balance) {
9085 stop_one_cpu_nowait(cpu_of(busiest),
9086 active_load_balance_cpu_stop, busiest,
9087 &busiest->active_balance_work);
9090 /* We've kicked active balancing, force task migration. */
9091 sd->nr_balance_failed = sd->cache_nice_tries+1;
9094 sd->nr_balance_failed = 0;
9096 if (likely(!active_balance) || voluntary_active_balance(&env)) {
9097 /* We were unbalanced, so reset the balancing interval */
9098 sd->balance_interval = sd->min_interval;
9101 * If we've begun active balancing, start to back off. This
9102 * case may not be covered by the all_pinned logic if there
9103 * is only 1 task on the busy runqueue (because we don't call
9106 if (sd->balance_interval < sd->max_interval)
9107 sd->balance_interval *= 2;
9114 * We reach balance although we may have faced some affinity
9115 * constraints. Clear the imbalance flag only if other tasks got
9116 * a chance to move and fix the imbalance.
9118 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
9119 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9121 if (*group_imbalance)
9122 *group_imbalance = 0;
9127 * We reach balance because all tasks are pinned at this level so
9128 * we can't migrate them. Let the imbalance flag set so parent level
9129 * can try to migrate them.
9131 schedstat_inc(sd->lb_balanced[idle]);
9133 sd->nr_balance_failed = 0;
9139 * newidle_balance() disregards balance intervals, so we could
9140 * repeatedly reach this code, which would lead to balance_interval
9141 * skyrocketting in a short amount of time. Skip the balance_interval
9142 * increase logic to avoid that.
9144 if (env.idle == CPU_NEWLY_IDLE)
9147 /* tune up the balancing interval */
9148 if ((env.flags & LBF_ALL_PINNED &&
9149 sd->balance_interval < MAX_PINNED_INTERVAL) ||
9150 sd->balance_interval < sd->max_interval)
9151 sd->balance_interval *= 2;
9156 static inline unsigned long
9157 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
9159 unsigned long interval = sd->balance_interval;
9162 interval *= sd->busy_factor;
9164 /* scale ms to jiffies */
9165 interval = msecs_to_jiffies(interval);
9166 interval = clamp(interval, 1UL, max_load_balance_interval);
9172 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
9174 unsigned long interval, next;
9176 /* used by idle balance, so cpu_busy = 0 */
9177 interval = get_sd_balance_interval(sd, 0);
9178 next = sd->last_balance + interval;
9180 if (time_after(*next_balance, next))
9181 *next_balance = next;
9185 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9186 * running tasks off the busiest CPU onto idle CPUs. It requires at
9187 * least 1 task to be running on each physical CPU where possible, and
9188 * avoids physical / logical imbalances.
9190 static int active_load_balance_cpu_stop(void *data)
9192 struct rq *busiest_rq = data;
9193 int busiest_cpu = cpu_of(busiest_rq);
9194 int target_cpu = busiest_rq->push_cpu;
9195 struct rq *target_rq = cpu_rq(target_cpu);
9196 struct sched_domain *sd;
9197 struct task_struct *p = NULL;
9200 rq_lock_irq(busiest_rq, &rf);
9202 * Between queueing the stop-work and running it is a hole in which
9203 * CPUs can become inactive. We should not move tasks from or to
9206 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
9209 /* Make sure the requested CPU hasn't gone down in the meantime: */
9210 if (unlikely(busiest_cpu != smp_processor_id() ||
9211 !busiest_rq->active_balance))
9214 /* Is there any task to move? */
9215 if (busiest_rq->nr_running <= 1)
9219 * This condition is "impossible", if it occurs
9220 * we need to fix it. Originally reported by
9221 * Bjorn Helgaas on a 128-CPU setup.
9223 BUG_ON(busiest_rq == target_rq);
9225 /* Search for an sd spanning us and the target CPU. */
9227 for_each_domain(target_cpu, sd) {
9228 if ((sd->flags & SD_LOAD_BALANCE) &&
9229 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9234 struct lb_env env = {
9236 .dst_cpu = target_cpu,
9237 .dst_rq = target_rq,
9238 .src_cpu = busiest_rq->cpu,
9239 .src_rq = busiest_rq,
9242 * can_migrate_task() doesn't need to compute new_dst_cpu
9243 * for active balancing. Since we have CPU_IDLE, but no
9244 * @dst_grpmask we need to make that test go away with lying
9247 .flags = LBF_DST_PINNED,
9250 schedstat_inc(sd->alb_count);
9251 update_rq_clock(busiest_rq);
9253 p = detach_one_task(&env);
9255 schedstat_inc(sd->alb_pushed);
9256 /* Active balancing done, reset the failure counter. */
9257 sd->nr_balance_failed = 0;
9259 schedstat_inc(sd->alb_failed);
9264 busiest_rq->active_balance = 0;
9265 rq_unlock(busiest_rq, &rf);
9268 attach_one_task(target_rq, p);
9275 static DEFINE_SPINLOCK(balancing);
9278 * Scale the max load_balance interval with the number of CPUs in the system.
