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
26 * Targeted preemption latency for CPU-bound tasks:
28 * NOTE: this latency value is not the same as the concept of
29 * 'timeslice length' - timeslices in CFS are of variable length
30 * and have no persistent notion like in traditional, time-slice
31 * based scheduling concepts.
33 * (to see the precise effective timeslice length of your workload,
34 * run vmstat and monitor the context-switches (cs) field)
36 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
38 unsigned int sysctl_sched_latency = 6000000ULL;
39 static unsigned int normalized_sysctl_sched_latency = 6000000ULL;
42 * The initial- and re-scaling of tunables is configurable
46 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
47 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
48 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
50 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
52 unsigned int sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
55 * Minimal preemption granularity for CPU-bound tasks:
57 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
59 unsigned int sysctl_sched_min_granularity = 750000ULL;
60 static unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
63 * Minimal preemption granularity for CPU-bound SCHED_IDLE tasks.
64 * Applies only when SCHED_IDLE tasks compete with normal tasks.
66 * (default: 0.75 msec)
68 unsigned int sysctl_sched_idle_min_granularity = 750000ULL;
71 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
73 static unsigned int sched_nr_latency = 8;
76 * After fork, child runs first. If set to 0 (default) then
77 * parent will (try to) run first.
79 unsigned int sysctl_sched_child_runs_first __read_mostly;
82 * SCHED_OTHER wake-up granularity.
84 * This option delays the preemption effects of decoupled workloads
85 * and reduces their over-scheduling. Synchronous workloads will still
86 * have immediate wakeup/sleep latencies.
88 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
90 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
91 static unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
93 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
95 int sched_thermal_decay_shift;
96 static int __init setup_sched_thermal_decay_shift(char *str)
100 if (kstrtoint(str, 0, &_shift))
101 pr_warn("Unable to set scheduler thermal pressure decay shift parameter\n");
103 sched_thermal_decay_shift = clamp(_shift, 0, 10);
106 __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift);
110 * For asym packing, by default the lower numbered CPU has higher priority.
112 int __weak arch_asym_cpu_priority(int cpu)
118 * The margin used when comparing utilization with CPU capacity.
122 #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024)
125 * The margin used when comparing CPU capacities.
126 * is 'cap1' noticeably greater than 'cap2'
130 #define capacity_greater(cap1, cap2) ((cap1) * 1024 > (cap2) * 1078)
133 #ifdef CONFIG_CFS_BANDWIDTH
135 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
136 * each time a cfs_rq requests quota.
138 * Note: in the case that the slice exceeds the runtime remaining (either due
139 * to consumption or the quota being specified to be smaller than the slice)
140 * we will always only issue the remaining available time.
142 * (default: 5 msec, units: microseconds)
144 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
147 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
153 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
159 static inline void update_load_set(struct load_weight *lw, unsigned long w)
166 * Increase the granularity value when there are more CPUs,
167 * because with more CPUs the 'effective latency' as visible
168 * to users decreases. But the relationship is not linear,
169 * so pick a second-best guess by going with the log2 of the
172 * This idea comes from the SD scheduler of Con Kolivas:
174 static unsigned int get_update_sysctl_factor(void)
176 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
179 switch (sysctl_sched_tunable_scaling) {
180 case SCHED_TUNABLESCALING_NONE:
183 case SCHED_TUNABLESCALING_LINEAR:
186 case SCHED_TUNABLESCALING_LOG:
188 factor = 1 + ilog2(cpus);
195 static void update_sysctl(void)
197 unsigned int factor = get_update_sysctl_factor();
199 #define SET_SYSCTL(name) \
200 (sysctl_##name = (factor) * normalized_sysctl_##name)
201 SET_SYSCTL(sched_min_granularity);
202 SET_SYSCTL(sched_latency);
203 SET_SYSCTL(sched_wakeup_granularity);
207 void __init sched_init_granularity(void)
212 #define WMULT_CONST (~0U)
213 #define WMULT_SHIFT 32
215 static void __update_inv_weight(struct load_weight *lw)
219 if (likely(lw->inv_weight))
222 w = scale_load_down(lw->weight);
224 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
226 else if (unlikely(!w))
227 lw->inv_weight = WMULT_CONST;
229 lw->inv_weight = WMULT_CONST / w;
233 * delta_exec * weight / lw.weight
235 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
237 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
238 * we're guaranteed shift stays positive because inv_weight is guaranteed to
239 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
241 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
242 * weight/lw.weight <= 1, and therefore our shift will also be positive.
244 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
246 u64 fact = scale_load_down(weight);
247 u32 fact_hi = (u32)(fact >> 32);
248 int shift = WMULT_SHIFT;
251 __update_inv_weight(lw);
253 if (unlikely(fact_hi)) {
259 fact = mul_u32_u32(fact, lw->inv_weight);
261 fact_hi = (u32)(fact >> 32);
268 return mul_u64_u32_shr(delta_exec, fact, shift);
272 const struct sched_class fair_sched_class;
274 /**************************************************************
275 * CFS operations on generic schedulable entities:
278 #ifdef CONFIG_FAIR_GROUP_SCHED
280 /* Walk up scheduling entities hierarchy */
281 #define for_each_sched_entity(se) \
282 for (; se; se = se->parent)
284 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
289 if (cfs_rq && task_group_is_autogroup(cfs_rq->tg))
290 autogroup_path(cfs_rq->tg, path, len);
291 else if (cfs_rq && cfs_rq->tg->css.cgroup)
292 cgroup_path(cfs_rq->tg->css.cgroup, path, len);
294 strlcpy(path, "(null)", len);
297 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
299 struct rq *rq = rq_of(cfs_rq);
300 int cpu = cpu_of(rq);
303 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
308 * Ensure we either appear before our parent (if already
309 * enqueued) or force our parent to appear after us when it is
310 * enqueued. The fact that we always enqueue bottom-up
311 * reduces this to two cases and a special case for the root
312 * cfs_rq. Furthermore, it also means that we will always reset
313 * tmp_alone_branch either when the branch is connected
314 * to a tree or when we reach the top of the tree
316 if (cfs_rq->tg->parent &&
317 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
319 * If parent is already on the list, we add the child
320 * just before. Thanks to circular linked property of
321 * the list, this means to put the child at the tail
322 * of the list that starts by parent.
324 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
325 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
327 * The branch is now connected to its tree so we can
328 * reset tmp_alone_branch to the beginning of the
331 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
335 if (!cfs_rq->tg->parent) {
337 * cfs rq without parent should be put
338 * at the tail of the list.
340 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
341 &rq->leaf_cfs_rq_list);
343 * We have reach the top of a tree so we can reset
344 * tmp_alone_branch to the beginning of the list.
346 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
351 * The parent has not already been added so we want to
352 * make sure that it will be put after us.
353 * tmp_alone_branch points to the begin of the branch
354 * where we will add parent.
356 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
358 * update tmp_alone_branch to points to the new begin
361 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
365 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
367 if (cfs_rq->on_list) {
368 struct rq *rq = rq_of(cfs_rq);
371 * With cfs_rq being unthrottled/throttled during an enqueue,
372 * it can happen the tmp_alone_branch points the a leaf that
373 * we finally want to del. In this case, tmp_alone_branch moves
374 * to the prev element but it will point to rq->leaf_cfs_rq_list
375 * at the end of the enqueue.
377 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
378 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
380 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
385 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
387 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
390 /* Iterate thr' all leaf cfs_rq's on a runqueue */
391 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
392 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
395 /* Do the two (enqueued) entities belong to the same group ? */
396 static inline struct cfs_rq *
397 is_same_group(struct sched_entity *se, struct sched_entity *pse)
399 if (se->cfs_rq == pse->cfs_rq)
405 static inline struct sched_entity *parent_entity(struct sched_entity *se)
411 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
413 int se_depth, pse_depth;
416 * preemption test can be made between sibling entities who are in the
417 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
418 * both tasks until we find their ancestors who are siblings of common
422 /* First walk up until both entities are at same depth */
423 se_depth = (*se)->depth;
424 pse_depth = (*pse)->depth;
426 while (se_depth > pse_depth) {
428 *se = parent_entity(*se);
431 while (pse_depth > se_depth) {
433 *pse = parent_entity(*pse);
436 while (!is_same_group(*se, *pse)) {
437 *se = parent_entity(*se);
438 *pse = parent_entity(*pse);
442 static int tg_is_idle(struct task_group *tg)
447 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
449 return cfs_rq->idle > 0;
452 static int se_is_idle(struct sched_entity *se)
454 if (entity_is_task(se))
455 return task_has_idle_policy(task_of(se));
456 return cfs_rq_is_idle(group_cfs_rq(se));
459 #else /* !CONFIG_FAIR_GROUP_SCHED */
461 #define for_each_sched_entity(se) \
462 for (; se; se = NULL)
464 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
467 strlcpy(path, "(null)", len);
470 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
475 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
479 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
483 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
484 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
486 static inline struct sched_entity *parent_entity(struct sched_entity *se)
492 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
496 static inline int tg_is_idle(struct task_group *tg)
501 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
506 static int se_is_idle(struct sched_entity *se)
511 #endif /* CONFIG_FAIR_GROUP_SCHED */
513 static __always_inline
514 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
516 /**************************************************************
517 * Scheduling class tree data structure manipulation methods:
520 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
522 s64 delta = (s64)(vruntime - max_vruntime);
524 max_vruntime = vruntime;
529 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
531 s64 delta = (s64)(vruntime - min_vruntime);
533 min_vruntime = vruntime;
538 static inline bool entity_before(struct sched_entity *a,
539 struct sched_entity *b)
541 return (s64)(a->vruntime - b->vruntime) < 0;
544 #define __node_2_se(node) \
545 rb_entry((node), struct sched_entity, run_node)
547 static void update_min_vruntime(struct cfs_rq *cfs_rq)
549 struct sched_entity *curr = cfs_rq->curr;
550 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
552 u64 vruntime = cfs_rq->min_vruntime;
556 vruntime = curr->vruntime;
561 if (leftmost) { /* non-empty tree */
562 struct sched_entity *se = __node_2_se(leftmost);
565 vruntime = se->vruntime;
567 vruntime = min_vruntime(vruntime, se->vruntime);
570 /* ensure we never gain time by being placed backwards. */
571 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
574 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
578 static inline bool __entity_less(struct rb_node *a, const struct rb_node *b)
580 return entity_before(__node_2_se(a), __node_2_se(b));
584 * Enqueue an entity into the rb-tree:
586 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
588 rb_add_cached(&se->run_node, &cfs_rq->tasks_timeline, __entity_less);
591 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
593 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
596 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
598 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
603 return __node_2_se(left);
606 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
608 struct rb_node *next = rb_next(&se->run_node);
613 return __node_2_se(next);
616 #ifdef CONFIG_SCHED_DEBUG
617 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
619 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
624 return __node_2_se(last);
627 /**************************************************************
628 * Scheduling class statistics methods:
631 int sched_update_scaling(void)
633 unsigned int factor = get_update_sysctl_factor();
635 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
636 sysctl_sched_min_granularity);
638 #define WRT_SYSCTL(name) \
639 (normalized_sysctl_##name = sysctl_##name / (factor))
640 WRT_SYSCTL(sched_min_granularity);
641 WRT_SYSCTL(sched_latency);
642 WRT_SYSCTL(sched_wakeup_granularity);
652 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
654 if (unlikely(se->load.weight != NICE_0_LOAD))
655 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
661 * The idea is to set a period in which each task runs once.
663 * When there are too many tasks (sched_nr_latency) we have to stretch
664 * this period because otherwise the slices get too small.
666 * p = (nr <= nl) ? l : l*nr/nl
668 static u64 __sched_period(unsigned long nr_running)
670 if (unlikely(nr_running > sched_nr_latency))
671 return nr_running * sysctl_sched_min_granularity;
673 return sysctl_sched_latency;
676 static bool sched_idle_cfs_rq(struct cfs_rq *cfs_rq);
679 * We calculate the wall-time slice from the period by taking a part
680 * proportional to the weight.
684 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
686 unsigned int nr_running = cfs_rq->nr_running;
687 struct sched_entity *init_se = se;
688 unsigned int min_gran;
691 if (sched_feat(ALT_PERIOD))
692 nr_running = rq_of(cfs_rq)->cfs.h_nr_running;
694 slice = __sched_period(nr_running + !se->on_rq);
696 for_each_sched_entity(se) {
697 struct load_weight *load;
698 struct load_weight lw;
699 struct cfs_rq *qcfs_rq;
701 qcfs_rq = cfs_rq_of(se);
702 load = &qcfs_rq->load;
704 if (unlikely(!se->on_rq)) {
707 update_load_add(&lw, se->load.weight);
710 slice = __calc_delta(slice, se->load.weight, load);
713 if (sched_feat(BASE_SLICE)) {
714 if (se_is_idle(init_se) && !sched_idle_cfs_rq(cfs_rq))
715 min_gran = sysctl_sched_idle_min_granularity;
717 min_gran = sysctl_sched_min_granularity;
719 slice = max_t(u64, slice, min_gran);
726 * We calculate the vruntime slice of a to-be-inserted task.
730 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
732 return calc_delta_fair(sched_slice(cfs_rq, se), se);
738 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
739 static unsigned long task_h_load(struct task_struct *p);
740 static unsigned long capacity_of(int cpu);
742 /* Give new sched_entity start runnable values to heavy its load in infant time */
743 void init_entity_runnable_average(struct sched_entity *se)
745 struct sched_avg *sa = &se->avg;
747 memset(sa, 0, sizeof(*sa));
750 * Tasks are initialized with full load to be seen as heavy tasks until
751 * they get a chance to stabilize to their real load level.
752 * Group entities are initialized with zero load to reflect the fact that
753 * nothing has been attached to the task group yet.
755 if (entity_is_task(se))
756 sa->load_avg = scale_load_down(se->load.weight);
758 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
761 static void attach_entity_cfs_rq(struct sched_entity *se);
764 * With new tasks being created, their initial util_avgs are extrapolated
765 * based on the cfs_rq's current util_avg:
767 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
769 * However, in many cases, the above util_avg does not give a desired
770 * value. Moreover, the sum of the util_avgs may be divergent, such
771 * as when the series is a harmonic series.
773 * To solve this problem, we also cap the util_avg of successive tasks to
774 * only 1/2 of the left utilization budget:
776 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
778 * where n denotes the nth task and cpu_scale the CPU capacity.
780 * For example, for a CPU with 1024 of capacity, a simplest series from
781 * the beginning would be like:
783 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
784 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
786 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
787 * if util_avg > util_avg_cap.
789 void post_init_entity_util_avg(struct task_struct *p)
791 struct sched_entity *se = &p->se;
792 struct cfs_rq *cfs_rq = cfs_rq_of(se);
793 struct sched_avg *sa = &se->avg;
794 long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
795 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
798 if (cfs_rq->avg.util_avg != 0) {
799 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
800 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
802 if (sa->util_avg > cap)
809 sa->runnable_avg = sa->util_avg;
811 if (p->sched_class != &fair_sched_class) {
813 * For !fair tasks do:
815 update_cfs_rq_load_avg(now, cfs_rq);
816 attach_entity_load_avg(cfs_rq, se);
817 switched_from_fair(rq, p);
819 * such that the next switched_to_fair() has the
822 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
826 attach_entity_cfs_rq(se);
829 #else /* !CONFIG_SMP */
830 void init_entity_runnable_average(struct sched_entity *se)
833 void post_init_entity_util_avg(struct task_struct *p)
836 static void update_tg_load_avg(struct cfs_rq *cfs_rq)
839 #endif /* CONFIG_SMP */
842 * Update the current task's runtime statistics.
844 static void update_curr(struct cfs_rq *cfs_rq)
846 struct sched_entity *curr = cfs_rq->curr;
847 u64 now = rq_clock_task(rq_of(cfs_rq));
853 delta_exec = now - curr->exec_start;
854 if (unlikely((s64)delta_exec <= 0))
857 curr->exec_start = now;
859 if (schedstat_enabled()) {
860 struct sched_statistics *stats;
862 stats = __schedstats_from_se(curr);
863 __schedstat_set(stats->exec_max,
864 max(delta_exec, stats->exec_max));
867 curr->sum_exec_runtime += delta_exec;
868 schedstat_add(cfs_rq->exec_clock, delta_exec);
870 curr->vruntime += calc_delta_fair(delta_exec, curr);
871 update_min_vruntime(cfs_rq);
873 if (entity_is_task(curr)) {
874 struct task_struct *curtask = task_of(curr);
876 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
877 cgroup_account_cputime(curtask, delta_exec);
878 account_group_exec_runtime(curtask, delta_exec);
881 account_cfs_rq_runtime(cfs_rq, delta_exec);
884 static void update_curr_fair(struct rq *rq)
886 update_curr(cfs_rq_of(&rq->curr->se));
890 update_stats_wait_start_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
892 struct sched_statistics *stats;
893 struct task_struct *p = NULL;
895 if (!schedstat_enabled())
898 stats = __schedstats_from_se(se);
900 if (entity_is_task(se))
903 __update_stats_wait_start(rq_of(cfs_rq), p, stats);
907 update_stats_wait_end_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
909 struct sched_statistics *stats;
910 struct task_struct *p = NULL;
912 if (!schedstat_enabled())
915 stats = __schedstats_from_se(se);
918 * When the sched_schedstat changes from 0 to 1, some sched se
919 * maybe already in the runqueue, the se->statistics.wait_start
920 * will be 0.So it will let the delta wrong. We need to avoid this
923 if (unlikely(!schedstat_val(stats->wait_start)))
926 if (entity_is_task(se))
929 __update_stats_wait_end(rq_of(cfs_rq), p, stats);
933 update_stats_enqueue_sleeper_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
935 struct sched_statistics *stats;
936 struct task_struct *tsk = NULL;
938 if (!schedstat_enabled())
941 stats = __schedstats_from_se(se);
943 if (entity_is_task(se))
946 __update_stats_enqueue_sleeper(rq_of(cfs_rq), tsk, stats);
950 * Task is being enqueued - update stats:
953 update_stats_enqueue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
955 if (!schedstat_enabled())
959 * Are we enqueueing a waiting task? (for current tasks
960 * a dequeue/enqueue event is a NOP)
962 if (se != cfs_rq->curr)
963 update_stats_wait_start_fair(cfs_rq, se);
965 if (flags & ENQUEUE_WAKEUP)
966 update_stats_enqueue_sleeper_fair(cfs_rq, se);
970 update_stats_dequeue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
973 if (!schedstat_enabled())
977 * Mark the end of the wait period if dequeueing a
980 if (se != cfs_rq->curr)
981 update_stats_wait_end_fair(cfs_rq, se);
983 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
984 struct task_struct *tsk = task_of(se);
987 /* XXX racy against TTWU */
988 state = READ_ONCE(tsk->__state);
989 if (state & TASK_INTERRUPTIBLE)
990 __schedstat_set(tsk->stats.sleep_start,
991 rq_clock(rq_of(cfs_rq)));
992 if (state & TASK_UNINTERRUPTIBLE)
993 __schedstat_set(tsk->stats.block_start,
994 rq_clock(rq_of(cfs_rq)));
999 * We are picking a new current task - update its stats:
1002 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1005 * We are starting a new run period:
1007 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1010 /**************************************************
1011 * Scheduling class queueing methods:
1014 #ifdef CONFIG_NUMA_BALANCING
1016 * Approximate time to scan a full NUMA task in ms. The task scan period is
1017 * calculated based on the tasks virtual memory size and
1018 * numa_balancing_scan_size.
1020 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1021 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1023 /* Portion of address space to scan in MB */
1024 unsigned int sysctl_numa_balancing_scan_size = 256;
1026 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1027 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1030 refcount_t refcount;
1032 spinlock_t lock; /* nr_tasks, tasks */
1037 struct rcu_head rcu;
1038 unsigned long total_faults;
1039 unsigned long max_faults_cpu;
1041 * faults[] array is split into two regions: faults_mem and faults_cpu.
1043 * Faults_cpu is used to decide whether memory should move
1044 * towards the CPU. As a consequence, these stats are weighted
1045 * more by CPU use than by memory faults.
1047 unsigned long faults[];
1051 * For functions that can be called in multiple contexts that permit reading
1052 * ->numa_group (see struct task_struct for locking rules).
1054 static struct numa_group *deref_task_numa_group(struct task_struct *p)
1056 return rcu_dereference_check(p->numa_group, p == current ||
1057 (lockdep_is_held(__rq_lockp(task_rq(p))) && !READ_ONCE(p->on_cpu)));
1060 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1062 return rcu_dereference_protected(p->numa_group, p == current);
1065 static inline unsigned long group_faults_priv(struct numa_group *ng);
1066 static inline unsigned long group_faults_shared(struct numa_group *ng);
1068 static unsigned int task_nr_scan_windows(struct task_struct *p)
1070 unsigned long rss = 0;
1071 unsigned long nr_scan_pages;
1074 * Calculations based on RSS as non-present and empty pages are skipped
1075 * by the PTE scanner and NUMA hinting faults should be trapped based
1078 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1079 rss = get_mm_rss(p->mm);
1081 rss = nr_scan_pages;
1083 rss = round_up(rss, nr_scan_pages);
1084 return rss / nr_scan_pages;
1087 /* For sanity's sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1088 #define MAX_SCAN_WINDOW 2560
1090 static unsigned int task_scan_min(struct task_struct *p)
1092 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1093 unsigned int scan, floor;
1094 unsigned int windows = 1;
1096 if (scan_size < MAX_SCAN_WINDOW)
1097 windows = MAX_SCAN_WINDOW / scan_size;
1098 floor = 1000 / windows;
1100 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1101 return max_t(unsigned int, floor, scan);
1104 static unsigned int task_scan_start(struct task_struct *p)
1106 unsigned long smin = task_scan_min(p);
1107 unsigned long period = smin;
1108 struct numa_group *ng;
1110 /* Scale the maximum scan period with the amount of shared memory. */
1112 ng = rcu_dereference(p->numa_group);
1114 unsigned long shared = group_faults_shared(ng);
1115 unsigned long private = group_faults_priv(ng);
1117 period *= refcount_read(&ng->refcount);
1118 period *= shared + 1;
1119 period /= private + shared + 1;
1123 return max(smin, period);
1126 static unsigned int task_scan_max(struct task_struct *p)
1128 unsigned long smin = task_scan_min(p);
1130 struct numa_group *ng;
1132 /* Watch for min being lower than max due to floor calculations */
1133 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1135 /* Scale the maximum scan period with the amount of shared memory. */
1136 ng = deref_curr_numa_group(p);
1138 unsigned long shared = group_faults_shared(ng);
1139 unsigned long private = group_faults_priv(ng);
1140 unsigned long period = smax;
1142 period *= refcount_read(&ng->refcount);
1143 period *= shared + 1;
1144 period /= private + shared + 1;
1146 smax = max(smax, period);
1149 return max(smin, smax);
1152 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1154 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1155 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1158 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1160 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1161 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1164 /* Shared or private faults. */
1165 #define NR_NUMA_HINT_FAULT_TYPES 2
1167 /* Memory and CPU locality */
1168 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1170 /* Averaged statistics, and temporary buffers. */
1171 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1173 pid_t task_numa_group_id(struct task_struct *p)
1175 struct numa_group *ng;
1179 ng = rcu_dereference(p->numa_group);
1188 * The averaged statistics, shared & private, memory & CPU,
1189 * occupy the first half of the array. The second half of the
1190 * array is for current counters, which are averaged into the
1191 * first set by task_numa_placement.
1193 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1195 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1198 static inline unsigned long task_faults(struct task_struct *p, int nid)
1200 if (!p->numa_faults)
1203 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1204 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1207 static inline unsigned long group_faults(struct task_struct *p, int nid)
1209 struct numa_group *ng = deref_task_numa_group(p);
1214 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1215 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1218 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1220 return group->faults[task_faults_idx(NUMA_CPU, nid, 0)] +
1221 group->faults[task_faults_idx(NUMA_CPU, nid, 1)];
1224 static inline unsigned long group_faults_priv(struct numa_group *ng)
1226 unsigned long faults = 0;
1229 for_each_online_node(node) {
1230 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1236 static inline unsigned long group_faults_shared(struct numa_group *ng)
1238 unsigned long faults = 0;
1241 for_each_online_node(node) {
1242 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1249 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1250 * considered part of a numa group's pseudo-interleaving set. Migrations
1251 * between these nodes are slowed down, to allow things to settle down.
1253 #define ACTIVE_NODE_FRACTION 3
1255 static bool numa_is_active_node(int nid, struct numa_group *ng)
1257 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1260 /* Handle placement on systems where not all nodes are directly connected. */
1261 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1262 int maxdist, bool task)
1264 unsigned long score = 0;
1268 * All nodes are directly connected, and the same distance
1269 * from each other. No need for fancy placement algorithms.
1271 if (sched_numa_topology_type == NUMA_DIRECT)
1275 * This code is called for each node, introducing N^2 complexity,
1276 * which should be ok given the number of nodes rarely exceeds 8.
1278 for_each_online_node(node) {
1279 unsigned long faults;
1280 int dist = node_distance(nid, node);
1283 * The furthest away nodes in the system are not interesting
1284 * for placement; nid was already counted.
1286 if (dist == sched_max_numa_distance || node == nid)
1290 * On systems with a backplane NUMA topology, compare groups
1291 * of nodes, and move tasks towards the group with the most
1292 * memory accesses. When comparing two nodes at distance
1293 * "hoplimit", only nodes closer by than "hoplimit" are part
1294 * of each group. Skip other nodes.
