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->mm || (curr->flags & PF_EXITING) || 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;
3508 cfs_rq_util_change(cfs_rq, 0);
3514 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3515 * @cfs_rq: cfs_rq to attach to
3516 * @se: sched_entity to attach
3517 * @flags: migration hints
3519 * Must call update_cfs_rq_load_avg() before this, since we rely on
3520 * cfs_rq->avg.last_update_time being current.
3522 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3524 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3527 * When we attach the @se to the @cfs_rq, we must align the decay
3528 * window because without that, really weird and wonderful things can
3533 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3534 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3537 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3538 * period_contrib. This isn't strictly correct, but since we're
3539 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3542 se->avg.util_sum = se->avg.util_avg * divider;
3544 se->avg.load_sum = divider;
3545 if (se_weight(se)) {
3547 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3550 se->avg.runnable_load_sum = se->avg.load_sum;
3552 enqueue_load_avg(cfs_rq, se);
3553 cfs_rq->avg.util_avg += se->avg.util_avg;
3554 cfs_rq->avg.util_sum += se->avg.util_sum;
3556 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3558 cfs_rq_util_change(cfs_rq, flags);
3560 trace_pelt_cfs_tp(cfs_rq);
3564 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3565 * @cfs_rq: cfs_rq to detach from
3566 * @se: sched_entity to detach
3568 * Must call update_cfs_rq_load_avg() before this, since we rely on
3569 * cfs_rq->avg.last_update_time being current.
3571 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3573 dequeue_load_avg(cfs_rq, se);
3574 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3575 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3577 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3579 cfs_rq_util_change(cfs_rq, 0);
3581 trace_pelt_cfs_tp(cfs_rq);
3585 * Optional action to be done while updating the load average
3587 #define UPDATE_TG 0x1
3588 #define SKIP_AGE_LOAD 0x2
3589 #define DO_ATTACH 0x4
3591 /* Update task and its cfs_rq load average */
3592 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3594 u64 now = cfs_rq_clock_pelt(cfs_rq);
3598 * Track task load average for carrying it to new CPU after migrated, and
3599 * track group sched_entity load average for task_h_load calc in migration
3601 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3602 __update_load_avg_se(now, cfs_rq, se);
3604 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3605 decayed |= propagate_entity_load_avg(se);
3607 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3610 * DO_ATTACH means we're here from enqueue_entity().
3611 * !last_update_time means we've passed through
3612 * migrate_task_rq_fair() indicating we migrated.
3614 * IOW we're enqueueing a task on a new CPU.
3616 attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
3617 update_tg_load_avg(cfs_rq, 0);
3619 } else if (decayed && (flags & UPDATE_TG))
3620 update_tg_load_avg(cfs_rq, 0);
3623 #ifndef CONFIG_64BIT
3624 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3626 u64 last_update_time_copy;
3627 u64 last_update_time;
3630 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3632 last_update_time = cfs_rq->avg.last_update_time;
3633 } while (last_update_time != last_update_time_copy);
3635 return last_update_time;
3638 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3640 return cfs_rq->avg.last_update_time;
3645 * Synchronize entity load avg of dequeued entity without locking
3648 static void sync_entity_load_avg(struct sched_entity *se)
3650 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3651 u64 last_update_time;
3653 last_update_time = cfs_rq_last_update_time(cfs_rq);
3654 __update_load_avg_blocked_se(last_update_time, se);
3658 * Task first catches up with cfs_rq, and then subtract
3659 * itself from the cfs_rq (task must be off the queue now).
3661 static void remove_entity_load_avg(struct sched_entity *se)
3663 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3664 unsigned long flags;
3667 * tasks cannot exit without having gone through wake_up_new_task() ->
3668 * post_init_entity_util_avg() which will have added things to the
3669 * cfs_rq, so we can remove unconditionally.
3672 sync_entity_load_avg(se);
3674 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3675 ++cfs_rq->removed.nr;
3676 cfs_rq->removed.util_avg += se->avg.util_avg;
3677 cfs_rq->removed.load_avg += se->avg.load_avg;
3678 cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */
3679 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3682 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3684 return cfs_rq->avg.runnable_load_avg;
3687 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3689 return cfs_rq->avg.load_avg;
3692 static inline unsigned long task_util(struct task_struct *p)
3694 return READ_ONCE(p->se.avg.util_avg);
3697 static inline unsigned long _task_util_est(struct task_struct *p)
3699 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3701 return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED);
3704 static inline unsigned long task_util_est(struct task_struct *p)
3706 return max(task_util(p), _task_util_est(p));
3709 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3710 struct task_struct *p)
3712 unsigned int enqueued;
3714 if (!sched_feat(UTIL_EST))
3717 /* Update root cfs_rq's estimated utilization */
3718 enqueued = cfs_rq->avg.util_est.enqueued;
3719 enqueued += _task_util_est(p);
3720 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3724 * Check if a (signed) value is within a specified (unsigned) margin,
3725 * based on the observation that:
3727 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3729 * NOTE: this only works when value + maring < INT_MAX.
3731 static inline bool within_margin(int value, int margin)
3733 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3737 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3739 long last_ewma_diff;
3743 if (!sched_feat(UTIL_EST))
3746 /* Update root cfs_rq's estimated utilization */
3747 ue.enqueued = cfs_rq->avg.util_est.enqueued;
3748 ue.enqueued -= min_t(unsigned int, ue.enqueued, _task_util_est(p));
3749 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3752 * Skip update of task's estimated utilization when the task has not
3753 * yet completed an activation, e.g. being migrated.
3759 * If the PELT values haven't changed since enqueue time,
3760 * skip the util_est update.
3762 ue = p->se.avg.util_est;
3763 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3767 * Skip update of task's estimated utilization when its EWMA is
3768 * already ~1% close to its last activation value.
3770 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3771 last_ewma_diff = ue.enqueued - ue.ewma;
3772 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3776 * To avoid overestimation of actual task utilization, skip updates if
3777 * we cannot grant there is idle time in this CPU.
3779 cpu = cpu_of(rq_of(cfs_rq));
3780 if (task_util(p) > capacity_orig_of(cpu))
3784 * Update Task's estimated utilization
3786 * When *p completes an activation we can consolidate another sample
3787 * of the task size. This is done by storing the current PELT value
3788 * as ue.enqueued and by using this value to update the Exponential
3789 * Weighted Moving Average (EWMA):
3791 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
3792 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
3793 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
3794 * = w * ( last_ewma_diff ) + ewma(t-1)
3795 * = w * (last_ewma_diff + ewma(t-1) / w)
3797 * Where 'w' is the weight of new samples, which is configured to be
3798 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3800 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
3801 ue.ewma += last_ewma_diff;
3802 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
3803 WRITE_ONCE(p->se.avg.util_est, ue);
3806 static inline int task_fits_capacity(struct task_struct *p, long capacity)
3808 return fits_capacity(task_util_est(p), capacity);
3811 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
3813 if (!static_branch_unlikely(&sched_asym_cpucapacity))
3817 rq->misfit_task_load = 0;
3821 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
3822 rq->misfit_task_load = 0;
3826 rq->misfit_task_load = task_h_load(p);
3829 #else /* CONFIG_SMP */
3831 #define UPDATE_TG 0x0
3832 #define SKIP_AGE_LOAD 0x0
3833 #define DO_ATTACH 0x0
3835 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3837 cfs_rq_util_change(cfs_rq, 0);
3840 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3843 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
3845 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3847 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3853 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
3856 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
3858 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
3860 #endif /* CONFIG_SMP */
3862 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3864 #ifdef CONFIG_SCHED_DEBUG
3865 s64 d = se->vruntime - cfs_rq->min_vruntime;
3870 if (d > 3*sysctl_sched_latency)
3871 schedstat_inc(cfs_rq->nr_spread_over);
3876 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3878 u64 vruntime = cfs_rq->min_vruntime;
3881 * The 'current' period is already promised to the current tasks,
3882 * however the extra weight of the new task will slow them down a
3883 * little, place the new task so that it fits in the slot that
3884 * stays open at the end.
3886 if (initial && sched_feat(START_DEBIT))
3887 vruntime += sched_vslice(cfs_rq, se);
3889 /* sleeps up to a single latency don't count. */
3891 unsigned long thresh = sysctl_sched_latency;
3894 * Halve their sleep time's effect, to allow
3895 * for a gentler effect of sleepers:
3897 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3903 /* ensure we never gain time by being placed backwards. */
3904 se->vruntime = max_vruntime(se->vruntime, vruntime);
3907 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3909 static inline void check_schedstat_required(void)
3911 #ifdef CONFIG_SCHEDSTATS
3912 if (schedstat_enabled())
3915 /* Force schedstat enabled if a dependent tracepoint is active */
3916 if (trace_sched_stat_wait_enabled() ||
3917 trace_sched_stat_sleep_enabled() ||
3918 trace_sched_stat_iowait_enabled() ||
3919 trace_sched_stat_blocked_enabled() ||
3920 trace_sched_stat_runtime_enabled()) {
3921 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3922 "stat_blocked and stat_runtime require the "
3923 "kernel parameter schedstats=enable or "
3924 "kernel.sched_schedstats=1\n");
3935 * update_min_vruntime()
3936 * vruntime -= min_vruntime
3940 * update_min_vruntime()
3941 * vruntime += min_vruntime
3943 * this way the vruntime transition between RQs is done when both
3944 * min_vruntime are up-to-date.
3948 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3949 * vruntime -= min_vruntime
3953 * update_min_vruntime()
3954 * vruntime += min_vruntime
3956 * this way we don't have the most up-to-date min_vruntime on the originating
3957 * CPU and an up-to-date min_vruntime on the destination CPU.
3961 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3963 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3964 bool curr = cfs_rq->curr == se;
3967 * If we're the current task, we must renormalise before calling
3971 se->vruntime += cfs_rq->min_vruntime;
3973 update_curr(cfs_rq);
3976 * Otherwise, renormalise after, such that we're placed at the current
3977 * moment in time, instead of some random moment in the past. Being
3978 * placed in the past could significantly boost this task to the
3979 * fairness detriment of existing tasks.
3981 if (renorm && !curr)
3982 se->vruntime += cfs_rq->min_vruntime;
3985 * When enqueuing a sched_entity, we must:
3986 * - Update loads to have both entity and cfs_rq synced with now.
3987 * - Add its load to cfs_rq->runnable_avg
3988 * - For group_entity, update its weight to reflect the new share of
3990 * - Add its new weight to cfs_rq->load.weight
3992 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
3993 update_cfs_group(se);
3994 enqueue_runnable_load_avg(cfs_rq, se);
3995 account_entity_enqueue(cfs_rq, se);
3997 if (flags & ENQUEUE_WAKEUP)
3998 place_entity(cfs_rq, se, 0);
4000 check_schedstat_required();
4001 update_stats_enqueue(cfs_rq, se, flags);
4002 check_spread(cfs_rq, se);
4004 __enqueue_entity(cfs_rq, se);
4007 if (cfs_rq->nr_running == 1) {
4008 list_add_leaf_cfs_rq(cfs_rq);
4009 check_enqueue_throttle(cfs_rq);
4013 static void __clear_buddies_last(struct sched_entity *se)
4015 for_each_sched_entity(se) {
4016 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4017 if (cfs_rq->last != se)
4020 cfs_rq->last = NULL;
4024 static void __clear_buddies_next(struct sched_entity *se)
4026 for_each_sched_entity(se) {
4027 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4028 if (cfs_rq->next != se)
4031 cfs_rq->next = NULL;
4035 static void __clear_buddies_skip(struct sched_entity *se)
4037 for_each_sched_entity(se) {
4038 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4039 if (cfs_rq->skip != se)
4042 cfs_rq->skip = NULL;
4046 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4048 if (cfs_rq->last == se)
4049 __clear_buddies_last(se);
4051 if (cfs_rq->next == se)
4052 __clear_buddies_next(se);
4054 if (cfs_rq->skip == se)
4055 __clear_buddies_skip(se);
4058 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4061 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4064 * Update run-time statistics of the 'current'.
4066 update_curr(cfs_rq);
4069 * When dequeuing a sched_entity, we must:
4070 * - Update loads to have both entity and cfs_rq synced with now.
4071 * - Subtract its load from the cfs_rq->runnable_avg.
4072 * - Subtract its previous weight from cfs_rq->load.weight.
4073 * - For group entity, update its weight to reflect the new share
4074 * of its group cfs_rq.
4076 update_load_avg(cfs_rq, se, UPDATE_TG);
4077 dequeue_runnable_load_avg(cfs_rq, se);
4079 update_stats_dequeue(cfs_rq, se, flags);
4081 clear_buddies(cfs_rq, se);
4083 if (se != cfs_rq->curr)
4084 __dequeue_entity(cfs_rq, se);
4086 account_entity_dequeue(cfs_rq, se);
4089 * Normalize after update_curr(); which will also have moved
4090 * min_vruntime if @se is the one holding it back. But before doing
4091 * update_min_vruntime() again, which will discount @se's position and
4092 * can move min_vruntime forward still more.
4094 if (!(flags & DEQUEUE_SLEEP))
4095 se->vruntime -= cfs_rq->min_vruntime;
4097 /* return excess runtime on last dequeue */
4098 return_cfs_rq_runtime(cfs_rq);
4100 update_cfs_group(se);
4103 * Now advance min_vruntime if @se was the entity holding it back,
4104 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4105 * put back on, and if we advance min_vruntime, we'll be placed back
4106 * further than we started -- ie. we'll be penalized.
4108 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4109 update_min_vruntime(cfs_rq);
4113 * Preempt the current task with a newly woken task if needed:
4116 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4118 unsigned long ideal_runtime, delta_exec;
4119 struct sched_entity *se;
4122 ideal_runtime = sched_slice(cfs_rq, curr);
4123 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4124 if (delta_exec > ideal_runtime) {
4125 resched_curr(rq_of(cfs_rq));
4127 * The current task ran long enough, ensure it doesn't get
4128 * re-elected due to buddy favours.
4130 clear_buddies(cfs_rq, curr);
4135 * Ensure that a task that missed wakeup preemption by a
4136 * narrow margin doesn't have to wait for a full slice.
4137 * This also mitigates buddy induced latencies under load.
4139 if (delta_exec < sysctl_sched_min_granularity)
4142 se = __pick_first_entity(cfs_rq);
4143 delta = curr->vruntime - se->vruntime;
4148 if (delta > ideal_runtime)
4149 resched_curr(rq_of(cfs_rq));
4153 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4155 /* 'current' is not kept within the tree. */
4158 * Any task has to be enqueued before it get to execute on
4159 * a CPU. So account for the time it spent waiting on the
4162 update_stats_wait_end(cfs_rq, se);
4163 __dequeue_entity(cfs_rq, se);
4164 update_load_avg(cfs_rq, se, UPDATE_TG);
4167 update_stats_curr_start(cfs_rq, se);
4171 * Track our maximum slice length, if the CPU's load is at
4172 * least twice that of our own weight (i.e. dont track it
4173 * when there are only lesser-weight tasks around):
4175 if (schedstat_enabled() &&
4176 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
4177 schedstat_set(se->statistics.slice_max,
4178 max((u64)schedstat_val(se->statistics.slice_max),
4179 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4182 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4186 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4189 * Pick the next process, keeping these things in mind, in this order:
4190 * 1) keep things fair between processes/task groups
4191 * 2) pick the "next" process, since someone really wants that to run
4192 * 3) pick the "last" process, for cache locality
4193 * 4) do not run the "skip" process, if something else is available
4195 static struct sched_entity *
4196 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4198 struct sched_entity *left = __pick_first_entity(cfs_rq);
4199 struct sched_entity *se;
4202 * If curr is set we have to see if its left of the leftmost entity
4203 * still in the tree, provided there was anything in the tree at all.
4205 if (!left || (curr && entity_before(curr, left)))
4208 se = left; /* ideally we run the leftmost entity */
4211 * Avoid running the skip buddy, if running something else can
4212 * be done without getting too unfair.
4214 if (cfs_rq->skip == se) {
4215 struct sched_entity *second;
4218 second = __pick_first_entity(cfs_rq);
4220 second = __pick_next_entity(se);
4221 if (!second || (curr && entity_before(curr, second)))
4225 if (second && wakeup_preempt_entity(second, left) < 1)
4230 * Prefer last buddy, try to return the CPU to a preempted task.
4232 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4236 * Someone really wants this to run. If it's not unfair, run it.
4238 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4241 clear_buddies(cfs_rq, se);
4246 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4248 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4251 * If still on the runqueue then deactivate_task()
4252 * was not called and update_curr() has to be done:
4255 update_curr(cfs_rq);
4257 /* throttle cfs_rqs exceeding runtime */
4258 check_cfs_rq_runtime(cfs_rq);
4260 check_spread(cfs_rq, prev);
4263 update_stats_wait_start(cfs_rq, prev);
4264 /* Put 'current' back into the tree. */
4265 __enqueue_entity(cfs_rq, prev);
4266 /* in !on_rq case, update occurred at dequeue */
4267 update_load_avg(cfs_rq, prev, 0);
4269 cfs_rq->curr = NULL;
4273 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4276 * Update run-time statistics of the 'current'.
4278 update_curr(cfs_rq);
4281 * Ensure that runnable average is periodically updated.
4283 update_load_avg(cfs_rq, curr, UPDATE_TG);
4284 update_cfs_group(curr);
4286 #ifdef CONFIG_SCHED_HRTICK
4288 * queued ticks are scheduled to match the slice, so don't bother
4289 * validating it and just reschedule.
4292 resched_curr(rq_of(cfs_rq));
4296 * don't let the period tick interfere with the hrtick preemption
4298 if (!sched_feat(DOUBLE_TICK) &&
4299 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4303 if (cfs_rq->nr_running > 1)
4304 check_preempt_tick(cfs_rq, curr);
4308 /**************************************************
4309 * CFS bandwidth control machinery
4312 #ifdef CONFIG_CFS_BANDWIDTH
4314 #ifdef CONFIG_JUMP_LABEL
4315 static struct static_key __cfs_bandwidth_used;
4317 static inline bool cfs_bandwidth_used(void)
4319 return static_key_false(&__cfs_bandwidth_used);
4322 void cfs_bandwidth_usage_inc(void)
4324 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4327 void cfs_bandwidth_usage_dec(void)
4329 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4331 #else /* CONFIG_JUMP_LABEL */
4332 static bool cfs_bandwidth_used(void)
4337 void cfs_bandwidth_usage_inc(void) {}
4338 void cfs_bandwidth_usage_dec(void) {}
4339 #endif /* CONFIG_JUMP_LABEL */
4342 * default period for cfs group bandwidth.
4343 * default: 0.1s, units: nanoseconds
4345 static inline u64 default_cfs_period(void)
4347 return 100000000ULL;
4350 static inline u64 sched_cfs_bandwidth_slice(void)
4352 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4356 * Replenish runtime according to assigned quota and update expiration time.
4357 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4358 * additional synchronization around rq->lock.
4360 * requires cfs_b->lock
4362 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4366 if (cfs_b->quota == RUNTIME_INF)
4369 now = sched_clock_cpu(smp_processor_id());
4370 cfs_b->runtime = cfs_b->quota;
4373 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4375 return &tg->cfs_bandwidth;
4378 /* returns 0 on failure to allocate runtime */
4379 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4381 struct task_group *tg = cfs_rq->tg;
4382 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4383 u64 amount = 0, min_amount;
4385 /* note: this is a positive sum as runtime_remaining <= 0 */
4386 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4388 raw_spin_lock(&cfs_b->lock);
4389 if (cfs_b->quota == RUNTIME_INF)
4390 amount = min_amount;
4392 start_cfs_bandwidth(cfs_b);
4394 if (cfs_b->runtime > 0) {
4395 amount = min(cfs_b->runtime, min_amount);
4396 cfs_b->runtime -= amount;
4400 raw_spin_unlock(&cfs_b->lock);
4402 cfs_rq->runtime_remaining += amount;
4404 return cfs_rq->runtime_remaining > 0;
4407 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4409 /* dock delta_exec before expiring quota (as it could span periods) */
4410 cfs_rq->runtime_remaining -= delta_exec;
4412 if (likely(cfs_rq->runtime_remaining > 0))
4415 if (cfs_rq->throttled)
4418 * if we're unable to extend our runtime we resched so that the active
4419 * hierarchy can be throttled
4421 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4422 resched_curr(rq_of(cfs_rq));
4425 static __always_inline
4426 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4428 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4431 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4434 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4436 return cfs_bandwidth_used() && cfs_rq->throttled;
4439 /* check whether cfs_rq, or any parent, is throttled */
4440 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4442 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4446 * Ensure that neither of the group entities corresponding to src_cpu or
4447 * dest_cpu are members of a throttled hierarchy when performing group
4448 * load-balance operations.