9279 * This trades load-balance latency on larger machines for less cross talk.
9281 void update_max_interval(void)
9283 max_load_balance_interval = HZ*num_online_cpus()/10;
9287 * It checks each scheduling domain to see if it is due to be balanced,
9288 * and initiates a balancing operation if so.
9290 * Balancing parameters are set up in init_sched_domains.
9292 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9294 int continue_balancing = 1;
9296 unsigned long interval;
9297 struct sched_domain *sd;
9298 /* Earliest time when we have to do rebalance again */
9299 unsigned long next_balance = jiffies + 60*HZ;
9300 int update_next_balance = 0;
9301 int need_serialize, need_decay = 0;
9305 for_each_domain(cpu, sd) {
9307 * Decay the newidle max times here because this is a regular
9308 * visit to all the domains. Decay ~1% per second.
9310 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9311 sd->max_newidle_lb_cost =
9312 (sd->max_newidle_lb_cost * 253) / 256;
9313 sd->next_decay_max_lb_cost = jiffies + HZ;
9316 max_cost += sd->max_newidle_lb_cost;
9318 if (!(sd->flags & SD_LOAD_BALANCE))
9322 * Stop the load balance at this level. There is another
9323 * CPU in our sched group which is doing load balancing more
9326 if (!continue_balancing) {
9332 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9334 need_serialize = sd->flags & SD_SERIALIZE;
9335 if (need_serialize) {
9336 if (!spin_trylock(&balancing))
9340 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9341 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9343 * The LBF_DST_PINNED logic could have changed
9344 * env->dst_cpu, so we can't know our idle
9345 * state even if we migrated tasks. Update it.
9347 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9349 sd->last_balance = jiffies;
9350 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9353 spin_unlock(&balancing);
9355 if (time_after(next_balance, sd->last_balance + interval)) {
9356 next_balance = sd->last_balance + interval;
9357 update_next_balance = 1;
9362 * Ensure the rq-wide value also decays but keep it at a
9363 * reasonable floor to avoid funnies with rq->avg_idle.
9365 rq->max_idle_balance_cost =
9366 max((u64)sysctl_sched_migration_cost, max_cost);
9371 * next_balance will be updated only when there is a need.
9372 * When the cpu is attached to null domain for ex, it will not be
9375 if (likely(update_next_balance)) {
9376 rq->next_balance = next_balance;
9378 #ifdef CONFIG_NO_HZ_COMMON
9380 * If this CPU has been elected to perform the nohz idle
9381 * balance. Other idle CPUs have already rebalanced with
9382 * nohz_idle_balance() and nohz.next_balance has been
9383 * updated accordingly. This CPU is now running the idle load
9384 * balance for itself and we need to update the
9385 * nohz.next_balance accordingly.
9387 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9388 nohz.next_balance = rq->next_balance;
9393 static inline int on_null_domain(struct rq *rq)
9395 return unlikely(!rcu_dereference_sched(rq->sd));
9398 #ifdef CONFIG_NO_HZ_COMMON
9400 * idle load balancing details
9401 * - When one of the busy CPUs notice that there may be an idle rebalancing
9402 * needed, they will kick the idle load balancer, which then does idle
9403 * load balancing for all the idle CPUs.
9404 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
9408 static inline int find_new_ilb(void)
9412 for_each_cpu_and(ilb, nohz.idle_cpus_mask,
9413 housekeeping_cpumask(HK_FLAG_MISC)) {
9422 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
9423 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
9425 static void kick_ilb(unsigned int flags)
9430 * Increase nohz.next_balance only when if full ilb is triggered but
9431 * not if we only update stats.
9433 if (flags & NOHZ_BALANCE_KICK)
9434 nohz.next_balance = jiffies+1;
9436 ilb_cpu = find_new_ilb();
9438 if (ilb_cpu >= nr_cpu_ids)
9441 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
9442 if (flags & NOHZ_KICK_MASK)
9446 * Use smp_send_reschedule() instead of resched_cpu().
9447 * This way we generate a sched IPI on the target CPU which
9448 * is idle. And the softirq performing nohz idle load balance
9449 * will be run before returning from the IPI.
9451 smp_send_reschedule(ilb_cpu);
9455 * Current decision point for kicking the idle load balancer in the presence
9456 * of idle CPUs in the system.