1296 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1300 /* Add up the faults from nearby nodes. */
1302 faults = task_faults(p, node);
1304 faults = group_faults(p, node);
1307 * On systems with a glueless mesh NUMA topology, there are
1308 * no fixed "groups of nodes". Instead, nodes that are not
1309 * directly connected bounce traffic through intermediate
1310 * nodes; a numa_group can occupy any set of nodes.
1311 * The further away a node is, the less the faults count.
1312 * This seems to result in good task placement.
1314 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1315 faults *= (sched_max_numa_distance - dist);
1316 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1326 * These return the fraction of accesses done by a particular task, or
1327 * task group, on a particular numa node. The group weight is given a
1328 * larger multiplier, in order to group tasks together that are almost
1329 * evenly spread out between numa nodes.
1331 static inline unsigned long task_weight(struct task_struct *p, int nid,
1334 unsigned long faults, total_faults;
1336 if (!p->numa_faults)
1339 total_faults = p->total_numa_faults;
1344 faults = task_faults(p, nid);
1345 faults += score_nearby_nodes(p, nid, dist, true);
1347 return 1000 * faults / total_faults;
1350 static inline unsigned long group_weight(struct task_struct *p, int nid,
1353 struct numa_group *ng = deref_task_numa_group(p);
1354 unsigned long faults, total_faults;
1359 total_faults = ng->total_faults;
1364 faults = group_faults(p, nid);
1365 faults += score_nearby_nodes(p, nid, dist, false);
1367 return 1000 * faults / total_faults;
1370 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1371 int src_nid, int dst_cpu)
1373 struct numa_group *ng = deref_curr_numa_group(p);
1374 int dst_nid = cpu_to_node(dst_cpu);
1375 int last_cpupid, this_cpupid;
1377 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1378 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1381 * Allow first faults or private faults to migrate immediately early in
1382 * the lifetime of a task. The magic number 4 is based on waiting for
1383 * two full passes of the "multi-stage node selection" test that is
1386 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1387 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1391 * Multi-stage node selection is used in conjunction with a periodic
1392 * migration fault to build a temporal task<->page relation. By using
1393 * a two-stage filter we remove short/unlikely relations.
1395 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1396 * a task's usage of a particular page (n_p) per total usage of this
1397 * page (n_t) (in a given time-span) to a probability.
1399 * Our periodic faults will sample this probability and getting the
1400 * same result twice in a row, given these samples are fully
1401 * independent, is then given by P(n)^2, provided our sample period
1402 * is sufficiently short compared to the usage pattern.
1404 * This quadric squishes small probabilities, making it less likely we
1405 * act on an unlikely task<->page relation.
1407 if (!cpupid_pid_unset(last_cpupid) &&
1408 cpupid_to_nid(last_cpupid) != dst_nid)
1411 /* Always allow migrate on private faults */
1412 if (cpupid_match_pid(p, last_cpupid))
1415 /* A shared fault, but p->numa_group has not been set up yet. */
1420 * Destination node is much more heavily used than the source
1421 * node? Allow migration.
1423 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1424 ACTIVE_NODE_FRACTION)
1428 * Distribute memory according to CPU & memory use on each node,
1429 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1431 * faults_cpu(dst) 3 faults_cpu(src)
1432 * --------------- * - > ---------------
1433 * faults_mem(dst) 4 faults_mem(src)
1435 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1436 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1440 * 'numa_type' describes the node at the moment of load balancing.
1443 /* The node has spare capacity that can be used to run more tasks. */
1446 * The node is fully used and the tasks don't compete for more CPU
1447 * cycles. Nevertheless, some tasks might wait before running.
1451 * The node is overloaded and can't provide expected CPU cycles to all
1457 /* Cached statistics for all CPUs within a node */
1460 unsigned long runnable;
1462 /* Total compute capacity of CPUs on a node */
1463 unsigned long compute_capacity;
1464 unsigned int nr_running;
1465 unsigned int weight;
1466 enum numa_type node_type;
1470 static inline bool is_core_idle(int cpu)
1472 #ifdef CONFIG_SCHED_SMT
1475 for_each_cpu(sibling, cpu_smt_mask(cpu)) {
1479 if (!idle_cpu(sibling))
1487 struct task_numa_env {
1488 struct task_struct *p;
1490 int src_cpu, src_nid;
1491 int dst_cpu, dst_nid;
1493 struct numa_stats src_stats, dst_stats;
1498 struct task_struct *best_task;
1503 static unsigned long cpu_load(struct rq *rq);
1504 static unsigned long cpu_runnable(struct rq *rq);
1505 static inline long adjust_numa_imbalance(int imbalance,
1506 int dst_running, int dst_weight);
1509 numa_type numa_classify(unsigned int imbalance_pct,
1510 struct numa_stats *ns)
1512 if ((ns->nr_running > ns->weight) &&
1513 (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) ||
1514 ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100))))
1515 return node_overloaded;
1517 if ((ns->nr_running < ns->weight) ||
1518 (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) &&
1519 ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100))))
1520 return node_has_spare;
1522 return node_fully_busy;
1525 #ifdef CONFIG_SCHED_SMT
1526 /* Forward declarations of select_idle_sibling helpers */
1527 static inline bool test_idle_cores(int cpu, bool def);
1528 static inline int numa_idle_core(int idle_core, int cpu)
1530 if (!static_branch_likely(&sched_smt_present) ||
1531 idle_core >= 0 || !test_idle_cores(cpu, false))
1535 * Prefer cores instead of packing HT siblings
1536 * and triggering future load balancing.
1538 if (is_core_idle(cpu))
1544 static inline int numa_idle_core(int idle_core, int cpu)
1551 * Gather all necessary information to make NUMA balancing placement
1552 * decisions that are compatible with standard load balancer. This
1553 * borrows code and logic from update_sg_lb_stats but sharing a
1554 * common implementation is impractical.
1556 static void update_numa_stats(struct task_numa_env *env,
1557 struct numa_stats *ns, int nid,
1560 int cpu, idle_core = -1;
1562 memset(ns, 0, sizeof(*ns));
1566 for_each_cpu(cpu, cpumask_of_node(nid)) {
1567 struct rq *rq = cpu_rq(cpu);
1569 ns->load += cpu_load(rq);
1570 ns->runnable += cpu_runnable(rq);
1571 ns->util += cpu_util_cfs(cpu);
1572 ns->nr_running += rq->cfs.h_nr_running;
1573 ns->compute_capacity += capacity_of(cpu);
1575 if (find_idle && !rq->nr_running && idle_cpu(cpu)) {
1576 if (READ_ONCE(rq->numa_migrate_on) ||
1577 !cpumask_test_cpu(cpu, env->p->cpus_ptr))
1580 if (ns->idle_cpu == -1)
1583 idle_core = numa_idle_core(idle_core, cpu);
1588 ns->weight = cpumask_weight(cpumask_of_node(nid));
1590 ns->node_type = numa_classify(env->imbalance_pct, ns);
1593 ns->idle_cpu = idle_core;
1596 static void task_numa_assign(struct task_numa_env *env,
1597 struct task_struct *p, long imp)
1599 struct rq *rq = cpu_rq(env->dst_cpu);
1601 /* Check if run-queue part of active NUMA balance. */
1602 if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) {
1604 int start = env->dst_cpu;
1606 /* Find alternative idle CPU. */
1607 for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start) {
1608 if (cpu == env->best_cpu || !idle_cpu(cpu) ||
1609 !cpumask_test_cpu(cpu, env->p->cpus_ptr)) {
1614 rq = cpu_rq(env->dst_cpu);
1615 if (!xchg(&rq->numa_migrate_on, 1))
1619 /* Failed to find an alternative idle CPU */
1625 * Clear previous best_cpu/rq numa-migrate flag, since task now
1626 * found a better CPU to move/swap.
1628 if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) {
1629 rq = cpu_rq(env->best_cpu);
1630 WRITE_ONCE(rq->numa_migrate_on, 0);
1634 put_task_struct(env->best_task);
1639 env->best_imp = imp;
1640 env->best_cpu = env->dst_cpu;
1643 static bool load_too_imbalanced(long src_load, long dst_load,
1644 struct task_numa_env *env)
1647 long orig_src_load, orig_dst_load;
1648 long src_capacity, dst_capacity;
1651 * The load is corrected for the CPU capacity available on each node.
1654 * ------------ vs ---------
1655 * src_capacity dst_capacity
1657 src_capacity = env->src_stats.compute_capacity;
1658 dst_capacity = env->dst_stats.compute_capacity;
1660 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1662 orig_src_load = env->src_stats.load;
1663 orig_dst_load = env->dst_stats.load;
1665 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1667 /* Would this change make things worse? */
1668 return (imb > old_imb);
1672 * Maximum NUMA importance can be 1998 (2*999);
1673 * SMALLIMP @ 30 would be close to 1998/64.
1674 * Used to deter task migration.
1679 * This checks if the overall compute and NUMA accesses of the system would
1680 * be improved if the source tasks was migrated to the target dst_cpu taking
1681 * into account that it might be best if task running on the dst_cpu should
1682 * be exchanged with the source task
1684 static bool task_numa_compare(struct task_numa_env *env,
1685 long taskimp, long groupimp, bool maymove)
1687 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
1688 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1689 long imp = p_ng ? groupimp : taskimp;
1690 struct task_struct *cur;
1691 long src_load, dst_load;
1692 int dist = env->dist;
1695 bool stopsearch = false;
1697 if (READ_ONCE(dst_rq->numa_migrate_on))
1701 cur = rcu_dereference(dst_rq->curr);
1702 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1706 * Because we have preemption enabled we can get migrated around and
1707 * end try selecting ourselves (current == env->p) as a swap candidate.
1709 if (cur == env->p) {
1715 if (maymove && moveimp >= env->best_imp)
1721 /* Skip this swap candidate if cannot move to the source cpu. */
1722 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
1726 * Skip this swap candidate if it is not moving to its preferred
1727 * node and the best task is.
1729 if (env->best_task &&
1730 env->best_task->numa_preferred_nid == env->src_nid &&
1731 cur->numa_preferred_nid != env->src_nid) {
1736 * "imp" is the fault differential for the source task between the
1737 * source and destination node. Calculate the total differential for
1738 * the source task and potential destination task. The more negative
1739 * the value is, the more remote accesses that would be expected to
1740 * be incurred if the tasks were swapped.
1742 * If dst and source tasks are in the same NUMA group, or not
1743 * in any group then look only at task weights.
1745 cur_ng = rcu_dereference(cur->numa_group);
1746 if (cur_ng == p_ng) {
1747 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1748 task_weight(cur, env->dst_nid, dist);
1750 * Add some hysteresis to prevent swapping the
1751 * tasks within a group over tiny differences.
1757 * Compare the group weights. If a task is all by itself
1758 * (not part of a group), use the task weight instead.
1761 imp += group_weight(cur, env->src_nid, dist) -
1762 group_weight(cur, env->dst_nid, dist);
1764 imp += task_weight(cur, env->src_nid, dist) -
1765 task_weight(cur, env->dst_nid, dist);
1768 /* Discourage picking a task already on its preferred node */
1769 if (cur->numa_preferred_nid == env->dst_nid)
1773 * Encourage picking a task that moves to its preferred node.
1774 * This potentially makes imp larger than it's maximum of
1775 * 1998 (see SMALLIMP and task_weight for why) but in this
1776 * case, it does not matter.
1778 if (cur->numa_preferred_nid == env->src_nid)
1781 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1788 * Prefer swapping with a task moving to its preferred node over a
1791 if (env->best_task && cur->numa_preferred_nid == env->src_nid &&
1792 env->best_task->numa_preferred_nid != env->src_nid) {
1797 * If the NUMA importance is less than SMALLIMP,
1798 * task migration might only result in ping pong
1799 * of tasks and also hurt performance due to cache
1802 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1806 * In the overloaded case, try and keep the load balanced.
1808 load = task_h_load(env->p) - task_h_load(cur);
1812 dst_load = env->dst_stats.load + load;
1813 src_load = env->src_stats.load - load;
1815 if (load_too_imbalanced(src_load, dst_load, env))
1819 /* Evaluate an idle CPU for a task numa move. */
1821 int cpu = env->dst_stats.idle_cpu;
1823 /* Nothing cached so current CPU went idle since the search. */
1828 * If the CPU is no longer truly idle and the previous best CPU
1829 * is, keep using it.
1831 if (!idle_cpu(cpu) && env->best_cpu >= 0 &&
1832 idle_cpu(env->best_cpu)) {
1833 cpu = env->best_cpu;
1839 task_numa_assign(env, cur, imp);
1842 * If a move to idle is allowed because there is capacity or load
1843 * balance improves then stop the search. While a better swap
1844 * candidate may exist, a search is not free.
1846 if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu))
1850 * If a swap candidate must be identified and the current best task
1851 * moves its preferred node then stop the search.
1853 if (!maymove && env->best_task &&
1854 env->best_task->numa_preferred_nid == env->src_nid) {
1863 static void task_numa_find_cpu(struct task_numa_env *env,
1864 long taskimp, long groupimp)
1866 bool maymove = false;
1870 * If dst node has spare capacity, then check if there is an
1871 * imbalance that would be overruled by the load balancer.
1873 if (env->dst_stats.node_type == node_has_spare) {
1874 unsigned int imbalance;
1875 int src_running, dst_running;
1878 * Would movement cause an imbalance? Note that if src has
1879 * more running tasks that the imbalance is ignored as the
1880 * move improves the imbalance from the perspective of the
1881 * CPU load balancer.
1883 src_running = env->src_stats.nr_running - 1;
1884 dst_running = env->dst_stats.nr_running + 1;
1885 imbalance = max(0, dst_running - src_running);
1886 imbalance = adjust_numa_imbalance(imbalance, dst_running,
1887 env->dst_stats.weight);
1889 /* Use idle CPU if there is no imbalance */
1892 if (env->dst_stats.idle_cpu >= 0) {
1893 env->dst_cpu = env->dst_stats.idle_cpu;
1894 task_numa_assign(env, NULL, 0);
1899 long src_load, dst_load, load;
1901 * If the improvement from just moving env->p direction is better
1902 * than swapping tasks around, check if a move is possible.
1904 load = task_h_load(env->p);
1905 dst_load = env->dst_stats.load + load;
1906 src_load = env->src_stats.load - load;
1907 maymove = !load_too_imbalanced(src_load, dst_load, env);
1910 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1911 /* Skip this CPU if the source task cannot migrate */
1912 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
1916 if (task_numa_compare(env, taskimp, groupimp, maymove))
1921 static int task_numa_migrate(struct task_struct *p)
1923 struct task_numa_env env = {
1926 .src_cpu = task_cpu(p),
1927 .src_nid = task_node(p),
1929 .imbalance_pct = 112,
1935 unsigned long taskweight, groupweight;
1936 struct sched_domain *sd;
1937 long taskimp, groupimp;
1938 struct numa_group *ng;
1943 * Pick the lowest SD_NUMA domain, as that would have the smallest
1944 * imbalance and would be the first to start moving tasks about.
1946 * And we want to avoid any moving of tasks about, as that would create
1947 * random movement of tasks -- counter the numa conditions we're trying
1951 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1953 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1957 * Cpusets can break the scheduler domain tree into smaller
1958 * balance domains, some of which do not cross NUMA boundaries.
1959 * Tasks that are "trapped" in such domains cannot be migrated
1960 * elsewhere, so there is no point in (re)trying.
1962 if (unlikely(!sd)) {
1963 sched_setnuma(p, task_node(p));
1967 env.dst_nid = p->numa_preferred_nid;
1968 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1969 taskweight = task_weight(p, env.src_nid, dist);
1970 groupweight = group_weight(p, env.src_nid, dist);
1971 update_numa_stats(&env, &env.src_stats, env.src_nid, false);
1972 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1973 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1974 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
1976 /* Try to find a spot on the preferred nid. */
1977 task_numa_find_cpu(&env, taskimp, groupimp);
1980 * Look at other nodes in these cases:
1981 * - there is no space available on the preferred_nid
1982 * - the task is part of a numa_group that is interleaved across
1983 * multiple NUMA nodes; in order to better consolidate the group,
1984 * we need to check other locations.
1986 ng = deref_curr_numa_group(p);
1987 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
1988 for_each_online_node(nid) {
1989 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1992 dist = node_distance(env.src_nid, env.dst_nid);
1993 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1995 taskweight = task_weight(p, env.src_nid, dist);
1996 groupweight = group_weight(p, env.src_nid, dist);
1999 /* Only consider nodes where both task and groups benefit */
2000 taskimp = task_weight(p, nid, dist) - taskweight;
2001 groupimp = group_weight(p, nid, dist) - groupweight;
2002 if (taskimp < 0 && groupimp < 0)
2007 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2008 task_numa_find_cpu(&env, taskimp, groupimp);
2013 * If the task is part of a workload that spans multiple NUMA nodes,
2014 * and is migrating into one of the workload's active nodes, remember
2015 * this node as the task's preferred numa node, so the workload can
2017 * A task that migrated to a second choice node will be better off
2018 * trying for a better one later. Do not set the preferred node here.
2021 if (env.best_cpu == -1)
2024 nid = cpu_to_node(env.best_cpu);
2026 if (nid != p->numa_preferred_nid)
2027 sched_setnuma(p, nid);
2030 /* No better CPU than the current one was found. */
2031 if (env.best_cpu == -1) {
2032 trace_sched_stick_numa(p, env.src_cpu, NULL, -1);
2036 best_rq = cpu_rq(env.best_cpu);
2037 if (env.best_task == NULL) {
2038 ret = migrate_task_to(p, env.best_cpu);
2039 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2041 trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu);
2045 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
2046 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2049 trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu);
2050 put_task_struct(env.best_task);
2054 /* Attempt to migrate a task to a CPU on the preferred node. */
2055 static void numa_migrate_preferred(struct task_struct *p)
2057 unsigned long interval = HZ;
2059 /* This task has no NUMA fault statistics yet */
2060 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
2063 /* Periodically retry migrating the task to the preferred node */
2064 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
2065 p->numa_migrate_retry = jiffies + interval;
2067 /* Success if task is already running on preferred CPU */
2068 if (task_node(p) == p->numa_preferred_nid)
2071 /* Otherwise, try migrate to a CPU on the preferred node */
2072 task_numa_migrate(p);
2076 * Find out how many nodes the workload is actively running on. Do this by
2077 * tracking the nodes from which NUMA hinting faults are triggered. This can
2078 * be different from the set of nodes where the workload's memory is currently
2081 static void numa_group_count_active_nodes(struct numa_group *numa_group)
2083 unsigned long faults, max_faults = 0;
2084 int nid, active_nodes = 0;
2086 for_each_online_node(nid) {
2087 faults = group_faults_cpu(numa_group, nid);
2088 if (faults > max_faults)
2089 max_faults = faults;
2092 for_each_online_node(nid) {
2093 faults = group_faults_cpu(numa_group, nid);
2094 if (faults * ACTIVE_NODE_FRACTION > max_faults)
2098 numa_group->max_faults_cpu = max_faults;
2099 numa_group->active_nodes = active_nodes;
2103 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
2104 * increments. The more local the fault statistics are, the higher the scan
2105 * period will be for the next scan window. If local/(local+remote) ratio is
2106 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
2107 * the scan period will decrease. Aim for 70% local accesses.
2109 #define NUMA_PERIOD_SLOTS 10
2110 #define NUMA_PERIOD_THRESHOLD 7
2113 * Increase the scan period (slow down scanning) if the majority of
2114 * our memory is already on our local node, or if the majority of
2115 * the page accesses are shared with other processes.
2116 * Otherwise, decrease the scan period.
2118 static void update_task_scan_period(struct task_struct *p,
2119 unsigned long shared, unsigned long private)
2121 unsigned int period_slot;
2122 int lr_ratio, ps_ratio;
2125 unsigned long remote = p->numa_faults_locality[0];
2126 unsigned long local = p->numa_faults_locality[1];
2129 * If there were no record hinting faults then either the task is
2130 * completely idle or all activity is in areas that are not of interest
2131 * to automatic numa balancing. Related to that, if there were failed
2132 * migration then it implies we are migrating too quickly or the local
2133 * node is overloaded. In either case, scan slower
2135 if (local + shared == 0 || p->numa_faults_locality[2]) {
2136 p->numa_scan_period = min(p->numa_scan_period_max,
2137 p->numa_scan_period << 1);
2139 p->mm->numa_next_scan = jiffies +
2140 msecs_to_jiffies(p->numa_scan_period);
2146 * Prepare to scale scan period relative to the current period.
2147 * == NUMA_PERIOD_THRESHOLD scan period stays the same
2148 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2149 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2151 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
2152 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
2153 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
2155 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
2157 * Most memory accesses are local. There is no need to
2158 * do fast NUMA scanning, since memory is already local.
2160 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2163 diff = slot * period_slot;
2164 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2166 * Most memory accesses are shared with other tasks.
2167 * There is no point in continuing fast NUMA scanning,
2168 * since other tasks may just move the memory elsewhere.
2170 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2173 diff = slot * period_slot;
2176 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2177 * yet they are not on the local NUMA node. Speed up
2178 * NUMA scanning to get the memory moved over.
2180 int ratio = max(lr_ratio, ps_ratio);
2181 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2184 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2185 task_scan_min(p), task_scan_max(p));
2186 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2190 * Get the fraction of time the task has been running since the last
2191 * NUMA placement cycle. The scheduler keeps similar statistics, but
2192 * decays those on a 32ms period, which is orders of magnitude off
2193 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2194 * stats only if the task is so new there are no NUMA statistics yet.
2196 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2198 u64 runtime, delta, now;
2199 /* Use the start of this time slice to avoid calculations. */
2200 now = p->se.exec_start;
2201 runtime = p->se.sum_exec_runtime;
2203 if (p->last_task_numa_placement) {
2204 delta = runtime - p->last_sum_exec_runtime;
2205 *period = now - p->last_task_numa_placement;
2207 /* Avoid time going backwards, prevent potential divide error: */
2208 if (unlikely((s64)*period < 0))
2211 delta = p->se.avg.load_sum;
2212 *period = LOAD_AVG_MAX;
2215 p->last_sum_exec_runtime = runtime;
2216 p->last_task_numa_placement = now;
2222 * Determine the preferred nid for a task in a numa_group. This needs to
2223 * be done in a way that produces consistent results with group_weight,
2224 * otherwise workloads might not converge.
2226 static int preferred_group_nid(struct task_struct *p, int nid)
2231 /* Direct connections between all NUMA nodes. */
2232 if (sched_numa_topology_type == NUMA_DIRECT)
2236 * On a system with glueless mesh NUMA topology, group_weight
2237 * scores nodes according to the number of NUMA hinting faults on
2238 * both the node itself, and on nearby nodes.
2240 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2241 unsigned long score, max_score = 0;
2242 int node, max_node = nid;
2244 dist = sched_max_numa_distance;
2246 for_each_online_node(node) {
2247 score = group_weight(p, node, dist);
2248 if (score > max_score) {
2257 * Finding the preferred nid in a system with NUMA backplane
2258 * interconnect topology is more involved. The goal is to locate
2259 * tasks from numa_groups near each other in the system, and
2260 * untangle workloads from different sides of the system. This requires
2261 * searching down the hierarchy of node groups, recursively searching
2262 * inside the highest scoring group of nodes. The nodemask tricks
2263 * keep the complexity of the search down.
2265 nodes = node_online_map;
2266 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2267 unsigned long max_faults = 0;
2268 nodemask_t max_group = NODE_MASK_NONE;
2271 /* Are there nodes at this distance from each other? */
2272 if (!find_numa_distance(dist))
2275 for_each_node_mask(a, nodes) {
2276 unsigned long faults = 0;
2277 nodemask_t this_group;
2278 nodes_clear(this_group);
2280 /* Sum group's NUMA faults; includes a==b case. */
2281 for_each_node_mask(b, nodes) {
2282 if (node_distance(a, b) < dist) {
2283 faults += group_faults(p, b);
2284 node_set(b, this_group);
2285 node_clear(b, nodes);
2289 /* Remember the top group. */
2290 if (faults > max_faults) {
2291 max_faults = faults;
2292 max_group = this_group;
2294 * subtle: at the smallest distance there is
2295 * just one node left in each "group", the
2296 * winner is the preferred nid.
2301 /* Next round, evaluate the nodes within max_group. */
2309 static void task_numa_placement(struct task_struct *p)
2311 int seq, nid, max_nid = NUMA_NO_NODE;
2312 unsigned long max_faults = 0;
2313 unsigned long fault_types[2] = { 0, 0 };
2314 unsigned long total_faults;
2315 u64 runtime, period;
2316 spinlock_t *group_lock = NULL;
2317 struct numa_group *ng;
2320 * The p->mm->numa_scan_seq field gets updated without
2321 * exclusive access. Use READ_ONCE() here to ensure
2322 * that the field is read in a single access:
2324 seq = READ_ONCE(p->mm->numa_scan_seq);
2325 if (p->numa_scan_seq == seq)
2327 p->numa_scan_seq = seq;
2328 p->numa_scan_period_max = task_scan_max(p);
2330 total_faults = p->numa_faults_locality[0] +
2331 p->numa_faults_locality[1];
2332 runtime = numa_get_avg_runtime(p, &period);
2334 /* If the task is part of a group prevent parallel updates to group stats */
2335 ng = deref_curr_numa_group(p);
2337 group_lock = &ng->lock;
2338 spin_lock_irq(group_lock);
2341 /* Find the node with the highest number of faults */
2342 for_each_online_node(nid) {
2343 /* Keep track of the offsets in numa_faults array */
2344 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2345 unsigned long faults = 0, group_faults = 0;
2348 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2349 long diff, f_diff, f_weight;
2351 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2352 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2353 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2354 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2356 /* Decay existing window, copy faults since last scan */
2357 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2358 fault_types[priv] += p->numa_faults[membuf_idx];
2359 p->numa_faults[membuf_idx] = 0;
2362 * Normalize the faults_from, so all tasks in a group
2363 * count according to CPU use, instead of by the raw
2364 * number of faults. Tasks with little runtime have
2365 * little over-all impact on throughput, and thus their
2366 * faults are less important.