4450 static inline int throttled_lb_pair(struct task_group *tg,
4451 int src_cpu, int dest_cpu)
4453 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4455 src_cfs_rq = tg->cfs_rq[src_cpu];
4456 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4458 return throttled_hierarchy(src_cfs_rq) ||
4459 throttled_hierarchy(dest_cfs_rq);
4462 static int tg_unthrottle_up(struct task_group *tg, void *data)
4464 struct rq *rq = data;
4465 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4467 cfs_rq->throttle_count--;
4468 if (!cfs_rq->throttle_count) {
4469 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4470 cfs_rq->throttled_clock_task;
4472 /* Add cfs_rq with already running entity in the list */
4473 if (cfs_rq->nr_running >= 1)
4474 list_add_leaf_cfs_rq(cfs_rq);
4480 static int tg_throttle_down(struct task_group *tg, void *data)
4482 struct rq *rq = data;
4483 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4485 /* group is entering throttled state, stop time */
4486 if (!cfs_rq->throttle_count) {
4487 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4488 list_del_leaf_cfs_rq(cfs_rq);
4490 cfs_rq->throttle_count++;
4495 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4497 struct rq *rq = rq_of(cfs_rq);
4498 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4499 struct sched_entity *se;
4500 long task_delta, idle_task_delta, dequeue = 1;
4503 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4505 /* freeze hierarchy runnable averages while throttled */
4507 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4510 task_delta = cfs_rq->h_nr_running;
4511 idle_task_delta = cfs_rq->idle_h_nr_running;
4512 for_each_sched_entity(se) {
4513 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4514 /* throttled entity or throttle-on-deactivate */
4519 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4520 qcfs_rq->h_nr_running -= task_delta;
4521 qcfs_rq->idle_h_nr_running -= idle_task_delta;
4523 if (qcfs_rq->load.weight)
4528 sub_nr_running(rq, task_delta);
4530 cfs_rq->throttled = 1;
4531 cfs_rq->throttled_clock = rq_clock(rq);
4532 raw_spin_lock(&cfs_b->lock);
4533 empty = list_empty(&cfs_b->throttled_cfs_rq);
4536 * Add to the _head_ of the list, so that an already-started
4537 * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
4538 * not running add to the tail so that later runqueues don't get starved.
4540 if (cfs_b->distribute_running)
4541 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4543 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4546 * If we're the first throttled task, make sure the bandwidth
4550 start_cfs_bandwidth(cfs_b);
4552 raw_spin_unlock(&cfs_b->lock);
4555 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4557 struct rq *rq = rq_of(cfs_rq);
4558 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4559 struct sched_entity *se;
4561 long task_delta, idle_task_delta;
4563 se = cfs_rq->tg->se[cpu_of(rq)];
4565 cfs_rq->throttled = 0;
4567 update_rq_clock(rq);
4569 raw_spin_lock(&cfs_b->lock);
4570 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4571 list_del_rcu(&cfs_rq->throttled_list);
4572 raw_spin_unlock(&cfs_b->lock);
4574 /* update hierarchical throttle state */
4575 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4577 if (!cfs_rq->load.weight)
4580 task_delta = cfs_rq->h_nr_running;
4581 idle_task_delta = cfs_rq->idle_h_nr_running;
4582 for_each_sched_entity(se) {
4586 cfs_rq = cfs_rq_of(se);
4588 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4589 cfs_rq->h_nr_running += task_delta;
4590 cfs_rq->idle_h_nr_running += idle_task_delta;
4592 if (cfs_rq_throttled(cfs_rq))
4596 assert_list_leaf_cfs_rq(rq);
4599 add_nr_running(rq, task_delta);
4601 /* Determine whether we need to wake up potentially idle CPU: */
4602 if (rq->curr == rq->idle && rq->cfs.nr_running)
4606 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, u64 remaining)
4608 struct cfs_rq *cfs_rq;
4610 u64 starting_runtime = remaining;
4613 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4615 struct rq *rq = rq_of(cfs_rq);
4618 rq_lock_irqsave(rq, &rf);
4619 if (!cfs_rq_throttled(cfs_rq))
4622 /* By the above check, this should never be true */
4623 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
4625 runtime = -cfs_rq->runtime_remaining + 1;
4626 if (runtime > remaining)
4627 runtime = remaining;
4628 remaining -= runtime;
4630 cfs_rq->runtime_remaining += runtime;
4632 /* we check whether we're throttled above */
4633 if (cfs_rq->runtime_remaining > 0)
4634 unthrottle_cfs_rq(cfs_rq);
4637 rq_unlock_irqrestore(rq, &rf);
4644 return starting_runtime - remaining;
4648 * Responsible for refilling a task_group's bandwidth and unthrottling its
4649 * cfs_rqs as appropriate. If there has been no activity within the last
4650 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4651 * used to track this state.
4653 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
4658 /* no need to continue the timer with no bandwidth constraint */
4659 if (cfs_b->quota == RUNTIME_INF)
4660 goto out_deactivate;
4662 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4663 cfs_b->nr_periods += overrun;
4666 * idle depends on !throttled (for the case of a large deficit), and if
4667 * we're going inactive then everything else can be deferred
4669 if (cfs_b->idle && !throttled)
4670 goto out_deactivate;
4672 __refill_cfs_bandwidth_runtime(cfs_b);
4675 /* mark as potentially idle for the upcoming period */
4680 /* account preceding periods in which throttling occurred */
4681 cfs_b->nr_throttled += overrun;
4684 * This check is repeated as we are holding onto the new bandwidth while
4685 * we unthrottle. This can potentially race with an unthrottled group
4686 * trying to acquire new bandwidth from the global pool. This can result
4687 * in us over-using our runtime if it is all used during this loop, but
4688 * only by limited amounts in that extreme case.
4690 while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
4691 runtime = cfs_b->runtime;
4692 cfs_b->distribute_running = 1;
4693 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4694 /* we can't nest cfs_b->lock while distributing bandwidth */
4695 runtime = distribute_cfs_runtime(cfs_b, runtime);
4696 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4698 cfs_b->distribute_running = 0;
4699 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4701 lsub_positive(&cfs_b->runtime, runtime);
4705 * While we are ensured activity in the period following an
4706 * unthrottle, this also covers the case in which the new bandwidth is
4707 * insufficient to cover the existing bandwidth deficit. (Forcing the
4708 * timer to remain active while there are any throttled entities.)
4718 /* a cfs_rq won't donate quota below this amount */
4719 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4720 /* minimum remaining period time to redistribute slack quota */
4721 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4722 /* how long we wait to gather additional slack before distributing */
4723 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4726 * Are we near the end of the current quota period?
4728 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4729 * hrtimer base being cleared by hrtimer_start. In the case of
4730 * migrate_hrtimers, base is never cleared, so we are fine.
4732 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4734 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4737 /* if the call-back is running a quota refresh is already occurring */
4738 if (hrtimer_callback_running(refresh_timer))
4741 /* is a quota refresh about to occur? */
4742 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4743 if (remaining < min_expire)
4749 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4751 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4753 /* if there's a quota refresh soon don't bother with slack */
4754 if (runtime_refresh_within(cfs_b, min_left))
4757 /* don't push forwards an existing deferred unthrottle */
4758 if (cfs_b->slack_started)
4760 cfs_b->slack_started = true;
4762 hrtimer_start(&cfs_b->slack_timer,
4763 ns_to_ktime(cfs_bandwidth_slack_period),
4767 /* we know any runtime found here is valid as update_curr() precedes return */
4768 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4770 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4771 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4773 if (slack_runtime <= 0)
4776 raw_spin_lock(&cfs_b->lock);
4777 if (cfs_b->quota != RUNTIME_INF) {
4778 cfs_b->runtime += slack_runtime;
4780 /* we are under rq->lock, defer unthrottling using a timer */
4781 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4782 !list_empty(&cfs_b->throttled_cfs_rq))
4783 start_cfs_slack_bandwidth(cfs_b);
4785 raw_spin_unlock(&cfs_b->lock);
4787 /* even if it's not valid for return we don't want to try again */
4788 cfs_rq->runtime_remaining -= slack_runtime;
4791 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4793 if (!cfs_bandwidth_used())
4796 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4799 __return_cfs_rq_runtime(cfs_rq);
4803 * This is done with a timer (instead of inline with bandwidth return) since
4804 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4806 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4808 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4809 unsigned long flags;
4811 /* confirm we're still not at a refresh boundary */
4812 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4813 cfs_b->slack_started = false;
4814 if (cfs_b->distribute_running) {
4815 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4819 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4820 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4824 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4825 runtime = cfs_b->runtime;
4828 cfs_b->distribute_running = 1;
4830 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4835 runtime = distribute_cfs_runtime(cfs_b, runtime);
4837 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4838 lsub_positive(&cfs_b->runtime, runtime);
4839 cfs_b->distribute_running = 0;
4840 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4844 * When a group wakes up we want to make sure that its quota is not already
4845 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4846 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4848 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4850 if (!cfs_bandwidth_used())
4853 /* an active group must be handled by the update_curr()->put() path */
4854 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4857 /* ensure the group is not already throttled */
4858 if (cfs_rq_throttled(cfs_rq))
4861 /* update runtime allocation */
4862 account_cfs_rq_runtime(cfs_rq, 0);
4863 if (cfs_rq->runtime_remaining <= 0)
4864 throttle_cfs_rq(cfs_rq);
4867 static void sync_throttle(struct task_group *tg, int cpu)
4869 struct cfs_rq *pcfs_rq, *cfs_rq;
4871 if (!cfs_bandwidth_used())
4877 cfs_rq = tg->cfs_rq[cpu];
4878 pcfs_rq = tg->parent->cfs_rq[cpu];
4880 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4881 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4884 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4885 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4887 if (!cfs_bandwidth_used())
4890 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4894 * it's possible for a throttled entity to be forced into a running
4895 * state (e.g. set_curr_task), in this case we're finished.
4897 if (cfs_rq_throttled(cfs_rq))
4900 throttle_cfs_rq(cfs_rq);
4904 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4906 struct cfs_bandwidth *cfs_b =
4907 container_of(timer, struct cfs_bandwidth, slack_timer);
4909 do_sched_cfs_slack_timer(cfs_b);
4911 return HRTIMER_NORESTART;
4914 extern const u64 max_cfs_quota_period;
4916 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4918 struct cfs_bandwidth *cfs_b =
4919 container_of(timer, struct cfs_bandwidth, period_timer);
4920 unsigned long flags;
4925 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4927 overrun = hrtimer_forward_now(timer, cfs_b->period);
4932 u64 new, old = ktime_to_ns(cfs_b->period);
4934 new = (old * 147) / 128; /* ~115% */
4935 new = min(new, max_cfs_quota_period);
4937 cfs_b->period = ns_to_ktime(new);
4939 /* since max is 1s, this is limited to 1e9^2, which fits in u64 */
4940 cfs_b->quota *= new;
4941 cfs_b->quota = div64_u64(cfs_b->quota, old);
4943 pr_warn_ratelimited(
4944 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us %lld, cfs_quota_us = %lld)\n",
4946 div_u64(new, NSEC_PER_USEC),
4947 div_u64(cfs_b->quota, NSEC_PER_USEC));
4949 /* reset count so we don't come right back in here */
4953 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
4956 cfs_b->period_active = 0;
4957 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4959 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4962 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4964 raw_spin_lock_init(&cfs_b->lock);
4966 cfs_b->quota = RUNTIME_INF;
4967 cfs_b->period = ns_to_ktime(default_cfs_period());
4969 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4970 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4971 cfs_b->period_timer.function = sched_cfs_period_timer;
4972 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4973 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4974 cfs_b->distribute_running = 0;
4975 cfs_b->slack_started = false;
4978 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4980 cfs_rq->runtime_enabled = 0;
4981 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4984 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4988 lockdep_assert_held(&cfs_b->lock);
4990 if (cfs_b->period_active)
4993 cfs_b->period_active = 1;
4994 overrun = hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4995 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4998 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5000 /* init_cfs_bandwidth() was not called */
5001 if (!cfs_b->throttled_cfs_rq.next)
5004 hrtimer_cancel(&cfs_b->period_timer);
5005 hrtimer_cancel(&cfs_b->slack_timer);
5009 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5011 * The race is harmless, since modifying bandwidth settings of unhooked group
5012 * bits doesn't do much.
5015 /* cpu online calback */
5016 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5018 struct task_group *tg;
5020 lockdep_assert_held(&rq->lock);
5023 list_for_each_entry_rcu(tg, &task_groups, list) {
5024 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5025 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5027 raw_spin_lock(&cfs_b->lock);
5028 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5029 raw_spin_unlock(&cfs_b->lock);
5034 /* cpu offline callback */
5035 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5037 struct task_group *tg;
5039 lockdep_assert_held(&rq->lock);
5042 list_for_each_entry_rcu(tg, &task_groups, list) {
5043 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5045 if (!cfs_rq->runtime_enabled)
5049 * clock_task is not advancing so we just need to make sure
5050 * there's some valid quota amount
5052 cfs_rq->runtime_remaining = 1;
5054 * Offline rq is schedulable till CPU is completely disabled
5055 * in take_cpu_down(), so we prevent new cfs throttling here.
5057 cfs_rq->runtime_enabled = 0;
5059 if (cfs_rq_throttled(cfs_rq))
5060 unthrottle_cfs_rq(cfs_rq);
5065 #else /* CONFIG_CFS_BANDWIDTH */
5067 static inline bool cfs_bandwidth_used(void)
5072 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5073 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5074 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5075 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5076 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5078 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5083 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5088 static inline int throttled_lb_pair(struct task_group *tg,
5089 int src_cpu, int dest_cpu)
5094 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5096 #ifdef CONFIG_FAIR_GROUP_SCHED
5097 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5100 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5104 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5105 static inline void update_runtime_enabled(struct rq *rq) {}
5106 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5108 #endif /* CONFIG_CFS_BANDWIDTH */
5110 /**************************************************
5111 * CFS operations on tasks:
5114 #ifdef CONFIG_SCHED_HRTICK
5115 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5117 struct sched_entity *se = &p->se;
5118 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5120 SCHED_WARN_ON(task_rq(p) != rq);
5122 if (rq->cfs.h_nr_running > 1) {
5123 u64 slice = sched_slice(cfs_rq, se);
5124 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5125 s64 delta = slice - ran;
5132 hrtick_start(rq, delta);
5137 * called from enqueue/dequeue and updates the hrtick when the
5138 * current task is from our class and nr_running is low enough
5141 static void hrtick_update(struct rq *rq)
5143 struct task_struct *curr = rq->curr;
5145 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5148 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5149 hrtick_start_fair(rq, curr);
5151 #else /* !CONFIG_SCHED_HRTICK */
5153 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5157 static inline void hrtick_update(struct rq *rq)
5163 static inline unsigned long cpu_util(int cpu);
5165 static inline bool cpu_overutilized(int cpu)
5167 return !fits_capacity(cpu_util(cpu), capacity_of(cpu));
5170 static inline void update_overutilized_status(struct rq *rq)
5172 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
5173 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5174 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
5178 static inline void update_overutilized_status(struct rq *rq) { }
5182 * The enqueue_task method is called before nr_running is
5183 * increased. Here we update the fair scheduling stats and
5184 * then put the task into the rbtree:
5187 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5189 struct cfs_rq *cfs_rq;
5190 struct sched_entity *se = &p->se;
5191 int idle_h_nr_running = task_has_idle_policy(p);
5194 * The code below (indirectly) updates schedutil which looks at
5195 * the cfs_rq utilization to select a frequency.
5196 * Let's add the task's estimated utilization to the cfs_rq's
5197 * estimated utilization, before we update schedutil.
5199 util_est_enqueue(&rq->cfs, p);
5202 * If in_iowait is set, the code below may not trigger any cpufreq
5203 * utilization updates, so do it here explicitly with the IOWAIT flag
5207 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5209 for_each_sched_entity(se) {
5212 cfs_rq = cfs_rq_of(se);
5213 enqueue_entity(cfs_rq, se, flags);
5216 * end evaluation on encountering a throttled cfs_rq
5218 * note: in the case of encountering a throttled cfs_rq we will
5219 * post the final h_nr_running increment below.
5221 if (cfs_rq_throttled(cfs_rq))
5223 cfs_rq->h_nr_running++;
5224 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5226 flags = ENQUEUE_WAKEUP;
5229 for_each_sched_entity(se) {
5230 cfs_rq = cfs_rq_of(se);
5231 cfs_rq->h_nr_running++;
5232 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5234 if (cfs_rq_throttled(cfs_rq))
5237 update_load_avg(cfs_rq, se, UPDATE_TG);
5238 update_cfs_group(se);
5242 add_nr_running(rq, 1);
5244 * Since new tasks are assigned an initial util_avg equal to
5245 * half of the spare capacity of their CPU, tiny tasks have the
5246 * ability to cross the overutilized threshold, which will
5247 * result in the load balancer ruining all the task placement
5248 * done by EAS. As a way to mitigate that effect, do not account
5249 * for the first enqueue operation of new tasks during the
5250 * overutilized flag detection.
5252 * A better way of solving this problem would be to wait for
5253 * the PELT signals of tasks to converge before taking them
5254 * into account, but that is not straightforward to implement,
5255 * and the following generally works well enough in practice.
5257 if (flags & ENQUEUE_WAKEUP)
5258 update_overutilized_status(rq);
5262 if (cfs_bandwidth_used()) {
5264 * When bandwidth control is enabled; the cfs_rq_throttled()
5265 * breaks in the above iteration can result in incomplete
5266 * leaf list maintenance, resulting in triggering the assertion
5269 for_each_sched_entity(se) {
5270 cfs_rq = cfs_rq_of(se);
5272 if (list_add_leaf_cfs_rq(cfs_rq))
5277 assert_list_leaf_cfs_rq(rq);
5282 static void set_next_buddy(struct sched_entity *se);
5285 * The dequeue_task method is called before nr_running is
5286 * decreased. We remove the task from the rbtree and
5287 * update the fair scheduling stats:
5289 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5291 struct cfs_rq *cfs_rq;
5292 struct sched_entity *se = &p->se;
5293 int task_sleep = flags & DEQUEUE_SLEEP;
5294 int idle_h_nr_running = task_has_idle_policy(p);
5296 for_each_sched_entity(se) {
5297 cfs_rq = cfs_rq_of(se);
5298 dequeue_entity(cfs_rq, se, flags);
5301 * end evaluation on encountering a throttled cfs_rq
5303 * note: in the case of encountering a throttled cfs_rq we will
5304 * post the final h_nr_running decrement below.
5306 if (cfs_rq_throttled(cfs_rq))
5308 cfs_rq->h_nr_running--;
5309 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5311 /* Don't dequeue parent if it has other entities besides us */
5312 if (cfs_rq->load.weight) {
5313 /* Avoid re-evaluating load for this entity: */
5314 se = parent_entity(se);
5316 * Bias pick_next to pick a task from this cfs_rq, as
5317 * p is sleeping when it is within its sched_slice.
5319 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5323 flags |= DEQUEUE_SLEEP;
5326 for_each_sched_entity(se) {
5327 cfs_rq = cfs_rq_of(se);
5328 cfs_rq->h_nr_running--;
5329 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5331 if (cfs_rq_throttled(cfs_rq))
5334 update_load_avg(cfs_rq, se, UPDATE_TG);
5335 update_cfs_group(se);
5339 sub_nr_running(rq, 1);
5341 util_est_dequeue(&rq->cfs, p, task_sleep);
5347 /* Working cpumask for: load_balance, load_balance_newidle. */
5348 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5349 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5351 #ifdef CONFIG_NO_HZ_COMMON
5354 cpumask_var_t idle_cpus_mask;
5356 int has_blocked; /* Idle CPUS has blocked load */
5357 unsigned long next_balance; /* in jiffy units */
5358 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5359 } nohz ____cacheline_aligned;
5361 #endif /* CONFIG_NO_HZ_COMMON */
5363 /* CPU only has SCHED_IDLE tasks enqueued */
5364 static int sched_idle_cpu(int cpu)
5366 struct rq *rq = cpu_rq(cpu);
5368 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
5372 static unsigned long cpu_runnable_load(struct rq *rq)
5374 return cfs_rq_runnable_load_avg(&rq->cfs);
5377 static unsigned long capacity_of(int cpu)
5379 return cpu_rq(cpu)->cpu_capacity;
5382 static unsigned long cpu_avg_load_per_task(int cpu)
5384 struct rq *rq = cpu_rq(cpu);
5385 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5386 unsigned long load_avg = cpu_runnable_load(rq);
5389 return load_avg / nr_running;
5394 static void record_wakee(struct task_struct *p)
5397 * Only decay a single time; tasks that have less then 1 wakeup per
5398 * jiffy will not have built up many flips.