9458 static void nohz_balancer_kick(struct rq *rq)
9460 unsigned long now = jiffies;
9461 struct sched_domain_shared *sds;
9462 struct sched_domain *sd;
9463 int nr_busy, i, cpu = rq->cpu;
9464 unsigned int flags = 0;
9466 if (unlikely(rq->idle_balance))
9470 * We may be recently in ticked or tickless idle mode. At the first
9471 * busy tick after returning from idle, we will update the busy stats.
9473 nohz_balance_exit_idle(rq);
9476 * None are in tickless mode and hence no need for NOHZ idle load
9479 if (likely(!atomic_read(&nohz.nr_cpus)))
9482 if (READ_ONCE(nohz.has_blocked) &&
9483 time_after(now, READ_ONCE(nohz.next_blocked)))
9484 flags = NOHZ_STATS_KICK;
9486 if (time_before(now, nohz.next_balance))
9489 if (rq->nr_running >= 2) {
9490 flags = NOHZ_KICK_MASK;
9496 sd = rcu_dereference(rq->sd);
9499 * If there's a CFS task and the current CPU has reduced
9500 * capacity; kick the ILB to see if there's a better CPU to run
9503 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
9504 flags = NOHZ_KICK_MASK;
9509 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
9512 * When ASYM_PACKING; see if there's a more preferred CPU
9513 * currently idle; in which case, kick the ILB to move tasks
9516 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
9517 if (sched_asym_prefer(i, cpu)) {
9518 flags = NOHZ_KICK_MASK;
9524 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
9527 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
9528 * to run the misfit task on.
9530 if (check_misfit_status(rq, sd)) {
9531 flags = NOHZ_KICK_MASK;
9536 * For asymmetric systems, we do not want to nicely balance
9537 * cache use, instead we want to embrace asymmetry and only
9538 * ensure tasks have enough CPU capacity.
9540 * Skip the LLC logic because it's not relevant in that case.
9545 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9548 * If there is an imbalance between LLC domains (IOW we could
9549 * increase the overall cache use), we need some less-loaded LLC
9550 * domain to pull some load. Likewise, we may need to spread
9551 * load within the current LLC domain (e.g. packed SMT cores but
9552 * other CPUs are idle). We can't really know from here how busy
9553 * the others are - so just get a nohz balance going if it looks
9554 * like this LLC domain has tasks we could move.
9556 nr_busy = atomic_read(&sds->nr_busy_cpus);
9558 flags = NOHZ_KICK_MASK;
9569 static void set_cpu_sd_state_busy(int cpu)
9571 struct sched_domain *sd;
9574 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9576 if (!sd || !sd->nohz_idle)
9580 atomic_inc(&sd->shared->nr_busy_cpus);
9585 void nohz_balance_exit_idle(struct rq *rq)
9587 SCHED_WARN_ON(rq != this_rq());
9589 if (likely(!rq->nohz_tick_stopped))
9592 rq->nohz_tick_stopped = 0;
9593 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
9594 atomic_dec(&nohz.nr_cpus);
9596 set_cpu_sd_state_busy(rq->cpu);
9599 static void set_cpu_sd_state_idle(int cpu)
9601 struct sched_domain *sd;
9604 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9606 if (!sd || sd->nohz_idle)
9610 atomic_dec(&sd->shared->nr_busy_cpus);
9616 * This routine will record that the CPU is going idle with tick stopped.
9617 * This info will be used in performing idle load balancing in the future.
9619 void nohz_balance_enter_idle(int cpu)
9621 struct rq *rq = cpu_rq(cpu);
9623 SCHED_WARN_ON(cpu != smp_processor_id());
9625 /* If this CPU is going down, then nothing needs to be done: */
9626 if (!cpu_active(cpu))
9629 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9630 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9634 * Can be set safely without rq->lock held
9635 * If a clear happens, it will have evaluated last additions because
9636 * rq->lock is held during the check and the clear
9638 rq->has_blocked_load = 1;
9641 * The tick is still stopped but load could have been added in the
9642 * meantime. We set the nohz.has_blocked flag to trig a check of the
9643 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
9644 * of nohz.has_blocked can only happen after checking the new load
9646 if (rq->nohz_tick_stopped)
9649 /* If we're a completely isolated CPU, we don't play: */
9650 if (on_null_domain(rq))
9653 rq->nohz_tick_stopped = 1;
9655 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9656 atomic_inc(&nohz.nr_cpus);
9659 * Ensures that if nohz_idle_balance() fails to observe our
9660 * @idle_cpus_mask store, it must observe the @has_blocked
9663 smp_mb__after_atomic();
9665 set_cpu_sd_state_idle(cpu);
9669 * Each time a cpu enter idle, we assume that it has blocked load and
9670 * enable the periodic update of the load of idle cpus
9672 WRITE_ONCE(nohz.has_blocked, 1);
9676 * Internal function that runs load balance for all idle cpus. The load balance
9677 * can be a simple update of blocked load or a complete load balance with
9678 * tasks movement depending of flags.