2368 f_weight = div64_u64(runtime << 16, period + 1);
2369 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2371 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2372 p->numa_faults[cpubuf_idx] = 0;
2374 p->numa_faults[mem_idx] += diff;
2375 p->numa_faults[cpu_idx] += f_diff;
2376 faults += p->numa_faults[mem_idx];
2377 p->total_numa_faults += diff;
2380 * safe because we can only change our own group
2382 * mem_idx represents the offset for a given
2383 * nid and priv in a specific region because it
2384 * is at the beginning of the numa_faults array.
2386 ng->faults[mem_idx] += diff;
2387 ng->faults[cpu_idx] += f_diff;
2388 ng->total_faults += diff;
2389 group_faults += ng->faults[mem_idx];
2394 if (faults > max_faults) {
2395 max_faults = faults;
2398 } else if (group_faults > max_faults) {
2399 max_faults = group_faults;
2405 numa_group_count_active_nodes(ng);
2406 spin_unlock_irq(group_lock);
2407 max_nid = preferred_group_nid(p, max_nid);
2411 /* Set the new preferred node */
2412 if (max_nid != p->numa_preferred_nid)
2413 sched_setnuma(p, max_nid);
2416 update_task_scan_period(p, fault_types[0], fault_types[1]);
2419 static inline int get_numa_group(struct numa_group *grp)
2421 return refcount_inc_not_zero(&grp->refcount);
2424 static inline void put_numa_group(struct numa_group *grp)
2426 if (refcount_dec_and_test(&grp->refcount))
2427 kfree_rcu(grp, rcu);
2430 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2433 struct numa_group *grp, *my_grp;
2434 struct task_struct *tsk;
2436 int cpu = cpupid_to_cpu(cpupid);
2439 if (unlikely(!deref_curr_numa_group(p))) {
2440 unsigned int size = sizeof(struct numa_group) +
2441 NR_NUMA_HINT_FAULT_STATS *
2442 nr_node_ids * sizeof(unsigned long);
2444 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2448 refcount_set(&grp->refcount, 1);
2449 grp->active_nodes = 1;
2450 grp->max_faults_cpu = 0;
2451 spin_lock_init(&grp->lock);
2454 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2455 grp->faults[i] = p->numa_faults[i];
2457 grp->total_faults = p->total_numa_faults;
2460 rcu_assign_pointer(p->numa_group, grp);
2464 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2466 if (!cpupid_match_pid(tsk, cpupid))
2469 grp = rcu_dereference(tsk->numa_group);
2473 my_grp = deref_curr_numa_group(p);
2478 * Only join the other group if its bigger; if we're the bigger group,
2479 * the other task will join us.
2481 if (my_grp->nr_tasks > grp->nr_tasks)
2485 * Tie-break on the grp address.
2487 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2490 /* Always join threads in the same process. */
2491 if (tsk->mm == current->mm)
2494 /* Simple filter to avoid false positives due to PID collisions */
2495 if (flags & TNF_SHARED)
2498 /* Update priv based on whether false sharing was detected */
2501 if (join && !get_numa_group(grp))
2509 BUG_ON(irqs_disabled());
2510 double_lock_irq(&my_grp->lock, &grp->lock);
2512 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2513 my_grp->faults[i] -= p->numa_faults[i];
2514 grp->faults[i] += p->numa_faults[i];
2516 my_grp->total_faults -= p->total_numa_faults;
2517 grp->total_faults += p->total_numa_faults;
2522 spin_unlock(&my_grp->lock);
2523 spin_unlock_irq(&grp->lock);
2525 rcu_assign_pointer(p->numa_group, grp);
2527 put_numa_group(my_grp);
2536 * Get rid of NUMA statistics associated with a task (either current or dead).
2537 * If @final is set, the task is dead and has reached refcount zero, so we can
2538 * safely free all relevant data structures. Otherwise, there might be
2539 * concurrent reads from places like load balancing and procfs, and we should
2540 * reset the data back to default state without freeing ->numa_faults.
2542 void task_numa_free(struct task_struct *p, bool final)
2544 /* safe: p either is current or is being freed by current */
2545 struct numa_group *grp = rcu_dereference_raw(p->numa_group);
2546 unsigned long *numa_faults = p->numa_faults;
2547 unsigned long flags;
2554 spin_lock_irqsave(&grp->lock, flags);
2555 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2556 grp->faults[i] -= p->numa_faults[i];
2557 grp->total_faults -= p->total_numa_faults;
2560 spin_unlock_irqrestore(&grp->lock, flags);
2561 RCU_INIT_POINTER(p->numa_group, NULL);
2562 put_numa_group(grp);
2566 p->numa_faults = NULL;
2569 p->total_numa_faults = 0;
2570 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2576 * Got a PROT_NONE fault for a page on @node.
2578 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2580 struct task_struct *p = current;
2581 bool migrated = flags & TNF_MIGRATED;
2582 int cpu_node = task_node(current);
2583 int local = !!(flags & TNF_FAULT_LOCAL);
2584 struct numa_group *ng;
2587 if (!static_branch_likely(&sched_numa_balancing))
2590 /* for example, ksmd faulting in a user's mm */
2594 /* Allocate buffer to track faults on a per-node basis */
2595 if (unlikely(!p->numa_faults)) {
2596 int size = sizeof(*p->numa_faults) *
2597 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2599 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2600 if (!p->numa_faults)
2603 p->total_numa_faults = 0;
2604 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2608 * First accesses are treated as private, otherwise consider accesses
2609 * to be private if the accessing pid has not changed
2611 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2614 priv = cpupid_match_pid(p, last_cpupid);
2615 if (!priv && !(flags & TNF_NO_GROUP))
2616 task_numa_group(p, last_cpupid, flags, &priv);
2620 * If a workload spans multiple NUMA nodes, a shared fault that
2621 * occurs wholly within the set of nodes that the workload is
2622 * actively using should be counted as local. This allows the
2623 * scan rate to slow down when a workload has settled down.
2625 ng = deref_curr_numa_group(p);
2626 if (!priv && !local && ng && ng->active_nodes > 1 &&
2627 numa_is_active_node(cpu_node, ng) &&
2628 numa_is_active_node(mem_node, ng))
2632 * Retry to migrate task to preferred node periodically, in case it
2633 * previously failed, or the scheduler moved us.
2635 if (time_after(jiffies, p->numa_migrate_retry)) {
2636 task_numa_placement(p);
2637 numa_migrate_preferred(p);
2641 p->numa_pages_migrated += pages;
2642 if (flags & TNF_MIGRATE_FAIL)
2643 p->numa_faults_locality[2] += pages;
2645 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2646 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2647 p->numa_faults_locality[local] += pages;
2650 static void reset_ptenuma_scan(struct task_struct *p)
2653 * We only did a read acquisition of the mmap sem, so
2654 * p->mm->numa_scan_seq is written to without exclusive access
2655 * and the update is not guaranteed to be atomic. That's not
2656 * much of an issue though, since this is just used for
2657 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2658 * expensive, to avoid any form of compiler optimizations:
2660 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2661 p->mm->numa_scan_offset = 0;
2665 * The expensive part of numa migration is done from task_work context.
2666 * Triggered from task_tick_numa().
2668 static void task_numa_work(struct callback_head *work)
2670 unsigned long migrate, next_scan, now = jiffies;
2671 struct task_struct *p = current;
2672 struct mm_struct *mm = p->mm;
2673 u64 runtime = p->se.sum_exec_runtime;
2674 struct vm_area_struct *vma;
2675 unsigned long start, end;
2676 unsigned long nr_pte_updates = 0;
2677 long pages, virtpages;
2679 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2683 * Who cares about NUMA placement when they're dying.
2685 * NOTE: make sure not to dereference p->mm before this check,
2686 * exit_task_work() happens _after_ exit_mm() so we could be called
2687 * without p->mm even though we still had it when we enqueued this
2690 if (p->flags & PF_EXITING)
2693 if (!mm->numa_next_scan) {
2694 mm->numa_next_scan = now +
2695 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2699 * Enforce maximal scan/migration frequency..
2701 migrate = mm->numa_next_scan;
2702 if (time_before(now, migrate))
2705 if (p->numa_scan_period == 0) {
2706 p->numa_scan_period_max = task_scan_max(p);
2707 p->numa_scan_period = task_scan_start(p);
2710 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2711 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2715 * Delay this task enough that another task of this mm will likely win
2716 * the next time around.
2718 p->node_stamp += 2 * TICK_NSEC;
2720 start = mm->numa_scan_offset;
2721 pages = sysctl_numa_balancing_scan_size;
2722 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2723 virtpages = pages * 8; /* Scan up to this much virtual space */
2728 if (!mmap_read_trylock(mm))
2730 vma = find_vma(mm, start);
2732 reset_ptenuma_scan(p);
2736 for (; vma; vma = vma->vm_next) {
2737 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2738 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2743 * Shared library pages mapped by multiple processes are not
2744 * migrated as it is expected they are cache replicated. Avoid
2745 * hinting faults in read-only file-backed mappings or the vdso
2746 * as migrating the pages will be of marginal benefit.
2749 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2753 * Skip inaccessible VMAs to avoid any confusion between
2754 * PROT_NONE and NUMA hinting ptes
2756 if (!vma_is_accessible(vma))
2760 start = max(start, vma->vm_start);
2761 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2762 end = min(end, vma->vm_end);
2763 nr_pte_updates = change_prot_numa(vma, start, end);
2766 * Try to scan sysctl_numa_balancing_size worth of
2767 * hpages that have at least one present PTE that
2768 * is not already pte-numa. If the VMA contains
2769 * areas that are unused or already full of prot_numa
2770 * PTEs, scan up to virtpages, to skip through those
2774 pages -= (end - start) >> PAGE_SHIFT;
2775 virtpages -= (end - start) >> PAGE_SHIFT;
2778 if (pages <= 0 || virtpages <= 0)
2782 } while (end != vma->vm_end);
2787 * It is possible to reach the end of the VMA list but the last few
2788 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2789 * would find the !migratable VMA on the next scan but not reset the
2790 * scanner to the start so check it now.
2793 mm->numa_scan_offset = start;
2795 reset_ptenuma_scan(p);
2796 mmap_read_unlock(mm);
2799 * Make sure tasks use at least 32x as much time to run other code
2800 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2801 * Usually update_task_scan_period slows down scanning enough; on an
2802 * overloaded system we need to limit overhead on a per task basis.
2804 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2805 u64 diff = p->se.sum_exec_runtime - runtime;
2806 p->node_stamp += 32 * diff;
2810 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
2813 struct mm_struct *mm = p->mm;
2816 mm_users = atomic_read(&mm->mm_users);
2817 if (mm_users == 1) {
2818 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2819 mm->numa_scan_seq = 0;
2823 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
2824 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2825 /* Protect against double add, see task_tick_numa and task_numa_work */
2826 p->numa_work.next = &p->numa_work;
2827 p->numa_faults = NULL;
2828 RCU_INIT_POINTER(p->numa_group, NULL);
2829 p->last_task_numa_placement = 0;
2830 p->last_sum_exec_runtime = 0;
2832 init_task_work(&p->numa_work, task_numa_work);
2834 /* New address space, reset the preferred nid */
2835 if (!(clone_flags & CLONE_VM)) {
2836 p->numa_preferred_nid = NUMA_NO_NODE;
2841 * New thread, keep existing numa_preferred_nid which should be copied
2842 * already by arch_dup_task_struct but stagger when scans start.
2847 delay = min_t(unsigned int, task_scan_max(current),
2848 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
2849 delay += 2 * TICK_NSEC;
2850 p->node_stamp = delay;
2855 * Drive the periodic memory faults..
2857 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2859 struct callback_head *work = &curr->numa_work;
2863 * We don't care about NUMA placement if we don't have memory.
2865 if ((curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
2869 * Using runtime rather than walltime has the dual advantage that
2870 * we (mostly) drive the selection from busy threads and that the
2871 * task needs to have done some actual work before we bother with
2874 now = curr->se.sum_exec_runtime;
2875 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2877 if (now > curr->node_stamp + period) {
2878 if (!curr->node_stamp)
2879 curr->numa_scan_period = task_scan_start(curr);
2880 curr->node_stamp += period;
2882 if (!time_before(jiffies, curr->mm->numa_next_scan))
2883 task_work_add(curr, work, TWA_RESUME);
2887 static void update_scan_period(struct task_struct *p, int new_cpu)
2889 int src_nid = cpu_to_node(task_cpu(p));
2890 int dst_nid = cpu_to_node(new_cpu);
2892 if (!static_branch_likely(&sched_numa_balancing))
2895 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2898 if (src_nid == dst_nid)
2902 * Allow resets if faults have been trapped before one scan
2903 * has completed. This is most likely due to a new task that
2904 * is pulled cross-node due to wakeups or load balancing.
2906 if (p->numa_scan_seq) {
2908 * Avoid scan adjustments if moving to the preferred
2909 * node or if the task was not previously running on
2910 * the preferred node.
2912 if (dst_nid == p->numa_preferred_nid ||
2913 (p->numa_preferred_nid != NUMA_NO_NODE &&
2914 src_nid != p->numa_preferred_nid))
2918 p->numa_scan_period = task_scan_start(p);
2922 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2926 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2930 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2934 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2938 #endif /* CONFIG_NUMA_BALANCING */
2941 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2943 update_load_add(&cfs_rq->load, se->load.weight);
2945 if (entity_is_task(se)) {
2946 struct rq *rq = rq_of(cfs_rq);
2948 account_numa_enqueue(rq, task_of(se));
2949 list_add(&se->group_node, &rq->cfs_tasks);
2952 cfs_rq->nr_running++;
2954 cfs_rq->idle_nr_running++;
2958 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2960 update_load_sub(&cfs_rq->load, se->load.weight);
2962 if (entity_is_task(se)) {
2963 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2964 list_del_init(&se->group_node);
2967 cfs_rq->nr_running--;
2969 cfs_rq->idle_nr_running--;
2973 * Signed add and clamp on underflow.
2975 * Explicitly do a load-store to ensure the intermediate value never hits
2976 * memory. This allows lockless observations without ever seeing the negative
2979 #define add_positive(_ptr, _val) do { \
2980 typeof(_ptr) ptr = (_ptr); \
2981 typeof(_val) val = (_val); \
2982 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2986 if (val < 0 && res > var) \
2989 WRITE_ONCE(*ptr, res); \
2993 * Unsigned subtract and clamp on underflow.
2995 * Explicitly do a load-store to ensure the intermediate value never hits
2996 * memory. This allows lockless observations without ever seeing the negative
2999 #define sub_positive(_ptr, _val) do { \
3000 typeof(_ptr) ptr = (_ptr); \
3001 typeof(*ptr) val = (_val); \
3002 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3006 WRITE_ONCE(*ptr, res); \
3010 * Remove and clamp on negative, from a local variable.
3012 * A variant of sub_positive(), which does not use explicit load-store
3013 * and is thus optimized for local variable updates.
3015 #define lsub_positive(_ptr, _val) do { \
3016 typeof(_ptr) ptr = (_ptr); \
3017 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
3022 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3024 cfs_rq->avg.load_avg += se->avg.load_avg;
3025 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
3029 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3031 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3032 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
3033 /* See update_cfs_rq_load_avg() */
3034 cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
3035 cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
3039 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3041 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3044 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
3045 unsigned long weight)
3048 /* commit outstanding execution time */
3049 if (cfs_rq->curr == se)
3050 update_curr(cfs_rq);
3051 update_load_sub(&cfs_rq->load, se->load.weight);
3053 dequeue_load_avg(cfs_rq, se);
3055 update_load_set(&se->load, weight);
3059 u32 divider = get_pelt_divider(&se->avg);
3061 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
3065 enqueue_load_avg(cfs_rq, se);
3067 update_load_add(&cfs_rq->load, se->load.weight);
3071 void reweight_task(struct task_struct *p, int prio)
3073 struct sched_entity *se = &p->se;
3074 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3075 struct load_weight *load = &se->load;
3076 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
3078 reweight_entity(cfs_rq, se, weight);
3079 load->inv_weight = sched_prio_to_wmult[prio];
3082 #ifdef CONFIG_FAIR_GROUP_SCHED
3085 * All this does is approximate the hierarchical proportion which includes that
3086 * global sum we all love to hate.
3088 * That is, the weight of a group entity, is the proportional share of the
3089 * group weight based on the group runqueue weights. That is:
3091 * tg->weight * grq->load.weight
3092 * ge->load.weight = ----------------------------- (1)
3093 * \Sum grq->load.weight
3095 * Now, because computing that sum is prohibitively expensive to compute (been
3096 * there, done that) we approximate it with this average stuff. The average
3097 * moves slower and therefore the approximation is cheaper and more stable.
3099 * So instead of the above, we substitute:
3101 * grq->load.weight -> grq->avg.load_avg (2)
3103 * which yields the following:
3105 * tg->weight * grq->avg.load_avg
3106 * ge->load.weight = ------------------------------ (3)
3109 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3111 * That is shares_avg, and it is right (given the approximation (2)).
3113 * The problem with it is that because the average is slow -- it was designed
3114 * to be exactly that of course -- this leads to transients in boundary
3115 * conditions. In specific, the case where the group was idle and we start the
3116 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3117 * yielding bad latency etc..
3119 * Now, in that special case (1) reduces to:
3121 * tg->weight * grq->load.weight
3122 * ge->load.weight = ----------------------------- = tg->weight (4)
3125 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3127 * So what we do is modify our approximation (3) to approach (4) in the (near)
3132 * tg->weight * grq->load.weight
3133 * --------------------------------------------------- (5)
3134 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3136 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3137 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3140 * tg->weight * grq->load.weight
3141 * ge->load.weight = ----------------------------- (6)
3146 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3147 * max(grq->load.weight, grq->avg.load_avg)
3149 * And that is shares_weight and is icky. In the (near) UP case it approaches
3150 * (4) while in the normal case it approaches (3). It consistently
3151 * overestimates the ge->load.weight and therefore:
3153 * \Sum ge->load.weight >= tg->weight
3157 static long calc_group_shares(struct cfs_rq *cfs_rq)
3159 long tg_weight, tg_shares, load, shares;
3160 struct task_group *tg = cfs_rq->tg;
3162 tg_shares = READ_ONCE(tg->shares);
3164 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3166 tg_weight = atomic_long_read(&tg->load_avg);
3168 /* Ensure tg_weight >= load */
3169 tg_weight -= cfs_rq->tg_load_avg_contrib;
3172 shares = (tg_shares * load);
3174 shares /= tg_weight;
3177 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3178 * of a group with small tg->shares value. It is a floor value which is
3179 * assigned as a minimum load.weight to the sched_entity representing
3180 * the group on a CPU.
3182 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3183 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3184 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3185 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3188 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3190 #endif /* CONFIG_SMP */
3192 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3195 * Recomputes the group entity based on the current state of its group
3198 static void update_cfs_group(struct sched_entity *se)
3200 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3206 if (throttled_hierarchy(gcfs_rq))
3210 shares = READ_ONCE(gcfs_rq->tg->shares);
3212 if (likely(se->load.weight == shares))
3215 shares = calc_group_shares(gcfs_rq);
3218 reweight_entity(cfs_rq_of(se), se, shares);
3221 #else /* CONFIG_FAIR_GROUP_SCHED */
3222 static inline void update_cfs_group(struct sched_entity *se)
3225 #endif /* CONFIG_FAIR_GROUP_SCHED */
3227 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3229 struct rq *rq = rq_of(cfs_rq);
3231 if (&rq->cfs == cfs_rq) {
3233 * There are a few boundary cases this might miss but it should
3234 * get called often enough that that should (hopefully) not be
3237 * It will not get called when we go idle, because the idle
3238 * thread is a different class (!fair), nor will the utilization
3239 * number include things like RT tasks.
3241 * As is, the util number is not freq-invariant (we'd have to
3242 * implement arch_scale_freq_capacity() for that).
3244 * See cpu_util_cfs().
3246 cpufreq_update_util(rq, flags);
3251 #ifdef CONFIG_FAIR_GROUP_SCHED
3253 * Because list_add_leaf_cfs_rq always places a child cfs_rq on the list
3254 * immediately before a parent cfs_rq, and cfs_rqs are removed from the list
3255 * bottom-up, we only have to test whether the cfs_rq before us on the list
3257 * If cfs_rq is not on the list, test whether a child needs its to be added to
3258 * connect a branch to the tree * (see list_add_leaf_cfs_rq() for details).
3260 static inline bool child_cfs_rq_on_list(struct cfs_rq *cfs_rq)
3262 struct cfs_rq *prev_cfs_rq;
3263 struct list_head *prev;
3265 if (cfs_rq->on_list) {
3266 prev = cfs_rq->leaf_cfs_rq_list.prev;
3268 struct rq *rq = rq_of(cfs_rq);
3270 prev = rq->tmp_alone_branch;
3273 prev_cfs_rq = container_of(prev, struct cfs_rq, leaf_cfs_rq_list);
3275 return (prev_cfs_rq->tg->parent == cfs_rq->tg);
3278 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
3280 if (cfs_rq->load.weight)
3283 if (cfs_rq->avg.load_sum)
3286 if (cfs_rq->avg.util_sum)
3289 if (cfs_rq->avg.runnable_sum)
3292 if (child_cfs_rq_on_list(cfs_rq))
3296 * _avg must be null when _sum are null because _avg = _sum / divider
3297 * Make sure that rounding and/or propagation of PELT values never
3300 SCHED_WARN_ON(cfs_rq->avg.load_avg ||
3301 cfs_rq->avg.util_avg ||
3302 cfs_rq->avg.runnable_avg);
3308 * update_tg_load_avg - update the tg's load avg
3309 * @cfs_rq: the cfs_rq whose avg changed
3311 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3312 * However, because tg->load_avg is a global value there are performance
3315 * In order to avoid having to look at the other cfs_rq's, we use a
3316 * differential update where we store the last value we propagated. This in
3317 * turn allows skipping updates if the differential is 'small'.
3319 * Updating tg's load_avg is necessary before update_cfs_share().
3321 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
3323 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3326 * No need to update load_avg for root_task_group as it is not used.
3328 if (cfs_rq->tg == &root_task_group)
3331 if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3332 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3333 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3338 * Called within set_task_rq() right before setting a task's CPU. The
3339 * caller only guarantees p->pi_lock is held; no other assumptions,
3340 * including the state of rq->lock, should be made.
3342 void set_task_rq_fair(struct sched_entity *se,
3343 struct cfs_rq *prev, struct cfs_rq *next)
3345 u64 p_last_update_time;
3346 u64 n_last_update_time;
3348 if (!sched_feat(ATTACH_AGE_LOAD))
3352 * We are supposed to update the task to "current" time, then its up to
3353 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3354 * getting what current time is, so simply throw away the out-of-date
3355 * time. This will result in the wakee task is less decayed, but giving
3356 * the wakee more load sounds not bad.
3358 if (!(se->avg.last_update_time && prev))
3361 #ifndef CONFIG_64BIT
3363 u64 p_last_update_time_copy;
3364 u64 n_last_update_time_copy;
3367 p_last_update_time_copy = prev->load_last_update_time_copy;
3368 n_last_update_time_copy = next->load_last_update_time_copy;
3372 p_last_update_time = prev->avg.last_update_time;
3373 n_last_update_time = next->avg.last_update_time;
3375 } while (p_last_update_time != p_last_update_time_copy ||
3376 n_last_update_time != n_last_update_time_copy);
3379 p_last_update_time = prev->avg.last_update_time;
3380 n_last_update_time = next->avg.last_update_time;
3382 __update_load_avg_blocked_se(p_last_update_time, se);
3383 se->avg.last_update_time = n_last_update_time;
3387 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3388 * propagate its contribution. The key to this propagation is the invariant
3389 * that for each group:
3391 * ge->avg == grq->avg (1)
3393 * _IFF_ we look at the pure running and runnable sums. Because they
3394 * represent the very same entity, just at different points in the hierarchy.
3396 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
3397 * and simply copies the running/runnable sum over (but still wrong, because
3398 * the group entity and group rq do not have their PELT windows aligned).
3400 * However, update_tg_cfs_load() is more complex. So we have:
3402 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3404 * And since, like util, the runnable part should be directly transferable,
3405 * the following would _appear_ to be the straight forward approach:
3407 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3409 * And per (1) we have:
3411 * ge->avg.runnable_avg == grq->avg.runnable_avg
3415 * ge->load.weight * grq->avg.load_avg
3416 * ge->avg.load_avg = ----------------------------------- (4)
3419 * Except that is wrong!
3421 * Because while for entities historical weight is not important and we
3422 * really only care about our future and therefore can consider a pure
3423 * runnable sum, runqueues can NOT do this.
3425 * We specifically want runqueues to have a load_avg that includes
3426 * historical weights. Those represent the blocked load, the load we expect
3427 * to (shortly) return to us. This only works by keeping the weights as
3428 * integral part of the sum. We therefore cannot decompose as per (3).
3430 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3431 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3432 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3433 * runnable section of these tasks overlap (or not). If they were to perfectly
3434 * align the rq as a whole would be runnable 2/3 of the time. If however we
3435 * always have at least 1 runnable task, the rq as a whole is always runnable.