5400 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5401 current->wakee_flips >>= 1;
5402 current->wakee_flip_decay_ts = jiffies;
5405 if (current->last_wakee != p) {
5406 current->last_wakee = p;
5407 current->wakee_flips++;
5412 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5414 * A waker of many should wake a different task than the one last awakened
5415 * at a frequency roughly N times higher than one of its wakees.
5417 * In order to determine whether we should let the load spread vs consolidating
5418 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5419 * partner, and a factor of lls_size higher frequency in the other.
5421 * With both conditions met, we can be relatively sure that the relationship is
5422 * non-monogamous, with partner count exceeding socket size.
5424 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5425 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5428 static int wake_wide(struct task_struct *p)
5430 unsigned int master = current->wakee_flips;
5431 unsigned int slave = p->wakee_flips;
5432 int factor = this_cpu_read(sd_llc_size);
5435 swap(master, slave);
5436 if (slave < factor || master < slave * factor)
5442 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5443 * soonest. For the purpose of speed we only consider the waking and previous
5446 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5447 * cache-affine and is (or will be) idle.
5449 * wake_affine_weight() - considers the weight to reflect the average
5450 * scheduling latency of the CPUs. This seems to work
5451 * for the overloaded case.
5454 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5457 * If this_cpu is idle, it implies the wakeup is from interrupt
5458 * context. Only allow the move if cache is shared. Otherwise an
5459 * interrupt intensive workload could force all tasks onto one
5460 * node depending on the IO topology or IRQ affinity settings.
5462 * If the prev_cpu is idle and cache affine then avoid a migration.
5463 * There is no guarantee that the cache hot data from an interrupt
5464 * is more important than cache hot data on the prev_cpu and from
5465 * a cpufreq perspective, it's better to have higher utilisation
5468 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5469 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5471 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5474 return nr_cpumask_bits;
5478 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5479 int this_cpu, int prev_cpu, int sync)
5481 s64 this_eff_load, prev_eff_load;
5482 unsigned long task_load;
5484 this_eff_load = cpu_runnable_load(cpu_rq(this_cpu));
5487 unsigned long current_load = task_h_load(current);
5489 if (current_load > this_eff_load)
5492 this_eff_load -= current_load;
5495 task_load = task_h_load(p);
5497 this_eff_load += task_load;
5498 if (sched_feat(WA_BIAS))
5499 this_eff_load *= 100;
5500 this_eff_load *= capacity_of(prev_cpu);
5502 prev_eff_load = cpu_runnable_load(cpu_rq(prev_cpu));
5503 prev_eff_load -= task_load;
5504 if (sched_feat(WA_BIAS))
5505 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5506 prev_eff_load *= capacity_of(this_cpu);
5509 * If sync, adjust the weight of prev_eff_load such that if
5510 * prev_eff == this_eff that select_idle_sibling() will consider
5511 * stacking the wakee on top of the waker if no other CPU is
5517 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5520 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5521 int this_cpu, int prev_cpu, int sync)
5523 int target = nr_cpumask_bits;
5525 if (sched_feat(WA_IDLE))
5526 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5528 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5529 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5531 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5532 if (target == nr_cpumask_bits)
5535 schedstat_inc(sd->ttwu_move_affine);
5536 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5540 static unsigned long cpu_util_without(int cpu, struct task_struct *p);
5542 static unsigned long capacity_spare_without(int cpu, struct task_struct *p)
5544 return max_t(long, capacity_of(cpu) - cpu_util_without(cpu, p), 0);
5548 * find_idlest_group finds and returns the least busy CPU group within the
5551 * Assumes p is allowed on at least one CPU in sd.
5553 static struct sched_group *
5554 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5555 int this_cpu, int sd_flag)
5557 struct sched_group *idlest = NULL, *group = sd->groups;
5558 struct sched_group *most_spare_sg = NULL;
5559 unsigned long min_runnable_load = ULONG_MAX;
5560 unsigned long this_runnable_load = ULONG_MAX;
5561 unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5562 unsigned long most_spare = 0, this_spare = 0;
5563 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5564 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5565 (sd->imbalance_pct-100) / 100;
5568 unsigned long load, avg_load, runnable_load;
5569 unsigned long spare_cap, max_spare_cap;
5573 /* Skip over this group if it has no CPUs allowed */
5574 if (!cpumask_intersects(sched_group_span(group),
5578 local_group = cpumask_test_cpu(this_cpu,
5579 sched_group_span(group));
5582 * Tally up the load of all CPUs in the group and find
5583 * the group containing the CPU with most spare capacity.
5589 for_each_cpu(i, sched_group_span(group)) {
5590 load = cpu_runnable_load(cpu_rq(i));
5591 runnable_load += load;
5593 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5595 spare_cap = capacity_spare_without(i, p);
5597 if (spare_cap > max_spare_cap)
5598 max_spare_cap = spare_cap;
5601 /* Adjust by relative CPU capacity of the group */
5602 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5603 group->sgc->capacity;
5604 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5605 group->sgc->capacity;
5608 this_runnable_load = runnable_load;
5609 this_avg_load = avg_load;
5610 this_spare = max_spare_cap;
5612 if (min_runnable_load > (runnable_load + imbalance)) {
5614 * The runnable load is significantly smaller
5615 * so we can pick this new CPU:
5617 min_runnable_load = runnable_load;
5618 min_avg_load = avg_load;
5620 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5621 (100*min_avg_load > imbalance_scale*avg_load)) {
5623 * The runnable loads are close so take the
5624 * blocked load into account through avg_load:
5626 min_avg_load = avg_load;
5630 if (most_spare < max_spare_cap) {
5631 most_spare = max_spare_cap;
5632 most_spare_sg = group;
5635 } while (group = group->next, group != sd->groups);
5638 * The cross-over point between using spare capacity or least load
5639 * is too conservative for high utilization tasks on partially
5640 * utilized systems if we require spare_capacity > task_util(p),
5641 * so we allow for some task stuffing by using
5642 * spare_capacity > task_util(p)/2.
5644 * Spare capacity can't be used for fork because the utilization has
5645 * not been set yet, we must first select a rq to compute the initial
5648 if (sd_flag & SD_BALANCE_FORK)
5651 if (this_spare > task_util(p) / 2 &&
5652 imbalance_scale*this_spare > 100*most_spare)
5655 if (most_spare > task_util(p) / 2)
5656 return most_spare_sg;
5663 * When comparing groups across NUMA domains, it's possible for the
5664 * local domain to be very lightly loaded relative to the remote
5665 * domains but "imbalance" skews the comparison making remote CPUs
5666 * look much more favourable. When considering cross-domain, add
5667 * imbalance to the runnable load on the remote node and consider
5670 if ((sd->flags & SD_NUMA) &&
5671 min_runnable_load + imbalance >= this_runnable_load)
5674 if (min_runnable_load > (this_runnable_load + imbalance))
5677 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5678 (100*this_avg_load < imbalance_scale*min_avg_load))
5685 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5688 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5690 unsigned long load, min_load = ULONG_MAX;
5691 unsigned int min_exit_latency = UINT_MAX;
5692 u64 latest_idle_timestamp = 0;
5693 int least_loaded_cpu = this_cpu;
5694 int shallowest_idle_cpu = -1, si_cpu = -1;
5697 /* Check if we have any choice: */
5698 if (group->group_weight == 1)
5699 return cpumask_first(sched_group_span(group));
5701 /* Traverse only the allowed CPUs */
5702 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
5703 if (available_idle_cpu(i)) {
5704 struct rq *rq = cpu_rq(i);
5705 struct cpuidle_state *idle = idle_get_state(rq);
5706 if (idle && idle->exit_latency < min_exit_latency) {
5708 * We give priority to a CPU whose idle state
5709 * has the smallest exit latency irrespective
5710 * of any idle timestamp.
5712 min_exit_latency = idle->exit_latency;
5713 latest_idle_timestamp = rq->idle_stamp;
5714 shallowest_idle_cpu = i;
5715 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5716 rq->idle_stamp > latest_idle_timestamp) {
5718 * If equal or no active idle state, then
5719 * the most recently idled CPU might have
5722 latest_idle_timestamp = rq->idle_stamp;
5723 shallowest_idle_cpu = i;
5725 } else if (shallowest_idle_cpu == -1 && si_cpu == -1) {
5726 if (sched_idle_cpu(i)) {
5731 load = cpu_runnable_load(cpu_rq(i));
5732 if (load < min_load) {
5734 least_loaded_cpu = i;
5739 if (shallowest_idle_cpu != -1)
5740 return shallowest_idle_cpu;
5743 return least_loaded_cpu;
5746 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5747 int cpu, int prev_cpu, int sd_flag)
5751 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
5755 * We need task's util for capacity_spare_without, sync it up to
5756 * prev_cpu's last_update_time.
5758 if (!(sd_flag & SD_BALANCE_FORK))
5759 sync_entity_load_avg(&p->se);
5762 struct sched_group *group;
5763 struct sched_domain *tmp;
5766 if (!(sd->flags & sd_flag)) {
5771 group = find_idlest_group(sd, p, cpu, sd_flag);
5777 new_cpu = find_idlest_group_cpu(group, p, cpu);
5778 if (new_cpu == cpu) {
5779 /* Now try balancing at a lower domain level of 'cpu': */
5784 /* Now try balancing at a lower domain level of 'new_cpu': */
5786 weight = sd->span_weight;
5788 for_each_domain(cpu, tmp) {
5789 if (weight <= tmp->span_weight)
5791 if (tmp->flags & sd_flag)
5799 #ifdef CONFIG_SCHED_SMT
5800 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5801 EXPORT_SYMBOL_GPL(sched_smt_present);
5803 static inline void set_idle_cores(int cpu, int val)
5805 struct sched_domain_shared *sds;
5807 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5809 WRITE_ONCE(sds->has_idle_cores, val);
5812 static inline bool test_idle_cores(int cpu, bool def)
5814 struct sched_domain_shared *sds;
5816 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5818 return READ_ONCE(sds->has_idle_cores);
5824 * Scans the local SMT mask to see if the entire core is idle, and records this
5825 * information in sd_llc_shared->has_idle_cores.
5827 * Since SMT siblings share all cache levels, inspecting this limited remote
5828 * state should be fairly cheap.
5830 void __update_idle_core(struct rq *rq)
5832 int core = cpu_of(rq);
5836 if (test_idle_cores(core, true))
5839 for_each_cpu(cpu, cpu_smt_mask(core)) {
5843 if (!available_idle_cpu(cpu))
5847 set_idle_cores(core, 1);
5853 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5854 * there are no idle cores left in the system; tracked through
5855 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5857 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5859 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5862 if (!static_branch_likely(&sched_smt_present))
5865 if (!test_idle_cores(target, false))
5868 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
5870 for_each_cpu_wrap(core, cpus, target) {
5873 for_each_cpu(cpu, cpu_smt_mask(core)) {
5874 __cpumask_clear_cpu(cpu, cpus);
5875 if (!available_idle_cpu(cpu))
5884 * Failed to find an idle core; stop looking for one.
5886 set_idle_cores(target, 0);
5892 * Scan the local SMT mask for idle CPUs.
5894 static int select_idle_smt(struct task_struct *p, int target)
5896 int cpu, si_cpu = -1;
5898 if (!static_branch_likely(&sched_smt_present))
5901 for_each_cpu(cpu, cpu_smt_mask(target)) {
5902 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
5904 if (available_idle_cpu(cpu))
5906 if (si_cpu == -1 && sched_idle_cpu(cpu))
5913 #else /* CONFIG_SCHED_SMT */
5915 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5920 static inline int select_idle_smt(struct task_struct *p, int target)
5925 #endif /* CONFIG_SCHED_SMT */
5928 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5929 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5930 * average idle time for this rq (as found in rq->avg_idle).
5932 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
5934 struct sched_domain *this_sd;
5935 u64 avg_cost, avg_idle;
5938 int this = smp_processor_id();
5939 int cpu, nr = INT_MAX, si_cpu = -1;
5941 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
5946 * Due to large variance we need a large fuzz factor; hackbench in
5947 * particularly is sensitive here.
5949 avg_idle = this_rq()->avg_idle / 512;
5950 avg_cost = this_sd->avg_scan_cost + 1;
5952 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
5955 if (sched_feat(SIS_PROP)) {
5956 u64 span_avg = sd->span_weight * avg_idle;
5957 if (span_avg > 4*avg_cost)
5958 nr = div_u64(span_avg, avg_cost);
5963 time = cpu_clock(this);
5965 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
5968 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
5970 if (available_idle_cpu(cpu))
5972 if (si_cpu == -1 && sched_idle_cpu(cpu))
5976 time = cpu_clock(this) - time;
5977 cost = this_sd->avg_scan_cost;
5978 delta = (s64)(time - cost) / 8;
5979 this_sd->avg_scan_cost += delta;
5985 * Try and locate an idle core/thread in the LLC cache domain.
5987 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5989 struct sched_domain *sd;
5990 int i, recent_used_cpu;
5992 if (available_idle_cpu(target) || sched_idle_cpu(target))
5996 * If the previous CPU is cache affine and idle, don't be stupid:
5998 if (prev != target && cpus_share_cache(prev, target) &&
5999 (available_idle_cpu(prev) || sched_idle_cpu(prev)))
6002 /* Check a recently used CPU as a potential idle candidate: */
6003 recent_used_cpu = p->recent_used_cpu;
6004 if (recent_used_cpu != prev &&
6005 recent_used_cpu != target &&
6006 cpus_share_cache(recent_used_cpu, target) &&
6007 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
6008 cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr)) {
6010 * Replace recent_used_cpu with prev as it is a potential
6011 * candidate for the next wake:
6013 p->recent_used_cpu = prev;
6014 return recent_used_cpu;
6017 sd = rcu_dereference(per_cpu(sd_llc, target));
6021 i = select_idle_core(p, sd, target);
6022 if ((unsigned)i < nr_cpumask_bits)
6025 i = select_idle_cpu(p, sd, target);
6026 if ((unsigned)i < nr_cpumask_bits)
6029 i = select_idle_smt(p, target);
6030 if ((unsigned)i < nr_cpumask_bits)
6037 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6038 * @cpu: the CPU to get the utilization of
6040 * The unit of the return value must be the one of capacity so we can compare
6041 * the utilization with the capacity of the CPU that is available for CFS task
6042 * (ie cpu_capacity).
6044 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6045 * recent utilization of currently non-runnable tasks on a CPU. It represents
6046 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6047 * capacity_orig is the cpu_capacity available at the highest frequency
6048 * (arch_scale_freq_capacity()).
6049 * The utilization of a CPU converges towards a sum equal to or less than the
6050 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6051 * the running time on this CPU scaled by capacity_curr.
6053 * The estimated utilization of a CPU is defined to be the maximum between its
6054 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6055 * currently RUNNABLE on that CPU.
6056 * This allows to properly represent the expected utilization of a CPU which
6057 * has just got a big task running since a long sleep period. At the same time
6058 * however it preserves the benefits of the "blocked utilization" in
6059 * describing the potential for other tasks waking up on the same CPU.
6061 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6062 * higher than capacity_orig because of unfortunate rounding in
6063 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6064 * the average stabilizes with the new running time. We need to check that the
6065 * utilization stays within the range of [0..capacity_orig] and cap it if
6066 * necessary. Without utilization capping, a group could be seen as overloaded
6067 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6068 * available capacity. We allow utilization to overshoot capacity_curr (but not
6069 * capacity_orig) as it useful for predicting the capacity required after task
6070 * migrations (scheduler-driven DVFS).
6072 * Return: the (estimated) utilization for the specified CPU
6074 static inline unsigned long cpu_util(int cpu)
6076 struct cfs_rq *cfs_rq;
6079 cfs_rq = &cpu_rq(cpu)->cfs;
6080 util = READ_ONCE(cfs_rq->avg.util_avg);
6082 if (sched_feat(UTIL_EST))
6083 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6085 return min_t(unsigned long, util, capacity_orig_of(cpu));
6089 * cpu_util_without: compute cpu utilization without any contributions from *p
6090 * @cpu: the CPU which utilization is requested
6091 * @p: the task which utilization should be discounted
6093 * The utilization of a CPU is defined by the utilization of tasks currently
6094 * enqueued on that CPU as well as tasks which are currently sleeping after an
6095 * execution on that CPU.
6097 * This method returns the utilization of the specified CPU by discounting the
6098 * utilization of the specified task, whenever the task is currently
6099 * contributing to the CPU utilization.
6101 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6103 struct cfs_rq *cfs_rq;
6106 /* Task has no contribution or is new */
6107 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6108 return cpu_util(cpu);
6110 cfs_rq = &cpu_rq(cpu)->cfs;
6111 util = READ_ONCE(cfs_rq->avg.util_avg);
6113 /* Discount task's util from CPU's util */
6114 lsub_positive(&util, task_util(p));
6119 * a) if *p is the only task sleeping on this CPU, then:
6120 * cpu_util (== task_util) > util_est (== 0)
6121 * and thus we return:
6122 * cpu_util_without = (cpu_util - task_util) = 0
6124 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6126 * cpu_util >= task_util
6127 * cpu_util > util_est (== 0)
6128 * and thus we discount *p's blocked utilization to return:
6129 * cpu_util_without = (cpu_util - task_util) >= 0
6131 * c) if other tasks are RUNNABLE on that CPU and
6132 * util_est > cpu_util
6133 * then we use util_est since it returns a more restrictive
6134 * estimation of the spare capacity on that CPU, by just
6135 * considering the expected utilization of tasks already
6136 * runnable on that CPU.
6138 * Cases a) and b) are covered by the above code, while case c) is
6139 * covered by the following code when estimated utilization is
6142 if (sched_feat(UTIL_EST)) {
6143 unsigned int estimated =
6144 READ_ONCE(cfs_rq->avg.util_est.enqueued);
6147 * Despite the following checks we still have a small window
6148 * for a possible race, when an execl's select_task_rq_fair()
6149 * races with LB's detach_task():
6152 * p->on_rq = TASK_ON_RQ_MIGRATING;
6153 * ---------------------------------- A
6154 * deactivate_task() \
6155 * dequeue_task() + RaceTime
6156 * util_est_dequeue() /
6157 * ---------------------------------- B
6159 * The additional check on "current == p" it's required to
6160 * properly fix the execl regression and it helps in further
6161 * reducing the chances for the above race.
6163 if (unlikely(task_on_rq_queued(p) || current == p))
6164 lsub_positive(&estimated, _task_util_est(p));
6166 util = max(util, estimated);
6170 * Utilization (estimated) can exceed the CPU capacity, thus let's
6171 * clamp to the maximum CPU capacity to ensure consistency with
6172 * the cpu_util call.
6174 return min_t(unsigned long, util, capacity_orig_of(cpu));
6178 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6179 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6181 * In that case WAKE_AFFINE doesn't make sense and we'll let
6182 * BALANCE_WAKE sort things out.
6184 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6186 long min_cap, max_cap;
6188 if (!static_branch_unlikely(&sched_asym_cpucapacity))
6191 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6192 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6194 /* Minimum capacity is close to max, no need to abort wake_affine */
6195 if (max_cap - min_cap < max_cap >> 3)
6198 /* Bring task utilization in sync with prev_cpu */
6199 sync_entity_load_avg(&p->se);
6201 return !task_fits_capacity(p, min_cap);
6205 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6208 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6210 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6211 unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
6214 * If @p migrates from @cpu to another, remove its contribution. Or,
6215 * if @p migrates from another CPU to @cpu, add its contribution. In
6216 * the other cases, @cpu is not impacted by the migration, so the
6217 * util_avg should already be correct.