9679 * The function returns false if the loop has stopped before running
9680 * through all idle CPUs.
9682 static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
9683 enum cpu_idle_type idle)
9685 /* Earliest time when we have to do rebalance again */
9686 unsigned long now = jiffies;
9687 unsigned long next_balance = now + 60*HZ;
9688 bool has_blocked_load = false;
9689 int update_next_balance = 0;
9690 int this_cpu = this_rq->cpu;
9695 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
9698 * We assume there will be no idle load after this update and clear
9699 * the has_blocked flag. If a cpu enters idle in the mean time, it will
9700 * set the has_blocked flag and trig another update of idle load.
9701 * Because a cpu that becomes idle, is added to idle_cpus_mask before
9702 * setting the flag, we are sure to not clear the state and not
9703 * check the load of an idle cpu.
9705 WRITE_ONCE(nohz.has_blocked, 0);
9708 * Ensures that if we miss the CPU, we must see the has_blocked
9709 * store from nohz_balance_enter_idle().
9713 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9714 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9718 * If this CPU gets work to do, stop the load balancing
9719 * work being done for other CPUs. Next load
9720 * balancing owner will pick it up.
9722 if (need_resched()) {
9723 has_blocked_load = true;
9727 rq = cpu_rq(balance_cpu);
9729 has_blocked_load |= update_nohz_stats(rq, true);
9732 * If time for next balance is due,
9735 if (time_after_eq(jiffies, rq->next_balance)) {
9738 rq_lock_irqsave(rq, &rf);
9739 update_rq_clock(rq);
9740 rq_unlock_irqrestore(rq, &rf);
9742 if (flags & NOHZ_BALANCE_KICK)
9743 rebalance_domains(rq, CPU_IDLE);
9746 if (time_after(next_balance, rq->next_balance)) {
9747 next_balance = rq->next_balance;
9748 update_next_balance = 1;
9753 * next_balance will be updated only when there is a need.
9754 * When the CPU is attached to null domain for ex, it will not be
9757 if (likely(update_next_balance))
9758 nohz.next_balance = next_balance;
9760 /* Newly idle CPU doesn't need an update */
9761 if (idle != CPU_NEWLY_IDLE) {
9762 update_blocked_averages(this_cpu);
9763 has_blocked_load |= this_rq->has_blocked_load;
9766 if (flags & NOHZ_BALANCE_KICK)
9767 rebalance_domains(this_rq, CPU_IDLE);
9769 WRITE_ONCE(nohz.next_blocked,
9770 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
9772 /* The full idle balance loop has been done */
9776 /* There is still blocked load, enable periodic update */
9777 if (has_blocked_load)
9778 WRITE_ONCE(nohz.has_blocked, 1);
9784 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9785 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9787 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9789 int this_cpu = this_rq->cpu;
9792 if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
9795 if (idle != CPU_IDLE) {
9796 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9800 /* could be _relaxed() */
9801 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9802 if (!(flags & NOHZ_KICK_MASK))
9805 _nohz_idle_balance(this_rq, flags, idle);
9810 static void nohz_newidle_balance(struct rq *this_rq)
9812 int this_cpu = this_rq->cpu;
9815 * This CPU doesn't want to be disturbed by scheduler
9818 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
9821 /* Will wake up very soon. No time for doing anything else*/
9822 if (this_rq->avg_idle < sysctl_sched_migration_cost)
9825 /* Don't need to update blocked load of idle CPUs*/
9826 if (!READ_ONCE(nohz.has_blocked) ||
9827 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
9830 raw_spin_unlock(&this_rq->lock);
9832 * This CPU is going to be idle and blocked load of idle CPUs
9833 * need to be updated. Run the ilb locally as it is a good
9834 * candidate for ilb instead of waking up another idle CPU.
9835 * Kick an normal ilb if we failed to do the update.
9837 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
9838 kick_ilb(NOHZ_STATS_KICK);
9839 raw_spin_lock(&this_rq->lock);
9842 #else /* !CONFIG_NO_HZ_COMMON */
9843 static inline void nohz_balancer_kick(struct rq *rq) { }
9845 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9850 static inline void nohz_newidle_balance(struct rq *this_rq) { }
9851 #endif /* CONFIG_NO_HZ_COMMON */
9854 * idle_balance is called by schedule() if this_cpu is about to become
9855 * idle. Attempts to pull tasks from other CPUs.