3437 * So we'll have to approximate.. :/
3439 * Given the constraint:
3441 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3443 * We can construct a rule that adds runnable to a rq by assuming minimal
3446 * On removal, we'll assume each task is equally runnable; which yields:
3448 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3450 * XXX: only do this for the part of runnable > running ?
3454 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3456 long delta_sum, delta_avg = gcfs_rq->avg.util_avg - se->avg.util_avg;
3457 u32 new_sum, divider;
3459 /* Nothing to update */
3464 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3465 * See ___update_load_avg() for details.
3467 divider = get_pelt_divider(&cfs_rq->avg);
3470 /* Set new sched_entity's utilization */
3471 se->avg.util_avg = gcfs_rq->avg.util_avg;
3472 new_sum = se->avg.util_avg * divider;
3473 delta_sum = (long)new_sum - (long)se->avg.util_sum;
3474 se->avg.util_sum = new_sum;
3476 /* Update parent cfs_rq utilization */
3477 add_positive(&cfs_rq->avg.util_avg, delta_avg);
3478 add_positive(&cfs_rq->avg.util_sum, delta_sum);
3480 /* See update_cfs_rq_load_avg() */
3481 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
3482 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
3486 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3488 long delta_sum, delta_avg = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
3489 u32 new_sum, divider;
3491 /* Nothing to update */
3496 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3497 * See ___update_load_avg() for details.
3499 divider = get_pelt_divider(&cfs_rq->avg);
3501 /* Set new sched_entity's runnable */
3502 se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
3503 new_sum = se->avg.runnable_avg * divider;
3504 delta_sum = (long)new_sum - (long)se->avg.runnable_sum;
3505 se->avg.runnable_sum = new_sum;
3507 /* Update parent cfs_rq runnable */
3508 add_positive(&cfs_rq->avg.runnable_avg, delta_avg);
3509 add_positive(&cfs_rq->avg.runnable_sum, delta_sum);
3510 /* See update_cfs_rq_load_avg() */
3511 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
3512 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
3516 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3518 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3519 unsigned long load_avg;
3527 gcfs_rq->prop_runnable_sum = 0;
3530 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3531 * See ___update_load_avg() for details.
3533 divider = get_pelt_divider(&cfs_rq->avg);
3535 if (runnable_sum >= 0) {
3537 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3538 * the CPU is saturated running == runnable.
3540 runnable_sum += se->avg.load_sum;
3541 runnable_sum = min_t(long, runnable_sum, divider);
3544 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3545 * assuming all tasks are equally runnable.
3547 if (scale_load_down(gcfs_rq->load.weight)) {
3548 load_sum = div_u64(gcfs_rq->avg.load_sum,
3549 scale_load_down(gcfs_rq->load.weight));
3552 /* But make sure to not inflate se's runnable */
3553 runnable_sum = min(se->avg.load_sum, load_sum);
3557 * runnable_sum can't be lower than running_sum
3558 * Rescale running sum to be in the same range as runnable sum
3559 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3560 * runnable_sum is in [0 : LOAD_AVG_MAX]
3562 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3563 runnable_sum = max(runnable_sum, running_sum);
3565 load_sum = se_weight(se) * runnable_sum;
3566 load_avg = div_u64(load_sum, divider);
3568 delta_avg = load_avg - se->avg.load_avg;
3572 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3574 se->avg.load_sum = runnable_sum;
3575 se->avg.load_avg = load_avg;
3576 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3577 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3578 /* See update_cfs_rq_load_avg() */
3579 cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
3580 cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
3583 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3585 cfs_rq->propagate = 1;
3586 cfs_rq->prop_runnable_sum += runnable_sum;
3589 /* Update task and its cfs_rq load average */
3590 static inline int propagate_entity_load_avg(struct sched_entity *se)
3592 struct cfs_rq *cfs_rq, *gcfs_rq;
3594 if (entity_is_task(se))
3597 gcfs_rq = group_cfs_rq(se);
3598 if (!gcfs_rq->propagate)
3601 gcfs_rq->propagate = 0;
3603 cfs_rq = cfs_rq_of(se);
3605 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3607 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3608 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3609 update_tg_cfs_load(cfs_rq, se, gcfs_rq);
3611 trace_pelt_cfs_tp(cfs_rq);
3612 trace_pelt_se_tp(se);
3618 * Check if we need to update the load and the utilization of a blocked
3621 static inline bool skip_blocked_update(struct sched_entity *se)
3623 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3626 * If sched_entity still have not zero load or utilization, we have to
3629 if (se->avg.load_avg || se->avg.util_avg)
3633 * If there is a pending propagation, we have to update the load and
3634 * the utilization of the sched_entity:
3636 if (gcfs_rq->propagate)
3640 * Otherwise, the load and the utilization of the sched_entity is
3641 * already zero and there is no pending propagation, so it will be a
3642 * waste of time to try to decay it:
3647 #else /* CONFIG_FAIR_GROUP_SCHED */
3649 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {}
3651 static inline int propagate_entity_load_avg(struct sched_entity *se)
3656 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3658 #endif /* CONFIG_FAIR_GROUP_SCHED */
3661 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3662 * @now: current time, as per cfs_rq_clock_pelt()
3663 * @cfs_rq: cfs_rq to update
3665 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3666 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3667 * post_init_entity_util_avg().
3669 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3671 * Return: true if the load decayed or we removed load.
3673 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3674 * call update_tg_load_avg() when this function returns true.
3677 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3679 unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
3680 struct sched_avg *sa = &cfs_rq->avg;
3683 if (cfs_rq->removed.nr) {
3685 u32 divider = get_pelt_divider(&cfs_rq->avg);
3687 raw_spin_lock(&cfs_rq->removed.lock);
3688 swap(cfs_rq->removed.util_avg, removed_util);
3689 swap(cfs_rq->removed.load_avg, removed_load);
3690 swap(cfs_rq->removed.runnable_avg, removed_runnable);
3691 cfs_rq->removed.nr = 0;
3692 raw_spin_unlock(&cfs_rq->removed.lock);
3695 sub_positive(&sa->load_avg, r);
3696 sub_positive(&sa->load_sum, r * divider);
3697 /* See sa->util_sum below */
3698 sa->load_sum = max_t(u32, sa->load_sum, sa->load_avg * PELT_MIN_DIVIDER);
3701 sub_positive(&sa->util_avg, r);
3702 sub_positive(&sa->util_sum, r * divider);
3704 * Because of rounding, se->util_sum might ends up being +1 more than
3705 * cfs->util_sum. Although this is not a problem by itself, detaching
3706 * a lot of tasks with the rounding problem between 2 updates of
3707 * util_avg (~1ms) can make cfs->util_sum becoming null whereas
3708 * cfs_util_avg is not.
3709 * Check that util_sum is still above its lower bound for the new
3710 * util_avg. Given that period_contrib might have moved since the last
3711 * sync, we are only sure that util_sum must be above or equal to
3712 * util_avg * minimum possible divider
3714 sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * PELT_MIN_DIVIDER);
3716 r = removed_runnable;
3717 sub_positive(&sa->runnable_avg, r);
3718 sub_positive(&sa->runnable_sum, r * divider);
3719 /* See sa->util_sum above */
3720 sa->runnable_sum = max_t(u32, sa->runnable_sum,
3721 sa->runnable_avg * PELT_MIN_DIVIDER);
3724 * removed_runnable is the unweighted version of removed_load so we
3725 * can use it to estimate removed_load_sum.
3727 add_tg_cfs_propagate(cfs_rq,
3728 -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
3733 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
3735 #ifndef CONFIG_64BIT
3737 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3744 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3745 * @cfs_rq: cfs_rq to attach to
3746 * @se: sched_entity to attach
3748 * Must call update_cfs_rq_load_avg() before this, since we rely on
3749 * cfs_rq->avg.last_update_time being current.
3751 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3754 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3755 * See ___update_load_avg() for details.
3757 u32 divider = get_pelt_divider(&cfs_rq->avg);
3760 * When we attach the @se to the @cfs_rq, we must align the decay
3761 * window because without that, really weird and wonderful things can
3766 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3767 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3770 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3771 * period_contrib. This isn't strictly correct, but since we're
3772 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3775 se->avg.util_sum = se->avg.util_avg * divider;
3777 se->avg.runnable_sum = se->avg.runnable_avg * divider;
3779 se->avg.load_sum = se->avg.load_avg * divider;
3780 if (se_weight(se) < se->avg.load_sum)
3781 se->avg.load_sum = div_u64(se->avg.load_sum, se_weight(se));
3783 se->avg.load_sum = 1;
3785 enqueue_load_avg(cfs_rq, se);
3786 cfs_rq->avg.util_avg += se->avg.util_avg;
3787 cfs_rq->avg.util_sum += se->avg.util_sum;
3788 cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
3789 cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
3791 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3793 cfs_rq_util_change(cfs_rq, 0);
3795 trace_pelt_cfs_tp(cfs_rq);
3799 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3800 * @cfs_rq: cfs_rq to detach from
3801 * @se: sched_entity to detach
3803 * Must call update_cfs_rq_load_avg() before this, since we rely on
3804 * cfs_rq->avg.last_update_time being current.
3806 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3808 dequeue_load_avg(cfs_rq, se);
3809 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3810 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3811 /* See update_cfs_rq_load_avg() */
3812 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
3813 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
3815 sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);
3816 sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum);
3817 /* See update_cfs_rq_load_avg() */
3818 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
3819 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
3821 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3823 cfs_rq_util_change(cfs_rq, 0);
3825 trace_pelt_cfs_tp(cfs_rq);
3829 * Optional action to be done while updating the load average
3831 #define UPDATE_TG 0x1
3832 #define SKIP_AGE_LOAD 0x2
3833 #define DO_ATTACH 0x4
3835 /* Update task and its cfs_rq load average */
3836 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3838 u64 now = cfs_rq_clock_pelt(cfs_rq);
3842 * Track task load average for carrying it to new CPU after migrated, and
3843 * track group sched_entity load average for task_h_load calc in migration
3845 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3846 __update_load_avg_se(now, cfs_rq, se);
3848 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3849 decayed |= propagate_entity_load_avg(se);
3851 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3854 * DO_ATTACH means we're here from enqueue_entity().
3855 * !last_update_time means we've passed through
3856 * migrate_task_rq_fair() indicating we migrated.
3858 * IOW we're enqueueing a task on a new CPU.
3860 attach_entity_load_avg(cfs_rq, se);
3861 update_tg_load_avg(cfs_rq);
3863 } else if (decayed) {
3864 cfs_rq_util_change(cfs_rq, 0);
3866 if (flags & UPDATE_TG)
3867 update_tg_load_avg(cfs_rq);
3871 #ifndef CONFIG_64BIT
3872 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3874 u64 last_update_time_copy;
3875 u64 last_update_time;
3878 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3880 last_update_time = cfs_rq->avg.last_update_time;
3881 } while (last_update_time != last_update_time_copy);
3883 return last_update_time;
3886 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3888 return cfs_rq->avg.last_update_time;
3893 * Synchronize entity load avg of dequeued entity without locking
3896 static void sync_entity_load_avg(struct sched_entity *se)
3898 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3899 u64 last_update_time;
3901 last_update_time = cfs_rq_last_update_time(cfs_rq);
3902 __update_load_avg_blocked_se(last_update_time, se);
3906 * Task first catches up with cfs_rq, and then subtract
3907 * itself from the cfs_rq (task must be off the queue now).
3909 static void remove_entity_load_avg(struct sched_entity *se)
3911 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3912 unsigned long flags;
3915 * tasks cannot exit without having gone through wake_up_new_task() ->
3916 * post_init_entity_util_avg() which will have added things to the
3917 * cfs_rq, so we can remove unconditionally.
3920 sync_entity_load_avg(se);
3922 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3923 ++cfs_rq->removed.nr;
3924 cfs_rq->removed.util_avg += se->avg.util_avg;
3925 cfs_rq->removed.load_avg += se->avg.load_avg;
3926 cfs_rq->removed.runnable_avg += se->avg.runnable_avg;
3927 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3930 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
3932 return cfs_rq->avg.runnable_avg;
3935 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3937 return cfs_rq->avg.load_avg;
3940 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf);
3942 static inline unsigned long task_util(struct task_struct *p)
3944 return READ_ONCE(p->se.avg.util_avg);
3947 static inline unsigned long _task_util_est(struct task_struct *p)
3949 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3951 return max(ue.ewma, (ue.enqueued & ~UTIL_AVG_UNCHANGED));
3954 static inline unsigned long task_util_est(struct task_struct *p)
3956 return max(task_util(p), _task_util_est(p));
3959 #ifdef CONFIG_UCLAMP_TASK
3960 static inline unsigned long uclamp_task_util(struct task_struct *p)
3962 return clamp(task_util_est(p),
3963 uclamp_eff_value(p, UCLAMP_MIN),
3964 uclamp_eff_value(p, UCLAMP_MAX));
3967 static inline unsigned long uclamp_task_util(struct task_struct *p)
3969 return task_util_est(p);
3973 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3974 struct task_struct *p)
3976 unsigned int enqueued;
3978 if (!sched_feat(UTIL_EST))
3981 /* Update root cfs_rq's estimated utilization */
3982 enqueued = cfs_rq->avg.util_est.enqueued;
3983 enqueued += _task_util_est(p);
3984 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3986 trace_sched_util_est_cfs_tp(cfs_rq);
3989 static inline void util_est_dequeue(struct cfs_rq *cfs_rq,
3990 struct task_struct *p)
3992 unsigned int enqueued;
3994 if (!sched_feat(UTIL_EST))
3997 /* Update root cfs_rq's estimated utilization */
3998 enqueued = cfs_rq->avg.util_est.enqueued;
3999 enqueued -= min_t(unsigned int, enqueued, _task_util_est(p));
4000 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
4002 trace_sched_util_est_cfs_tp(cfs_rq);
4005 #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
4008 * Check if a (signed) value is within a specified (unsigned) margin,
4009 * based on the observation that:
4011 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
4013 * NOTE: this only works when value + margin < INT_MAX.
4015 static inline bool within_margin(int value, int margin)
4017 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
4020 static inline void util_est_update(struct cfs_rq *cfs_rq,
4021 struct task_struct *p,
4024 long last_ewma_diff, last_enqueued_diff;
4027 if (!sched_feat(UTIL_EST))
4031 * Skip update of task's estimated utilization when the task has not
4032 * yet completed an activation, e.g. being migrated.
4038 * If the PELT values haven't changed since enqueue time,
4039 * skip the util_est update.
4041 ue = p->se.avg.util_est;
4042 if (ue.enqueued & UTIL_AVG_UNCHANGED)
4045 last_enqueued_diff = ue.enqueued;
4048 * Reset EWMA on utilization increases, the moving average is used only
4049 * to smooth utilization decreases.
4051 ue.enqueued = task_util(p);
4052 if (sched_feat(UTIL_EST_FASTUP)) {
4053 if (ue.ewma < ue.enqueued) {
4054 ue.ewma = ue.enqueued;
4060 * Skip update of task's estimated utilization when its members are
4061 * already ~1% close to its last activation value.
4063 last_ewma_diff = ue.enqueued - ue.ewma;
4064 last_enqueued_diff -= ue.enqueued;
4065 if (within_margin(last_ewma_diff, UTIL_EST_MARGIN)) {
4066 if (!within_margin(last_enqueued_diff, UTIL_EST_MARGIN))
4073 * To avoid overestimation of actual task utilization, skip updates if
4074 * we cannot grant there is idle time in this CPU.
4076 if (task_util(p) > capacity_orig_of(cpu_of(rq_of(cfs_rq))))
4080 * Update Task's estimated utilization
4082 * When *p completes an activation we can consolidate another sample
4083 * of the task size. This is done by storing the current PELT value
4084 * as ue.enqueued and by using this value to update the Exponential
4085 * Weighted Moving Average (EWMA):
4087 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4088 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4089 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4090 * = w * ( last_ewma_diff ) + ewma(t-1)
4091 * = w * (last_ewma_diff + ewma(t-1) / w)
4093 * Where 'w' is the weight of new samples, which is configured to be
4094 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4096 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
4097 ue.ewma += last_ewma_diff;
4098 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
4100 ue.enqueued |= UTIL_AVG_UNCHANGED;
4101 WRITE_ONCE(p->se.avg.util_est, ue);
4103 trace_sched_util_est_se_tp(&p->se);
4106 static inline int task_fits_capacity(struct task_struct *p,
4107 unsigned long capacity)
4109 return fits_capacity(uclamp_task_util(p), capacity);
4112 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
4114 if (!static_branch_unlikely(&sched_asym_cpucapacity))
4117 if (!p || p->nr_cpus_allowed == 1) {
4118 rq->misfit_task_load = 0;
4122 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
4123 rq->misfit_task_load = 0;
4128 * Make sure that misfit_task_load will not be null even if
4129 * task_h_load() returns 0.
4131 rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1);
4134 #else /* CONFIG_SMP */
4136 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
4141 #define UPDATE_TG 0x0
4142 #define SKIP_AGE_LOAD 0x0
4143 #define DO_ATTACH 0x0
4145 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
4147 cfs_rq_util_change(cfs_rq, 0);
4150 static inline void remove_entity_load_avg(struct sched_entity *se) {}
4153 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4155 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4157 static inline int newidle_balance(struct rq *rq, struct rq_flags *rf)
4163 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4166 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4169 util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p,
4171 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
4173 #endif /* CONFIG_SMP */
4175 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
4177 #ifdef CONFIG_SCHED_DEBUG
4178 s64 d = se->vruntime - cfs_rq->min_vruntime;
4183 if (d > 3*sysctl_sched_latency)
4184 schedstat_inc(cfs_rq->nr_spread_over);
4189 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
4191 u64 vruntime = cfs_rq->min_vruntime;
4194 * The 'current' period is already promised to the current tasks,
4195 * however the extra weight of the new task will slow them down a
4196 * little, place the new task so that it fits in the slot that
4197 * stays open at the end.
4199 if (initial && sched_feat(START_DEBIT))
4200 vruntime += sched_vslice(cfs_rq, se);
4202 /* sleeps up to a single latency don't count. */
4204 unsigned long thresh;
4207 thresh = sysctl_sched_min_granularity;
4209 thresh = sysctl_sched_latency;
4212 * Halve their sleep time's effect, to allow
4213 * for a gentler effect of sleepers:
4215 if (sched_feat(GENTLE_FAIR_SLEEPERS))
4221 /* ensure we never gain time by being placed backwards. */
4222 se->vruntime = max_vruntime(se->vruntime, vruntime);
4225 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
4227 static inline bool cfs_bandwidth_used(void);
4234 * update_min_vruntime()
4235 * vruntime -= min_vruntime
4239 * update_min_vruntime()
4240 * vruntime += min_vruntime
4242 * this way the vruntime transition between RQs is done when both
4243 * min_vruntime are up-to-date.
4247 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
4248 * vruntime -= min_vruntime
4252 * update_min_vruntime()
4253 * vruntime += min_vruntime
4255 * this way we don't have the most up-to-date min_vruntime on the originating
4256 * CPU and an up-to-date min_vruntime on the destination CPU.
4260 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4262 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
4263 bool curr = cfs_rq->curr == se;
4266 * If we're the current task, we must renormalise before calling
4270 se->vruntime += cfs_rq->min_vruntime;
4272 update_curr(cfs_rq);
4275 * Otherwise, renormalise after, such that we're placed at the current
4276 * moment in time, instead of some random moment in the past. Being
4277 * placed in the past could significantly boost this task to the
4278 * fairness detriment of existing tasks.
4280 if (renorm && !curr)
4281 se->vruntime += cfs_rq->min_vruntime;
4284 * When enqueuing a sched_entity, we must:
4285 * - Update loads to have both entity and cfs_rq synced with now.
4286 * - Add its load to cfs_rq->runnable_avg
4287 * - For group_entity, update its weight to reflect the new share of
4289 * - Add its new weight to cfs_rq->load.weight
4291 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
4292 se_update_runnable(se);
4293 update_cfs_group(se);
4294 account_entity_enqueue(cfs_rq, se);
4296 if (flags & ENQUEUE_WAKEUP)
4297 place_entity(cfs_rq, se, 0);
4299 check_schedstat_required();
4300 update_stats_enqueue_fair(cfs_rq, se, flags);
4301 check_spread(cfs_rq, se);
4303 __enqueue_entity(cfs_rq, se);
4307 * When bandwidth control is enabled, cfs might have been removed
4308 * because of a parent been throttled but cfs->nr_running > 1. Try to
4309 * add it unconditionally.
4311 if (cfs_rq->nr_running == 1 || cfs_bandwidth_used())
4312 list_add_leaf_cfs_rq(cfs_rq);
4314 if (cfs_rq->nr_running == 1)
4315 check_enqueue_throttle(cfs_rq);
4318 static void __clear_buddies_last(struct sched_entity *se)
4320 for_each_sched_entity(se) {
4321 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4322 if (cfs_rq->last != se)
4325 cfs_rq->last = NULL;
4329 static void __clear_buddies_next(struct sched_entity *se)
4331 for_each_sched_entity(se) {
4332 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4333 if (cfs_rq->next != se)
4336 cfs_rq->next = NULL;
4340 static void __clear_buddies_skip(struct sched_entity *se)
4342 for_each_sched_entity(se) {
4343 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4344 if (cfs_rq->skip != se)
4347 cfs_rq->skip = NULL;
4351 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4353 if (cfs_rq->last == se)
4354 __clear_buddies_last(se);
4356 if (cfs_rq->next == se)
4357 __clear_buddies_next(se);
4359 if (cfs_rq->skip == se)
4360 __clear_buddies_skip(se);
4363 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4366 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4369 * Update run-time statistics of the 'current'.
4371 update_curr(cfs_rq);
4374 * When dequeuing a sched_entity, we must:
4375 * - Update loads to have both entity and cfs_rq synced with now.
4376 * - Subtract its load from the cfs_rq->runnable_avg.
4377 * - Subtract its previous weight from cfs_rq->load.weight.
4378 * - For group entity, update its weight to reflect the new share
4379 * of its group cfs_rq.
4381 update_load_avg(cfs_rq, se, UPDATE_TG);
4382 se_update_runnable(se);
4384 update_stats_dequeue_fair(cfs_rq, se, flags);
4386 clear_buddies(cfs_rq, se);
4388 if (se != cfs_rq->curr)
4389 __dequeue_entity(cfs_rq, se);
4391 account_entity_dequeue(cfs_rq, se);
4394 * Normalize after update_curr(); which will also have moved
4395 * min_vruntime if @se is the one holding it back. But before doing
4396 * update_min_vruntime() again, which will discount @se's position and
4397 * can move min_vruntime forward still more.
4399 if (!(flags & DEQUEUE_SLEEP))
4400 se->vruntime -= cfs_rq->min_vruntime;
4402 /* return excess runtime on last dequeue */
4403 return_cfs_rq_runtime(cfs_rq);
4405 update_cfs_group(se);
4408 * Now advance min_vruntime if @se was the entity holding it back,
4409 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4410 * put back on, and if we advance min_vruntime, we'll be placed back
4411 * further than we started -- ie. we'll be penalized.
4413 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4414 update_min_vruntime(cfs_rq);
4418 * Preempt the current task with a newly woken task if needed:
4421 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4423 unsigned long ideal_runtime, delta_exec;
4424 struct sched_entity *se;
4427 ideal_runtime = sched_slice(cfs_rq, curr);
4428 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4429 if (delta_exec > ideal_runtime) {
4430 resched_curr(rq_of(cfs_rq));
4432 * The current task ran long enough, ensure it doesn't get
4433 * re-elected due to buddy favours.
4435 clear_buddies(cfs_rq, curr);
4440 * Ensure that a task that missed wakeup preemption by a
4441 * narrow margin doesn't have to wait for a full slice.
4442 * This also mitigates buddy induced latencies under load.
4444 if (delta_exec < sysctl_sched_min_granularity)
4447 se = __pick_first_entity(cfs_rq);
4448 delta = curr->vruntime - se->vruntime;
4453 if (delta > ideal_runtime)
4454 resched_curr(rq_of(cfs_rq));
4458 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4460 clear_buddies(cfs_rq, se);
4462 /* 'current' is not kept within the tree. */
4465 * Any task has to be enqueued before it get to execute on
4466 * a CPU. So account for the time it spent waiting on the
4469 update_stats_wait_end_fair(cfs_rq, se);
4470 __dequeue_entity(cfs_rq, se);
4471 update_load_avg(cfs_rq, se, UPDATE_TG);
4474 update_stats_curr_start(cfs_rq, se);
4478 * Track our maximum slice length, if the CPU's load is at
4479 * least twice that of our own weight (i.e. dont track it
4480 * when there are only lesser-weight tasks around):
4482 if (schedstat_enabled() &&
4483 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
4484 struct sched_statistics *stats;
4486 stats = __schedstats_from_se(se);
4487 __schedstat_set(stats->slice_max,
4488 max((u64)stats->slice_max,
4489 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4492 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4496 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4499 * Pick the next process, keeping these things in mind, in this order:
4500 * 1) keep things fair between processes/task groups
4501 * 2) pick the "next" process, since someone really wants that to run
4502 * 3) pick the "last" process, for cache locality
4503 * 4) do not run the "skip" process, if something else is available
4505 static struct sched_entity *
4506 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4508 struct sched_entity *left = __pick_first_entity(cfs_rq);
4509 struct sched_entity *se;
4512 * If curr is set we have to see if its left of the leftmost entity
4513 * still in the tree, provided there was anything in the tree at all.