6219 if (task_cpu(p) == cpu && dst_cpu != cpu)
6220 sub_positive(&util, task_util(p));
6221 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6222 util += task_util(p);
6224 if (sched_feat(UTIL_EST)) {
6225 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6228 * During wake-up, the task isn't enqueued yet and doesn't
6229 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6230 * so just add it (if needed) to "simulate" what will be
6231 * cpu_util() after the task has been enqueued.
6234 util_est += _task_util_est(p);
6236 util = max(util, util_est);
6239 return min(util, capacity_orig_of(cpu));
6243 * compute_energy(): Estimates the energy that @pd would consume if @p was
6244 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6245 * landscape of @pd's CPUs after the task migration, and uses the Energy Model
6246 * to compute what would be the energy if we decided to actually migrate that
6250 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
6252 struct cpumask *pd_mask = perf_domain_span(pd);
6253 unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask));
6254 unsigned long max_util = 0, sum_util = 0;
6258 * The capacity state of CPUs of the current rd can be driven by CPUs
6259 * of another rd if they belong to the same pd. So, account for the
6260 * utilization of these CPUs too by masking pd with cpu_online_mask
6261 * instead of the rd span.
6263 * If an entire pd is outside of the current rd, it will not appear in
6264 * its pd list and will not be accounted by compute_energy().
6266 for_each_cpu_and(cpu, pd_mask, cpu_online_mask) {
6267 unsigned long cpu_util, util_cfs = cpu_util_next(cpu, p, dst_cpu);
6268 struct task_struct *tsk = cpu == dst_cpu ? p : NULL;
6271 * Busy time computation: utilization clamping is not
6272 * required since the ratio (sum_util / cpu_capacity)
6273 * is already enough to scale the EM reported power
6274 * consumption at the (eventually clamped) cpu_capacity.
6276 sum_util += schedutil_cpu_util(cpu, util_cfs, cpu_cap,
6280 * Performance domain frequency: utilization clamping
6281 * must be considered since it affects the selection
6282 * of the performance domain frequency.
6283 * NOTE: in case RT tasks are running, by default the
6284 * FREQUENCY_UTIL's utilization can be max OPP.
6286 cpu_util = schedutil_cpu_util(cpu, util_cfs, cpu_cap,
6287 FREQUENCY_UTIL, tsk);
6288 max_util = max(max_util, cpu_util);
6291 return em_pd_energy(pd->em_pd, max_util, sum_util);
6295 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6296 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6297 * spare capacity in each performance domain and uses it as a potential
6298 * candidate to execute the task. Then, it uses the Energy Model to figure
6299 * out which of the CPU candidates is the most energy-efficient.
6301 * The rationale for this heuristic is as follows. In a performance domain,
6302 * all the most energy efficient CPU candidates (according to the Energy
6303 * Model) are those for which we'll request a low frequency. When there are
6304 * several CPUs for which the frequency request will be the same, we don't
6305 * have enough data to break the tie between them, because the Energy Model
6306 * only includes active power costs. With this model, if we assume that
6307 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6308 * the maximum spare capacity in a performance domain is guaranteed to be among
6309 * the best candidates of the performance domain.
6311 * In practice, it could be preferable from an energy standpoint to pack
6312 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6313 * but that could also hurt our chances to go cluster idle, and we have no
6314 * ways to tell with the current Energy Model if this is actually a good
6315 * idea or not. So, find_energy_efficient_cpu() basically favors
6316 * cluster-packing, and spreading inside a cluster. That should at least be
6317 * a good thing for latency, and this is consistent with the idea that most
6318 * of the energy savings of EAS come from the asymmetry of the system, and
6319 * not so much from breaking the tie between identical CPUs. That's also the
6320 * reason why EAS is enabled in the topology code only for systems where
6321 * SD_ASYM_CPUCAPACITY is set.
6323 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6324 * they don't have any useful utilization data yet and it's not possible to
6325 * forecast their impact on energy consumption. Consequently, they will be
6326 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6327 * to be energy-inefficient in some use-cases. The alternative would be to
6328 * bias new tasks towards specific types of CPUs first, or to try to infer
6329 * their util_avg from the parent task, but those heuristics could hurt
6330 * other use-cases too. So, until someone finds a better way to solve this,
6331 * let's keep things simple by re-using the existing slow path.
6333 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6335 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
6336 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6337 unsigned long cpu_cap, util, base_energy = 0;
6338 int cpu, best_energy_cpu = prev_cpu;
6339 struct sched_domain *sd;
6340 struct perf_domain *pd;
6343 pd = rcu_dereference(rd->pd);
6344 if (!pd || READ_ONCE(rd->overutilized))
6348 * Energy-aware wake-up happens on the lowest sched_domain starting
6349 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6351 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6352 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6357 sync_entity_load_avg(&p->se);
6358 if (!task_util_est(p))
6361 for (; pd; pd = pd->next) {
6362 unsigned long cur_delta, spare_cap, max_spare_cap = 0;
6363 unsigned long base_energy_pd;
6364 int max_spare_cap_cpu = -1;
6366 /* Compute the 'base' energy of the pd, without @p */
6367 base_energy_pd = compute_energy(p, -1, pd);
6368 base_energy += base_energy_pd;
6370 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
6371 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6374 /* Skip CPUs that will be overutilized. */
6375 util = cpu_util_next(cpu, p, cpu);
6376 cpu_cap = capacity_of(cpu);
6377 if (!fits_capacity(util, cpu_cap))
6380 /* Always use prev_cpu as a candidate. */
6381 if (cpu == prev_cpu) {
6382 prev_delta = compute_energy(p, prev_cpu, pd);
6383 prev_delta -= base_energy_pd;
6384 best_delta = min(best_delta, prev_delta);
6388 * Find the CPU with the maximum spare capacity in
6389 * the performance domain
6391 spare_cap = cpu_cap - util;
6392 if (spare_cap > max_spare_cap) {
6393 max_spare_cap = spare_cap;
6394 max_spare_cap_cpu = cpu;
6398 /* Evaluate the energy impact of using this CPU. */
6399 if (max_spare_cap_cpu >= 0) {
6400 cur_delta = compute_energy(p, max_spare_cap_cpu, pd);
6401 cur_delta -= base_energy_pd;
6402 if (cur_delta < best_delta) {
6403 best_delta = cur_delta;
6404 best_energy_cpu = max_spare_cap_cpu;
6412 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6413 * least 6% of the energy used by prev_cpu.
6415 if (prev_delta == ULONG_MAX)
6416 return best_energy_cpu;
6418 if ((prev_delta - best_delta) > ((prev_delta + base_energy) >> 4))
6419 return best_energy_cpu;
6430 * select_task_rq_fair: Select target runqueue for the waking task in domains
6431 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6432 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6434 * Balances load by selecting the idlest CPU in the idlest group, or under
6435 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6437 * Returns the target CPU number.
6439 * preempt must be disabled.
6442 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6444 struct sched_domain *tmp, *sd = NULL;
6445 int cpu = smp_processor_id();
6446 int new_cpu = prev_cpu;
6447 int want_affine = 0;
6448 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6450 if (sd_flag & SD_BALANCE_WAKE) {
6453 if (sched_energy_enabled()) {
6454 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
6460 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu) &&
6461 cpumask_test_cpu(cpu, p->cpus_ptr);
6465 for_each_domain(cpu, tmp) {
6466 if (!(tmp->flags & SD_LOAD_BALANCE))
6470 * If both 'cpu' and 'prev_cpu' are part of this domain,
6471 * cpu is a valid SD_WAKE_AFFINE target.
6473 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6474 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6475 if (cpu != prev_cpu)
6476 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6478 sd = NULL; /* Prefer wake_affine over balance flags */
6482 if (tmp->flags & sd_flag)
6484 else if (!want_affine)
6490 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6491 } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6494 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6497 current->recent_used_cpu = cpu;
6504 static void detach_entity_cfs_rq(struct sched_entity *se);
6507 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6508 * cfs_rq_of(p) references at time of call are still valid and identify the
6509 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6511 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6514 * As blocked tasks retain absolute vruntime the migration needs to
6515 * deal with this by subtracting the old and adding the new
6516 * min_vruntime -- the latter is done by enqueue_entity() when placing
6517 * the task on the new runqueue.
6519 if (p->state == TASK_WAKING) {
6520 struct sched_entity *se = &p->se;
6521 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6524 #ifndef CONFIG_64BIT
6525 u64 min_vruntime_copy;
6528 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6530 min_vruntime = cfs_rq->min_vruntime;
6531 } while (min_vruntime != min_vruntime_copy);
6533 min_vruntime = cfs_rq->min_vruntime;
6536 se->vruntime -= min_vruntime;
6539 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6541 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6542 * rq->lock and can modify state directly.
6544 lockdep_assert_held(&task_rq(p)->lock);
6545 detach_entity_cfs_rq(&p->se);
6549 * We are supposed to update the task to "current" time, then
6550 * its up to date and ready to go to new CPU/cfs_rq. But we
6551 * have difficulty in getting what current time is, so simply
6552 * throw away the out-of-date time. This will result in the
6553 * wakee task is less decayed, but giving the wakee more load
6556 remove_entity_load_avg(&p->se);
6559 /* Tell new CPU we are migrated */
6560 p->se.avg.last_update_time = 0;
6562 /* We have migrated, no longer consider this task hot */
6563 p->se.exec_start = 0;
6565 update_scan_period(p, new_cpu);
6568 static void task_dead_fair(struct task_struct *p)
6570 remove_entity_load_avg(&p->se);
6572 #endif /* CONFIG_SMP */
6574 static unsigned long wakeup_gran(struct sched_entity *se)
6576 unsigned long gran = sysctl_sched_wakeup_granularity;
6579 * Since its curr running now, convert the gran from real-time
6580 * to virtual-time in his units.
6582 * By using 'se' instead of 'curr' we penalize light tasks, so
6583 * they get preempted easier. That is, if 'se' < 'curr' then
6584 * the resulting gran will be larger, therefore penalizing the
6585 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6586 * be smaller, again penalizing the lighter task.
6588 * This is especially important for buddies when the leftmost
6589 * task is higher priority than the buddy.
6591 return calc_delta_fair(gran, se);
6595 * Should 'se' preempt 'curr'.
6609 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6611 s64 gran, vdiff = curr->vruntime - se->vruntime;
6616 gran = wakeup_gran(se);
6623 static void set_last_buddy(struct sched_entity *se)
6625 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6628 for_each_sched_entity(se) {
6629 if (SCHED_WARN_ON(!se->on_rq))
6631 cfs_rq_of(se)->last = se;
6635 static void set_next_buddy(struct sched_entity *se)
6637 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6640 for_each_sched_entity(se) {
6641 if (SCHED_WARN_ON(!se->on_rq))
6643 cfs_rq_of(se)->next = se;
6647 static void set_skip_buddy(struct sched_entity *se)
6649 for_each_sched_entity(se)
6650 cfs_rq_of(se)->skip = se;
6654 * Preempt the current task with a newly woken task if needed:
6656 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6658 struct task_struct *curr = rq->curr;
6659 struct sched_entity *se = &curr->se, *pse = &p->se;
6660 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6661 int scale = cfs_rq->nr_running >= sched_nr_latency;
6662 int next_buddy_marked = 0;
6664 if (unlikely(se == pse))
6668 * This is possible from callers such as attach_tasks(), in which we
6669 * unconditionally check_prempt_curr() after an enqueue (which may have
6670 * lead to a throttle). This both saves work and prevents false
6671 * next-buddy nomination below.
6673 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6676 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6677 set_next_buddy(pse);
6678 next_buddy_marked = 1;
6682 * We can come here with TIF_NEED_RESCHED already set from new task
6685 * Note: this also catches the edge-case of curr being in a throttled
6686 * group (e.g. via set_curr_task), since update_curr() (in the
6687 * enqueue of curr) will have resulted in resched being set. This
6688 * prevents us from potentially nominating it as a false LAST_BUDDY
6691 if (test_tsk_need_resched(curr))
6694 /* Idle tasks are by definition preempted by non-idle tasks. */
6695 if (unlikely(task_has_idle_policy(curr)) &&
6696 likely(!task_has_idle_policy(p)))
6700 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6701 * is driven by the tick):
6703 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6706 find_matching_se(&se, &pse);
6707 update_curr(cfs_rq_of(se));
6709 if (wakeup_preempt_entity(se, pse) == 1) {
6711 * Bias pick_next to pick the sched entity that is
6712 * triggering this preemption.
6714 if (!next_buddy_marked)
6715 set_next_buddy(pse);
6724 * Only set the backward buddy when the current task is still
6725 * on the rq. This can happen when a wakeup gets interleaved
6726 * with schedule on the ->pre_schedule() or idle_balance()
6727 * point, either of which can * drop the rq lock.
6729 * Also, during early boot the idle thread is in the fair class,
6730 * for obvious reasons its a bad idea to schedule back to it.
6732 if (unlikely(!se->on_rq || curr == rq->idle))
6735 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6739 static struct task_struct *
6740 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6742 struct cfs_rq *cfs_rq = &rq->cfs;
6743 struct sched_entity *se;
6744 struct task_struct *p;
6748 if (!cfs_rq->nr_running)
6751 #ifdef CONFIG_FAIR_GROUP_SCHED
6752 if (!prev || prev->sched_class != &fair_sched_class)
6756 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6757 * likely that a next task is from the same cgroup as the current.
6759 * Therefore attempt to avoid putting and setting the entire cgroup
6760 * hierarchy, only change the part that actually changes.
6764 struct sched_entity *curr = cfs_rq->curr;
6767 * Since we got here without doing put_prev_entity() we also
6768 * have to consider cfs_rq->curr. If it is still a runnable
6769 * entity, update_curr() will update its vruntime, otherwise
6770 * forget we've ever seen it.
6774 update_curr(cfs_rq);
6779 * This call to check_cfs_rq_runtime() will do the
6780 * throttle and dequeue its entity in the parent(s).
6781 * Therefore the nr_running test will indeed
6784 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6787 if (!cfs_rq->nr_running)
6794 se = pick_next_entity(cfs_rq, curr);
6795 cfs_rq = group_cfs_rq(se);
6801 * Since we haven't yet done put_prev_entity and if the selected task
6802 * is a different task than we started out with, try and touch the
6803 * least amount of cfs_rqs.
6806 struct sched_entity *pse = &prev->se;
6808 while (!(cfs_rq = is_same_group(se, pse))) {
6809 int se_depth = se->depth;
6810 int pse_depth = pse->depth;
6812 if (se_depth <= pse_depth) {
6813 put_prev_entity(cfs_rq_of(pse), pse);
6814 pse = parent_entity(pse);
6816 if (se_depth >= pse_depth) {
6817 set_next_entity(cfs_rq_of(se), se);
6818 se = parent_entity(se);
6822 put_prev_entity(cfs_rq, pse);
6823 set_next_entity(cfs_rq, se);
6830 put_prev_task(rq, prev);
6833 se = pick_next_entity(cfs_rq, NULL);
6834 set_next_entity(cfs_rq, se);
6835 cfs_rq = group_cfs_rq(se);
6840 done: __maybe_unused;
6843 * Move the next running task to the front of
6844 * the list, so our cfs_tasks list becomes MRU
6847 list_move(&p->se.group_node, &rq->cfs_tasks);
6850 if (hrtick_enabled(rq))
6851 hrtick_start_fair(rq, p);
6853 update_misfit_status(p, rq);
6861 new_tasks = newidle_balance(rq, rf);
6864 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
6865 * possible for any higher priority task to appear. In that case we
6866 * must re-start the pick_next_entity() loop.
6875 * rq is about to be idle, check if we need to update the
6876 * lost_idle_time of clock_pelt
6878 update_idle_rq_clock_pelt(rq);
6884 * Account for a descheduled task:
6886 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6888 struct sched_entity *se = &prev->se;
6889 struct cfs_rq *cfs_rq;
6891 for_each_sched_entity(se) {
6892 cfs_rq = cfs_rq_of(se);
6893 put_prev_entity(cfs_rq, se);
6898 * sched_yield() is very simple
6900 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6902 static void yield_task_fair(struct rq *rq)
6904 struct task_struct *curr = rq->curr;
6905 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6906 struct sched_entity *se = &curr->se;
6909 * Are we the only task in the tree?
6911 if (unlikely(rq->nr_running == 1))
6914 clear_buddies(cfs_rq, se);
6916 if (curr->policy != SCHED_BATCH) {
6917 update_rq_clock(rq);
6919 * Update run-time statistics of the 'current'.
6921 update_curr(cfs_rq);
6923 * Tell update_rq_clock() that we've just updated,
6924 * so we don't do microscopic update in schedule()
6925 * and double the fastpath cost.
6927 rq_clock_skip_update(rq);
6933 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6935 struct sched_entity *se = &p->se;
6937 /* throttled hierarchies are not runnable */
6938 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6941 /* Tell the scheduler that we'd really like pse to run next. */
6944 yield_task_fair(rq);
6950 /**************************************************
6951 * Fair scheduling class load-balancing methods.
6955 * The purpose of load-balancing is to achieve the same basic fairness the
6956 * per-CPU scheduler provides, namely provide a proportional amount of compute
6957 * time to each task. This is expressed in the following equation:
6959 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6961 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
6962 * W_i,0 is defined as:
6964 * W_i,0 = \Sum_j w_i,j (2)
6966 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
6967 * is derived from the nice value as per sched_prio_to_weight[].
6969 * The weight average is an exponential decay average of the instantaneous
6972 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6974 * C_i is the compute capacity of CPU i, typically it is the
6975 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6976 * can also include other factors [XXX].
6978 * To achieve this balance we define a measure of imbalance which follows
6979 * directly from (1):
6981 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6983 * We them move tasks around to minimize the imbalance. In the continuous
6984 * function space it is obvious this converges, in the discrete case we get
6985 * a few fun cases generally called infeasible weight scenarios.
6988 * - infeasible weights;
6989 * - local vs global optima in the discrete case. ]
6994 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6995 * for all i,j solution, we create a tree of CPUs that follows the hardware
6996 * topology where each level pairs two lower groups (or better). This results
6997 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
6998 * tree to only the first of the previous level and we decrease the frequency
6999 * of load-balance at each level inv. proportional to the number of CPUs in
7005 * \Sum { --- * --- * 2^i } = O(n) (5)
7007 * `- size of each group
7008 * | | `- number of CPUs doing load-balance
7010 * `- sum over all levels
7012 * Coupled with a limit on how many tasks we can migrate every balance pass,
7013 * this makes (5) the runtime complexity of the balancer.
7015 * An important property here is that each CPU is still (indirectly) connected
7016 * to every other CPU in at most O(log n) steps:
7018 * The adjacency matrix of the resulting graph is given by:
7021 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7024 * And you'll find that:
7026 * A^(log_2 n)_i,j != 0 for all i,j (7)
7028 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7029 * The task movement gives a factor of O(m), giving a convergence complexity
7032 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7037 * In order to avoid CPUs going idle while there's still work to do, new idle
7038 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7039 * tree itself instead of relying on other CPUs to bring it work.
7041 * This adds some complexity to both (5) and (8) but it reduces the total idle
7049 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7052 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7057 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7059 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7061 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7064 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7065 * rewrite all of this once again.]
7068 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7070 enum fbq_type { regular, remote, all };
7079 #define LBF_ALL_PINNED 0x01
7080 #define LBF_NEED_BREAK 0x02
7081 #define LBF_DST_PINNED 0x04
7082 #define LBF_SOME_PINNED 0x08
7083 #define LBF_NOHZ_STATS 0x10
7084 #define LBF_NOHZ_AGAIN 0x20
7087 struct sched_domain *sd;
7095 struct cpumask *dst_grpmask;
7097 enum cpu_idle_type idle;
7099 /* The set of CPUs under consideration for load-balancing */
7100 struct cpumask *cpus;
7105 unsigned int loop_break;
7106 unsigned int loop_max;
7108 enum fbq_type fbq_type;
7109 enum group_type src_grp_type;
7110 struct list_head tasks;
7114 * Is this task likely cache-hot:
7116 static int task_hot(struct task_struct *p, struct lb_env *env)
7120 lockdep_assert_held(&env->src_rq->lock);
7122 if (p->sched_class != &fair_sched_class)
7125 if (unlikely(task_has_idle_policy(p)))
7129 * Buddy candidates are cache hot:
7131 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7132 (&p->se == cfs_rq_of(&p->se)->next ||
7133 &p->se == cfs_rq_of(&p->se)->last))
7136 if (sysctl_sched_migration_cost == -1)
7138 if (sysctl_sched_migration_cost == 0)
7141 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7143 return delta < (s64)sysctl_sched_migration_cost;
7146 #ifdef CONFIG_NUMA_BALANCING
7148 * Returns 1, if task migration degrades locality
7149 * Returns 0, if task migration improves locality i.e migration preferred.