9857 int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
9859 unsigned long next_balance = jiffies + HZ;
9860 int this_cpu = this_rq->cpu;
9861 struct sched_domain *sd;
9862 int pulled_task = 0;
9865 update_misfit_status(NULL, this_rq);
9867 * We must set idle_stamp _before_ calling idle_balance(), such that we
9868 * measure the duration of idle_balance() as idle time.
9870 this_rq->idle_stamp = rq_clock(this_rq);
9873 * Do not pull tasks towards !active CPUs...
9875 if (!cpu_active(this_cpu))
9879 * This is OK, because current is on_cpu, which avoids it being picked
9880 * for load-balance and preemption/IRQs are still disabled avoiding
9881 * further scheduler activity on it and we're being very careful to
9882 * re-start the picking loop.
9884 rq_unpin_lock(this_rq, rf);
9886 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
9887 !READ_ONCE(this_rq->rd->overload)) {
9890 sd = rcu_dereference_check_sched_domain(this_rq->sd);
9892 update_next_balance(sd, &next_balance);
9895 nohz_newidle_balance(this_rq);
9900 raw_spin_unlock(&this_rq->lock);
9902 update_blocked_averages(this_cpu);
9904 for_each_domain(this_cpu, sd) {
9905 int continue_balancing = 1;
9906 u64 t0, domain_cost;
9908 if (!(sd->flags & SD_LOAD_BALANCE))
9911 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
9912 update_next_balance(sd, &next_balance);
9916 if (sd->flags & SD_BALANCE_NEWIDLE) {
9917 t0 = sched_clock_cpu(this_cpu);
9919 pulled_task = load_balance(this_cpu, this_rq,
9921 &continue_balancing);
9923 domain_cost = sched_clock_cpu(this_cpu) - t0;
9924 if (domain_cost > sd->max_newidle_lb_cost)
9925 sd->max_newidle_lb_cost = domain_cost;
9927 curr_cost += domain_cost;
9930 update_next_balance(sd, &next_balance);
9933 * Stop searching for tasks to pull if there are
9934 * now runnable tasks on this rq.
9936 if (pulled_task || this_rq->nr_running > 0)
9941 raw_spin_lock(&this_rq->lock);
9943 if (curr_cost > this_rq->max_idle_balance_cost)
9944 this_rq->max_idle_balance_cost = curr_cost;
9948 * While browsing the domains, we released the rq lock, a task could
9949 * have been enqueued in the meantime. Since we're not going idle,
9950 * pretend we pulled a task.
9952 if (this_rq->cfs.h_nr_running && !pulled_task)
9955 /* Move the next balance forward */
9956 if (time_after(this_rq->next_balance, next_balance))
9957 this_rq->next_balance = next_balance;
9959 /* Is there a task of a high priority class? */
9960 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
9964 this_rq->idle_stamp = 0;
9966 rq_repin_lock(this_rq, rf);
9972 * run_rebalance_domains is triggered when needed from the scheduler tick.
9973 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9975 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9977 struct rq *this_rq = this_rq();
9978 enum cpu_idle_type idle = this_rq->idle_balance ?
9979 CPU_IDLE : CPU_NOT_IDLE;
9982 * If this CPU has a pending nohz_balance_kick, then do the
9983 * balancing on behalf of the other idle CPUs whose ticks are
9984 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9985 * give the idle CPUs a chance to load balance. Else we may
9986 * load balance only within the local sched_domain hierarchy
9987 * and abort nohz_idle_balance altogether if we pull some load.
9989 if (nohz_idle_balance(this_rq, idle))
9992 /* normal load balance */
9993 update_blocked_averages(this_rq->cpu);
9994 rebalance_domains(this_rq, idle);
9998 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
10000 void trigger_load_balance(struct rq *rq)
10002 /* Don't need to rebalance while attached to NULL domain */
10003 if (unlikely(on_null_domain(rq)))
10006 if (time_after_eq(jiffies, rq->next_balance))
10007 raise_softirq(SCHED_SOFTIRQ);
10009 nohz_balancer_kick(rq);
10012 static void rq_online_fair(struct rq *rq)
10016 update_runtime_enabled(rq);
10019 static void rq_offline_fair(struct rq *rq)
10023 /* Ensure any throttled groups are reachable by pick_next_task */
10024 unthrottle_offline_cfs_rqs(rq);
10027 #endif /* CONFIG_SMP */
10030 * scheduler tick hitting a task of our scheduling class.