4515 if (!left || (curr && entity_before(curr, left)))
4518 se = left; /* ideally we run the leftmost entity */
4521 * Avoid running the skip buddy, if running something else can
4522 * be done without getting too unfair.
4524 if (cfs_rq->skip && cfs_rq->skip == se) {
4525 struct sched_entity *second;
4528 second = __pick_first_entity(cfs_rq);
4530 second = __pick_next_entity(se);
4531 if (!second || (curr && entity_before(curr, second)))
4535 if (second && wakeup_preempt_entity(second, left) < 1)
4539 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) {
4541 * Someone really wants this to run. If it's not unfair, run it.
4544 } else if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) {
4546 * Prefer last buddy, try to return the CPU to a preempted task.
4554 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4556 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4559 * If still on the runqueue then deactivate_task()
4560 * was not called and update_curr() has to be done:
4563 update_curr(cfs_rq);
4565 /* throttle cfs_rqs exceeding runtime */
4566 check_cfs_rq_runtime(cfs_rq);
4568 check_spread(cfs_rq, prev);
4571 update_stats_wait_start_fair(cfs_rq, prev);
4572 /* Put 'current' back into the tree. */
4573 __enqueue_entity(cfs_rq, prev);
4574 /* in !on_rq case, update occurred at dequeue */
4575 update_load_avg(cfs_rq, prev, 0);
4577 cfs_rq->curr = NULL;
4581 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4584 * Update run-time statistics of the 'current'.
4586 update_curr(cfs_rq);
4589 * Ensure that runnable average is periodically updated.
4591 update_load_avg(cfs_rq, curr, UPDATE_TG);
4592 update_cfs_group(curr);
4594 #ifdef CONFIG_SCHED_HRTICK
4596 * queued ticks are scheduled to match the slice, so don't bother
4597 * validating it and just reschedule.
4600 resched_curr(rq_of(cfs_rq));
4604 * don't let the period tick interfere with the hrtick preemption
4606 if (!sched_feat(DOUBLE_TICK) &&
4607 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4611 if (cfs_rq->nr_running > 1)
4612 check_preempt_tick(cfs_rq, curr);
4616 /**************************************************
4617 * CFS bandwidth control machinery
4620 #ifdef CONFIG_CFS_BANDWIDTH
4622 #ifdef CONFIG_JUMP_LABEL
4623 static struct static_key __cfs_bandwidth_used;
4625 static inline bool cfs_bandwidth_used(void)
4627 return static_key_false(&__cfs_bandwidth_used);
4630 void cfs_bandwidth_usage_inc(void)
4632 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4635 void cfs_bandwidth_usage_dec(void)
4637 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4639 #else /* CONFIG_JUMP_LABEL */
4640 static bool cfs_bandwidth_used(void)
4645 void cfs_bandwidth_usage_inc(void) {}
4646 void cfs_bandwidth_usage_dec(void) {}
4647 #endif /* CONFIG_JUMP_LABEL */
4650 * default period for cfs group bandwidth.
4651 * default: 0.1s, units: nanoseconds
4653 static inline u64 default_cfs_period(void)
4655 return 100000000ULL;
4658 static inline u64 sched_cfs_bandwidth_slice(void)
4660 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4664 * Replenish runtime according to assigned quota. We use sched_clock_cpu
4665 * directly instead of rq->clock to avoid adding additional synchronization
4668 * requires cfs_b->lock
4670 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4674 if (unlikely(cfs_b->quota == RUNTIME_INF))
4677 cfs_b->runtime += cfs_b->quota;
4678 runtime = cfs_b->runtime_snap - cfs_b->runtime;
4680 cfs_b->burst_time += runtime;
4684 cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst);
4685 cfs_b->runtime_snap = cfs_b->runtime;
4688 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4690 return &tg->cfs_bandwidth;
4693 /* returns 0 on failure to allocate runtime */
4694 static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b,
4695 struct cfs_rq *cfs_rq, u64 target_runtime)
4697 u64 min_amount, amount = 0;
4699 lockdep_assert_held(&cfs_b->lock);
4701 /* note: this is a positive sum as runtime_remaining <= 0 */
4702 min_amount = target_runtime - cfs_rq->runtime_remaining;
4704 if (cfs_b->quota == RUNTIME_INF)
4705 amount = min_amount;
4707 start_cfs_bandwidth(cfs_b);
4709 if (cfs_b->runtime > 0) {
4710 amount = min(cfs_b->runtime, min_amount);
4711 cfs_b->runtime -= amount;
4716 cfs_rq->runtime_remaining += amount;
4718 return cfs_rq->runtime_remaining > 0;
4721 /* returns 0 on failure to allocate runtime */
4722 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4724 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4727 raw_spin_lock(&cfs_b->lock);
4728 ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice());
4729 raw_spin_unlock(&cfs_b->lock);
4734 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4736 /* dock delta_exec before expiring quota (as it could span periods) */
4737 cfs_rq->runtime_remaining -= delta_exec;
4739 if (likely(cfs_rq->runtime_remaining > 0))
4742 if (cfs_rq->throttled)
4745 * if we're unable to extend our runtime we resched so that the active
4746 * hierarchy can be throttled
4748 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4749 resched_curr(rq_of(cfs_rq));
4752 static __always_inline
4753 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4755 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4758 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4761 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4763 return cfs_bandwidth_used() && cfs_rq->throttled;
4766 /* check whether cfs_rq, or any parent, is throttled */
4767 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4769 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4773 * Ensure that neither of the group entities corresponding to src_cpu or
4774 * dest_cpu are members of a throttled hierarchy when performing group
4775 * load-balance operations.
4777 static inline int throttled_lb_pair(struct task_group *tg,
4778 int src_cpu, int dest_cpu)
4780 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4782 src_cfs_rq = tg->cfs_rq[src_cpu];
4783 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4785 return throttled_hierarchy(src_cfs_rq) ||
4786 throttled_hierarchy(dest_cfs_rq);
4789 static int tg_unthrottle_up(struct task_group *tg, void *data)
4791 struct rq *rq = data;
4792 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4794 cfs_rq->throttle_count--;
4795 if (!cfs_rq->throttle_count) {
4796 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4797 cfs_rq->throttled_clock_task;
4799 /* Add cfs_rq with load or one or more already running entities to the list */
4800 if (!cfs_rq_is_decayed(cfs_rq) || cfs_rq->nr_running)
4801 list_add_leaf_cfs_rq(cfs_rq);
4807 static int tg_throttle_down(struct task_group *tg, void *data)
4809 struct rq *rq = data;
4810 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4812 /* group is entering throttled state, stop time */
4813 if (!cfs_rq->throttle_count) {
4814 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4815 list_del_leaf_cfs_rq(cfs_rq);
4817 cfs_rq->throttle_count++;
4822 static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
4824 struct rq *rq = rq_of(cfs_rq);
4825 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4826 struct sched_entity *se;
4827 long task_delta, idle_task_delta, dequeue = 1;
4829 raw_spin_lock(&cfs_b->lock);
4830 /* This will start the period timer if necessary */
4831 if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) {
4833 * We have raced with bandwidth becoming available, and if we
4834 * actually throttled the timer might not unthrottle us for an
4835 * entire period. We additionally needed to make sure that any
4836 * subsequent check_cfs_rq_runtime calls agree not to throttle
4837 * us, as we may commit to do cfs put_prev+pick_next, so we ask
4838 * for 1ns of runtime rather than just check cfs_b.
4842 list_add_tail_rcu(&cfs_rq->throttled_list,
4843 &cfs_b->throttled_cfs_rq);
4845 raw_spin_unlock(&cfs_b->lock);
4848 return false; /* Throttle no longer required. */
4850 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4852 /* freeze hierarchy runnable averages while throttled */
4854 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4857 task_delta = cfs_rq->h_nr_running;
4858 idle_task_delta = cfs_rq->idle_h_nr_running;
4859 for_each_sched_entity(se) {
4860 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4861 /* throttled entity or throttle-on-deactivate */
4865 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4867 if (cfs_rq_is_idle(group_cfs_rq(se)))
4868 idle_task_delta = cfs_rq->h_nr_running;
4870 qcfs_rq->h_nr_running -= task_delta;
4871 qcfs_rq->idle_h_nr_running -= idle_task_delta;
4873 if (qcfs_rq->load.weight) {
4874 /* Avoid re-evaluating load for this entity: */
4875 se = parent_entity(se);
4880 for_each_sched_entity(se) {
4881 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4882 /* throttled entity or throttle-on-deactivate */
4886 update_load_avg(qcfs_rq, se, 0);
4887 se_update_runnable(se);
4889 if (cfs_rq_is_idle(group_cfs_rq(se)))
4890 idle_task_delta = cfs_rq->h_nr_running;
4892 qcfs_rq->h_nr_running -= task_delta;
4893 qcfs_rq->idle_h_nr_running -= idle_task_delta;
4896 /* At this point se is NULL and we are at root level*/
4897 sub_nr_running(rq, task_delta);
4901 * Note: distribution will already see us throttled via the
4902 * throttled-list. rq->lock protects completion.
4904 cfs_rq->throttled = 1;
4905 cfs_rq->throttled_clock = rq_clock(rq);
4909 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4911 struct rq *rq = rq_of(cfs_rq);
4912 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4913 struct sched_entity *se;
4914 long task_delta, idle_task_delta;
4916 se = cfs_rq->tg->se[cpu_of(rq)];
4918 cfs_rq->throttled = 0;
4920 update_rq_clock(rq);
4922 raw_spin_lock(&cfs_b->lock);
4923 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4924 list_del_rcu(&cfs_rq->throttled_list);
4925 raw_spin_unlock(&cfs_b->lock);
4927 /* update hierarchical throttle state */
4928 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4930 /* Nothing to run but something to decay (on_list)? Complete the branch */
4931 if (!cfs_rq->load.weight) {
4932 if (cfs_rq->on_list)
4933 goto unthrottle_throttle;
4937 task_delta = cfs_rq->h_nr_running;
4938 idle_task_delta = cfs_rq->idle_h_nr_running;
4939 for_each_sched_entity(se) {
4940 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4944 enqueue_entity(qcfs_rq, se, ENQUEUE_WAKEUP);
4946 if (cfs_rq_is_idle(group_cfs_rq(se)))
4947 idle_task_delta = cfs_rq->h_nr_running;
4949 qcfs_rq->h_nr_running += task_delta;
4950 qcfs_rq->idle_h_nr_running += idle_task_delta;
4952 /* end evaluation on encountering a throttled cfs_rq */
4953 if (cfs_rq_throttled(qcfs_rq))
4954 goto unthrottle_throttle;
4957 for_each_sched_entity(se) {
4958 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4960 update_load_avg(qcfs_rq, se, UPDATE_TG);
4961 se_update_runnable(se);
4963 if (cfs_rq_is_idle(group_cfs_rq(se)))
4964 idle_task_delta = cfs_rq->h_nr_running;
4966 qcfs_rq->h_nr_running += task_delta;
4967 qcfs_rq->idle_h_nr_running += idle_task_delta;
4969 /* end evaluation on encountering a throttled cfs_rq */
4970 if (cfs_rq_throttled(qcfs_rq))
4971 goto unthrottle_throttle;
4974 * One parent has been throttled and cfs_rq removed from the
4975 * list. Add it back to not break the leaf list.
4977 if (throttled_hierarchy(qcfs_rq))
4978 list_add_leaf_cfs_rq(qcfs_rq);
4981 /* At this point se is NULL and we are at root level*/
4982 add_nr_running(rq, task_delta);
4984 unthrottle_throttle:
4986 * The cfs_rq_throttled() breaks in the above iteration can result in
4987 * incomplete leaf list maintenance, resulting in triggering the
4990 for_each_sched_entity(se) {
4991 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4993 if (list_add_leaf_cfs_rq(qcfs_rq))
4997 assert_list_leaf_cfs_rq(rq);
4999 /* Determine whether we need to wake up potentially idle CPU: */
5000 if (rq->curr == rq->idle && rq->cfs.nr_running)
5004 static void distribute_cfs_runtime(struct cfs_bandwidth *cfs_b)
5006 struct cfs_rq *cfs_rq;
5007 u64 runtime, remaining = 1;
5010 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
5012 struct rq *rq = rq_of(cfs_rq);
5015 rq_lock_irqsave(rq, &rf);
5016 if (!cfs_rq_throttled(cfs_rq))
5019 /* By the above check, this should never be true */
5020 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
5022 raw_spin_lock(&cfs_b->lock);
5023 runtime = -cfs_rq->runtime_remaining + 1;
5024 if (runtime > cfs_b->runtime)
5025 runtime = cfs_b->runtime;
5026 cfs_b->runtime -= runtime;
5027 remaining = cfs_b->runtime;
5028 raw_spin_unlock(&cfs_b->lock);
5030 cfs_rq->runtime_remaining += runtime;
5032 /* we check whether we're throttled above */
5033 if (cfs_rq->runtime_remaining > 0)
5034 unthrottle_cfs_rq(cfs_rq);
5037 rq_unlock_irqrestore(rq, &rf);
5046 * Responsible for refilling a task_group's bandwidth and unthrottling its
5047 * cfs_rqs as appropriate. If there has been no activity within the last
5048 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
5049 * used to track this state.
5051 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
5055 /* no need to continue the timer with no bandwidth constraint */
5056 if (cfs_b->quota == RUNTIME_INF)
5057 goto out_deactivate;
5059 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
5060 cfs_b->nr_periods += overrun;
5062 /* Refill extra burst quota even if cfs_b->idle */
5063 __refill_cfs_bandwidth_runtime(cfs_b);
5066 * idle depends on !throttled (for the case of a large deficit), and if
5067 * we're going inactive then everything else can be deferred
5069 if (cfs_b->idle && !throttled)
5070 goto out_deactivate;
5073 /* mark as potentially idle for the upcoming period */
5078 /* account preceding periods in which throttling occurred */
5079 cfs_b->nr_throttled += overrun;
5082 * This check is repeated as we release cfs_b->lock while we unthrottle.
5084 while (throttled && cfs_b->runtime > 0) {
5085 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5086 /* we can't nest cfs_b->lock while distributing bandwidth */
5087 distribute_cfs_runtime(cfs_b);
5088 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5090 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
5094 * While we are ensured activity in the period following an
5095 * unthrottle, this also covers the case in which the new bandwidth is
5096 * insufficient to cover the existing bandwidth deficit. (Forcing the
5097 * timer to remain active while there are any throttled entities.)
5107 /* a cfs_rq won't donate quota below this amount */
5108 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
5109 /* minimum remaining period time to redistribute slack quota */
5110 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
5111 /* how long we wait to gather additional slack before distributing */
5112 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
5115 * Are we near the end of the current quota period?
5117 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
5118 * hrtimer base being cleared by hrtimer_start. In the case of
5119 * migrate_hrtimers, base is never cleared, so we are fine.
5121 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
5123 struct hrtimer *refresh_timer = &cfs_b->period_timer;
5126 /* if the call-back is running a quota refresh is already occurring */
5127 if (hrtimer_callback_running(refresh_timer))
5130 /* is a quota refresh about to occur? */
5131 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
5132 if (remaining < (s64)min_expire)
5138 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
5140 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
5142 /* if there's a quota refresh soon don't bother with slack */
5143 if (runtime_refresh_within(cfs_b, min_left))
5146 /* don't push forwards an existing deferred unthrottle */
5147 if (cfs_b->slack_started)
5149 cfs_b->slack_started = true;
5151 hrtimer_start(&cfs_b->slack_timer,
5152 ns_to_ktime(cfs_bandwidth_slack_period),
5156 /* we know any runtime found here is valid as update_curr() precedes return */
5157 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5159 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5160 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
5162 if (slack_runtime <= 0)
5165 raw_spin_lock(&cfs_b->lock);
5166 if (cfs_b->quota != RUNTIME_INF) {
5167 cfs_b->runtime += slack_runtime;
5169 /* we are under rq->lock, defer unthrottling using a timer */
5170 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
5171 !list_empty(&cfs_b->throttled_cfs_rq))
5172 start_cfs_slack_bandwidth(cfs_b);
5174 raw_spin_unlock(&cfs_b->lock);
5176 /* even if it's not valid for return we don't want to try again */
5177 cfs_rq->runtime_remaining -= slack_runtime;
5180 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5182 if (!cfs_bandwidth_used())
5185 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
5188 __return_cfs_rq_runtime(cfs_rq);
5192 * This is done with a timer (instead of inline with bandwidth return) since
5193 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
5195 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
5197 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
5198 unsigned long flags;
5200 /* confirm we're still not at a refresh boundary */
5201 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5202 cfs_b->slack_started = false;
5204 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
5205 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5209 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
5210 runtime = cfs_b->runtime;
5212 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5217 distribute_cfs_runtime(cfs_b);
5221 * When a group wakes up we want to make sure that its quota is not already
5222 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
5223 * runtime as update_curr() throttling can not trigger until it's on-rq.
5225 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
5227 if (!cfs_bandwidth_used())
5230 /* an active group must be handled by the update_curr()->put() path */
5231 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
5234 /* ensure the group is not already throttled */
5235 if (cfs_rq_throttled(cfs_rq))
5238 /* update runtime allocation */
5239 account_cfs_rq_runtime(cfs_rq, 0);
5240 if (cfs_rq->runtime_remaining <= 0)
5241 throttle_cfs_rq(cfs_rq);
5244 static void sync_throttle(struct task_group *tg, int cpu)
5246 struct cfs_rq *pcfs_rq, *cfs_rq;
5248 if (!cfs_bandwidth_used())
5254 cfs_rq = tg->cfs_rq[cpu];
5255 pcfs_rq = tg->parent->cfs_rq[cpu];
5257 cfs_rq->throttle_count = pcfs_rq->throttle_count;
5258 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
5261 /* conditionally throttle active cfs_rq's from put_prev_entity() */
5262 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5264 if (!cfs_bandwidth_used())
5267 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
5271 * it's possible for a throttled entity to be forced into a running
5272 * state (e.g. set_curr_task), in this case we're finished.
5274 if (cfs_rq_throttled(cfs_rq))
5277 return throttle_cfs_rq(cfs_rq);
5280 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
5282 struct cfs_bandwidth *cfs_b =
5283 container_of(timer, struct cfs_bandwidth, slack_timer);
5285 do_sched_cfs_slack_timer(cfs_b);
5287 return HRTIMER_NORESTART;
5290 extern const u64 max_cfs_quota_period;
5292 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
5294 struct cfs_bandwidth *cfs_b =
5295 container_of(timer, struct cfs_bandwidth, period_timer);
5296 unsigned long flags;
5301 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5303 overrun = hrtimer_forward_now(timer, cfs_b->period);
5307 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
5310 u64 new, old = ktime_to_ns(cfs_b->period);
5313 * Grow period by a factor of 2 to avoid losing precision.
5314 * Precision loss in the quota/period ratio can cause __cfs_schedulable
5318 if (new < max_cfs_quota_period) {
5319 cfs_b->period = ns_to_ktime(new);
5323 pr_warn_ratelimited(
5324 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5326 div_u64(new, NSEC_PER_USEC),
5327 div_u64(cfs_b->quota, NSEC_PER_USEC));
5329 pr_warn_ratelimited(
5330 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5332 div_u64(old, NSEC_PER_USEC),
5333 div_u64(cfs_b->quota, NSEC_PER_USEC));
5336 /* reset count so we don't come right back in here */
5341 cfs_b->period_active = 0;
5342 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5344 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
5347 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5349 raw_spin_lock_init(&cfs_b->lock);
5351 cfs_b->quota = RUNTIME_INF;
5352 cfs_b->period = ns_to_ktime(default_cfs_period());
5355 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
5356 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
5357 cfs_b->period_timer.function = sched_cfs_period_timer;
5358 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5359 cfs_b->slack_timer.function = sched_cfs_slack_timer;
5360 cfs_b->slack_started = false;
5363 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5365 cfs_rq->runtime_enabled = 0;
5366 INIT_LIST_HEAD(&cfs_rq->throttled_list);
5369 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5371 lockdep_assert_held(&cfs_b->lock);
5373 if (cfs_b->period_active)
5376 cfs_b->period_active = 1;
5377 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5378 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5381 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5383 /* init_cfs_bandwidth() was not called */
5384 if (!cfs_b->throttled_cfs_rq.next)
5387 hrtimer_cancel(&cfs_b->period_timer);
5388 hrtimer_cancel(&cfs_b->slack_timer);
5392 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5394 * The race is harmless, since modifying bandwidth settings of unhooked group
5395 * bits doesn't do much.
5398 /* cpu online callback */
5399 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5401 struct task_group *tg;
5403 lockdep_assert_rq_held(rq);
5406 list_for_each_entry_rcu(tg, &task_groups, list) {
5407 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5408 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5410 raw_spin_lock(&cfs_b->lock);
5411 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5412 raw_spin_unlock(&cfs_b->lock);
5417 /* cpu offline callback */
5418 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5420 struct task_group *tg;
5422 lockdep_assert_rq_held(rq);
5425 list_for_each_entry_rcu(tg, &task_groups, list) {
5426 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5428 if (!cfs_rq->runtime_enabled)
5432 * clock_task is not advancing so we just need to make sure
5433 * there's some valid quota amount
5435 cfs_rq->runtime_remaining = 1;
5437 * Offline rq is schedulable till CPU is completely disabled
5438 * in take_cpu_down(), so we prevent new cfs throttling here.
5440 cfs_rq->runtime_enabled = 0;
5442 if (cfs_rq_throttled(cfs_rq))
5443 unthrottle_cfs_rq(cfs_rq);
5448 #else /* CONFIG_CFS_BANDWIDTH */
5450 static inline bool cfs_bandwidth_used(void)
5455 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5456 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5457 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5458 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5459 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5461 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5466 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5471 static inline int throttled_lb_pair(struct task_group *tg,
5472 int src_cpu, int dest_cpu)
5477 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5479 #ifdef CONFIG_FAIR_GROUP_SCHED
5480 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5483 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5487 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5488 static inline void update_runtime_enabled(struct rq *rq) {}
5489 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5491 #endif /* CONFIG_CFS_BANDWIDTH */
5493 /**************************************************
5494 * CFS operations on tasks:
5497 #ifdef CONFIG_SCHED_HRTICK
5498 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5500 struct sched_entity *se = &p->se;
5501 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5503 SCHED_WARN_ON(task_rq(p) != rq);
5505 if (rq->cfs.h_nr_running > 1) {
5506 u64 slice = sched_slice(cfs_rq, se);
5507 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5508 s64 delta = slice - ran;
5511 if (task_current(rq, p))
5515 hrtick_start(rq, delta);
5520 * called from enqueue/dequeue and updates the hrtick when the
5521 * current task is from our class and nr_running is low enough
5524 static void hrtick_update(struct rq *rq)
5526 struct task_struct *curr = rq->curr;
5528 if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class)
5531 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5532 hrtick_start_fair(rq, curr);
5534 #else /* !CONFIG_SCHED_HRTICK */
5536 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5540 static inline void hrtick_update(struct rq *rq)
5546 static inline bool cpu_overutilized(int cpu)
5548 return !fits_capacity(cpu_util_cfs(cpu), capacity_of(cpu));
5551 static inline void update_overutilized_status(struct rq *rq)
5553 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
5554 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5555 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
5559 static inline void update_overutilized_status(struct rq *rq) { }
5562 /* Runqueue only has SCHED_IDLE tasks enqueued */
5563 static int sched_idle_rq(struct rq *rq)
5565 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
5570 * Returns true if cfs_rq only has SCHED_IDLE entities enqueued. Note the use
5571 * of idle_nr_running, which does not consider idle descendants of normal
5574 static bool sched_idle_cfs_rq(struct cfs_rq *cfs_rq)
5576 return cfs_rq->nr_running &&
5577 cfs_rq->nr_running == cfs_rq->idle_nr_running;
5581 static int sched_idle_cpu(int cpu)
5583 return sched_idle_rq(cpu_rq(cpu));
5588 * The enqueue_task method is called before nr_running is
5589 * increased. Here we update the fair scheduling stats and
5590 * then put the task into the rbtree:
5593 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5595 struct cfs_rq *cfs_rq;
5596 struct sched_entity *se = &p->se;
5597 int idle_h_nr_running = task_has_idle_policy(p);
5598 int task_new = !(flags & ENQUEUE_WAKEUP);
5601 * The code below (indirectly) updates schedutil which looks at
5602 * the cfs_rq utilization to select a frequency.
5603 * Let's add the task's estimated utilization to the cfs_rq's
5604 * estimated utilization, before we update schedutil.
5606 util_est_enqueue(&rq->cfs, p);
5609 * If in_iowait is set, the code below may not trigger any cpufreq
5610 * utilization updates, so do it here explicitly with the IOWAIT flag
5614 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5616 for_each_sched_entity(se) {
5619 cfs_rq = cfs_rq_of(se);
5620 enqueue_entity(cfs_rq, se, flags);
5622 cfs_rq->h_nr_running++;
5623 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5625 if (cfs_rq_is_idle(cfs_rq))
5626 idle_h_nr_running = 1;
5628 /* end evaluation on encountering a throttled cfs_rq */
5629 if (cfs_rq_throttled(cfs_rq))
5630 goto enqueue_throttle;
5632 flags = ENQUEUE_WAKEUP;
5635 for_each_sched_entity(se) {
5636 cfs_rq = cfs_rq_of(se);
5638 update_load_avg(cfs_rq, se, UPDATE_TG);
5639 se_update_runnable(se);
5640 update_cfs_group(se);
5642 cfs_rq->h_nr_running++;
5643 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5645 if (cfs_rq_is_idle(cfs_rq))
5646 idle_h_nr_running = 1;
5648 /* end evaluation on encountering a throttled cfs_rq */
5649 if (cfs_rq_throttled(cfs_rq))
5650 goto enqueue_throttle;
5653 * One parent has been throttled and cfs_rq removed from the
5654 * list. Add it back to not break the leaf list.