7150 * Returns -1, if task migration is not affected by locality.
7152 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7154 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7155 unsigned long src_weight, dst_weight;
7156 int src_nid, dst_nid, dist;
7158 if (!static_branch_likely(&sched_numa_balancing))
7161 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7164 src_nid = cpu_to_node(env->src_cpu);
7165 dst_nid = cpu_to_node(env->dst_cpu);
7167 if (src_nid == dst_nid)
7170 /* Migrating away from the preferred node is always bad. */
7171 if (src_nid == p->numa_preferred_nid) {
7172 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7178 /* Encourage migration to the preferred node. */
7179 if (dst_nid == p->numa_preferred_nid)
7182 /* Leaving a core idle is often worse than degrading locality. */
7183 if (env->idle == CPU_IDLE)
7186 dist = node_distance(src_nid, dst_nid);
7188 src_weight = group_weight(p, src_nid, dist);
7189 dst_weight = group_weight(p, dst_nid, dist);
7191 src_weight = task_weight(p, src_nid, dist);
7192 dst_weight = task_weight(p, dst_nid, dist);
7195 return dst_weight < src_weight;
7199 static inline int migrate_degrades_locality(struct task_struct *p,
7207 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7210 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7214 lockdep_assert_held(&env->src_rq->lock);
7217 * We do not migrate tasks that are:
7218 * 1) throttled_lb_pair, or
7219 * 2) cannot be migrated to this CPU due to cpus_ptr, or
7220 * 3) running (obviously), or
7221 * 4) are cache-hot on their current CPU.
7223 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7226 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
7229 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7231 env->flags |= LBF_SOME_PINNED;
7234 * Remember if this task can be migrated to any other CPU in
7235 * our sched_group. We may want to revisit it if we couldn't
7236 * meet load balance goals by pulling other tasks on src_cpu.
7238 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7239 * already computed one in current iteration.
7241 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7244 /* Prevent to re-select dst_cpu via env's CPUs: */
7245 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7246 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
7247 env->flags |= LBF_DST_PINNED;
7248 env->new_dst_cpu = cpu;
7256 /* Record that we found atleast one task that could run on dst_cpu */
7257 env->flags &= ~LBF_ALL_PINNED;
7259 if (task_running(env->src_rq, p)) {
7260 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7265 * Aggressive migration if:
7266 * 1) destination numa is preferred
7267 * 2) task is cache cold, or
7268 * 3) too many balance attempts have failed.
7270 tsk_cache_hot = migrate_degrades_locality(p, env);
7271 if (tsk_cache_hot == -1)
7272 tsk_cache_hot = task_hot(p, env);
7274 if (tsk_cache_hot <= 0 ||
7275 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7276 if (tsk_cache_hot == 1) {
7277 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7278 schedstat_inc(p->se.statistics.nr_forced_migrations);
7283 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7288 * detach_task() -- detach the task for the migration specified in env
7290 static void detach_task(struct task_struct *p, struct lb_env *env)
7292 lockdep_assert_held(&env->src_rq->lock);
7294 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7295 set_task_cpu(p, env->dst_cpu);
7299 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7300 * part of active balancing operations within "domain".
7302 * Returns a task if successful and NULL otherwise.
7304 static struct task_struct *detach_one_task(struct lb_env *env)
7306 struct task_struct *p;
7308 lockdep_assert_held(&env->src_rq->lock);
7310 list_for_each_entry_reverse(p,
7311 &env->src_rq->cfs_tasks, se.group_node) {
7312 if (!can_migrate_task(p, env))
7315 detach_task(p, env);
7318 * Right now, this is only the second place where
7319 * lb_gained[env->idle] is updated (other is detach_tasks)
7320 * so we can safely collect stats here rather than
7321 * inside detach_tasks().
7323 schedstat_inc(env->sd->lb_gained[env->idle]);
7329 static const unsigned int sched_nr_migrate_break = 32;
7332 * detach_tasks() -- tries to detach up to imbalance runnable load from
7333 * busiest_rq, as part of a balancing operation within domain "sd".
7335 * Returns number of detached tasks if successful and 0 otherwise.
7337 static int detach_tasks(struct lb_env *env)
7339 struct list_head *tasks = &env->src_rq->cfs_tasks;
7340 struct task_struct *p;
7344 lockdep_assert_held(&env->src_rq->lock);
7346 if (env->imbalance <= 0)
7349 while (!list_empty(tasks)) {
7351 * We don't want to steal all, otherwise we may be treated likewise,
7352 * which could at worst lead to a livelock crash.
7354 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7357 p = list_last_entry(tasks, struct task_struct, se.group_node);
7360 /* We've more or less seen every task there is, call it quits */
7361 if (env->loop > env->loop_max)
7364 /* take a breather every nr_migrate tasks */
7365 if (env->loop > env->loop_break) {
7366 env->loop_break += sched_nr_migrate_break;
7367 env->flags |= LBF_NEED_BREAK;
7371 if (!can_migrate_task(p, env))
7374 load = task_h_load(p);
7376 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7379 if ((load / 2) > env->imbalance)
7382 detach_task(p, env);
7383 list_add(&p->se.group_node, &env->tasks);
7386 env->imbalance -= load;
7388 #ifdef CONFIG_PREEMPTION
7390 * NEWIDLE balancing is a source of latency, so preemptible
7391 * kernels will stop after the first task is detached to minimize
7392 * the critical section.
7394 if (env->idle == CPU_NEWLY_IDLE)
7399 * We only want to steal up to the prescribed amount of
7402 if (env->imbalance <= 0)
7407 list_move(&p->se.group_node, tasks);
7411 * Right now, this is one of only two places we collect this stat
7412 * so we can safely collect detach_one_task() stats here rather
7413 * than inside detach_one_task().
7415 schedstat_add(env->sd->lb_gained[env->idle], detached);
7421 * attach_task() -- attach the task detached by detach_task() to its new rq.
7423 static void attach_task(struct rq *rq, struct task_struct *p)
7425 lockdep_assert_held(&rq->lock);
7427 BUG_ON(task_rq(p) != rq);
7428 activate_task(rq, p, ENQUEUE_NOCLOCK);
7429 check_preempt_curr(rq, p, 0);
7433 * attach_one_task() -- attaches the task returned from detach_one_task() to
7436 static void attach_one_task(struct rq *rq, struct task_struct *p)
7441 update_rq_clock(rq);
7447 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7450 static void attach_tasks(struct lb_env *env)
7452 struct list_head *tasks = &env->tasks;
7453 struct task_struct *p;
7456 rq_lock(env->dst_rq, &rf);
7457 update_rq_clock(env->dst_rq);
7459 while (!list_empty(tasks)) {
7460 p = list_first_entry(tasks, struct task_struct, se.group_node);
7461 list_del_init(&p->se.group_node);
7463 attach_task(env->dst_rq, p);
7466 rq_unlock(env->dst_rq, &rf);
7469 #ifdef CONFIG_NO_HZ_COMMON
7470 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7472 if (cfs_rq->avg.load_avg)
7475 if (cfs_rq->avg.util_avg)
7481 static inline bool others_have_blocked(struct rq *rq)
7483 if (READ_ONCE(rq->avg_rt.util_avg))
7486 if (READ_ONCE(rq->avg_dl.util_avg))
7489 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7490 if (READ_ONCE(rq->avg_irq.util_avg))
7497 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
7499 rq->last_blocked_load_update_tick = jiffies;
7502 rq->has_blocked_load = 0;
7505 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
7506 static inline bool others_have_blocked(struct rq *rq) { return false; }
7507 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
7510 #ifdef CONFIG_FAIR_GROUP_SCHED
7512 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7514 if (cfs_rq->load.weight)
7517 if (cfs_rq->avg.load_sum)
7520 if (cfs_rq->avg.util_sum)
7523 if (cfs_rq->avg.runnable_load_sum)
7529 static void update_blocked_averages(int cpu)
7531 struct rq *rq = cpu_rq(cpu);
7532 struct cfs_rq *cfs_rq, *pos;
7533 const struct sched_class *curr_class;
7537 rq_lock_irqsave(rq, &rf);
7538 update_rq_clock(rq);
7541 * Iterates the task_group tree in a bottom up fashion, see
7542 * list_add_leaf_cfs_rq() for details.
7544 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7545 struct sched_entity *se;
7547 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq))
7548 update_tg_load_avg(cfs_rq, 0);
7550 /* Propagate pending load changes to the parent, if any: */
7551 se = cfs_rq->tg->se[cpu];
7552 if (se && !skip_blocked_update(se))
7553 update_load_avg(cfs_rq_of(se), se, 0);
7556 * There can be a lot of idle CPU cgroups. Don't let fully
7557 * decayed cfs_rqs linger on the list.
7559 if (cfs_rq_is_decayed(cfs_rq))
7560 list_del_leaf_cfs_rq(cfs_rq);
7562 /* Don't need periodic decay once load/util_avg are null */
7563 if (cfs_rq_has_blocked(cfs_rq))
7567 curr_class = rq->curr->sched_class;
7568 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class);
7569 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class);
7570 update_irq_load_avg(rq, 0);
7571 /* Don't need periodic decay once load/util_avg are null */
7572 if (others_have_blocked(rq))
7575 update_blocked_load_status(rq, !done);
7576 rq_unlock_irqrestore(rq, &rf);
7580 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7581 * This needs to be done in a top-down fashion because the load of a child
7582 * group is a fraction of its parents load.
7584 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7586 struct rq *rq = rq_of(cfs_rq);
7587 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7588 unsigned long now = jiffies;
7591 if (cfs_rq->last_h_load_update == now)
7594 WRITE_ONCE(cfs_rq->h_load_next, NULL);
7595 for_each_sched_entity(se) {
7596 cfs_rq = cfs_rq_of(se);
7597 WRITE_ONCE(cfs_rq->h_load_next, se);
7598 if (cfs_rq->last_h_load_update == now)
7603 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7604 cfs_rq->last_h_load_update = now;
7607 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
7608 load = cfs_rq->h_load;
7609 load = div64_ul(load * se->avg.load_avg,
7610 cfs_rq_load_avg(cfs_rq) + 1);
7611 cfs_rq = group_cfs_rq(se);
7612 cfs_rq->h_load = load;
7613 cfs_rq->last_h_load_update = now;
7617 static unsigned long task_h_load(struct task_struct *p)
7619 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7621 update_cfs_rq_h_load(cfs_rq);
7622 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7623 cfs_rq_load_avg(cfs_rq) + 1);
7626 static inline void update_blocked_averages(int cpu)
7628 struct rq *rq = cpu_rq(cpu);
7629 struct cfs_rq *cfs_rq = &rq->cfs;
7630 const struct sched_class *curr_class;
7633 rq_lock_irqsave(rq, &rf);
7634 update_rq_clock(rq);
7635 update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
7637 curr_class = rq->curr->sched_class;
7638 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class);
7639 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class);
7640 update_irq_load_avg(rq, 0);
7641 update_blocked_load_status(rq, cfs_rq_has_blocked(cfs_rq) || others_have_blocked(rq));
7642 rq_unlock_irqrestore(rq, &rf);
7645 static unsigned long task_h_load(struct task_struct *p)
7647 return p->se.avg.load_avg;
7651 /********** Helpers for find_busiest_group ************************/
7654 * sg_lb_stats - stats of a sched_group required for load_balancing
7656 struct sg_lb_stats {
7657 unsigned long avg_load; /*Avg load across the CPUs of the group */
7658 unsigned long group_load; /* Total load over the CPUs of the group */
7659 unsigned long load_per_task;
7660 unsigned long group_capacity;
7661 unsigned long group_util; /* Total utilization of the group */
7662 unsigned int sum_nr_running; /* Nr tasks running in the group */
7663 unsigned int idle_cpus;
7664 unsigned int group_weight;
7665 enum group_type group_type;
7666 int group_no_capacity;
7667 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
7668 #ifdef CONFIG_NUMA_BALANCING
7669 unsigned int nr_numa_running;
7670 unsigned int nr_preferred_running;
7675 * sd_lb_stats - Structure to store the statistics of a sched_domain
7676 * during load balancing.
7678 struct sd_lb_stats {
7679 struct sched_group *busiest; /* Busiest group in this sd */
7680 struct sched_group *local; /* Local group in this sd */
7681 unsigned long total_running;
7682 unsigned long total_load; /* Total load of all groups in sd */
7683 unsigned long total_capacity; /* Total capacity of all groups in sd */
7684 unsigned long avg_load; /* Average load across all groups in sd */
7686 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7687 struct sg_lb_stats local_stat; /* Statistics of the local group */
7690 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7693 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7694 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7695 * We must however clear busiest_stat::avg_load because
7696 * update_sd_pick_busiest() reads this before assignment.
7698 *sds = (struct sd_lb_stats){
7701 .total_running = 0UL,
7703 .total_capacity = 0UL,
7706 .sum_nr_running = 0,
7707 .group_type = group_other,
7712 static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu)
7714 struct rq *rq = cpu_rq(cpu);
7715 unsigned long max = arch_scale_cpu_capacity(cpu);
7716 unsigned long used, free;
7719 irq = cpu_util_irq(rq);
7721 if (unlikely(irq >= max))
7724 used = READ_ONCE(rq->avg_rt.util_avg);
7725 used += READ_ONCE(rq->avg_dl.util_avg);
7727 if (unlikely(used >= max))
7732 return scale_irq_capacity(free, irq, max);
7735 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7737 unsigned long capacity = scale_rt_capacity(sd, cpu);
7738 struct sched_group *sdg = sd->groups;
7740 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
7745 cpu_rq(cpu)->cpu_capacity = capacity;
7746 sdg->sgc->capacity = capacity;
7747 sdg->sgc->min_capacity = capacity;
7748 sdg->sgc->max_capacity = capacity;
7751 void update_group_capacity(struct sched_domain *sd, int cpu)
7753 struct sched_domain *child = sd->child;
7754 struct sched_group *group, *sdg = sd->groups;
7755 unsigned long capacity, min_capacity, max_capacity;
7756 unsigned long interval;
7758 interval = msecs_to_jiffies(sd->balance_interval);
7759 interval = clamp(interval, 1UL, max_load_balance_interval);
7760 sdg->sgc->next_update = jiffies + interval;
7763 update_cpu_capacity(sd, cpu);
7768 min_capacity = ULONG_MAX;
7771 if (child->flags & SD_OVERLAP) {
7773 * SD_OVERLAP domains cannot assume that child groups
7774 * span the current group.
7777 for_each_cpu(cpu, sched_group_span(sdg)) {
7778 struct sched_group_capacity *sgc;
7779 struct rq *rq = cpu_rq(cpu);
7782 * build_sched_domains() -> init_sched_groups_capacity()
7783 * gets here before we've attached the domains to the
7786 * Use capacity_of(), which is set irrespective of domains
7787 * in update_cpu_capacity().
7789 * This avoids capacity from being 0 and
7790 * causing divide-by-zero issues on boot.
7792 if (unlikely(!rq->sd)) {
7793 capacity += capacity_of(cpu);
7795 sgc = rq->sd->groups->sgc;
7796 capacity += sgc->capacity;
7799 min_capacity = min(capacity, min_capacity);
7800 max_capacity = max(capacity, max_capacity);
7804 * !SD_OVERLAP domains can assume that child groups
7805 * span the current group.
7808 group = child->groups;
7810 struct sched_group_capacity *sgc = group->sgc;
7812 capacity += sgc->capacity;
7813 min_capacity = min(sgc->min_capacity, min_capacity);
7814 max_capacity = max(sgc->max_capacity, max_capacity);
7815 group = group->next;
7816 } while (group != child->groups);
7819 sdg->sgc->capacity = capacity;
7820 sdg->sgc->min_capacity = min_capacity;
7821 sdg->sgc->max_capacity = max_capacity;
7825 * Check whether the capacity of the rq has been noticeably reduced by side
7826 * activity. The imbalance_pct is used for the threshold.
7827 * Return true is the capacity is reduced
7830 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7832 return ((rq->cpu_capacity * sd->imbalance_pct) <
7833 (rq->cpu_capacity_orig * 100));
7837 * Check whether a rq has a misfit task and if it looks like we can actually
7838 * help that task: we can migrate the task to a CPU of higher capacity, or
7839 * the task's current CPU is heavily pressured.
7841 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
7843 return rq->misfit_task_load &&
7844 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
7845 check_cpu_capacity(rq, sd));
7849 * Group imbalance indicates (and tries to solve) the problem where balancing
7850 * groups is inadequate due to ->cpus_ptr constraints.
7852 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
7853 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
7856 * { 0 1 2 3 } { 4 5 6 7 }
7859 * If we were to balance group-wise we'd place two tasks in the first group and
7860 * two tasks in the second group. Clearly this is undesired as it will overload
7861 * cpu 3 and leave one of the CPUs in the second group unused.
7863 * The current solution to this issue is detecting the skew in the first group
7864 * by noticing the lower domain failed to reach balance and had difficulty
7865 * moving tasks due to affinity constraints.
7867 * When this is so detected; this group becomes a candidate for busiest; see
7868 * update_sd_pick_busiest(). And calculate_imbalance() and
7869 * find_busiest_group() avoid some of the usual balance conditions to allow it
7870 * to create an effective group imbalance.
7872 * This is a somewhat tricky proposition since the next run might not find the
7873 * group imbalance and decide the groups need to be balanced again. A most
7874 * subtle and fragile situation.
7877 static inline int sg_imbalanced(struct sched_group *group)
7879 return group->sgc->imbalance;
7883 * group_has_capacity returns true if the group has spare capacity that could
7884 * be used by some tasks.
7885 * We consider that a group has spare capacity if the * number of task is
7886 * smaller than the number of CPUs or if the utilization is lower than the
7887 * available capacity for CFS tasks.
7888 * For the latter, we use a threshold to stabilize the state, to take into
7889 * account the variance of the tasks' load and to return true if the available
7890 * capacity in meaningful for the load balancer.
7891 * As an example, an available capacity of 1% can appear but it doesn't make
7892 * any benefit for the load balance.
7895 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7897 if (sgs->sum_nr_running < sgs->group_weight)
7900 if ((sgs->group_capacity * 100) >
7901 (sgs->group_util * env->sd->imbalance_pct))
7908 * group_is_overloaded returns true if the group has more tasks than it can
7910 * group_is_overloaded is not equals to !group_has_capacity because a group
7911 * with the exact right number of tasks, has no more spare capacity but is not
7912 * overloaded so both group_has_capacity and group_is_overloaded return
7916 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7918 if (sgs->sum_nr_running <= sgs->group_weight)
7921 if ((sgs->group_capacity * 100) <
7922 (sgs->group_util * env->sd->imbalance_pct))
7929 * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller
7930 * per-CPU capacity than sched_group ref.
7933 group_smaller_min_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7935 return fits_capacity(sg->sgc->min_capacity, ref->sgc->min_capacity);
7939 * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller
7940 * per-CPU capacity_orig than sched_group ref.
7943 group_smaller_max_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7945 return fits_capacity(sg->sgc->max_capacity, ref->sgc->max_capacity);
7949 group_type group_classify(struct sched_group *group,
7950 struct sg_lb_stats *sgs)
7952 if (sgs->group_no_capacity)
7953 return group_overloaded;
7955 if (sg_imbalanced(group))
7956 return group_imbalanced;
7958 if (sgs->group_misfit_task_load)
7959 return group_misfit_task;
7964 static bool update_nohz_stats(struct rq *rq, bool force)
7966 #ifdef CONFIG_NO_HZ_COMMON
7967 unsigned int cpu = rq->cpu;
7969 if (!rq->has_blocked_load)
7972 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
7975 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
7978 update_blocked_averages(cpu);
7980 return rq->has_blocked_load;
7987 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7988 * @env: The load balancing environment.
7989 * @group: sched_group whose statistics are to be updated.
7990 * @sgs: variable to hold the statistics for this group.