10032 * NOTE: This function can be called remotely by the tick offload that
10033 * goes along full dynticks. Therefore no local assumption can be made
10034 * and everything must be accessed through the @rq and @curr passed in
10037 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
10039 struct cfs_rq *cfs_rq;
10040 struct sched_entity *se = &curr->se;
10042 for_each_sched_entity(se) {
10043 cfs_rq = cfs_rq_of(se);
10044 entity_tick(cfs_rq, se, queued);
10047 if (static_branch_unlikely(&sched_numa_balancing))
10048 task_tick_numa(rq, curr);
10050 update_misfit_status(curr, rq);
10051 update_overutilized_status(task_rq(curr));
10055 * called on fork with the child task as argument from the parent's context
10056 * - child not yet on the tasklist
10057 * - preemption disabled
10059 static void task_fork_fair(struct task_struct *p)
10061 struct cfs_rq *cfs_rq;
10062 struct sched_entity *se = &p->se, *curr;
10063 struct rq *rq = this_rq();
10064 struct rq_flags rf;
10067 update_rq_clock(rq);
10069 cfs_rq = task_cfs_rq(current);
10070 curr = cfs_rq->curr;
10072 update_curr(cfs_rq);
10073 se->vruntime = curr->vruntime;
10075 place_entity(cfs_rq, se, 1);
10077 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
10079 * Upon rescheduling, sched_class::put_prev_task() will place
10080 * 'current' within the tree based on its new key value.
10082 swap(curr->vruntime, se->vruntime);
10086 se->vruntime -= cfs_rq->min_vruntime;
10087 rq_unlock(rq, &rf);
10091 * Priority of the task has changed. Check to see if we preempt
10092 * the current task.
10095 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
10097 if (!task_on_rq_queued(p))
10101 * Reschedule if we are currently running on this runqueue and
10102 * our priority decreased, or if we are not currently running on
10103 * this runqueue and our priority is higher than the current's
10105 if (rq->curr == p) {
10106 if (p->prio > oldprio)
10109 check_preempt_curr(rq, p, 0);
10112 static inline bool vruntime_normalized(struct task_struct *p)
10114 struct sched_entity *se = &p->se;
10117 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
10118 * the dequeue_entity(.flags=0) will already have normalized the
10125 * When !on_rq, vruntime of the task has usually NOT been normalized.
10126 * But there are some cases where it has already been normalized:
10128 * - A forked child which is waiting for being woken up by
10129 * wake_up_new_task().
10130 * - A task which has been woken up by try_to_wake_up() and
10131 * waiting for actually being woken up by sched_ttwu_pending().
10133 if (!se->sum_exec_runtime ||
10134 (p->state == TASK_WAKING && p->sched_remote_wakeup))
10140 #ifdef CONFIG_FAIR_GROUP_SCHED
10142 * Propagate the changes of the sched_entity across the tg tree to make it
10143 * visible to the root
10145 static void propagate_entity_cfs_rq(struct sched_entity *se)
10147 struct cfs_rq *cfs_rq;
10149 list_add_leaf_cfs_rq(cfs_rq_of(se));
10151 /* Start to propagate at parent */
10154 for_each_sched_entity(se) {
10155 cfs_rq = cfs_rq_of(se);
10157 if (!cfs_rq_throttled(cfs_rq)){
10158 update_load_avg(cfs_rq, se, UPDATE_TG);
10159 list_add_leaf_cfs_rq(cfs_rq);
10163 if (list_add_leaf_cfs_rq(cfs_rq))
10168 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
10171 static void detach_entity_cfs_rq(struct sched_entity *se)
10173 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10175 /* Catch up with the cfs_rq and remove our load when we leave */
10176 update_load_avg(cfs_rq, se, 0);
10177 detach_entity_load_avg(cfs_rq, se);
10178 update_tg_load_avg(cfs_rq, false);
10179 propagate_entity_cfs_rq(se);
10182 static void attach_entity_cfs_rq(struct sched_entity *se)
10184 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10186 #ifdef CONFIG_FAIR_GROUP_SCHED
10188 * Since the real-depth could have been changed (only FAIR
10189 * class maintain depth value), reset depth properly.
10191 se->depth = se->parent ? se->parent->depth + 1 : 0;
10194 /* Synchronize entity with its cfs_rq */
10195 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
10196 attach_entity_load_avg(cfs_rq, se, 0);
10197 update_tg_load_avg(cfs_rq, false);
10198 propagate_entity_cfs_rq(se);
10201 static void detach_task_cfs_rq(struct task_struct *p)
10203 struct sched_entity *se = &p->se;
10204 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10206 if (!vruntime_normalized(p)) {
10208 * Fix up our vruntime so that the current sleep doesn't
10209 * cause 'unlimited' sleep bonus.