5656 if (throttled_hierarchy(cfs_rq))
5657 list_add_leaf_cfs_rq(cfs_rq);
5660 /* At this point se is NULL and we are at root level*/
5661 add_nr_running(rq, 1);
5664 * Since new tasks are assigned an initial util_avg equal to
5665 * half of the spare capacity of their CPU, tiny tasks have the
5666 * ability to cross the overutilized threshold, which will
5667 * result in the load balancer ruining all the task placement
5668 * done by EAS. As a way to mitigate that effect, do not account
5669 * for the first enqueue operation of new tasks during the
5670 * overutilized flag detection.
5672 * A better way of solving this problem would be to wait for
5673 * the PELT signals of tasks to converge before taking them
5674 * into account, but that is not straightforward to implement,
5675 * and the following generally works well enough in practice.
5678 update_overutilized_status(rq);
5681 if (cfs_bandwidth_used()) {
5683 * When bandwidth control is enabled; the cfs_rq_throttled()
5684 * breaks in the above iteration can result in incomplete
5685 * leaf list maintenance, resulting in triggering the assertion
5688 for_each_sched_entity(se) {
5689 cfs_rq = cfs_rq_of(se);
5691 if (list_add_leaf_cfs_rq(cfs_rq))
5696 assert_list_leaf_cfs_rq(rq);
5701 static void set_next_buddy(struct sched_entity *se);
5704 * The dequeue_task method is called before nr_running is
5705 * decreased. We remove the task from the rbtree and
5706 * update the fair scheduling stats:
5708 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5710 struct cfs_rq *cfs_rq;
5711 struct sched_entity *se = &p->se;
5712 int task_sleep = flags & DEQUEUE_SLEEP;
5713 int idle_h_nr_running = task_has_idle_policy(p);
5714 bool was_sched_idle = sched_idle_rq(rq);
5716 util_est_dequeue(&rq->cfs, p);
5718 for_each_sched_entity(se) {
5719 cfs_rq = cfs_rq_of(se);
5720 dequeue_entity(cfs_rq, se, flags);
5722 cfs_rq->h_nr_running--;
5723 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5725 if (cfs_rq_is_idle(cfs_rq))
5726 idle_h_nr_running = 1;
5728 /* end evaluation on encountering a throttled cfs_rq */
5729 if (cfs_rq_throttled(cfs_rq))
5730 goto dequeue_throttle;
5732 /* Don't dequeue parent if it has other entities besides us */
5733 if (cfs_rq->load.weight) {
5734 /* Avoid re-evaluating load for this entity: */
5735 se = parent_entity(se);
5737 * Bias pick_next to pick a task from this cfs_rq, as
5738 * p is sleeping when it is within its sched_slice.
5740 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5744 flags |= DEQUEUE_SLEEP;
5747 for_each_sched_entity(se) {
5748 cfs_rq = cfs_rq_of(se);
5750 update_load_avg(cfs_rq, se, UPDATE_TG);
5751 se_update_runnable(se);
5752 update_cfs_group(se);
5754 cfs_rq->h_nr_running--;
5755 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5757 if (cfs_rq_is_idle(cfs_rq))
5758 idle_h_nr_running = 1;
5760 /* end evaluation on encountering a throttled cfs_rq */
5761 if (cfs_rq_throttled(cfs_rq))
5762 goto dequeue_throttle;
5766 /* At this point se is NULL and we are at root level*/
5767 sub_nr_running(rq, 1);
5769 /* balance early to pull high priority tasks */
5770 if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
5771 rq->next_balance = jiffies;
5774 util_est_update(&rq->cfs, p, task_sleep);
5780 /* Working cpumask for: load_balance, load_balance_newidle. */
5781 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5782 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5784 #ifdef CONFIG_NO_HZ_COMMON
5787 cpumask_var_t idle_cpus_mask;
5789 int has_blocked; /* Idle CPUS has blocked load */
5790 int needs_update; /* Newly idle CPUs need their next_balance collated */
5791 unsigned long next_balance; /* in jiffy units */
5792 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5793 } nohz ____cacheline_aligned;
5795 #endif /* CONFIG_NO_HZ_COMMON */
5797 static unsigned long cpu_load(struct rq *rq)
5799 return cfs_rq_load_avg(&rq->cfs);
5803 * cpu_load_without - compute CPU load without any contributions from *p
5804 * @cpu: the CPU which load is requested
5805 * @p: the task which load should be discounted
5807 * The load of a CPU is defined by the load of tasks currently enqueued on that
5808 * CPU as well as tasks which are currently sleeping after an execution on that
5811 * This method returns the load of the specified CPU by discounting the load of
5812 * the specified task, whenever the task is currently contributing to the CPU
5815 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
5817 struct cfs_rq *cfs_rq;
5820 /* Task has no contribution or is new */
5821 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
5822 return cpu_load(rq);
5825 load = READ_ONCE(cfs_rq->avg.load_avg);
5827 /* Discount task's util from CPU's util */
5828 lsub_positive(&load, task_h_load(p));
5833 static unsigned long cpu_runnable(struct rq *rq)
5835 return cfs_rq_runnable_avg(&rq->cfs);
5838 static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
5840 struct cfs_rq *cfs_rq;
5841 unsigned int runnable;
5843 /* Task has no contribution or is new */
5844 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
5845 return cpu_runnable(rq);
5848 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
5850 /* Discount task's runnable from CPU's runnable */
5851 lsub_positive(&runnable, p->se.avg.runnable_avg);
5856 static unsigned long capacity_of(int cpu)
5858 return cpu_rq(cpu)->cpu_capacity;
5861 static void record_wakee(struct task_struct *p)
5864 * Only decay a single time; tasks that have less then 1 wakeup per
5865 * jiffy will not have built up many flips.
5867 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5868 current->wakee_flips >>= 1;
5869 current->wakee_flip_decay_ts = jiffies;
5872 if (current->last_wakee != p) {
5873 current->last_wakee = p;
5874 current->wakee_flips++;
5879 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5881 * A waker of many should wake a different task than the one last awakened
5882 * at a frequency roughly N times higher than one of its wakees.
5884 * In order to determine whether we should let the load spread vs consolidating
5885 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5886 * partner, and a factor of lls_size higher frequency in the other.
5888 * With both conditions met, we can be relatively sure that the relationship is
5889 * non-monogamous, with partner count exceeding socket size.
5891 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5892 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5895 static int wake_wide(struct task_struct *p)
5897 unsigned int master = current->wakee_flips;
5898 unsigned int slave = p->wakee_flips;
5899 int factor = __this_cpu_read(sd_llc_size);
5902 swap(master, slave);
5903 if (slave < factor || master < slave * factor)
5909 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5910 * soonest. For the purpose of speed we only consider the waking and previous
5913 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5914 * cache-affine and is (or will be) idle.
5916 * wake_affine_weight() - considers the weight to reflect the average
5917 * scheduling latency of the CPUs. This seems to work
5918 * for the overloaded case.
5921 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5924 * If this_cpu is idle, it implies the wakeup is from interrupt
5925 * context. Only allow the move if cache is shared. Otherwise an
5926 * interrupt intensive workload could force all tasks onto one
5927 * node depending on the IO topology or IRQ affinity settings.
5929 * If the prev_cpu is idle and cache affine then avoid a migration.
5930 * There is no guarantee that the cache hot data from an interrupt
5931 * is more important than cache hot data on the prev_cpu and from
5932 * a cpufreq perspective, it's better to have higher utilisation
5935 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5936 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5938 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5941 if (available_idle_cpu(prev_cpu))
5944 return nr_cpumask_bits;
5948 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5949 int this_cpu, int prev_cpu, int sync)
5951 s64 this_eff_load, prev_eff_load;
5952 unsigned long task_load;
5954 this_eff_load = cpu_load(cpu_rq(this_cpu));
5957 unsigned long current_load = task_h_load(current);
5959 if (current_load > this_eff_load)
5962 this_eff_load -= current_load;
5965 task_load = task_h_load(p);
5967 this_eff_load += task_load;
5968 if (sched_feat(WA_BIAS))
5969 this_eff_load *= 100;
5970 this_eff_load *= capacity_of(prev_cpu);
5972 prev_eff_load = cpu_load(cpu_rq(prev_cpu));
5973 prev_eff_load -= task_load;
5974 if (sched_feat(WA_BIAS))
5975 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5976 prev_eff_load *= capacity_of(this_cpu);
5979 * If sync, adjust the weight of prev_eff_load such that if
5980 * prev_eff == this_eff that select_idle_sibling() will consider
5981 * stacking the wakee on top of the waker if no other CPU is
5987 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5990 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5991 int this_cpu, int prev_cpu, int sync)
5993 int target = nr_cpumask_bits;
5995 if (sched_feat(WA_IDLE))
5996 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5998 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5999 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
6001 schedstat_inc(p->stats.nr_wakeups_affine_attempts);
6002 if (target == nr_cpumask_bits)
6005 schedstat_inc(sd->ttwu_move_affine);
6006 schedstat_inc(p->stats.nr_wakeups_affine);
6010 static struct sched_group *
6011 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu);
6014 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
6017 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
6019 unsigned long load, min_load = ULONG_MAX;
6020 unsigned int min_exit_latency = UINT_MAX;
6021 u64 latest_idle_timestamp = 0;
6022 int least_loaded_cpu = this_cpu;
6023 int shallowest_idle_cpu = -1;
6026 /* Check if we have any choice: */
6027 if (group->group_weight == 1)
6028 return cpumask_first(sched_group_span(group));
6030 /* Traverse only the allowed CPUs */
6031 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
6032 struct rq *rq = cpu_rq(i);
6034 if (!sched_core_cookie_match(rq, p))
6037 if (sched_idle_cpu(i))
6040 if (available_idle_cpu(i)) {
6041 struct cpuidle_state *idle = idle_get_state(rq);
6042 if (idle && idle->exit_latency < min_exit_latency) {
6044 * We give priority to a CPU whose idle state
6045 * has the smallest exit latency irrespective
6046 * of any idle timestamp.
6048 min_exit_latency = idle->exit_latency;
6049 latest_idle_timestamp = rq->idle_stamp;
6050 shallowest_idle_cpu = i;
6051 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
6052 rq->idle_stamp > latest_idle_timestamp) {
6054 * If equal or no active idle state, then
6055 * the most recently idled CPU might have
6058 latest_idle_timestamp = rq->idle_stamp;
6059 shallowest_idle_cpu = i;
6061 } else if (shallowest_idle_cpu == -1) {
6062 load = cpu_load(cpu_rq(i));
6063 if (load < min_load) {
6065 least_loaded_cpu = i;
6070 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
6073 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
6074 int cpu, int prev_cpu, int sd_flag)
6078 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
6082 * We need task's util for cpu_util_without, sync it up to
6083 * prev_cpu's last_update_time.
6085 if (!(sd_flag & SD_BALANCE_FORK))
6086 sync_entity_load_avg(&p->se);
6089 struct sched_group *group;
6090 struct sched_domain *tmp;
6093 if (!(sd->flags & sd_flag)) {
6098 group = find_idlest_group(sd, p, cpu);
6104 new_cpu = find_idlest_group_cpu(group, p, cpu);
6105 if (new_cpu == cpu) {
6106 /* Now try balancing at a lower domain level of 'cpu': */
6111 /* Now try balancing at a lower domain level of 'new_cpu': */
6113 weight = sd->span_weight;
6115 for_each_domain(cpu, tmp) {
6116 if (weight <= tmp->span_weight)
6118 if (tmp->flags & sd_flag)
6126 static inline int __select_idle_cpu(int cpu, struct task_struct *p)
6128 if ((available_idle_cpu(cpu) || sched_idle_cpu(cpu)) &&
6129 sched_cpu_cookie_match(cpu_rq(cpu), p))
6135 #ifdef CONFIG_SCHED_SMT
6136 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
6137 EXPORT_SYMBOL_GPL(sched_smt_present);
6139 static inline void set_idle_cores(int cpu, int val)
6141 struct sched_domain_shared *sds;
6143 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6145 WRITE_ONCE(sds->has_idle_cores, val);
6148 static inline bool test_idle_cores(int cpu, bool def)
6150 struct sched_domain_shared *sds;
6152 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6154 return READ_ONCE(sds->has_idle_cores);
6160 * Scans the local SMT mask to see if the entire core is idle, and records this
6161 * information in sd_llc_shared->has_idle_cores.
6163 * Since SMT siblings share all cache levels, inspecting this limited remote
6164 * state should be fairly cheap.
6166 void __update_idle_core(struct rq *rq)
6168 int core = cpu_of(rq);
6172 if (test_idle_cores(core, true))
6175 for_each_cpu(cpu, cpu_smt_mask(core)) {
6179 if (!available_idle_cpu(cpu))
6183 set_idle_cores(core, 1);
6189 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6190 * there are no idle cores left in the system; tracked through
6191 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6193 static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
6198 if (!static_branch_likely(&sched_smt_present))
6199 return __select_idle_cpu(core, p);
6201 for_each_cpu(cpu, cpu_smt_mask(core)) {
6202 if (!available_idle_cpu(cpu)) {
6204 if (*idle_cpu == -1) {
6205 if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, p->cpus_ptr)) {
6213 if (*idle_cpu == -1 && cpumask_test_cpu(cpu, p->cpus_ptr))
6220 cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
6225 * Scan the local SMT mask for idle CPUs.
6227 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6231 for_each_cpu(cpu, cpu_smt_mask(target)) {
6232 if (!cpumask_test_cpu(cpu, p->cpus_ptr) ||
6233 !cpumask_test_cpu(cpu, sched_domain_span(sd)))
6235 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
6242 #else /* CONFIG_SCHED_SMT */
6244 static inline void set_idle_cores(int cpu, int val)
6248 static inline bool test_idle_cores(int cpu, bool def)
6253 static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
6255 return __select_idle_cpu(core, p);
6258 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6263 #endif /* CONFIG_SCHED_SMT */
6266 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6267 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6268 * average idle time for this rq (as found in rq->avg_idle).
6270 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target)
6272 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6273 int i, cpu, idle_cpu = -1, nr = INT_MAX;
6274 struct rq *this_rq = this_rq();
6275 int this = smp_processor_id();
6276 struct sched_domain *this_sd;
6279 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6283 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6285 if (sched_feat(SIS_PROP) && !has_idle_core) {
6286 u64 avg_cost, avg_idle, span_avg;
6287 unsigned long now = jiffies;
6290 * If we're busy, the assumption that the last idle period
6291 * predicts the future is flawed; age away the remaining
6292 * predicted idle time.
6294 if (unlikely(this_rq->wake_stamp < now)) {
6295 while (this_rq->wake_stamp < now && this_rq->wake_avg_idle) {
6296 this_rq->wake_stamp++;
6297 this_rq->wake_avg_idle >>= 1;
6301 avg_idle = this_rq->wake_avg_idle;
6302 avg_cost = this_sd->avg_scan_cost + 1;
6304 span_avg = sd->span_weight * avg_idle;
6305 if (span_avg > 4*avg_cost)
6306 nr = div_u64(span_avg, avg_cost);
6310 time = cpu_clock(this);
6313 for_each_cpu_wrap(cpu, cpus, target + 1) {
6314 if (has_idle_core) {
6315 i = select_idle_core(p, cpu, cpus, &idle_cpu);
6316 if ((unsigned int)i < nr_cpumask_bits)
6322 idle_cpu = __select_idle_cpu(cpu, p);
6323 if ((unsigned int)idle_cpu < nr_cpumask_bits)
6329 set_idle_cores(target, false);
6331 if (sched_feat(SIS_PROP) && !has_idle_core) {
6332 time = cpu_clock(this) - time;
6335 * Account for the scan cost of wakeups against the average
6338 this_rq->wake_avg_idle -= min(this_rq->wake_avg_idle, time);
6340 update_avg(&this_sd->avg_scan_cost, time);
6347 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
6348 * the task fits. If no CPU is big enough, but there are idle ones, try to
6349 * maximize capacity.
6352 select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
6354 unsigned long task_util, best_cap = 0;
6355 int cpu, best_cpu = -1;
6356 struct cpumask *cpus;
6358 cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6359 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6361 task_util = uclamp_task_util(p);
6363 for_each_cpu_wrap(cpu, cpus, target) {
6364 unsigned long cpu_cap = capacity_of(cpu);
6366 if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
6368 if (fits_capacity(task_util, cpu_cap))
6371 if (cpu_cap > best_cap) {
6380 static inline bool asym_fits_capacity(unsigned long task_util, int cpu)
6382 if (static_branch_unlikely(&sched_asym_cpucapacity))
6383 return fits_capacity(task_util, capacity_of(cpu));
6389 * Try and locate an idle core/thread in the LLC cache domain.
6391 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6393 bool has_idle_core = false;
6394 struct sched_domain *sd;
6395 unsigned long task_util;
6396 int i, recent_used_cpu;
6399 * On asymmetric system, update task utilization because we will check
6400 * that the task fits with cpu's capacity.
6402 if (static_branch_unlikely(&sched_asym_cpucapacity)) {
6403 sync_entity_load_avg(&p->se);
6404 task_util = uclamp_task_util(p);
6408 * per-cpu select_idle_mask usage
6410 lockdep_assert_irqs_disabled();
6412 if ((available_idle_cpu(target) || sched_idle_cpu(target)) &&
6413 asym_fits_capacity(task_util, target))
6417 * If the previous CPU is cache affine and idle, don't be stupid:
6419 if (prev != target && cpus_share_cache(prev, target) &&
6420 (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
6421 asym_fits_capacity(task_util, prev))
6425 * Allow a per-cpu kthread to stack with the wakee if the
6426 * kworker thread and the tasks previous CPUs are the same.
6427 * The assumption is that the wakee queued work for the
6428 * per-cpu kthread that is now complete and the wakeup is
6429 * essentially a sync wakeup. An obvious example of this
6430 * pattern is IO completions.
6432 if (is_per_cpu_kthread(current) &&
6434 prev == smp_processor_id() &&
6435 this_rq()->nr_running <= 1 &&
6436 asym_fits_capacity(task_util, prev)) {
6440 /* Check a recently used CPU as a potential idle candidate: */
6441 recent_used_cpu = p->recent_used_cpu;
6442 p->recent_used_cpu = prev;
6443 if (recent_used_cpu != prev &&
6444 recent_used_cpu != target &&
6445 cpus_share_cache(recent_used_cpu, target) &&
6446 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
6447 cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr) &&
6448 asym_fits_capacity(task_util, recent_used_cpu)) {
6449 return recent_used_cpu;
6453 * For asymmetric CPU capacity systems, our domain of interest is
6454 * sd_asym_cpucapacity rather than sd_llc.
6456 if (static_branch_unlikely(&sched_asym_cpucapacity)) {
6457 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
6459 * On an asymmetric CPU capacity system where an exclusive
6460 * cpuset defines a symmetric island (i.e. one unique
6461 * capacity_orig value through the cpuset), the key will be set
6462 * but the CPUs within that cpuset will not have a domain with
6463 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
6467 i = select_idle_capacity(p, sd, target);
6468 return ((unsigned)i < nr_cpumask_bits) ? i : target;
6472 sd = rcu_dereference(per_cpu(sd_llc, target));
6476 if (sched_smt_active()) {
6477 has_idle_core = test_idle_cores(target, false);
6479 if (!has_idle_core && cpus_share_cache(prev, target)) {
6480 i = select_idle_smt(p, sd, prev);
6481 if ((unsigned int)i < nr_cpumask_bits)
6486 i = select_idle_cpu(p, sd, has_idle_core, target);
6487 if ((unsigned)i < nr_cpumask_bits)
6494 * cpu_util_without: compute cpu utilization without any contributions from *p
6495 * @cpu: the CPU which utilization is requested
6496 * @p: the task which utilization should be discounted
6498 * The utilization of a CPU is defined by the utilization of tasks currently
6499 * enqueued on that CPU as well as tasks which are currently sleeping after an
6500 * execution on that CPU.
6502 * This method returns the utilization of the specified CPU by discounting the
6503 * utilization of the specified task, whenever the task is currently
6504 * contributing to the CPU utilization.
6506 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6508 struct cfs_rq *cfs_rq;
6511 /* Task has no contribution or is new */
6512 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6513 return cpu_util_cfs(cpu);
6515 cfs_rq = &cpu_rq(cpu)->cfs;
6516 util = READ_ONCE(cfs_rq->avg.util_avg);
6518 /* Discount task's util from CPU's util */
6519 lsub_positive(&util, task_util(p));
6524 * a) if *p is the only task sleeping on this CPU, then:
6525 * cpu_util (== task_util) > util_est (== 0)
6526 * and thus we return:
6527 * cpu_util_without = (cpu_util - task_util) = 0
6529 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6531 * cpu_util >= task_util
6532 * cpu_util > util_est (== 0)
6533 * and thus we discount *p's blocked utilization to return:
6534 * cpu_util_without = (cpu_util - task_util) >= 0
6536 * c) if other tasks are RUNNABLE on that CPU and
6537 * util_est > cpu_util
6538 * then we use util_est since it returns a more restrictive
6539 * estimation of the spare capacity on that CPU, by just
6540 * considering the expected utilization of tasks already
6541 * runnable on that CPU.
6543 * Cases a) and b) are covered by the above code, while case c) is
6544 * covered by the following code when estimated utilization is
6547 if (sched_feat(UTIL_EST)) {
6548 unsigned int estimated =
6549 READ_ONCE(cfs_rq->avg.util_est.enqueued);
6552 * Despite the following checks we still have a small window
6553 * for a possible race, when an execl's select_task_rq_fair()
6554 * races with LB's detach_task():
6557 * p->on_rq = TASK_ON_RQ_MIGRATING;
6558 * ---------------------------------- A
6559 * deactivate_task() \
6560 * dequeue_task() + RaceTime
6561 * util_est_dequeue() /
6562 * ---------------------------------- B
6564 * The additional check on "current == p" it's required to
6565 * properly fix the execl regression and it helps in further
6566 * reducing the chances for the above race.
6568 if (unlikely(task_on_rq_queued(p) || current == p))
6569 lsub_positive(&estimated, _task_util_est(p));
6571 util = max(util, estimated);
6575 * Utilization (estimated) can exceed the CPU capacity, thus let's
6576 * clamp to the maximum CPU capacity to ensure consistency with
6579 return min_t(unsigned long, util, capacity_orig_of(cpu));
6583 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6586 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6588 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6589 unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
6592 * If @p migrates from @cpu to another, remove its contribution. Or,
6593 * if @p migrates from another CPU to @cpu, add its contribution. In
6594 * the other cases, @cpu is not impacted by the migration, so the
6595 * util_avg should already be correct.
6597 if (task_cpu(p) == cpu && dst_cpu != cpu)
6598 lsub_positive(&util, task_util(p));
6599 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6600 util += task_util(p);
6602 if (sched_feat(UTIL_EST)) {
6603 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6606 * During wake-up, the task isn't enqueued yet and doesn't
6607 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6608 * so just add it (if needed) to "simulate" what will be
6609 * cpu_util after the task has been enqueued.
6612 util_est += _task_util_est(p);
6614 util = max(util, util_est);
6617 return min(util, capacity_orig_of(cpu));
6621 * compute_energy(): Estimates the energy that @pd would consume if @p was
6622 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6623 * landscape of @pd's CPUs after the task migration, and uses the Energy Model
6624 * to compute what would be the energy if we decided to actually migrate that
6628 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
6630 struct cpumask *pd_mask = perf_domain_span(pd);
6631 unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask));
6632 unsigned long max_util = 0, sum_util = 0;
6633 unsigned long _cpu_cap = cpu_cap;
6636 _cpu_cap -= arch_scale_thermal_pressure(cpumask_first(pd_mask));
6639 * The capacity state of CPUs of the current rd can be driven by CPUs
6640 * of another rd if they belong to the same pd. So, account for the
6641 * utilization of these CPUs too by masking pd with cpu_online_mask
6642 * instead of the rd span.
6644 * If an entire pd is outside of the current rd, it will not appear in
6645 * its pd list and will not be accounted by compute_energy().
6647 for_each_cpu_and(cpu, pd_mask, cpu_online_mask) {
6648 unsigned long util_freq = cpu_util_next(cpu, p, dst_cpu);
6649 unsigned long cpu_util, util_running = util_freq;
6650 struct task_struct *tsk = NULL;
6653 * When @p is placed on @cpu:
6655 * util_running = max(cpu_util, cpu_util_est) +
6656 * max(task_util, _task_util_est)
6658 * while cpu_util_next is: max(cpu_util + task_util,
6659 * cpu_util_est + _task_util_est)
6661 if (cpu == dst_cpu) {
6664 cpu_util_next(cpu, p, -1) + task_util_est(p);
6668 * Busy time computation: utilization clamping is not
6669 * required since the ratio (sum_util / cpu_capacity)
6670 * is already enough to scale the EM reported power
6671 * consumption at the (eventually clamped) cpu_capacity.
6673 cpu_util = effective_cpu_util(cpu, util_running, cpu_cap,
6676 sum_util += min(cpu_util, _cpu_cap);
6679 * Performance domain frequency: utilization clamping
6680 * must be considered since it affects the selection
6681 * of the performance domain frequency.
6682 * NOTE: in case RT tasks are running, by default the
6683 * FREQUENCY_UTIL's utilization can be max OPP.