7991 * @sg_status: Holds flag indicating the status of the sched_group
7993 static inline void update_sg_lb_stats(struct lb_env *env,
7994 struct sched_group *group,
7995 struct sg_lb_stats *sgs,
8000 memset(sgs, 0, sizeof(*sgs));
8002 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8003 struct rq *rq = cpu_rq(i);
8005 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
8006 env->flags |= LBF_NOHZ_AGAIN;
8008 sgs->group_load += cpu_runnable_load(rq);
8009 sgs->group_util += cpu_util(i);
8010 sgs->sum_nr_running += rq->cfs.h_nr_running;
8012 nr_running = rq->nr_running;
8014 *sg_status |= SG_OVERLOAD;
8016 if (cpu_overutilized(i))
8017 *sg_status |= SG_OVERUTILIZED;
8019 #ifdef CONFIG_NUMA_BALANCING
8020 sgs->nr_numa_running += rq->nr_numa_running;
8021 sgs->nr_preferred_running += rq->nr_preferred_running;
8024 * No need to call idle_cpu() if nr_running is not 0
8026 if (!nr_running && idle_cpu(i))
8029 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8030 sgs->group_misfit_task_load < rq->misfit_task_load) {
8031 sgs->group_misfit_task_load = rq->misfit_task_load;
8032 *sg_status |= SG_OVERLOAD;
8036 /* Adjust by relative CPU capacity of the group */
8037 sgs->group_capacity = group->sgc->capacity;
8038 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
8040 if (sgs->sum_nr_running)
8041 sgs->load_per_task = sgs->group_load / sgs->sum_nr_running;
8043 sgs->group_weight = group->group_weight;
8045 sgs->group_no_capacity = group_is_overloaded(env, sgs);
8046 sgs->group_type = group_classify(group, sgs);
8050 * update_sd_pick_busiest - return 1 on busiest group
8051 * @env: The load balancing environment.
8052 * @sds: sched_domain statistics
8053 * @sg: sched_group candidate to be checked for being the busiest
8054 * @sgs: sched_group statistics
8056 * Determine if @sg is a busier group than the previously selected
8059 * Return: %true if @sg is a busier group than the previously selected
8060 * busiest group. %false otherwise.
8062 static bool update_sd_pick_busiest(struct lb_env *env,
8063 struct sd_lb_stats *sds,
8064 struct sched_group *sg,
8065 struct sg_lb_stats *sgs)
8067 struct sg_lb_stats *busiest = &sds->busiest_stat;
8070 * Don't try to pull misfit tasks we can't help.
8071 * We can use max_capacity here as reduction in capacity on some
8072 * CPUs in the group should either be possible to resolve
8073 * internally or be covered by avg_load imbalance (eventually).
8075 if (sgs->group_type == group_misfit_task &&
8076 (!group_smaller_max_cpu_capacity(sg, sds->local) ||
8077 !group_has_capacity(env, &sds->local_stat)))
8080 if (sgs->group_type > busiest->group_type)
8083 if (sgs->group_type < busiest->group_type)
8086 if (sgs->avg_load <= busiest->avg_load)
8089 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
8093 * Candidate sg has no more than one task per CPU and
8094 * has higher per-CPU capacity. Migrating tasks to less
8095 * capable CPUs may harm throughput. Maximize throughput,
8096 * power/energy consequences are not considered.
8098 if (sgs->sum_nr_running <= sgs->group_weight &&
8099 group_smaller_min_cpu_capacity(sds->local, sg))
8103 * If we have more than one misfit sg go with the biggest misfit.
8105 if (sgs->group_type == group_misfit_task &&
8106 sgs->group_misfit_task_load < busiest->group_misfit_task_load)
8110 /* This is the busiest node in its class. */
8111 if (!(env->sd->flags & SD_ASYM_PACKING))
8114 /* No ASYM_PACKING if target CPU is already busy */
8115 if (env->idle == CPU_NOT_IDLE)
8118 * ASYM_PACKING needs to move all the work to the highest
8119 * prority CPUs in the group, therefore mark all groups
8120 * of lower priority than ourself as busy.
8122 if (sgs->sum_nr_running &&
8123 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
8127 /* Prefer to move from lowest priority CPU's work */
8128 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
8129 sg->asym_prefer_cpu))
8136 #ifdef CONFIG_NUMA_BALANCING
8137 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8139 if (sgs->sum_nr_running > sgs->nr_numa_running)
8141 if (sgs->sum_nr_running > sgs->nr_preferred_running)
8146 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8148 if (rq->nr_running > rq->nr_numa_running)
8150 if (rq->nr_running > rq->nr_preferred_running)
8155 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8160 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8164 #endif /* CONFIG_NUMA_BALANCING */
8167 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8168 * @env: The load balancing environment.
8169 * @sds: variable to hold the statistics for this sched_domain.
8171 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8173 struct sched_domain *child = env->sd->child;
8174 struct sched_group *sg = env->sd->groups;
8175 struct sg_lb_stats *local = &sds->local_stat;
8176 struct sg_lb_stats tmp_sgs;
8177 bool prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
8180 #ifdef CONFIG_NO_HZ_COMMON
8181 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
8182 env->flags |= LBF_NOHZ_STATS;
8186 struct sg_lb_stats *sgs = &tmp_sgs;
8189 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
8194 if (env->idle != CPU_NEWLY_IDLE ||
8195 time_after_eq(jiffies, sg->sgc->next_update))
8196 update_group_capacity(env->sd, env->dst_cpu);
8199 update_sg_lb_stats(env, sg, sgs, &sg_status);
8205 * In case the child domain prefers tasks go to siblings
8206 * first, lower the sg capacity so that we'll try
8207 * and move all the excess tasks away. We lower the capacity
8208 * of a group only if the local group has the capacity to fit
8209 * these excess tasks. The extra check prevents the case where
8210 * you always pull from the heaviest group when it is already
8211 * under-utilized (possible with a large weight task outweighs
8212 * the tasks on the system).
8214 if (prefer_sibling && sds->local &&
8215 group_has_capacity(env, local) &&
8216 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
8217 sgs->group_no_capacity = 1;
8218 sgs->group_type = group_classify(sg, sgs);
8221 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8223 sds->busiest_stat = *sgs;
8227 /* Now, start updating sd_lb_stats */
8228 sds->total_running += sgs->sum_nr_running;
8229 sds->total_load += sgs->group_load;
8230 sds->total_capacity += sgs->group_capacity;
8233 } while (sg != env->sd->groups);
8235 #ifdef CONFIG_NO_HZ_COMMON
8236 if ((env->flags & LBF_NOHZ_AGAIN) &&
8237 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8239 WRITE_ONCE(nohz.next_blocked,
8240 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8244 if (env->sd->flags & SD_NUMA)
8245 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8247 if (!env->sd->parent) {
8248 struct root_domain *rd = env->dst_rq->rd;
8250 /* update overload indicator if we are at root domain */
8251 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
8253 /* Update over-utilization (tipping point, U >= 0) indicator */
8254 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
8255 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
8256 } else if (sg_status & SG_OVERUTILIZED) {
8257 struct root_domain *rd = env->dst_rq->rd;
8259 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
8260 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
8265 * check_asym_packing - Check to see if the group is packed into the
8268 * This is primarily intended to used at the sibling level. Some
8269 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8270 * case of POWER7, it can move to lower SMT modes only when higher
8271 * threads are idle. When in lower SMT modes, the threads will
8272 * perform better since they share less core resources. Hence when we
8273 * have idle threads, we want them to be the higher ones.
8275 * This packing function is run on idle threads. It checks to see if
8276 * the busiest CPU in this domain (core in the P7 case) has a higher
8277 * CPU number than the packing function is being run on. Here we are
8278 * assuming lower CPU number will be equivalent to lower a SMT thread
8281 * Return: 1 when packing is required and a task should be moved to
8282 * this CPU. The amount of the imbalance is returned in env->imbalance.
8284 * @env: The load balancing environment.
8285 * @sds: Statistics of the sched_domain which is to be packed
8287 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8291 if (!(env->sd->flags & SD_ASYM_PACKING))
8294 if (env->idle == CPU_NOT_IDLE)
8300 busiest_cpu = sds->busiest->asym_prefer_cpu;
8301 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
8304 env->imbalance = sds->busiest_stat.group_load;
8310 * fix_small_imbalance - Calculate the minor imbalance that exists
8311 * amongst the groups of a sched_domain, during
8313 * @env: The load balancing environment.
8314 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8317 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8319 unsigned long tmp, capa_now = 0, capa_move = 0;
8320 unsigned int imbn = 2;
8321 unsigned long scaled_busy_load_per_task;
8322 struct sg_lb_stats *local, *busiest;
8324 local = &sds->local_stat;
8325 busiest = &sds->busiest_stat;
8327 if (!local->sum_nr_running)
8328 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8329 else if (busiest->load_per_task > local->load_per_task)
8332 scaled_busy_load_per_task =
8333 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8334 busiest->group_capacity;
8336 if (busiest->avg_load + scaled_busy_load_per_task >=
8337 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8338 env->imbalance = busiest->load_per_task;
8343 * OK, we don't have enough imbalance to justify moving tasks,
8344 * however we may be able to increase total CPU capacity used by
8348 capa_now += busiest->group_capacity *
8349 min(busiest->load_per_task, busiest->avg_load);
8350 capa_now += local->group_capacity *
8351 min(local->load_per_task, local->avg_load);
8352 capa_now /= SCHED_CAPACITY_SCALE;
8354 /* Amount of load we'd subtract */
8355 if (busiest->avg_load > scaled_busy_load_per_task) {
8356 capa_move += busiest->group_capacity *
8357 min(busiest->load_per_task,
8358 busiest->avg_load - scaled_busy_load_per_task);
8361 /* Amount of load we'd add */
8362 if (busiest->avg_load * busiest->group_capacity <
8363 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8364 tmp = (busiest->avg_load * busiest->group_capacity) /
8365 local->group_capacity;
8367 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8368 local->group_capacity;
8370 capa_move += local->group_capacity *
8371 min(local->load_per_task, local->avg_load + tmp);
8372 capa_move /= SCHED_CAPACITY_SCALE;
8374 /* Move if we gain throughput */
8375 if (capa_move > capa_now)
8376 env->imbalance = busiest->load_per_task;
8380 * calculate_imbalance - Calculate the amount of imbalance present within the
8381 * groups of a given sched_domain during load balance.
8382 * @env: load balance environment
8383 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8385 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8387 unsigned long max_pull, load_above_capacity = ~0UL;
8388 struct sg_lb_stats *local, *busiest;
8390 local = &sds->local_stat;
8391 busiest = &sds->busiest_stat;
8393 if (busiest->group_type == group_imbalanced) {
8395 * In the group_imb case we cannot rely on group-wide averages
8396 * to ensure CPU-load equilibrium, look at wider averages. XXX
8398 busiest->load_per_task =
8399 min(busiest->load_per_task, sds->avg_load);
8403 * Avg load of busiest sg can be less and avg load of local sg can
8404 * be greater than avg load across all sgs of sd because avg load
8405 * factors in sg capacity and sgs with smaller group_type are
8406 * skipped when updating the busiest sg:
8408 if (busiest->group_type != group_misfit_task &&
8409 (busiest->avg_load <= sds->avg_load ||
8410 local->avg_load >= sds->avg_load)) {
8412 return fix_small_imbalance(env, sds);
8416 * If there aren't any idle CPUs, avoid creating some.
8418 if (busiest->group_type == group_overloaded &&
8419 local->group_type == group_overloaded) {
8420 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
8421 if (load_above_capacity > busiest->group_capacity) {
8422 load_above_capacity -= busiest->group_capacity;
8423 load_above_capacity *= scale_load_down(NICE_0_LOAD);
8424 load_above_capacity /= busiest->group_capacity;
8426 load_above_capacity = ~0UL;
8430 * We're trying to get all the CPUs to the average_load, so we don't
8431 * want to push ourselves above the average load, nor do we wish to
8432 * reduce the max loaded CPU below the average load. At the same time,
8433 * we also don't want to reduce the group load below the group
8434 * capacity. Thus we look for the minimum possible imbalance.
8436 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8438 /* How much load to actually move to equalise the imbalance */
8439 env->imbalance = min(
8440 max_pull * busiest->group_capacity,
8441 (sds->avg_load - local->avg_load) * local->group_capacity
8442 ) / SCHED_CAPACITY_SCALE;
8444 /* Boost imbalance to allow misfit task to be balanced. */
8445 if (busiest->group_type == group_misfit_task) {
8446 env->imbalance = max_t(long, env->imbalance,
8447 busiest->group_misfit_task_load);
8451 * if *imbalance is less than the average load per runnable task
8452 * there is no guarantee that any tasks will be moved so we'll have
8453 * a think about bumping its value to force at least one task to be
8456 if (env->imbalance < busiest->load_per_task)
8457 return fix_small_imbalance(env, sds);
8460 /******* find_busiest_group() helpers end here *********************/
8463 * find_busiest_group - Returns the busiest group within the sched_domain
8464 * if there is an imbalance.
8466 * Also calculates the amount of runnable load which should be moved
8467 * to restore balance.
8469 * @env: The load balancing environment.
8471 * Return: - The busiest group if imbalance exists.
8473 static struct sched_group *find_busiest_group(struct lb_env *env)
8475 struct sg_lb_stats *local, *busiest;
8476 struct sd_lb_stats sds;
8478 init_sd_lb_stats(&sds);
8481 * Compute the various statistics relavent for load balancing at
8484 update_sd_lb_stats(env, &sds);
8486 if (sched_energy_enabled()) {
8487 struct root_domain *rd = env->dst_rq->rd;
8489 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
8493 local = &sds.local_stat;
8494 busiest = &sds.busiest_stat;
8496 /* ASYM feature bypasses nice load balance check */
8497 if (check_asym_packing(env, &sds))
8500 /* There is no busy sibling group to pull tasks from */
8501 if (!sds.busiest || busiest->sum_nr_running == 0)
8504 /* XXX broken for overlapping NUMA groups */
8505 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8506 / sds.total_capacity;
8509 * If the busiest group is imbalanced the below checks don't
8510 * work because they assume all things are equal, which typically
8511 * isn't true due to cpus_ptr constraints and the like.
8513 if (busiest->group_type == group_imbalanced)
8517 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8518 * capacities from resulting in underutilization due to avg_load.
8520 if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8521 busiest->group_no_capacity)
8524 /* Misfit tasks should be dealt with regardless of the avg load */
8525 if (busiest->group_type == group_misfit_task)
8529 * If the local group is busier than the selected busiest group
8530 * don't try and pull any tasks.
8532 if (local->avg_load >= busiest->avg_load)
8536 * Don't pull any tasks if this group is already above the domain
8539 if (local->avg_load >= sds.avg_load)
8542 if (env->idle == CPU_IDLE) {
8544 * This CPU is idle. If the busiest group is not overloaded
8545 * and there is no imbalance between this and busiest group
8546 * wrt idle CPUs, it is balanced. The imbalance becomes
8547 * significant if the diff is greater than 1 otherwise we
8548 * might end up to just move the imbalance on another group
8550 if ((busiest->group_type != group_overloaded) &&
8551 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8555 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8556 * imbalance_pct to be conservative.
8558 if (100 * busiest->avg_load <=
8559 env->sd->imbalance_pct * local->avg_load)
8564 /* Looks like there is an imbalance. Compute it */
8565 env->src_grp_type = busiest->group_type;
8566 calculate_imbalance(env, &sds);
8567 return env->imbalance ? sds.busiest : NULL;
8575 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
8577 static struct rq *find_busiest_queue(struct lb_env *env,
8578 struct sched_group *group)
8580 struct rq *busiest = NULL, *rq;
8581 unsigned long busiest_load = 0, busiest_capacity = 1;
8584 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8585 unsigned long capacity, load;
8589 rt = fbq_classify_rq(rq);
8592 * We classify groups/runqueues into three groups:
8593 * - regular: there are !numa tasks
8594 * - remote: there are numa tasks that run on the 'wrong' node
8595 * - all: there is no distinction
8597 * In order to avoid migrating ideally placed numa tasks,
8598 * ignore those when there's better options.
8600 * If we ignore the actual busiest queue to migrate another
8601 * task, the next balance pass can still reduce the busiest
8602 * queue by moving tasks around inside the node.
8604 * If we cannot move enough load due to this classification
8605 * the next pass will adjust the group classification and
8606 * allow migration of more tasks.
8608 * Both cases only affect the total convergence complexity.
8610 if (rt > env->fbq_type)
8614 * For ASYM_CPUCAPACITY domains with misfit tasks we simply
8615 * seek the "biggest" misfit task.
8617 if (env->src_grp_type == group_misfit_task) {
8618 if (rq->misfit_task_load > busiest_load) {
8619 busiest_load = rq->misfit_task_load;
8626 capacity = capacity_of(i);
8629 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
8630 * eventually lead to active_balancing high->low capacity.
8631 * Higher per-CPU capacity is considered better than balancing
8634 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8635 capacity_of(env->dst_cpu) < capacity &&
8636 rq->nr_running == 1)
8639 load = cpu_runnable_load(rq);
8642 * When comparing with imbalance, use cpu_runnable_load()
8643 * which is not scaled with the CPU capacity.
8646 if (rq->nr_running == 1 && load > env->imbalance &&
8647 !check_cpu_capacity(rq, env->sd))
8651 * For the load comparisons with the other CPU's, consider
8652 * the cpu_runnable_load() scaled with the CPU capacity, so
8653 * that the load can be moved away from the CPU that is
8654 * potentially running at a lower capacity.
8656 * Thus we're looking for max(load_i / capacity_i), crosswise
8657 * multiplication to rid ourselves of the division works out
8658 * to: load_i * capacity_j > load_j * capacity_i; where j is
8659 * our previous maximum.
8661 if (load * busiest_capacity > busiest_load * capacity) {
8662 busiest_load = load;
8663 busiest_capacity = capacity;
8672 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8673 * so long as it is large enough.
8675 #define MAX_PINNED_INTERVAL 512
8678 asym_active_balance(struct lb_env *env)
8681 * ASYM_PACKING needs to force migrate tasks from busy but
8682 * lower priority CPUs in order to pack all tasks in the
8683 * highest priority CPUs.
8685 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
8686 sched_asym_prefer(env->dst_cpu, env->src_cpu);
8690 voluntary_active_balance(struct lb_env *env)
8692 struct sched_domain *sd = env->sd;
8694 if (asym_active_balance(env))
8698 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8699 * It's worth migrating the task if the src_cpu's capacity is reduced
8700 * because of other sched_class or IRQs if more capacity stays
8701 * available on dst_cpu.
8703 if ((env->idle != CPU_NOT_IDLE) &&
8704 (env->src_rq->cfs.h_nr_running == 1)) {
8705 if ((check_cpu_capacity(env->src_rq, sd)) &&
8706 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8710 if (env->src_grp_type == group_misfit_task)
8716 static int need_active_balance(struct lb_env *env)
8718 struct sched_domain *sd = env->sd;
8720 if (voluntary_active_balance(env))
8723 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8726 static int active_load_balance_cpu_stop(void *data);
8728 static int should_we_balance(struct lb_env *env)
8730 struct sched_group *sg = env->sd->groups;
8731 int cpu, balance_cpu = -1;
8734 * Ensure the balancing environment is consistent; can happen
8735 * when the softirq triggers 'during' hotplug.
8737 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8741 * In the newly idle case, we will allow all the CPUs
8742 * to do the newly idle load balance.
8744 if (env->idle == CPU_NEWLY_IDLE)
8747 /* Try to find first idle CPU */
8748 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8756 if (balance_cpu == -1)
8757 balance_cpu = group_balance_cpu(sg);
8760 * First idle CPU or the first CPU(busiest) in this sched group
8761 * is eligible for doing load balancing at this and above domains.
8763 return balance_cpu == env->dst_cpu;
8767 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8768 * tasks if there is an imbalance.