10211 place_entity(cfs_rq, se, 0);
10212 se->vruntime -= cfs_rq->min_vruntime;
10215 detach_entity_cfs_rq(se);
10218 static void attach_task_cfs_rq(struct task_struct *p)
10220 struct sched_entity *se = &p->se;
10221 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10223 attach_entity_cfs_rq(se);
10225 if (!vruntime_normalized(p))
10226 se->vruntime += cfs_rq->min_vruntime;
10229 static void switched_from_fair(struct rq *rq, struct task_struct *p)
10231 detach_task_cfs_rq(p);
10234 static void switched_to_fair(struct rq *rq, struct task_struct *p)
10236 attach_task_cfs_rq(p);
10238 if (task_on_rq_queued(p)) {
10240 * We were most likely switched from sched_rt, so
10241 * kick off the schedule if running, otherwise just see
10242 * if we can still preempt the current task.
10247 check_preempt_curr(rq, p, 0);
10251 /* Account for a task changing its policy or group.
10253 * This routine is mostly called to set cfs_rq->curr field when a task
10254 * migrates between groups/classes.
10256 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
10258 struct sched_entity *se = &p->se;
10261 if (task_on_rq_queued(p)) {
10263 * Move the next running task to the front of the list, so our
10264 * cfs_tasks list becomes MRU one.
10266 list_move(&se->group_node, &rq->cfs_tasks);
10270 for_each_sched_entity(se) {
10271 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10273 set_next_entity(cfs_rq, se);
10274 /* ensure bandwidth has been allocated on our new cfs_rq */
10275 account_cfs_rq_runtime(cfs_rq, 0);
10279 void init_cfs_rq(struct cfs_rq *cfs_rq)
10281 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
10282 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
10283 #ifndef CONFIG_64BIT
10284 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
10287 raw_spin_lock_init(&cfs_rq->removed.lock);
10291 #ifdef CONFIG_FAIR_GROUP_SCHED
10292 static void task_set_group_fair(struct task_struct *p)
10294 struct sched_entity *se = &p->se;
10296 set_task_rq(p, task_cpu(p));
10297 se->depth = se->parent ? se->parent->depth + 1 : 0;
10300 static void task_move_group_fair(struct task_struct *p)
10302 detach_task_cfs_rq(p);
10303 set_task_rq(p, task_cpu(p));
10306 /* Tell se's cfs_rq has been changed -- migrated */
10307 p->se.avg.last_update_time = 0;
10309 attach_task_cfs_rq(p);
10312 static void task_change_group_fair(struct task_struct *p, int type)
10315 case TASK_SET_GROUP:
10316 task_set_group_fair(p);
10319 case TASK_MOVE_GROUP:
10320 task_move_group_fair(p);
10325 void free_fair_sched_group(struct task_group *tg)
10329 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
10331 for_each_possible_cpu(i) {
10333 kfree(tg->cfs_rq[i]);
10342 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10344 struct sched_entity *se;
10345 struct cfs_rq *cfs_rq;
10348 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
10351 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
10355 tg->shares = NICE_0_LOAD;
10357 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
10359 for_each_possible_cpu(i) {
10360 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
10361 GFP_KERNEL, cpu_to_node(i));
10365 se = kzalloc_node(sizeof(struct sched_entity),
10366 GFP_KERNEL, cpu_to_node(i));
10370 init_cfs_rq(cfs_rq);
10371 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
10372 init_entity_runnable_average(se);
10383 void online_fair_sched_group(struct task_group *tg)
10385 struct sched_entity *se;
10386 struct rq_flags rf;
10390 for_each_possible_cpu(i) {
10393 rq_lock_irq(rq, &rf);
10394 update_rq_clock(rq);
10395 attach_entity_cfs_rq(se);
10396 sync_throttle(tg, i);
10397 rq_unlock_irq(rq, &rf);
10401 void unregister_fair_sched_group(struct task_group *tg)
10403 unsigned long flags;
10407 for_each_possible_cpu(cpu) {
10409 remove_entity_load_avg(tg->se[cpu]);
10412 * Only empty task groups can be destroyed; so we can speculatively
10413 * check on_list without danger of it being re-added.
10415 if (!tg->cfs_rq[cpu]->on_list)
10420 raw_spin_lock_irqsave(&rq->lock, flags);
10421 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
10422 raw_spin_unlock_irqrestore(&rq->lock, flags);
10426 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
10427 struct sched_entity *se, int cpu,
10428 struct sched_entity *parent)
10430 struct rq *rq = cpu_rq(cpu);
10434 init_cfs_rq_runtime(cfs_rq);
10436 tg->cfs_rq[cpu] = cfs_rq;
10439 /* se could be NULL for root_task_group */
10444 se->cfs_rq = &rq->cfs;
10447 se->cfs_rq = parent->my_q;
10448 se->depth = parent->depth + 1;
10452 /* guarantee group entities always have weight */
10453 update_load_set(&se->load, NICE_0_LOAD);
10454 se->parent = parent;
10457 static DEFINE_MUTEX(shares_mutex);
10459 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10464 * We can't change the weight of the root cgroup.