6685 cpu_util = effective_cpu_util(cpu, util_freq, cpu_cap,
6686 FREQUENCY_UTIL, tsk);
6687 max_util = max(max_util, min(cpu_util, _cpu_cap));
6690 return em_cpu_energy(pd->em_pd, max_util, sum_util, _cpu_cap);
6694 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6695 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6696 * spare capacity in each performance domain and uses it as a potential
6697 * candidate to execute the task. Then, it uses the Energy Model to figure
6698 * out which of the CPU candidates is the most energy-efficient.
6700 * The rationale for this heuristic is as follows. In a performance domain,
6701 * all the most energy efficient CPU candidates (according to the Energy
6702 * Model) are those for which we'll request a low frequency. When there are
6703 * several CPUs for which the frequency request will be the same, we don't
6704 * have enough data to break the tie between them, because the Energy Model
6705 * only includes active power costs. With this model, if we assume that
6706 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6707 * the maximum spare capacity in a performance domain is guaranteed to be among
6708 * the best candidates of the performance domain.
6710 * In practice, it could be preferable from an energy standpoint to pack
6711 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6712 * but that could also hurt our chances to go cluster idle, and we have no
6713 * ways to tell with the current Energy Model if this is actually a good
6714 * idea or not. So, find_energy_efficient_cpu() basically favors
6715 * cluster-packing, and spreading inside a cluster. That should at least be
6716 * a good thing for latency, and this is consistent with the idea that most
6717 * of the energy savings of EAS come from the asymmetry of the system, and
6718 * not so much from breaking the tie between identical CPUs. That's also the
6719 * reason why EAS is enabled in the topology code only for systems where
6720 * SD_ASYM_CPUCAPACITY is set.
6722 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6723 * they don't have any useful utilization data yet and it's not possible to
6724 * forecast their impact on energy consumption. Consequently, they will be
6725 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6726 * to be energy-inefficient in some use-cases. The alternative would be to
6727 * bias new tasks towards specific types of CPUs first, or to try to infer
6728 * their util_avg from the parent task, but those heuristics could hurt
6729 * other use-cases too. So, until someone finds a better way to solve this,
6730 * let's keep things simple by re-using the existing slow path.
6732 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6734 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
6735 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6736 int cpu, best_energy_cpu = prev_cpu, target = -1;
6737 unsigned long cpu_cap, util, base_energy = 0;
6738 struct sched_domain *sd;
6739 struct perf_domain *pd;
6742 pd = rcu_dereference(rd->pd);
6743 if (!pd || READ_ONCE(rd->overutilized))
6747 * Energy-aware wake-up happens on the lowest sched_domain starting
6748 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6750 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6751 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6758 sync_entity_load_avg(&p->se);
6759 if (!task_util_est(p))
6762 for (; pd; pd = pd->next) {
6763 unsigned long cur_delta, spare_cap, max_spare_cap = 0;
6764 bool compute_prev_delta = false;
6765 unsigned long base_energy_pd;
6766 int max_spare_cap_cpu = -1;
6768 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
6769 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6772 util = cpu_util_next(cpu, p, cpu);
6773 cpu_cap = capacity_of(cpu);
6774 spare_cap = cpu_cap;
6775 lsub_positive(&spare_cap, util);
6778 * Skip CPUs that cannot satisfy the capacity request.
6779 * IOW, placing the task there would make the CPU
6780 * overutilized. Take uclamp into account to see how
6781 * much capacity we can get out of the CPU; this is
6782 * aligned with sched_cpu_util().
6784 util = uclamp_rq_util_with(cpu_rq(cpu), util, p);
6785 if (!fits_capacity(util, cpu_cap))
6788 if (cpu == prev_cpu) {
6789 /* Always use prev_cpu as a candidate. */
6790 compute_prev_delta = true;
6791 } else if (spare_cap > max_spare_cap) {
6793 * Find the CPU with the maximum spare capacity
6794 * in the performance domain.
6796 max_spare_cap = spare_cap;
6797 max_spare_cap_cpu = cpu;
6801 if (max_spare_cap_cpu < 0 && !compute_prev_delta)
6804 /* Compute the 'base' energy of the pd, without @p */
6805 base_energy_pd = compute_energy(p, -1, pd);
6806 base_energy += base_energy_pd;
6808 /* Evaluate the energy impact of using prev_cpu. */
6809 if (compute_prev_delta) {
6810 prev_delta = compute_energy(p, prev_cpu, pd);
6811 if (prev_delta < base_energy_pd)
6813 prev_delta -= base_energy_pd;
6814 best_delta = min(best_delta, prev_delta);
6817 /* Evaluate the energy impact of using max_spare_cap_cpu. */
6818 if (max_spare_cap_cpu >= 0) {
6819 cur_delta = compute_energy(p, max_spare_cap_cpu, pd);
6820 if (cur_delta < base_energy_pd)
6822 cur_delta -= base_energy_pd;
6823 if (cur_delta < best_delta) {
6824 best_delta = cur_delta;
6825 best_energy_cpu = max_spare_cap_cpu;
6832 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6833 * least 6% of the energy used by prev_cpu.
6835 if ((prev_delta == ULONG_MAX) ||
6836 (prev_delta - best_delta) > ((prev_delta + base_energy) >> 4))
6837 target = best_energy_cpu;
6848 * select_task_rq_fair: Select target runqueue for the waking task in domains
6849 * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
6850 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6852 * Balances load by selecting the idlest CPU in the idlest group, or under
6853 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6855 * Returns the target CPU number.
6858 select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
6860 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6861 struct sched_domain *tmp, *sd = NULL;
6862 int cpu = smp_processor_id();
6863 int new_cpu = prev_cpu;
6864 int want_affine = 0;
6865 /* SD_flags and WF_flags share the first nibble */
6866 int sd_flag = wake_flags & 0xF;
6869 * required for stable ->cpus_allowed
6871 lockdep_assert_held(&p->pi_lock);
6872 if (wake_flags & WF_TTWU) {
6875 if (sched_energy_enabled()) {
6876 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
6882 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
6886 for_each_domain(cpu, tmp) {
6888 * If both 'cpu' and 'prev_cpu' are part of this domain,
6889 * cpu is a valid SD_WAKE_AFFINE target.
6891 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6892 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6893 if (cpu != prev_cpu)
6894 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6896 sd = NULL; /* Prefer wake_affine over balance flags */
6901 * Usually only true for WF_EXEC and WF_FORK, as sched_domains
6902 * usually do not have SD_BALANCE_WAKE set. That means wakeup
6903 * will usually go to the fast path.
6905 if (tmp->flags & sd_flag)
6907 else if (!want_affine)
6913 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6914 } else if (wake_flags & WF_TTWU) { /* XXX always ? */
6916 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6923 static void detach_entity_cfs_rq(struct sched_entity *se);
6926 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6927 * cfs_rq_of(p) references at time of call are still valid and identify the
6928 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6930 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6933 * As blocked tasks retain absolute vruntime the migration needs to
6934 * deal with this by subtracting the old and adding the new
6935 * min_vruntime -- the latter is done by enqueue_entity() when placing
6936 * the task on the new runqueue.
6938 if (READ_ONCE(p->__state) == TASK_WAKING) {
6939 struct sched_entity *se = &p->se;
6940 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6943 #ifndef CONFIG_64BIT
6944 u64 min_vruntime_copy;
6947 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6949 min_vruntime = cfs_rq->min_vruntime;
6950 } while (min_vruntime != min_vruntime_copy);
6952 min_vruntime = cfs_rq->min_vruntime;
6955 se->vruntime -= min_vruntime;
6958 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6960 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6961 * rq->lock and can modify state directly.
6963 lockdep_assert_rq_held(task_rq(p));
6964 detach_entity_cfs_rq(&p->se);
6968 * We are supposed to update the task to "current" time, then
6969 * its up to date and ready to go to new CPU/cfs_rq. But we
6970 * have difficulty in getting what current time is, so simply
6971 * throw away the out-of-date time. This will result in the
6972 * wakee task is less decayed, but giving the wakee more load
6975 remove_entity_load_avg(&p->se);
6978 /* Tell new CPU we are migrated */
6979 p->se.avg.last_update_time = 0;
6981 /* We have migrated, no longer consider this task hot */
6982 p->se.exec_start = 0;
6984 update_scan_period(p, new_cpu);
6987 static void task_dead_fair(struct task_struct *p)
6989 remove_entity_load_avg(&p->se);
6993 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6998 return newidle_balance(rq, rf) != 0;
7000 #endif /* CONFIG_SMP */
7002 static unsigned long wakeup_gran(struct sched_entity *se)
7004 unsigned long gran = sysctl_sched_wakeup_granularity;
7007 * Since its curr running now, convert the gran from real-time
7008 * to virtual-time in his units.
7010 * By using 'se' instead of 'curr' we penalize light tasks, so
7011 * they get preempted easier. That is, if 'se' < 'curr' then
7012 * the resulting gran will be larger, therefore penalizing the
7013 * lighter, if otoh 'se' > 'curr' then the resulting gran will
7014 * be smaller, again penalizing the lighter task.
7016 * This is especially important for buddies when the leftmost
7017 * task is higher priority than the buddy.
7019 return calc_delta_fair(gran, se);
7023 * Should 'se' preempt 'curr'.
7037 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
7039 s64 gran, vdiff = curr->vruntime - se->vruntime;
7044 gran = wakeup_gran(se);
7051 static void set_last_buddy(struct sched_entity *se)
7053 for_each_sched_entity(se) {
7054 if (SCHED_WARN_ON(!se->on_rq))
7058 cfs_rq_of(se)->last = se;
7062 static void set_next_buddy(struct sched_entity *se)
7064 for_each_sched_entity(se) {
7065 if (SCHED_WARN_ON(!se->on_rq))
7069 cfs_rq_of(se)->next = se;
7073 static void set_skip_buddy(struct sched_entity *se)
7075 for_each_sched_entity(se)
7076 cfs_rq_of(se)->skip = se;
7080 * Preempt the current task with a newly woken task if needed:
7082 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
7084 struct task_struct *curr = rq->curr;
7085 struct sched_entity *se = &curr->se, *pse = &p->se;
7086 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7087 int scale = cfs_rq->nr_running >= sched_nr_latency;
7088 int next_buddy_marked = 0;
7089 int cse_is_idle, pse_is_idle;
7091 if (unlikely(se == pse))
7095 * This is possible from callers such as attach_tasks(), in which we
7096 * unconditionally check_preempt_curr() after an enqueue (which may have
7097 * lead to a throttle). This both saves work and prevents false
7098 * next-buddy nomination below.
7100 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
7103 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
7104 set_next_buddy(pse);
7105 next_buddy_marked = 1;
7109 * We can come here with TIF_NEED_RESCHED already set from new task
7112 * Note: this also catches the edge-case of curr being in a throttled
7113 * group (e.g. via set_curr_task), since update_curr() (in the
7114 * enqueue of curr) will have resulted in resched being set. This
7115 * prevents us from potentially nominating it as a false LAST_BUDDY
7118 if (test_tsk_need_resched(curr))
7121 /* Idle tasks are by definition preempted by non-idle tasks. */
7122 if (unlikely(task_has_idle_policy(curr)) &&
7123 likely(!task_has_idle_policy(p)))
7127 * Batch and idle tasks do not preempt non-idle tasks (their preemption
7128 * is driven by the tick):
7130 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
7133 find_matching_se(&se, &pse);
7136 cse_is_idle = se_is_idle(se);
7137 pse_is_idle = se_is_idle(pse);
7140 * Preempt an idle group in favor of a non-idle group (and don't preempt
7141 * in the inverse case).
7143 if (cse_is_idle && !pse_is_idle)
7145 if (cse_is_idle != pse_is_idle)
7148 update_curr(cfs_rq_of(se));
7149 if (wakeup_preempt_entity(se, pse) == 1) {
7151 * Bias pick_next to pick the sched entity that is
7152 * triggering this preemption.
7154 if (!next_buddy_marked)
7155 set_next_buddy(pse);
7164 * Only set the backward buddy when the current task is still
7165 * on the rq. This can happen when a wakeup gets interleaved
7166 * with schedule on the ->pre_schedule() or idle_balance()
7167 * point, either of which can * drop the rq lock.
7169 * Also, during early boot the idle thread is in the fair class,
7170 * for obvious reasons its a bad idea to schedule back to it.
7172 if (unlikely(!se->on_rq || curr == rq->idle))
7175 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
7180 static struct task_struct *pick_task_fair(struct rq *rq)
7182 struct sched_entity *se;
7183 struct cfs_rq *cfs_rq;
7187 if (!cfs_rq->nr_running)
7191 struct sched_entity *curr = cfs_rq->curr;
7193 /* When we pick for a remote RQ, we'll not have done put_prev_entity() */
7196 update_curr(cfs_rq);
7200 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
7204 se = pick_next_entity(cfs_rq, curr);
7205 cfs_rq = group_cfs_rq(se);
7212 struct task_struct *
7213 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
7215 struct cfs_rq *cfs_rq = &rq->cfs;
7216 struct sched_entity *se;
7217 struct task_struct *p;
7221 if (!sched_fair_runnable(rq))
7224 #ifdef CONFIG_FAIR_GROUP_SCHED
7225 if (!prev || prev->sched_class != &fair_sched_class)
7229 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
7230 * likely that a next task is from the same cgroup as the current.
7232 * Therefore attempt to avoid putting and setting the entire cgroup
7233 * hierarchy, only change the part that actually changes.
7237 struct sched_entity *curr = cfs_rq->curr;
7240 * Since we got here without doing put_prev_entity() we also
7241 * have to consider cfs_rq->curr. If it is still a runnable
7242 * entity, update_curr() will update its vruntime, otherwise
7243 * forget we've ever seen it.
7247 update_curr(cfs_rq);
7252 * This call to check_cfs_rq_runtime() will do the
7253 * throttle and dequeue its entity in the parent(s).
7254 * Therefore the nr_running test will indeed
7257 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
7260 if (!cfs_rq->nr_running)
7267 se = pick_next_entity(cfs_rq, curr);
7268 cfs_rq = group_cfs_rq(se);
7274 * Since we haven't yet done put_prev_entity and if the selected task
7275 * is a different task than we started out with, try and touch the
7276 * least amount of cfs_rqs.
7279 struct sched_entity *pse = &prev->se;
7281 while (!(cfs_rq = is_same_group(se, pse))) {
7282 int se_depth = se->depth;
7283 int pse_depth = pse->depth;
7285 if (se_depth <= pse_depth) {
7286 put_prev_entity(cfs_rq_of(pse), pse);
7287 pse = parent_entity(pse);
7289 if (se_depth >= pse_depth) {
7290 set_next_entity(cfs_rq_of(se), se);
7291 se = parent_entity(se);
7295 put_prev_entity(cfs_rq, pse);
7296 set_next_entity(cfs_rq, se);
7303 put_prev_task(rq, prev);
7306 se = pick_next_entity(cfs_rq, NULL);
7307 set_next_entity(cfs_rq, se);
7308 cfs_rq = group_cfs_rq(se);
7313 done: __maybe_unused;
7316 * Move the next running task to the front of
7317 * the list, so our cfs_tasks list becomes MRU
7320 list_move(&p->se.group_node, &rq->cfs_tasks);
7323 if (hrtick_enabled_fair(rq))
7324 hrtick_start_fair(rq, p);
7326 update_misfit_status(p, rq);
7334 new_tasks = newidle_balance(rq, rf);
7337 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
7338 * possible for any higher priority task to appear. In that case we
7339 * must re-start the pick_next_entity() loop.
7348 * rq is about to be idle, check if we need to update the
7349 * lost_idle_time of clock_pelt
7351 update_idle_rq_clock_pelt(rq);
7356 static struct task_struct *__pick_next_task_fair(struct rq *rq)
7358 return pick_next_task_fair(rq, NULL, NULL);
7362 * Account for a descheduled task:
7364 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
7366 struct sched_entity *se = &prev->se;
7367 struct cfs_rq *cfs_rq;
7369 for_each_sched_entity(se) {
7370 cfs_rq = cfs_rq_of(se);
7371 put_prev_entity(cfs_rq, se);
7376 * sched_yield() is very simple
7378 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7380 static void yield_task_fair(struct rq *rq)
7382 struct task_struct *curr = rq->curr;
7383 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7384 struct sched_entity *se = &curr->se;
7387 * Are we the only task in the tree?
7389 if (unlikely(rq->nr_running == 1))
7392 clear_buddies(cfs_rq, se);
7394 if (curr->policy != SCHED_BATCH) {
7395 update_rq_clock(rq);
7397 * Update run-time statistics of the 'current'.
7399 update_curr(cfs_rq);
7401 * Tell update_rq_clock() that we've just updated,
7402 * so we don't do microscopic update in schedule()
7403 * and double the fastpath cost.
7405 rq_clock_skip_update(rq);
7411 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p)
7413 struct sched_entity *se = &p->se;
7415 /* throttled hierarchies are not runnable */
7416 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
7419 /* Tell the scheduler that we'd really like pse to run next. */
7422 yield_task_fair(rq);
7428 /**************************************************
7429 * Fair scheduling class load-balancing methods.
7433 * The purpose of load-balancing is to achieve the same basic fairness the
7434 * per-CPU scheduler provides, namely provide a proportional amount of compute
7435 * time to each task. This is expressed in the following equation:
7437 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7439 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7440 * W_i,0 is defined as:
7442 * W_i,0 = \Sum_j w_i,j (2)
7444 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7445 * is derived from the nice value as per sched_prio_to_weight[].
7447 * The weight average is an exponential decay average of the instantaneous
7450 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7452 * C_i is the compute capacity of CPU i, typically it is the
7453 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7454 * can also include other factors [XXX].
7456 * To achieve this balance we define a measure of imbalance which follows
7457 * directly from (1):
7459 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7461 * We them move tasks around to minimize the imbalance. In the continuous
7462 * function space it is obvious this converges, in the discrete case we get
7463 * a few fun cases generally called infeasible weight scenarios.
7466 * - infeasible weights;
7467 * - local vs global optima in the discrete case. ]
7472 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7473 * for all i,j solution, we create a tree of CPUs that follows the hardware
7474 * topology where each level pairs two lower groups (or better). This results
7475 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7476 * tree to only the first of the previous level and we decrease the frequency
7477 * of load-balance at each level inv. proportional to the number of CPUs in
7483 * \Sum { --- * --- * 2^i } = O(n) (5)
7485 * `- size of each group
7486 * | | `- number of CPUs doing load-balance
7488 * `- sum over all levels
7490 * Coupled with a limit on how many tasks we can migrate every balance pass,
7491 * this makes (5) the runtime complexity of the balancer.
7493 * An important property here is that each CPU is still (indirectly) connected
7494 * to every other CPU in at most O(log n) steps:
7496 * The adjacency matrix of the resulting graph is given by:
7499 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7502 * And you'll find that:
7504 * A^(log_2 n)_i,j != 0 for all i,j (7)
7506 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7507 * The task movement gives a factor of O(m), giving a convergence complexity
7510 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7515 * In order to avoid CPUs going idle while there's still work to do, new idle
7516 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7517 * tree itself instead of relying on other CPUs to bring it work.
7519 * This adds some complexity to both (5) and (8) but it reduces the total idle
7527 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7530 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7535 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7537 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7539 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7542 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7543 * rewrite all of this once again.]
7546 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7548 enum fbq_type { regular, remote, all };
7551 * 'group_type' describes the group of CPUs at the moment of load balancing.
7553 * The enum is ordered by pulling priority, with the group with lowest priority
7554 * first so the group_type can simply be compared when selecting the busiest
7555 * group. See update_sd_pick_busiest().
7558 /* The group has spare capacity that can be used to run more tasks. */
7559 group_has_spare = 0,
7561 * The group is fully used and the tasks don't compete for more CPU
7562 * cycles. Nevertheless, some tasks might wait before running.
7566 * SD_ASYM_CPUCAPACITY only: One task doesn't fit with CPU's capacity
7567 * and must be migrated to a more powerful CPU.
7571 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
7572 * and the task should be migrated to it instead of running on the
7577 * The tasks' affinity constraints previously prevented the scheduler
7578 * from balancing the load across the system.
7582 * The CPU is overloaded and can't provide expected CPU cycles to all
7588 enum migration_type {
7595 #define LBF_ALL_PINNED 0x01
7596 #define LBF_NEED_BREAK 0x02
7597 #define LBF_DST_PINNED 0x04
7598 #define LBF_SOME_PINNED 0x08
7599 #define LBF_ACTIVE_LB 0x10
7602 struct sched_domain *sd;
7610 struct cpumask *dst_grpmask;
7612 enum cpu_idle_type idle;
7614 /* The set of CPUs under consideration for load-balancing */
7615 struct cpumask *cpus;
7620 unsigned int loop_break;
7621 unsigned int loop_max;
7623 enum fbq_type fbq_type;
7624 enum migration_type migration_type;
7625 struct list_head tasks;
7629 * Is this task likely cache-hot:
7631 static int task_hot(struct task_struct *p, struct lb_env *env)
7635 lockdep_assert_rq_held(env->src_rq);
7637 if (p->sched_class != &fair_sched_class)
7640 if (unlikely(task_has_idle_policy(p)))
7643 /* SMT siblings share cache */
7644 if (env->sd->flags & SD_SHARE_CPUCAPACITY)
7648 * Buddy candidates are cache hot:
7650 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7651 (&p->se == cfs_rq_of(&p->se)->next ||
7652 &p->se == cfs_rq_of(&p->se)->last))
7655 if (sysctl_sched_migration_cost == -1)
7659 * Don't migrate task if the task's cookie does not match
7660 * with the destination CPU's core cookie.
7662 if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p))
7665 if (sysctl_sched_migration_cost == 0)
7668 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7670 return delta < (s64)sysctl_sched_migration_cost;
7673 #ifdef CONFIG_NUMA_BALANCING
7675 * Returns 1, if task migration degrades locality
7676 * Returns 0, if task migration improves locality i.e migration preferred.
7677 * Returns -1, if task migration is not affected by locality.
7679 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7681 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7682 unsigned long src_weight, dst_weight;
7683 int src_nid, dst_nid, dist;
7685 if (!static_branch_likely(&sched_numa_balancing))
7688 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7691 src_nid = cpu_to_node(env->src_cpu);
7692 dst_nid = cpu_to_node(env->dst_cpu);
7694 if (src_nid == dst_nid)
7697 /* Migrating away from the preferred node is always bad. */
7698 if (src_nid == p->numa_preferred_nid) {
7699 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7705 /* Encourage migration to the preferred node. */
7706 if (dst_nid == p->numa_preferred_nid)
7709 /* Leaving a core idle is often worse than degrading locality. */
7710 if (env->idle == CPU_IDLE)
7713 dist = node_distance(src_nid, dst_nid);
7715 src_weight = group_weight(p, src_nid, dist);
7716 dst_weight = group_weight(p, dst_nid, dist);
7718 src_weight = task_weight(p, src_nid, dist);
7719 dst_weight = task_weight(p, dst_nid, dist);
7722 return dst_weight < src_weight;
7726 static inline int migrate_degrades_locality(struct task_struct *p,
7734 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7737 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7741 lockdep_assert_rq_held(env->src_rq);
7744 * We do not migrate tasks that are:
7745 * 1) throttled_lb_pair, or
7746 * 2) cannot be migrated to this CPU due to cpus_ptr, or
7747 * 3) running (obviously), or
7748 * 4) are cache-hot on their current CPU.
7750 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7753 /* Disregard pcpu kthreads; they are where they need to be. */
7754 if (kthread_is_per_cpu(p))
7757 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
7760 schedstat_inc(p->stats.nr_failed_migrations_affine);
7762 env->flags |= LBF_SOME_PINNED;
7765 * Remember if this task can be migrated to any other CPU in
7766 * our sched_group. We may want to revisit it if we couldn't
7767 * meet load balance goals by pulling other tasks on src_cpu.
7769 * Avoid computing new_dst_cpu
7771 * - if we have already computed one in current iteration
7772 * - if it's an active balance
7774 if (env->idle == CPU_NEWLY_IDLE ||
7775 env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB))
7778 /* Prevent to re-select dst_cpu via env's CPUs: */
7779 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7780 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
7781 env->flags |= LBF_DST_PINNED;
7782 env->new_dst_cpu = cpu;
7790 /* Record that we found at least one task that could run on dst_cpu */
7791 env->flags &= ~LBF_ALL_PINNED;
7793 if (task_running(env->src_rq, p)) {
7794 schedstat_inc(p->stats.nr_failed_migrations_running);
7799 * Aggressive migration if:
7801 * 2) destination numa is preferred
7802 * 3) task is cache cold, or
7803 * 4) too many balance attempts have failed.
7805 if (env->flags & LBF_ACTIVE_LB)
7808 tsk_cache_hot = migrate_degrades_locality(p, env);
7809 if (tsk_cache_hot == -1)
7810 tsk_cache_hot = task_hot(p, env);
7812 if (tsk_cache_hot <= 0 ||
7813 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7814 if (tsk_cache_hot == 1) {
7815 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7816 schedstat_inc(p->stats.nr_forced_migrations);
7821 schedstat_inc(p->stats.nr_failed_migrations_hot);
7826 * detach_task() -- detach the task for the migration specified in env
7828 static void detach_task(struct task_struct *p, struct lb_env *env)
7830 lockdep_assert_rq_held(env->src_rq);
7832 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7833 set_task_cpu(p, env->dst_cpu);
7837 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7838 * part of active balancing operations within "domain".
7840 * Returns a task if successful and NULL otherwise.