8770 static int load_balance(int this_cpu, struct rq *this_rq,
8771 struct sched_domain *sd, enum cpu_idle_type idle,
8772 int *continue_balancing)
8774 int ld_moved, cur_ld_moved, active_balance = 0;
8775 struct sched_domain *sd_parent = sd->parent;
8776 struct sched_group *group;
8779 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8781 struct lb_env env = {
8783 .dst_cpu = this_cpu,
8785 .dst_grpmask = sched_group_span(sd->groups),
8787 .loop_break = sched_nr_migrate_break,
8790 .tasks = LIST_HEAD_INIT(env.tasks),
8793 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8795 schedstat_inc(sd->lb_count[idle]);
8798 if (!should_we_balance(&env)) {
8799 *continue_balancing = 0;
8803 group = find_busiest_group(&env);
8805 schedstat_inc(sd->lb_nobusyg[idle]);
8809 busiest = find_busiest_queue(&env, group);
8811 schedstat_inc(sd->lb_nobusyq[idle]);
8815 BUG_ON(busiest == env.dst_rq);
8817 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8819 env.src_cpu = busiest->cpu;
8820 env.src_rq = busiest;
8823 if (busiest->nr_running > 1) {
8825 * Attempt to move tasks. If find_busiest_group has found
8826 * an imbalance but busiest->nr_running <= 1, the group is
8827 * still unbalanced. ld_moved simply stays zero, so it is
8828 * correctly treated as an imbalance.
8830 env.flags |= LBF_ALL_PINNED;
8831 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8834 rq_lock_irqsave(busiest, &rf);
8835 update_rq_clock(busiest);
8838 * cur_ld_moved - load moved in current iteration
8839 * ld_moved - cumulative load moved across iterations
8841 cur_ld_moved = detach_tasks(&env);
8844 * We've detached some tasks from busiest_rq. Every
8845 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8846 * unlock busiest->lock, and we are able to be sure
8847 * that nobody can manipulate the tasks in parallel.
8848 * See task_rq_lock() family for the details.
8851 rq_unlock(busiest, &rf);
8855 ld_moved += cur_ld_moved;
8858 local_irq_restore(rf.flags);
8860 if (env.flags & LBF_NEED_BREAK) {
8861 env.flags &= ~LBF_NEED_BREAK;
8866 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8867 * us and move them to an alternate dst_cpu in our sched_group
8868 * where they can run. The upper limit on how many times we
8869 * iterate on same src_cpu is dependent on number of CPUs in our
8872 * This changes load balance semantics a bit on who can move
8873 * load to a given_cpu. In addition to the given_cpu itself
8874 * (or a ilb_cpu acting on its behalf where given_cpu is
8875 * nohz-idle), we now have balance_cpu in a position to move
8876 * load to given_cpu. In rare situations, this may cause
8877 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8878 * _independently_ and at _same_ time to move some load to
8879 * given_cpu) causing exceess load to be moved to given_cpu.
8880 * This however should not happen so much in practice and
8881 * moreover subsequent load balance cycles should correct the
8882 * excess load moved.
8884 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8886 /* Prevent to re-select dst_cpu via env's CPUs */
8887 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
8889 env.dst_rq = cpu_rq(env.new_dst_cpu);
8890 env.dst_cpu = env.new_dst_cpu;
8891 env.flags &= ~LBF_DST_PINNED;
8893 env.loop_break = sched_nr_migrate_break;
8896 * Go back to "more_balance" rather than "redo" since we
8897 * need to continue with same src_cpu.
8903 * We failed to reach balance because of affinity.
8906 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8908 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8909 *group_imbalance = 1;
8912 /* All tasks on this runqueue were pinned by CPU affinity */
8913 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8914 __cpumask_clear_cpu(cpu_of(busiest), cpus);
8916 * Attempting to continue load balancing at the current
8917 * sched_domain level only makes sense if there are
8918 * active CPUs remaining as possible busiest CPUs to
8919 * pull load from which are not contained within the
8920 * destination group that is receiving any migrated
8923 if (!cpumask_subset(cpus, env.dst_grpmask)) {
8925 env.loop_break = sched_nr_migrate_break;
8928 goto out_all_pinned;
8933 schedstat_inc(sd->lb_failed[idle]);
8935 * Increment the failure counter only on periodic balance.
8936 * We do not want newidle balance, which can be very
8937 * frequent, pollute the failure counter causing
8938 * excessive cache_hot migrations and active balances.
8940 if (idle != CPU_NEWLY_IDLE)
8941 sd->nr_balance_failed++;
8943 if (need_active_balance(&env)) {
8944 unsigned long flags;
8946 raw_spin_lock_irqsave(&busiest->lock, flags);
8949 * Don't kick the active_load_balance_cpu_stop,
8950 * if the curr task on busiest CPU can't be
8951 * moved to this_cpu:
8953 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
8954 raw_spin_unlock_irqrestore(&busiest->lock,
8956 env.flags |= LBF_ALL_PINNED;
8957 goto out_one_pinned;
8961 * ->active_balance synchronizes accesses to
8962 * ->active_balance_work. Once set, it's cleared
8963 * only after active load balance is finished.
8965 if (!busiest->active_balance) {
8966 busiest->active_balance = 1;
8967 busiest->push_cpu = this_cpu;
8970 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8972 if (active_balance) {
8973 stop_one_cpu_nowait(cpu_of(busiest),
8974 active_load_balance_cpu_stop, busiest,
8975 &busiest->active_balance_work);
8978 /* We've kicked active balancing, force task migration. */
8979 sd->nr_balance_failed = sd->cache_nice_tries+1;
8982 sd->nr_balance_failed = 0;
8984 if (likely(!active_balance) || voluntary_active_balance(&env)) {
8985 /* We were unbalanced, so reset the balancing interval */
8986 sd->balance_interval = sd->min_interval;
8989 * If we've begun active balancing, start to back off. This
8990 * case may not be covered by the all_pinned logic if there
8991 * is only 1 task on the busy runqueue (because we don't call
8994 if (sd->balance_interval < sd->max_interval)
8995 sd->balance_interval *= 2;
9002 * We reach balance although we may have faced some affinity
9003 * constraints. Clear the imbalance flag only if other tasks got
9004 * a chance to move and fix the imbalance.
9006 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
9007 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9009 if (*group_imbalance)
9010 *group_imbalance = 0;
9015 * We reach balance because all tasks are pinned at this level so
9016 * we can't migrate them. Let the imbalance flag set so parent level
9017 * can try to migrate them.
9019 schedstat_inc(sd->lb_balanced[idle]);
9021 sd->nr_balance_failed = 0;
9027 * newidle_balance() disregards balance intervals, so we could
9028 * repeatedly reach this code, which would lead to balance_interval
9029 * skyrocketting in a short amount of time. Skip the balance_interval
9030 * increase logic to avoid that.
9032 if (env.idle == CPU_NEWLY_IDLE)
9035 /* tune up the balancing interval */
9036 if ((env.flags & LBF_ALL_PINNED &&
9037 sd->balance_interval < MAX_PINNED_INTERVAL) ||
9038 sd->balance_interval < sd->max_interval)
9039 sd->balance_interval *= 2;
9044 static inline unsigned long
9045 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
9047 unsigned long interval = sd->balance_interval;
9050 interval *= sd->busy_factor;
9052 /* scale ms to jiffies */
9053 interval = msecs_to_jiffies(interval);
9054 interval = clamp(interval, 1UL, max_load_balance_interval);
9060 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
9062 unsigned long interval, next;
9064 /* used by idle balance, so cpu_busy = 0 */
9065 interval = get_sd_balance_interval(sd, 0);
9066 next = sd->last_balance + interval;
9068 if (time_after(*next_balance, next))
9069 *next_balance = next;
9073 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9074 * running tasks off the busiest CPU onto idle CPUs. It requires at
9075 * least 1 task to be running on each physical CPU where possible, and
9076 * avoids physical / logical imbalances.
9078 static int active_load_balance_cpu_stop(void *data)
9080 struct rq *busiest_rq = data;
9081 int busiest_cpu = cpu_of(busiest_rq);
9082 int target_cpu = busiest_rq->push_cpu;
9083 struct rq *target_rq = cpu_rq(target_cpu);
9084 struct sched_domain *sd;
9085 struct task_struct *p = NULL;
9088 rq_lock_irq(busiest_rq, &rf);
9090 * Between queueing the stop-work and running it is a hole in which
9091 * CPUs can become inactive. We should not move tasks from or to
9094 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
9097 /* Make sure the requested CPU hasn't gone down in the meantime: */
9098 if (unlikely(busiest_cpu != smp_processor_id() ||
9099 !busiest_rq->active_balance))
9102 /* Is there any task to move? */
9103 if (busiest_rq->nr_running <= 1)
9107 * This condition is "impossible", if it occurs
9108 * we need to fix it. Originally reported by
9109 * Bjorn Helgaas on a 128-CPU setup.
9111 BUG_ON(busiest_rq == target_rq);
9113 /* Search for an sd spanning us and the target CPU. */
9115 for_each_domain(target_cpu, sd) {
9116 if ((sd->flags & SD_LOAD_BALANCE) &&
9117 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9122 struct lb_env env = {
9124 .dst_cpu = target_cpu,
9125 .dst_rq = target_rq,
9126 .src_cpu = busiest_rq->cpu,
9127 .src_rq = busiest_rq,
9130 * can_migrate_task() doesn't need to compute new_dst_cpu
9131 * for active balancing. Since we have CPU_IDLE, but no
9132 * @dst_grpmask we need to make that test go away with lying
9135 .flags = LBF_DST_PINNED,
9138 schedstat_inc(sd->alb_count);
9139 update_rq_clock(busiest_rq);
9141 p = detach_one_task(&env);
9143 schedstat_inc(sd->alb_pushed);
9144 /* Active balancing done, reset the failure counter. */
9145 sd->nr_balance_failed = 0;
9147 schedstat_inc(sd->alb_failed);
9152 busiest_rq->active_balance = 0;
9153 rq_unlock(busiest_rq, &rf);
9156 attach_one_task(target_rq, p);
9163 static DEFINE_SPINLOCK(balancing);
9166 * Scale the max load_balance interval with the number of CPUs in the system.
9167 * This trades load-balance latency on larger machines for less cross talk.
9169 void update_max_interval(void)
9171 max_load_balance_interval = HZ*num_online_cpus()/10;
9175 * It checks each scheduling domain to see if it is due to be balanced,
9176 * and initiates a balancing operation if so.
9178 * Balancing parameters are set up in init_sched_domains.
9180 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9182 int continue_balancing = 1;
9184 unsigned long interval;
9185 struct sched_domain *sd;
9186 /* Earliest time when we have to do rebalance again */
9187 unsigned long next_balance = jiffies + 60*HZ;
9188 int update_next_balance = 0;
9189 int need_serialize, need_decay = 0;
9193 for_each_domain(cpu, sd) {
9195 * Decay the newidle max times here because this is a regular
9196 * visit to all the domains. Decay ~1% per second.
9198 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9199 sd->max_newidle_lb_cost =
9200 (sd->max_newidle_lb_cost * 253) / 256;
9201 sd->next_decay_max_lb_cost = jiffies + HZ;
9204 max_cost += sd->max_newidle_lb_cost;
9206 if (!(sd->flags & SD_LOAD_BALANCE))
9210 * Stop the load balance at this level. There is another
9211 * CPU in our sched group which is doing load balancing more
9214 if (!continue_balancing) {
9220 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9222 need_serialize = sd->flags & SD_SERIALIZE;
9223 if (need_serialize) {
9224 if (!spin_trylock(&balancing))
9228 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9229 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9231 * The LBF_DST_PINNED logic could have changed
9232 * env->dst_cpu, so we can't know our idle
9233 * state even if we migrated tasks. Update it.
9235 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9237 sd->last_balance = jiffies;
9238 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9241 spin_unlock(&balancing);
9243 if (time_after(next_balance, sd->last_balance + interval)) {
9244 next_balance = sd->last_balance + interval;
9245 update_next_balance = 1;
9250 * Ensure the rq-wide value also decays but keep it at a
9251 * reasonable floor to avoid funnies with rq->avg_idle.
9253 rq->max_idle_balance_cost =
9254 max((u64)sysctl_sched_migration_cost, max_cost);
9259 * next_balance will be updated only when there is a need.
9260 * When the cpu is attached to null domain for ex, it will not be
9263 if (likely(update_next_balance)) {
9264 rq->next_balance = next_balance;
9266 #ifdef CONFIG_NO_HZ_COMMON
9268 * If this CPU has been elected to perform the nohz idle
9269 * balance. Other idle CPUs have already rebalanced with
9270 * nohz_idle_balance() and nohz.next_balance has been
9271 * updated accordingly. This CPU is now running the idle load
9272 * balance for itself and we need to update the
9273 * nohz.next_balance accordingly.
9275 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9276 nohz.next_balance = rq->next_balance;
9281 static inline int on_null_domain(struct rq *rq)
9283 return unlikely(!rcu_dereference_sched(rq->sd));
9286 #ifdef CONFIG_NO_HZ_COMMON
9288 * idle load balancing details
9289 * - When one of the busy CPUs notice that there may be an idle rebalancing
9290 * needed, they will kick the idle load balancer, which then does idle
9291 * load balancing for all the idle CPUs.
9292 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
9296 static inline int find_new_ilb(void)
9300 for_each_cpu_and(ilb, nohz.idle_cpus_mask,
9301 housekeeping_cpumask(HK_FLAG_MISC)) {
9310 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
9311 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
9313 static void kick_ilb(unsigned int flags)
9317 nohz.next_balance++;
9319 ilb_cpu = find_new_ilb();
9321 if (ilb_cpu >= nr_cpu_ids)
9324 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
9325 if (flags & NOHZ_KICK_MASK)
9329 * Use smp_send_reschedule() instead of resched_cpu().
9330 * This way we generate a sched IPI on the target CPU which
9331 * is idle. And the softirq performing nohz idle load balance
9332 * will be run before returning from the IPI.
9334 smp_send_reschedule(ilb_cpu);
9338 * Current decision point for kicking the idle load balancer in the presence
9339 * of idle CPUs in the system.
9341 static void nohz_balancer_kick(struct rq *rq)
9343 unsigned long now = jiffies;
9344 struct sched_domain_shared *sds;
9345 struct sched_domain *sd;
9346 int nr_busy, i, cpu = rq->cpu;
9347 unsigned int flags = 0;
9349 if (unlikely(rq->idle_balance))
9353 * We may be recently in ticked or tickless idle mode. At the first
9354 * busy tick after returning from idle, we will update the busy stats.
9356 nohz_balance_exit_idle(rq);
9359 * None are in tickless mode and hence no need for NOHZ idle load
9362 if (likely(!atomic_read(&nohz.nr_cpus)))
9365 if (READ_ONCE(nohz.has_blocked) &&
9366 time_after(now, READ_ONCE(nohz.next_blocked)))
9367 flags = NOHZ_STATS_KICK;
9369 if (time_before(now, nohz.next_balance))
9372 if (rq->nr_running >= 2) {
9373 flags = NOHZ_KICK_MASK;
9379 sd = rcu_dereference(rq->sd);
9382 * If there's a CFS task and the current CPU has reduced
9383 * capacity; kick the ILB to see if there's a better CPU to run
9386 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
9387 flags = NOHZ_KICK_MASK;
9392 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
9395 * When ASYM_PACKING; see if there's a more preferred CPU
9396 * currently idle; in which case, kick the ILB to move tasks
9399 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
9400 if (sched_asym_prefer(i, cpu)) {
9401 flags = NOHZ_KICK_MASK;
9407 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
9410 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
9411 * to run the misfit task on.
9413 if (check_misfit_status(rq, sd)) {
9414 flags = NOHZ_KICK_MASK;
9419 * For asymmetric systems, we do not want to nicely balance
9420 * cache use, instead we want to embrace asymmetry and only
9421 * ensure tasks have enough CPU capacity.
9423 * Skip the LLC logic because it's not relevant in that case.
9428 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9431 * If there is an imbalance between LLC domains (IOW we could
9432 * increase the overall cache use), we need some less-loaded LLC
9433 * domain to pull some load. Likewise, we may need to spread
9434 * load within the current LLC domain (e.g. packed SMT cores but
9435 * other CPUs are idle). We can't really know from here how busy
9436 * the others are - so just get a nohz balance going if it looks
9437 * like this LLC domain has tasks we could move.
9439 nr_busy = atomic_read(&sds->nr_busy_cpus);
9441 flags = NOHZ_KICK_MASK;
9452 static void set_cpu_sd_state_busy(int cpu)
9454 struct sched_domain *sd;
9457 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9459 if (!sd || !sd->nohz_idle)
9463 atomic_inc(&sd->shared->nr_busy_cpus);
9468 void nohz_balance_exit_idle(struct rq *rq)
9470 SCHED_WARN_ON(rq != this_rq());
9472 if (likely(!rq->nohz_tick_stopped))
9475 rq->nohz_tick_stopped = 0;
9476 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
9477 atomic_dec(&nohz.nr_cpus);
9479 set_cpu_sd_state_busy(rq->cpu);
9482 static void set_cpu_sd_state_idle(int cpu)
9484 struct sched_domain *sd;
9487 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9489 if (!sd || sd->nohz_idle)
9493 atomic_dec(&sd->shared->nr_busy_cpus);
9499 * This routine will record that the CPU is going idle with tick stopped.
9500 * This info will be used in performing idle load balancing in the future.
9502 void nohz_balance_enter_idle(int cpu)
9504 struct rq *rq = cpu_rq(cpu);
9506 SCHED_WARN_ON(cpu != smp_processor_id());
9508 /* If this CPU is going down, then nothing needs to be done: */
9509 if (!cpu_active(cpu))
9512 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9513 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9517 * Can be set safely without rq->lock held
9518 * If a clear happens, it will have evaluated last additions because
9519 * rq->lock is held during the check and the clear
9521 rq->has_blocked_load = 1;
9524 * The tick is still stopped but load could have been added in the
9525 * meantime. We set the nohz.has_blocked flag to trig a check of the
9526 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
9527 * of nohz.has_blocked can only happen after checking the new load
9529 if (rq->nohz_tick_stopped)
9532 /* If we're a completely isolated CPU, we don't play: */
9533 if (on_null_domain(rq))
9536 rq->nohz_tick_stopped = 1;
9538 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9539 atomic_inc(&nohz.nr_cpus);
9542 * Ensures that if nohz_idle_balance() fails to observe our
9543 * @idle_cpus_mask store, it must observe the @has_blocked
9546 smp_mb__after_atomic();
9548 set_cpu_sd_state_idle(cpu);
9552 * Each time a cpu enter idle, we assume that it has blocked load and
9553 * enable the periodic update of the load of idle cpus
9555 WRITE_ONCE(nohz.has_blocked, 1);
9559 * Internal function that runs load balance for all idle cpus. The load balance
9560 * can be a simple update of blocked load or a complete load balance with
9561 * tasks movement depending of flags.
9562 * The function returns false if the loop has stopped before running
9563 * through all idle CPUs.
9565 static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
9566 enum cpu_idle_type idle)
9568 /* Earliest time when we have to do rebalance again */
9569 unsigned long now = jiffies;
9570 unsigned long next_balance = now + 60*HZ;
9571 bool has_blocked_load = false;
9572 int update_next_balance = 0;
9573 int this_cpu = this_rq->cpu;
9578 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
9581 * We assume there will be no idle load after this update and clear
9582 * the has_blocked flag. If a cpu enters idle in the mean time, it will
9583 * set the has_blocked flag and trig another update of idle load.
9584 * Because a cpu that becomes idle, is added to idle_cpus_mask before
9585 * setting the flag, we are sure to not clear the state and not
9586 * check the load of an idle cpu.
9588 WRITE_ONCE(nohz.has_blocked, 0);
9591 * Ensures that if we miss the CPU, we must see the has_blocked
9592 * store from nohz_balance_enter_idle().
9596 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9597 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9601 * If this CPU gets work to do, stop the load balancing
9602 * work being done for other CPUs. Next load
9603 * balancing owner will pick it up.
9605 if (need_resched()) {
9606 has_blocked_load = true;
9610 rq = cpu_rq(balance_cpu);
9612 has_blocked_load |= update_nohz_stats(rq, true);
9615 * If time for next balance is due,
9618 if (time_after_eq(jiffies, rq->next_balance)) {
9621 rq_lock_irqsave(rq, &rf);
9622 update_rq_clock(rq);
9623 rq_unlock_irqrestore(rq, &rf);
9625 if (flags & NOHZ_BALANCE_KICK)
9626 rebalance_domains(rq, CPU_IDLE);
9629 if (time_after(next_balance, rq->next_balance)) {
9630 next_balance = rq->next_balance;
9631 update_next_balance = 1;
9635 /* Newly idle CPU doesn't need an update */
9636 if (idle != CPU_NEWLY_IDLE) {
9637 update_blocked_averages(this_cpu);
9638 has_blocked_load |= this_rq->has_blocked_load;
9641 if (flags & NOHZ_BALANCE_KICK)
9642 rebalance_domains(this_rq, CPU_IDLE);
9644 WRITE_ONCE(nohz.next_blocked,
9645 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
9647 /* The full idle balance loop has been done */
9651 /* There is still blocked load, enable periodic update */
9652 if (has_blocked_load)
9653 WRITE_ONCE(nohz.has_blocked, 1);
9656 * next_balance will be updated only when there is a need.