10469 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10471 mutex_lock(&shares_mutex);
10472 if (tg->shares == shares)
10475 tg->shares = shares;
10476 for_each_possible_cpu(i) {
10477 struct rq *rq = cpu_rq(i);
10478 struct sched_entity *se = tg->se[i];
10479 struct rq_flags rf;
10481 /* Propagate contribution to hierarchy */
10482 rq_lock_irqsave(rq, &rf);
10483 update_rq_clock(rq);
10484 for_each_sched_entity(se) {
10485 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
10486 update_cfs_group(se);
10488 rq_unlock_irqrestore(rq, &rf);
10492 mutex_unlock(&shares_mutex);
10495 #else /* CONFIG_FAIR_GROUP_SCHED */
10497 void free_fair_sched_group(struct task_group *tg) { }
10499 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10504 void online_fair_sched_group(struct task_group *tg) { }
10506 void unregister_fair_sched_group(struct task_group *tg) { }
10508 #endif /* CONFIG_FAIR_GROUP_SCHED */
10511 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10513 struct sched_entity *se = &task->se;
10514 unsigned int rr_interval = 0;
10517 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10520 if (rq->cfs.load.weight)
10521 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10523 return rr_interval;
10527 * All the scheduling class methods:
10529 const struct sched_class fair_sched_class = {
10530 .next = &idle_sched_class,
10531 .enqueue_task = enqueue_task_fair,
10532 .dequeue_task = dequeue_task_fair,
10533 .yield_task = yield_task_fair,
10534 .yield_to_task = yield_to_task_fair,
10536 .check_preempt_curr = check_preempt_wakeup,
10538 .pick_next_task = pick_next_task_fair,
10539 .put_prev_task = put_prev_task_fair,
10540 .set_next_task = set_next_task_fair,
10543 .balance = balance_fair,
10544 .select_task_rq = select_task_rq_fair,
10545 .migrate_task_rq = migrate_task_rq_fair,
10547 .rq_online = rq_online_fair,
10548 .rq_offline = rq_offline_fair,
10550 .task_dead = task_dead_fair,
10551 .set_cpus_allowed = set_cpus_allowed_common,
10554 .task_tick = task_tick_fair,
10555 .task_fork = task_fork_fair,
10557 .prio_changed = prio_changed_fair,
10558 .switched_from = switched_from_fair,
10559 .switched_to = switched_to_fair,
10561 .get_rr_interval = get_rr_interval_fair,
10563 .update_curr = update_curr_fair,
10565 #ifdef CONFIG_FAIR_GROUP_SCHED
10566 .task_change_group = task_change_group_fair,
10569 #ifdef CONFIG_UCLAMP_TASK
10570 .uclamp_enabled = 1,
10574 #ifdef CONFIG_SCHED_DEBUG
10575 void print_cfs_stats(struct seq_file *m, int cpu)
10577 struct cfs_rq *cfs_rq, *pos;
10580 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
10581 print_cfs_rq(m, cpu, cfs_rq);
10585 #ifdef CONFIG_NUMA_BALANCING
10586 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10589 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10590 struct numa_group *ng;
10593 ng = rcu_dereference(p->numa_group);
10594 for_each_online_node(node) {
10595 if (p->numa_faults) {
10596 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10597 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10600 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
10601 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
10603 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10607 #endif /* CONFIG_NUMA_BALANCING */
10608 #endif /* CONFIG_SCHED_DEBUG */
10610 __init void init_sched_fair_class(void)
10613 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10615 #ifdef CONFIG_NO_HZ_COMMON
10616 nohz.next_balance = jiffies;
10617 nohz.next_blocked = jiffies;
10618 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
10625 * Helper functions to facilitate extracting info from tracepoints.
10628 const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq)
10631 return cfs_rq ? &cfs_rq->avg : NULL;
10636 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg);
10638 char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len)
10642 strlcpy(str, "(null)", len);
10647 cfs_rq_tg_path(cfs_rq, str, len);
10650 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path);
10652 int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq)
10654 return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1;
10656 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu);
10658 const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq)
10661 return rq ? &rq->avg_rt : NULL;
10666 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt);
10668 const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq)
10671 return rq ? &rq->avg_dl : NULL;
10676 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl);
10678 const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq)
10680 #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ)
10681 return rq ? &rq->avg_irq : NULL;
10686 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq);
10688 int sched_trace_rq_cpu(struct rq *rq)
10690 return rq ? cpu_of(rq) : -1;
10692 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu);
10694 const struct cpumask *sched_trace_rd_span(struct root_domain *rd)
10697 return rd ? rd->span : NULL;
10702 EXPORT_SYMBOL_GPL(sched_trace_rd_span);