9657 * When the CPU is attached to null domain for ex, it will not be
9660 if (likely(update_next_balance))
9661 nohz.next_balance = next_balance;
9667 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9668 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9670 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9672 int this_cpu = this_rq->cpu;
9675 if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
9678 if (idle != CPU_IDLE) {
9679 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9683 /* could be _relaxed() */
9684 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9685 if (!(flags & NOHZ_KICK_MASK))
9688 _nohz_idle_balance(this_rq, flags, idle);
9693 static void nohz_newidle_balance(struct rq *this_rq)
9695 int this_cpu = this_rq->cpu;
9698 * This CPU doesn't want to be disturbed by scheduler
9701 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
9704 /* Will wake up very soon. No time for doing anything else*/
9705 if (this_rq->avg_idle < sysctl_sched_migration_cost)
9708 /* Don't need to update blocked load of idle CPUs*/
9709 if (!READ_ONCE(nohz.has_blocked) ||
9710 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
9713 raw_spin_unlock(&this_rq->lock);
9715 * This CPU is going to be idle and blocked load of idle CPUs
9716 * need to be updated. Run the ilb locally as it is a good
9717 * candidate for ilb instead of waking up another idle CPU.
9718 * Kick an normal ilb if we failed to do the update.
9720 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
9721 kick_ilb(NOHZ_STATS_KICK);
9722 raw_spin_lock(&this_rq->lock);
9725 #else /* !CONFIG_NO_HZ_COMMON */
9726 static inline void nohz_balancer_kick(struct rq *rq) { }
9728 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9733 static inline void nohz_newidle_balance(struct rq *this_rq) { }
9734 #endif /* CONFIG_NO_HZ_COMMON */
9737 * idle_balance is called by schedule() if this_cpu is about to become
9738 * idle. Attempts to pull tasks from other CPUs.
9740 int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
9742 unsigned long next_balance = jiffies + HZ;
9743 int this_cpu = this_rq->cpu;
9744 struct sched_domain *sd;
9745 int pulled_task = 0;
9748 update_misfit_status(NULL, this_rq);
9750 * We must set idle_stamp _before_ calling idle_balance(), such that we
9751 * measure the duration of idle_balance() as idle time.
9753 this_rq->idle_stamp = rq_clock(this_rq);
9756 * Do not pull tasks towards !active CPUs...
9758 if (!cpu_active(this_cpu))
9762 * This is OK, because current is on_cpu, which avoids it being picked
9763 * for load-balance and preemption/IRQs are still disabled avoiding
9764 * further scheduler activity on it and we're being very careful to
9765 * re-start the picking loop.
9767 rq_unpin_lock(this_rq, rf);
9769 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
9770 !READ_ONCE(this_rq->rd->overload)) {
9773 sd = rcu_dereference_check_sched_domain(this_rq->sd);
9775 update_next_balance(sd, &next_balance);
9778 nohz_newidle_balance(this_rq);
9783 raw_spin_unlock(&this_rq->lock);
9785 update_blocked_averages(this_cpu);
9787 for_each_domain(this_cpu, sd) {
9788 int continue_balancing = 1;
9789 u64 t0, domain_cost;
9791 if (!(sd->flags & SD_LOAD_BALANCE))
9794 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
9795 update_next_balance(sd, &next_balance);
9799 if (sd->flags & SD_BALANCE_NEWIDLE) {
9800 t0 = sched_clock_cpu(this_cpu);
9802 pulled_task = load_balance(this_cpu, this_rq,
9804 &continue_balancing);
9806 domain_cost = sched_clock_cpu(this_cpu) - t0;
9807 if (domain_cost > sd->max_newidle_lb_cost)
9808 sd->max_newidle_lb_cost = domain_cost;
9810 curr_cost += domain_cost;
9813 update_next_balance(sd, &next_balance);
9816 * Stop searching for tasks to pull if there are
9817 * now runnable tasks on this rq.
9819 if (pulled_task || this_rq->nr_running > 0)
9824 raw_spin_lock(&this_rq->lock);
9826 if (curr_cost > this_rq->max_idle_balance_cost)
9827 this_rq->max_idle_balance_cost = curr_cost;
9831 * While browsing the domains, we released the rq lock, a task could
9832 * have been enqueued in the meantime. Since we're not going idle,
9833 * pretend we pulled a task.
9835 if (this_rq->cfs.h_nr_running && !pulled_task)
9838 /* Move the next balance forward */
9839 if (time_after(this_rq->next_balance, next_balance))
9840 this_rq->next_balance = next_balance;
9842 /* Is there a task of a high priority class? */
9843 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
9847 this_rq->idle_stamp = 0;
9849 rq_repin_lock(this_rq, rf);
9855 * run_rebalance_domains is triggered when needed from the scheduler tick.
9856 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9858 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9860 struct rq *this_rq = this_rq();
9861 enum cpu_idle_type idle = this_rq->idle_balance ?
9862 CPU_IDLE : CPU_NOT_IDLE;
9865 * If this CPU has a pending nohz_balance_kick, then do the
9866 * balancing on behalf of the other idle CPUs whose ticks are
9867 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9868 * give the idle CPUs a chance to load balance. Else we may
9869 * load balance only within the local sched_domain hierarchy
9870 * and abort nohz_idle_balance altogether if we pull some load.
9872 if (nohz_idle_balance(this_rq, idle))
9875 /* normal load balance */
9876 update_blocked_averages(this_rq->cpu);
9877 rebalance_domains(this_rq, idle);
9881 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9883 void trigger_load_balance(struct rq *rq)
9885 /* Don't need to rebalance while attached to NULL domain */
9886 if (unlikely(on_null_domain(rq)))
9889 if (time_after_eq(jiffies, rq->next_balance))
9890 raise_softirq(SCHED_SOFTIRQ);
9892 nohz_balancer_kick(rq);
9895 static void rq_online_fair(struct rq *rq)
9899 update_runtime_enabled(rq);
9902 static void rq_offline_fair(struct rq *rq)
9906 /* Ensure any throttled groups are reachable by pick_next_task */
9907 unthrottle_offline_cfs_rqs(rq);
9910 #endif /* CONFIG_SMP */
9913 * scheduler tick hitting a task of our scheduling class.
9915 * NOTE: This function can be called remotely by the tick offload that
9916 * goes along full dynticks. Therefore no local assumption can be made
9917 * and everything must be accessed through the @rq and @curr passed in
9920 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9922 struct cfs_rq *cfs_rq;
9923 struct sched_entity *se = &curr->se;
9925 for_each_sched_entity(se) {
9926 cfs_rq = cfs_rq_of(se);
9927 entity_tick(cfs_rq, se, queued);
9930 if (static_branch_unlikely(&sched_numa_balancing))
9931 task_tick_numa(rq, curr);
9933 update_misfit_status(curr, rq);
9934 update_overutilized_status(task_rq(curr));
9938 * called on fork with the child task as argument from the parent's context
9939 * - child not yet on the tasklist
9940 * - preemption disabled
9942 static void task_fork_fair(struct task_struct *p)
9944 struct cfs_rq *cfs_rq;
9945 struct sched_entity *se = &p->se, *curr;
9946 struct rq *rq = this_rq();
9950 update_rq_clock(rq);
9952 cfs_rq = task_cfs_rq(current);
9953 curr = cfs_rq->curr;
9955 update_curr(cfs_rq);
9956 se->vruntime = curr->vruntime;
9958 place_entity(cfs_rq, se, 1);
9960 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9962 * Upon rescheduling, sched_class::put_prev_task() will place
9963 * 'current' within the tree based on its new key value.
9965 swap(curr->vruntime, se->vruntime);
9969 se->vruntime -= cfs_rq->min_vruntime;
9974 * Priority of the task has changed. Check to see if we preempt
9978 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9980 if (!task_on_rq_queued(p))
9984 * Reschedule if we are currently running on this runqueue and
9985 * our priority decreased, or if we are not currently running on
9986 * this runqueue and our priority is higher than the current's
9988 if (rq->curr == p) {
9989 if (p->prio > oldprio)
9992 check_preempt_curr(rq, p, 0);
9995 static inline bool vruntime_normalized(struct task_struct *p)
9997 struct sched_entity *se = &p->se;
10000 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
10001 * the dequeue_entity(.flags=0) will already have normalized the
10008 * When !on_rq, vruntime of the task has usually NOT been normalized.
10009 * But there are some cases where it has already been normalized:
10011 * - A forked child which is waiting for being woken up by
10012 * wake_up_new_task().
10013 * - A task which has been woken up by try_to_wake_up() and
10014 * waiting for actually being woken up by sched_ttwu_pending().
10016 if (!se->sum_exec_runtime ||
10017 (p->state == TASK_WAKING && p->sched_remote_wakeup))
10023 #ifdef CONFIG_FAIR_GROUP_SCHED
10025 * Propagate the changes of the sched_entity across the tg tree to make it
10026 * visible to the root
10028 static void propagate_entity_cfs_rq(struct sched_entity *se)
10030 struct cfs_rq *cfs_rq;
10032 /* Start to propagate at parent */
10035 for_each_sched_entity(se) {
10036 cfs_rq = cfs_rq_of(se);
10038 if (cfs_rq_throttled(cfs_rq))
10041 update_load_avg(cfs_rq, se, UPDATE_TG);
10045 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
10048 static void detach_entity_cfs_rq(struct sched_entity *se)
10050 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10052 /* Catch up with the cfs_rq and remove our load when we leave */
10053 update_load_avg(cfs_rq, se, 0);
10054 detach_entity_load_avg(cfs_rq, se);
10055 update_tg_load_avg(cfs_rq, false);
10056 propagate_entity_cfs_rq(se);
10059 static void attach_entity_cfs_rq(struct sched_entity *se)
10061 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10063 #ifdef CONFIG_FAIR_GROUP_SCHED
10065 * Since the real-depth could have been changed (only FAIR
10066 * class maintain depth value), reset depth properly.
10068 se->depth = se->parent ? se->parent->depth + 1 : 0;
10071 /* Synchronize entity with its cfs_rq */
10072 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
10073 attach_entity_load_avg(cfs_rq, se, 0);
10074 update_tg_load_avg(cfs_rq, false);
10075 propagate_entity_cfs_rq(se);
10078 static void detach_task_cfs_rq(struct task_struct *p)
10080 struct sched_entity *se = &p->se;
10081 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10083 if (!vruntime_normalized(p)) {
10085 * Fix up our vruntime so that the current sleep doesn't
10086 * cause 'unlimited' sleep bonus.
10088 place_entity(cfs_rq, se, 0);
10089 se->vruntime -= cfs_rq->min_vruntime;
10092 detach_entity_cfs_rq(se);
10095 static void attach_task_cfs_rq(struct task_struct *p)
10097 struct sched_entity *se = &p->se;
10098 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10100 attach_entity_cfs_rq(se);
10102 if (!vruntime_normalized(p))
10103 se->vruntime += cfs_rq->min_vruntime;
10106 static void switched_from_fair(struct rq *rq, struct task_struct *p)
10108 detach_task_cfs_rq(p);
10111 static void switched_to_fair(struct rq *rq, struct task_struct *p)
10113 attach_task_cfs_rq(p);
10115 if (task_on_rq_queued(p)) {
10117 * We were most likely switched from sched_rt, so
10118 * kick off the schedule if running, otherwise just see
10119 * if we can still preempt the current task.
10124 check_preempt_curr(rq, p, 0);
10128 /* Account for a task changing its policy or group.
10130 * This routine is mostly called to set cfs_rq->curr field when a task
10131 * migrates between groups/classes.
10133 static void set_next_task_fair(struct rq *rq, struct task_struct *p)
10135 struct sched_entity *se = &p->se;
10138 if (task_on_rq_queued(p)) {
10140 * Move the next running task to the front of the list, so our
10141 * cfs_tasks list becomes MRU one.
10143 list_move(&se->group_node, &rq->cfs_tasks);
10147 for_each_sched_entity(se) {
10148 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10150 set_next_entity(cfs_rq, se);
10151 /* ensure bandwidth has been allocated on our new cfs_rq */
10152 account_cfs_rq_runtime(cfs_rq, 0);
10156 void init_cfs_rq(struct cfs_rq *cfs_rq)
10158 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
10159 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
10160 #ifndef CONFIG_64BIT
10161 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
10164 raw_spin_lock_init(&cfs_rq->removed.lock);
10168 #ifdef CONFIG_FAIR_GROUP_SCHED
10169 static void task_set_group_fair(struct task_struct *p)
10171 struct sched_entity *se = &p->se;
10173 set_task_rq(p, task_cpu(p));
10174 se->depth = se->parent ? se->parent->depth + 1 : 0;
10177 static void task_move_group_fair(struct task_struct *p)
10179 detach_task_cfs_rq(p);
10180 set_task_rq(p, task_cpu(p));
10183 /* Tell se's cfs_rq has been changed -- migrated */
10184 p->se.avg.last_update_time = 0;
10186 attach_task_cfs_rq(p);
10189 static void task_change_group_fair(struct task_struct *p, int type)
10192 case TASK_SET_GROUP:
10193 task_set_group_fair(p);
10196 case TASK_MOVE_GROUP:
10197 task_move_group_fair(p);
10202 void free_fair_sched_group(struct task_group *tg)
10206 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
10208 for_each_possible_cpu(i) {
10210 kfree(tg->cfs_rq[i]);
10219 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10221 struct sched_entity *se;
10222 struct cfs_rq *cfs_rq;
10225 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
10228 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
10232 tg->shares = NICE_0_LOAD;
10234 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
10236 for_each_possible_cpu(i) {
10237 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
10238 GFP_KERNEL, cpu_to_node(i));
10242 se = kzalloc_node(sizeof(struct sched_entity),
10243 GFP_KERNEL, cpu_to_node(i));
10247 init_cfs_rq(cfs_rq);
10248 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
10249 init_entity_runnable_average(se);
10260 void online_fair_sched_group(struct task_group *tg)
10262 struct sched_entity *se;
10263 struct rq_flags rf;
10267 for_each_possible_cpu(i) {
10270 rq_lock_irq(rq, &rf);
10271 update_rq_clock(rq);
10272 attach_entity_cfs_rq(se);
10273 sync_throttle(tg, i);
10274 rq_unlock_irq(rq, &rf);
10278 void unregister_fair_sched_group(struct task_group *tg)
10280 unsigned long flags;
10284 for_each_possible_cpu(cpu) {
10286 remove_entity_load_avg(tg->se[cpu]);
10289 * Only empty task groups can be destroyed; so we can speculatively
10290 * check on_list without danger of it being re-added.
10292 if (!tg->cfs_rq[cpu]->on_list)
10297 raw_spin_lock_irqsave(&rq->lock, flags);
10298 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
10299 raw_spin_unlock_irqrestore(&rq->lock, flags);
10303 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
10304 struct sched_entity *se, int cpu,
10305 struct sched_entity *parent)
10307 struct rq *rq = cpu_rq(cpu);
10311 init_cfs_rq_runtime(cfs_rq);
10313 tg->cfs_rq[cpu] = cfs_rq;
10316 /* se could be NULL for root_task_group */
10321 se->cfs_rq = &rq->cfs;
10324 se->cfs_rq = parent->my_q;
10325 se->depth = parent->depth + 1;
10329 /* guarantee group entities always have weight */
10330 update_load_set(&se->load, NICE_0_LOAD);
10331 se->parent = parent;
10334 static DEFINE_MUTEX(shares_mutex);
10336 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10341 * We can't change the weight of the root cgroup.
10346 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10348 mutex_lock(&shares_mutex);
10349 if (tg->shares == shares)
10352 tg->shares = shares;
10353 for_each_possible_cpu(i) {
10354 struct rq *rq = cpu_rq(i);
10355 struct sched_entity *se = tg->se[i];
10356 struct rq_flags rf;
10358 /* Propagate contribution to hierarchy */
10359 rq_lock_irqsave(rq, &rf);
10360 update_rq_clock(rq);
10361 for_each_sched_entity(se) {
10362 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
10363 update_cfs_group(se);
10365 rq_unlock_irqrestore(rq, &rf);
10369 mutex_unlock(&shares_mutex);
10372 #else /* CONFIG_FAIR_GROUP_SCHED */
10374 void free_fair_sched_group(struct task_group *tg) { }
10376 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10381 void online_fair_sched_group(struct task_group *tg) { }
10383 void unregister_fair_sched_group(struct task_group *tg) { }
10385 #endif /* CONFIG_FAIR_GROUP_SCHED */
10388 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10390 struct sched_entity *se = &task->se;
10391 unsigned int rr_interval = 0;
10394 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10397 if (rq->cfs.load.weight)
10398 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10400 return rr_interval;
10404 * All the scheduling class methods:
10406 const struct sched_class fair_sched_class = {
10407 .next = &idle_sched_class,
10408 .enqueue_task = enqueue_task_fair,
10409 .dequeue_task = dequeue_task_fair,
10410 .yield_task = yield_task_fair,
10411 .yield_to_task = yield_to_task_fair,
10413 .check_preempt_curr = check_preempt_wakeup,
10415 .pick_next_task = pick_next_task_fair,
10417 .put_prev_task = put_prev_task_fair,
10418 .set_next_task = set_next_task_fair,
10421 .select_task_rq = select_task_rq_fair,
10422 .migrate_task_rq = migrate_task_rq_fair,
10424 .rq_online = rq_online_fair,
10425 .rq_offline = rq_offline_fair,
10427 .task_dead = task_dead_fair,
10428 .set_cpus_allowed = set_cpus_allowed_common,
10431 .task_tick = task_tick_fair,
10432 .task_fork = task_fork_fair,
10434 .prio_changed = prio_changed_fair,
10435 .switched_from = switched_from_fair,
10436 .switched_to = switched_to_fair,
10438 .get_rr_interval = get_rr_interval_fair,
10440 .update_curr = update_curr_fair,
10442 #ifdef CONFIG_FAIR_GROUP_SCHED
10443 .task_change_group = task_change_group_fair,
10446 #ifdef CONFIG_UCLAMP_TASK
10447 .uclamp_enabled = 1,
10451 #ifdef CONFIG_SCHED_DEBUG
10452 void print_cfs_stats(struct seq_file *m, int cpu)
10454 struct cfs_rq *cfs_rq, *pos;
10457 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
10458 print_cfs_rq(m, cpu, cfs_rq);
10462 #ifdef CONFIG_NUMA_BALANCING
10463 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10466 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10467 struct numa_group *ng;
10470 ng = rcu_dereference(p->numa_group);
10471 for_each_online_node(node) {
10472 if (p->numa_faults) {
10473 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10474 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10477 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
10478 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
10480 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10484 #endif /* CONFIG_NUMA_BALANCING */
10485 #endif /* CONFIG_SCHED_DEBUG */
10487 __init void init_sched_fair_class(void)
10490 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10492 #ifdef CONFIG_NO_HZ_COMMON
10493 nohz.next_balance = jiffies;
10494 nohz.next_blocked = jiffies;
10495 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
10502 * Helper functions to facilitate extracting info from tracepoints.
10505 const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq)
10508 return cfs_rq ? &cfs_rq->avg : NULL;
10513 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg);
10515 char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len)
10519 strlcpy(str, "(null)", len);
10524 cfs_rq_tg_path(cfs_rq, str, len);
10527 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path);
10529 int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq)
10531 return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1;
10533 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu);
10535 const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq)
10538 return rq ? &rq->avg_rt : NULL;
10543 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt);
10545 const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq)
10548 return rq ? &rq->avg_dl : NULL;
10553 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl);
10555 const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq)
10557 #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ)
10558 return rq ? &rq->avg_irq : NULL;
10563 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq);
10565 int sched_trace_rq_cpu(struct rq *rq)
10567 return rq ? cpu_of(rq) : -1;
10569 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu);
10571 const struct cpumask *sched_trace_rd_span(struct root_domain *rd)
10574 return rd ? rd->span : NULL;
10579 EXPORT_SYMBOL_GPL(sched_trace_rd_span);