7842 static struct task_struct *detach_one_task(struct lb_env *env)
7844 struct task_struct *p;
7846 lockdep_assert_rq_held(env->src_rq);
7848 list_for_each_entry_reverse(p,
7849 &env->src_rq->cfs_tasks, se.group_node) {
7850 if (!can_migrate_task(p, env))
7853 detach_task(p, env);
7856 * Right now, this is only the second place where
7857 * lb_gained[env->idle] is updated (other is detach_tasks)
7858 * so we can safely collect stats here rather than
7859 * inside detach_tasks().
7861 schedstat_inc(env->sd->lb_gained[env->idle]);
7867 static const unsigned int sched_nr_migrate_break = 32;
7870 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
7871 * busiest_rq, as part of a balancing operation within domain "sd".
7873 * Returns number of detached tasks if successful and 0 otherwise.
7875 static int detach_tasks(struct lb_env *env)
7877 struct list_head *tasks = &env->src_rq->cfs_tasks;
7878 unsigned long util, load;
7879 struct task_struct *p;
7882 lockdep_assert_rq_held(env->src_rq);
7885 * Source run queue has been emptied by another CPU, clear
7886 * LBF_ALL_PINNED flag as we will not test any task.
7888 if (env->src_rq->nr_running <= 1) {
7889 env->flags &= ~LBF_ALL_PINNED;
7893 if (env->imbalance <= 0)
7896 while (!list_empty(tasks)) {
7898 * We don't want to steal all, otherwise we may be treated likewise,
7899 * which could at worst lead to a livelock crash.
7901 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7904 p = list_last_entry(tasks, struct task_struct, se.group_node);
7907 /* We've more or less seen every task there is, call it quits */
7908 if (env->loop > env->loop_max)
7911 /* take a breather every nr_migrate tasks */
7912 if (env->loop > env->loop_break) {
7913 env->loop_break += sched_nr_migrate_break;
7914 env->flags |= LBF_NEED_BREAK;
7918 if (!can_migrate_task(p, env))
7921 switch (env->migration_type) {
7924 * Depending of the number of CPUs and tasks and the
7925 * cgroup hierarchy, task_h_load() can return a null
7926 * value. Make sure that env->imbalance decreases
7927 * otherwise detach_tasks() will stop only after
7928 * detaching up to loop_max tasks.
7930 load = max_t(unsigned long, task_h_load(p), 1);
7932 if (sched_feat(LB_MIN) &&
7933 load < 16 && !env->sd->nr_balance_failed)
7937 * Make sure that we don't migrate too much load.
7938 * Nevertheless, let relax the constraint if
7939 * scheduler fails to find a good waiting task to
7942 if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance)
7945 env->imbalance -= load;
7949 util = task_util_est(p);
7951 if (util > env->imbalance)
7954 env->imbalance -= util;
7961 case migrate_misfit:
7962 /* This is not a misfit task */
7963 if (task_fits_capacity(p, capacity_of(env->src_cpu)))
7970 detach_task(p, env);
7971 list_add(&p->se.group_node, &env->tasks);
7975 #ifdef CONFIG_PREEMPTION
7977 * NEWIDLE balancing is a source of latency, so preemptible
7978 * kernels will stop after the first task is detached to minimize
7979 * the critical section.
7981 if (env->idle == CPU_NEWLY_IDLE)
7986 * We only want to steal up to the prescribed amount of
7989 if (env->imbalance <= 0)
7994 list_move(&p->se.group_node, tasks);
7998 * Right now, this is one of only two places we collect this stat
7999 * so we can safely collect detach_one_task() stats here rather
8000 * than inside detach_one_task().
8002 schedstat_add(env->sd->lb_gained[env->idle], detached);
8008 * attach_task() -- attach the task detached by detach_task() to its new rq.
8010 static void attach_task(struct rq *rq, struct task_struct *p)
8012 lockdep_assert_rq_held(rq);
8014 BUG_ON(task_rq(p) != rq);
8015 activate_task(rq, p, ENQUEUE_NOCLOCK);
8016 check_preempt_curr(rq, p, 0);
8020 * attach_one_task() -- attaches the task returned from detach_one_task() to
8023 static void attach_one_task(struct rq *rq, struct task_struct *p)
8028 update_rq_clock(rq);
8034 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
8037 static void attach_tasks(struct lb_env *env)
8039 struct list_head *tasks = &env->tasks;
8040 struct task_struct *p;
8043 rq_lock(env->dst_rq, &rf);
8044 update_rq_clock(env->dst_rq);
8046 while (!list_empty(tasks)) {
8047 p = list_first_entry(tasks, struct task_struct, se.group_node);
8048 list_del_init(&p->se.group_node);
8050 attach_task(env->dst_rq, p);
8053 rq_unlock(env->dst_rq, &rf);
8056 #ifdef CONFIG_NO_HZ_COMMON
8057 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
8059 if (cfs_rq->avg.load_avg)
8062 if (cfs_rq->avg.util_avg)
8068 static inline bool others_have_blocked(struct rq *rq)
8070 if (READ_ONCE(rq->avg_rt.util_avg))
8073 if (READ_ONCE(rq->avg_dl.util_avg))
8076 if (thermal_load_avg(rq))
8079 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
8080 if (READ_ONCE(rq->avg_irq.util_avg))
8087 static inline void update_blocked_load_tick(struct rq *rq)
8089 WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies);
8092 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
8095 rq->has_blocked_load = 0;
8098 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
8099 static inline bool others_have_blocked(struct rq *rq) { return false; }
8100 static inline void update_blocked_load_tick(struct rq *rq) {}
8101 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
8104 static bool __update_blocked_others(struct rq *rq, bool *done)
8106 const struct sched_class *curr_class;
8107 u64 now = rq_clock_pelt(rq);
8108 unsigned long thermal_pressure;
8112 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
8113 * DL and IRQ signals have been updated before updating CFS.
8115 curr_class = rq->curr->sched_class;
8117 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
8119 decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
8120 update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
8121 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) |
8122 update_irq_load_avg(rq, 0);
8124 if (others_have_blocked(rq))
8130 #ifdef CONFIG_FAIR_GROUP_SCHED
8132 static bool __update_blocked_fair(struct rq *rq, bool *done)
8134 struct cfs_rq *cfs_rq, *pos;
8135 bool decayed = false;
8136 int cpu = cpu_of(rq);
8139 * Iterates the task_group tree in a bottom up fashion, see
8140 * list_add_leaf_cfs_rq() for details.
8142 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
8143 struct sched_entity *se;
8145 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
8146 update_tg_load_avg(cfs_rq);
8148 if (cfs_rq == &rq->cfs)
8152 /* Propagate pending load changes to the parent, if any: */
8153 se = cfs_rq->tg->se[cpu];
8154 if (se && !skip_blocked_update(se))
8155 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
8158 * There can be a lot of idle CPU cgroups. Don't let fully
8159 * decayed cfs_rqs linger on the list.
8161 if (cfs_rq_is_decayed(cfs_rq))
8162 list_del_leaf_cfs_rq(cfs_rq);
8164 /* Don't need periodic decay once load/util_avg are null */
8165 if (cfs_rq_has_blocked(cfs_rq))
8173 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
8174 * This needs to be done in a top-down fashion because the load of a child
8175 * group is a fraction of its parents load.
8177 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
8179 struct rq *rq = rq_of(cfs_rq);
8180 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
8181 unsigned long now = jiffies;
8184 if (cfs_rq->last_h_load_update == now)
8187 WRITE_ONCE(cfs_rq->h_load_next, NULL);
8188 for_each_sched_entity(se) {
8189 cfs_rq = cfs_rq_of(se);
8190 WRITE_ONCE(cfs_rq->h_load_next, se);
8191 if (cfs_rq->last_h_load_update == now)
8196 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
8197 cfs_rq->last_h_load_update = now;
8200 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
8201 load = cfs_rq->h_load;
8202 load = div64_ul(load * se->avg.load_avg,
8203 cfs_rq_load_avg(cfs_rq) + 1);
8204 cfs_rq = group_cfs_rq(se);
8205 cfs_rq->h_load = load;
8206 cfs_rq->last_h_load_update = now;
8210 static unsigned long task_h_load(struct task_struct *p)
8212 struct cfs_rq *cfs_rq = task_cfs_rq(p);
8214 update_cfs_rq_h_load(cfs_rq);
8215 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
8216 cfs_rq_load_avg(cfs_rq) + 1);
8219 static bool __update_blocked_fair(struct rq *rq, bool *done)
8221 struct cfs_rq *cfs_rq = &rq->cfs;
8224 decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
8225 if (cfs_rq_has_blocked(cfs_rq))
8231 static unsigned long task_h_load(struct task_struct *p)
8233 return p->se.avg.load_avg;
8237 static void update_blocked_averages(int cpu)
8239 bool decayed = false, done = true;
8240 struct rq *rq = cpu_rq(cpu);
8243 rq_lock_irqsave(rq, &rf);
8244 update_blocked_load_tick(rq);
8245 update_rq_clock(rq);
8247 decayed |= __update_blocked_others(rq, &done);
8248 decayed |= __update_blocked_fair(rq, &done);
8250 update_blocked_load_status(rq, !done);
8252 cpufreq_update_util(rq, 0);
8253 rq_unlock_irqrestore(rq, &rf);
8256 /********** Helpers for find_busiest_group ************************/
8259 * sg_lb_stats - stats of a sched_group required for load_balancing
8261 struct sg_lb_stats {
8262 unsigned long avg_load; /*Avg load across the CPUs of the group */
8263 unsigned long group_load; /* Total load over the CPUs of the group */
8264 unsigned long group_capacity;
8265 unsigned long group_util; /* Total utilization over the CPUs of the group */
8266 unsigned long group_runnable; /* Total runnable time over the CPUs of the group */
8267 unsigned int sum_nr_running; /* Nr of tasks running in the group */
8268 unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
8269 unsigned int idle_cpus;
8270 unsigned int group_weight;
8271 enum group_type group_type;
8272 unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
8273 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
8274 #ifdef CONFIG_NUMA_BALANCING
8275 unsigned int nr_numa_running;
8276 unsigned int nr_preferred_running;
8281 * sd_lb_stats - Structure to store the statistics of a sched_domain
8282 * during load balancing.
8284 struct sd_lb_stats {
8285 struct sched_group *busiest; /* Busiest group in this sd */
8286 struct sched_group *local; /* Local group in this sd */
8287 unsigned long total_load; /* Total load of all groups in sd */
8288 unsigned long total_capacity; /* Total capacity of all groups in sd */
8289 unsigned long avg_load; /* Average load across all groups in sd */
8290 unsigned int prefer_sibling; /* tasks should go to sibling first */
8292 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
8293 struct sg_lb_stats local_stat; /* Statistics of the local group */
8296 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
8299 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
8300 * local_stat because update_sg_lb_stats() does a full clear/assignment.
8301 * We must however set busiest_stat::group_type and
8302 * busiest_stat::idle_cpus to the worst busiest group because
8303 * update_sd_pick_busiest() reads these before assignment.
8305 *sds = (struct sd_lb_stats){
8309 .total_capacity = 0UL,
8311 .idle_cpus = UINT_MAX,
8312 .group_type = group_has_spare,
8317 static unsigned long scale_rt_capacity(int cpu)
8319 struct rq *rq = cpu_rq(cpu);
8320 unsigned long max = arch_scale_cpu_capacity(cpu);
8321 unsigned long used, free;
8324 irq = cpu_util_irq(rq);
8326 if (unlikely(irq >= max))
8330 * avg_rt.util_avg and avg_dl.util_avg track binary signals
8331 * (running and not running) with weights 0 and 1024 respectively.
8332 * avg_thermal.load_avg tracks thermal pressure and the weighted
8333 * average uses the actual delta max capacity(load).
8335 used = READ_ONCE(rq->avg_rt.util_avg);
8336 used += READ_ONCE(rq->avg_dl.util_avg);
8337 used += thermal_load_avg(rq);
8339 if (unlikely(used >= max))
8344 return scale_irq_capacity(free, irq, max);
8347 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
8349 unsigned long capacity = scale_rt_capacity(cpu);
8350 struct sched_group *sdg = sd->groups;
8352 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
8357 cpu_rq(cpu)->cpu_capacity = capacity;
8358 trace_sched_cpu_capacity_tp(cpu_rq(cpu));
8360 sdg->sgc->capacity = capacity;
8361 sdg->sgc->min_capacity = capacity;
8362 sdg->sgc->max_capacity = capacity;
8365 void update_group_capacity(struct sched_domain *sd, int cpu)
8367 struct sched_domain *child = sd->child;
8368 struct sched_group *group, *sdg = sd->groups;
8369 unsigned long capacity, min_capacity, max_capacity;
8370 unsigned long interval;
8372 interval = msecs_to_jiffies(sd->balance_interval);
8373 interval = clamp(interval, 1UL, max_load_balance_interval);
8374 sdg->sgc->next_update = jiffies + interval;
8377 update_cpu_capacity(sd, cpu);
8382 min_capacity = ULONG_MAX;
8385 if (child->flags & SD_OVERLAP) {
8387 * SD_OVERLAP domains cannot assume that child groups
8388 * span the current group.
8391 for_each_cpu(cpu, sched_group_span(sdg)) {
8392 unsigned long cpu_cap = capacity_of(cpu);
8394 capacity += cpu_cap;
8395 min_capacity = min(cpu_cap, min_capacity);
8396 max_capacity = max(cpu_cap, max_capacity);
8400 * !SD_OVERLAP domains can assume that child groups
8401 * span the current group.
8404 group = child->groups;
8406 struct sched_group_capacity *sgc = group->sgc;
8408 capacity += sgc->capacity;
8409 min_capacity = min(sgc->min_capacity, min_capacity);
8410 max_capacity = max(sgc->max_capacity, max_capacity);
8411 group = group->next;
8412 } while (group != child->groups);
8415 sdg->sgc->capacity = capacity;
8416 sdg->sgc->min_capacity = min_capacity;
8417 sdg->sgc->max_capacity = max_capacity;
8421 * Check whether the capacity of the rq has been noticeably reduced by side
8422 * activity. The imbalance_pct is used for the threshold.
8423 * Return true is the capacity is reduced
8426 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
8428 return ((rq->cpu_capacity * sd->imbalance_pct) <
8429 (rq->cpu_capacity_orig * 100));
8433 * Check whether a rq has a misfit task and if it looks like we can actually
8434 * help that task: we can migrate the task to a CPU of higher capacity, or
8435 * the task's current CPU is heavily pressured.
8437 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
8439 return rq->misfit_task_load &&
8440 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
8441 check_cpu_capacity(rq, sd));
8445 * Group imbalance indicates (and tries to solve) the problem where balancing
8446 * groups is inadequate due to ->cpus_ptr constraints.
8448 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
8449 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
8452 * { 0 1 2 3 } { 4 5 6 7 }
8455 * If we were to balance group-wise we'd place two tasks in the first group and
8456 * two tasks in the second group. Clearly this is undesired as it will overload
8457 * cpu 3 and leave one of the CPUs in the second group unused.
8459 * The current solution to this issue is detecting the skew in the first group
8460 * by noticing the lower domain failed to reach balance and had difficulty
8461 * moving tasks due to affinity constraints.
8463 * When this is so detected; this group becomes a candidate for busiest; see
8464 * update_sd_pick_busiest(). And calculate_imbalance() and
8465 * find_busiest_group() avoid some of the usual balance conditions to allow it
8466 * to create an effective group imbalance.
8468 * This is a somewhat tricky proposition since the next run might not find the
8469 * group imbalance and decide the groups need to be balanced again. A most
8470 * subtle and fragile situation.
8473 static inline int sg_imbalanced(struct sched_group *group)
8475 return group->sgc->imbalance;
8479 * group_has_capacity returns true if the group has spare capacity that could
8480 * be used by some tasks.
8481 * We consider that a group has spare capacity if the * number of task is
8482 * smaller than the number of CPUs or if the utilization is lower than the
8483 * available capacity for CFS tasks.
8484 * For the latter, we use a threshold to stabilize the state, to take into
8485 * account the variance of the tasks' load and to return true if the available
8486 * capacity in meaningful for the load balancer.
8487 * As an example, an available capacity of 1% can appear but it doesn't make
8488 * any benefit for the load balance.
8491 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
8493 if (sgs->sum_nr_running < sgs->group_weight)
8496 if ((sgs->group_capacity * imbalance_pct) <
8497 (sgs->group_runnable * 100))
8500 if ((sgs->group_capacity * 100) >
8501 (sgs->group_util * imbalance_pct))
8508 * group_is_overloaded returns true if the group has more tasks than it can
8510 * group_is_overloaded is not equals to !group_has_capacity because a group
8511 * with the exact right number of tasks, has no more spare capacity but is not
8512 * overloaded so both group_has_capacity and group_is_overloaded return
8516 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
8518 if (sgs->sum_nr_running <= sgs->group_weight)
8521 if ((sgs->group_capacity * 100) <
8522 (sgs->group_util * imbalance_pct))
8525 if ((sgs->group_capacity * imbalance_pct) <
8526 (sgs->group_runnable * 100))
8533 group_type group_classify(unsigned int imbalance_pct,
8534 struct sched_group *group,
8535 struct sg_lb_stats *sgs)
8537 if (group_is_overloaded(imbalance_pct, sgs))
8538 return group_overloaded;
8540 if (sg_imbalanced(group))
8541 return group_imbalanced;
8543 if (sgs->group_asym_packing)
8544 return group_asym_packing;
8546 if (sgs->group_misfit_task_load)
8547 return group_misfit_task;
8549 if (!group_has_capacity(imbalance_pct, sgs))
8550 return group_fully_busy;
8552 return group_has_spare;
8556 * asym_smt_can_pull_tasks - Check whether the load balancing CPU can pull tasks
8557 * @dst_cpu: Destination CPU of the load balancing
8558 * @sds: Load-balancing data with statistics of the local group
8559 * @sgs: Load-balancing statistics of the candidate busiest group
8560 * @sg: The candidate busiest group
8562 * Check the state of the SMT siblings of both @sds::local and @sg and decide
8563 * if @dst_cpu can pull tasks.
8565 * If @dst_cpu does not have SMT siblings, it can pull tasks if two or more of
8566 * the SMT siblings of @sg are busy. If only one CPU in @sg is busy, pull tasks
8567 * only if @dst_cpu has higher priority.
8569 * If both @dst_cpu and @sg have SMT siblings, and @sg has exactly one more
8570 * busy CPU than @sds::local, let @dst_cpu pull tasks if it has higher priority.
8571 * Bigger imbalances in the number of busy CPUs will be dealt with in
8572 * update_sd_pick_busiest().
8574 * If @sg does not have SMT siblings, only pull tasks if all of the SMT siblings
8575 * of @dst_cpu are idle and @sg has lower priority.
8577 * Return: true if @dst_cpu can pull tasks, false otherwise.
8579 static bool asym_smt_can_pull_tasks(int dst_cpu, struct sd_lb_stats *sds,
8580 struct sg_lb_stats *sgs,
8581 struct sched_group *sg)
8583 #ifdef CONFIG_SCHED_SMT
8584 bool local_is_smt, sg_is_smt;
8587 local_is_smt = sds->local->flags & SD_SHARE_CPUCAPACITY;
8588 sg_is_smt = sg->flags & SD_SHARE_CPUCAPACITY;
8590 sg_busy_cpus = sgs->group_weight - sgs->idle_cpus;
8592 if (!local_is_smt) {
8594 * If we are here, @dst_cpu is idle and does not have SMT
8595 * siblings. Pull tasks if candidate group has two or more
8598 if (sg_busy_cpus >= 2) /* implies sg_is_smt */
8602 * @dst_cpu does not have SMT siblings. @sg may have SMT
8603 * siblings and only one is busy. In such case, @dst_cpu
8604 * can help if it has higher priority and is idle (i.e.,
8605 * it has no running tasks).
8607 return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu);
8610 /* @dst_cpu has SMT siblings. */
8613 int local_busy_cpus = sds->local->group_weight -
8614 sds->local_stat.idle_cpus;
8615 int busy_cpus_delta = sg_busy_cpus - local_busy_cpus;
8617 if (busy_cpus_delta == 1)
8618 return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu);
8624 * @sg does not have SMT siblings. Ensure that @sds::local does not end
8625 * up with more than one busy SMT sibling and only pull tasks if there
8626 * are not busy CPUs (i.e., no CPU has running tasks).
8628 if (!sds->local_stat.sum_nr_running)
8629 return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu);
8633 /* Always return false so that callers deal with non-SMT cases. */
8639 sched_asym(struct lb_env *env, struct sd_lb_stats *sds, struct sg_lb_stats *sgs,
8640 struct sched_group *group)
8642 /* Only do SMT checks if either local or candidate have SMT siblings */
8643 if ((sds->local->flags & SD_SHARE_CPUCAPACITY) ||
8644 (group->flags & SD_SHARE_CPUCAPACITY))
8645 return asym_smt_can_pull_tasks(env->dst_cpu, sds, sgs, group);
8647 return sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu);
8651 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8652 * @env: The load balancing environment.
8653 * @sds: Load-balancing data with statistics of the local group.
8654 * @group: sched_group whose statistics are to be updated.
8655 * @sgs: variable to hold the statistics for this group.
8656 * @sg_status: Holds flag indicating the status of the sched_group
8658 static inline void update_sg_lb_stats(struct lb_env *env,
8659 struct sd_lb_stats *sds,
8660 struct sched_group *group,
8661 struct sg_lb_stats *sgs,
8664 int i, nr_running, local_group;
8666 memset(sgs, 0, sizeof(*sgs));
8668 local_group = group == sds->local;
8670 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8671 struct rq *rq = cpu_rq(i);
8673 sgs->group_load += cpu_load(rq);
8674 sgs->group_util += cpu_util_cfs(i);
8675 sgs->group_runnable += cpu_runnable(rq);
8676 sgs->sum_h_nr_running += rq->cfs.h_nr_running;
8678 nr_running = rq->nr_running;
8679 sgs->sum_nr_running += nr_running;
8682 *sg_status |= SG_OVERLOAD;
8684 if (cpu_overutilized(i))
8685 *sg_status |= SG_OVERUTILIZED;
8687 #ifdef CONFIG_NUMA_BALANCING
8688 sgs->nr_numa_running += rq->nr_numa_running;
8689 sgs->nr_preferred_running += rq->nr_preferred_running;
8692 * No need to call idle_cpu() if nr_running is not 0
8694 if (!nr_running && idle_cpu(i)) {
8696 /* Idle cpu can't have misfit task */
8703 /* Check for a misfit task on the cpu */
8704 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8705 sgs->group_misfit_task_load < rq->misfit_task_load) {
8706 sgs->group_misfit_task_load = rq->misfit_task_load;
8707 *sg_status |= SG_OVERLOAD;
8711 sgs->group_capacity = group->sgc->capacity;
8713 sgs->group_weight = group->group_weight;
8715 /* Check if dst CPU is idle and preferred to this group */
8716 if (!local_group && env->sd->flags & SD_ASYM_PACKING &&
8717 env->idle != CPU_NOT_IDLE && sgs->sum_h_nr_running &&
8718 sched_asym(env, sds, sgs, group)) {
8719 sgs->group_asym_packing = 1;
8722 sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
8724 /* Computing avg_load makes sense only when group is overloaded */
8725 if (sgs->group_type == group_overloaded)
8726 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
8727 sgs->group_capacity;
8731 * update_sd_pick_busiest - return 1 on busiest group
8732 * @env: The load balancing environment.
8733 * @sds: sched_domain statistics
8734 * @sg: sched_group candidate to be checked for being the busiest
8735 * @sgs: sched_group statistics
8737 * Determine if @sg is a busier group than the previously selected
8740 * Return: %true if @sg is a busier group than the previously selected
8741 * busiest group. %false otherwise.
8743 static bool update_sd_pick_busiest(struct lb_env *env,
8744 struct sd_lb_stats *sds,
8745 struct sched_group *sg,
8746 struct sg_lb_stats *sgs)
8748 struct sg_lb_stats *busiest = &sds->busiest_stat;
8750 /* Make sure that there is at least one task to pull */
8751 if (!sgs->sum_h_nr_running)
8755 * Don't try to pull misfit tasks we can't help.
8756 * We can use max_capacity here as reduction in capacity on some
8757 * CPUs in the group should either be possible to resolve
8758 * internally or be covered by avg_load imbalance (eventually).
8760 if (sgs->group_type == group_misfit_task &&
8761 (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) ||
8762 sds->local_stat.group_type != group_has_spare))
8765 if (sgs->group_type > busiest->group_type)
8768 if (sgs->group_type < busiest->group_type)
8772 * The candidate and the current busiest group are the same type of
8773 * group. Let check which one is the busiest according to the type.
8776 switch (sgs->group_type) {
8777 case group_overloaded:
8778 /* Select the overloaded group with highest avg_load. */
8779 if (sgs->avg_load <= busiest->avg_load)
8783 case group_imbalanced:
8785 * Select the 1st imbalanced group as we don't have any way to
8786 * choose one more than another.
8790 case group_asym_packing:
8791 /* Prefer to move from lowest priority CPU's work */
8792 if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu))
8796 case group_misfit_task:
8798 * If we have more than one misfit sg go with the biggest
8801 if (sgs->group_misfit_task_load < busiest->group_misfit_task_load)
8805 case group_fully_busy:
8807 * Select the fully busy group with highest avg_load. In
8808 * theory, there is no need to pull task from such kind of
8809 * group because tasks have all compute capacity that they need
8810 * but we can still improve the overall throughput by reducing
8811 * contention when accessing shared HW resources.
8813 * XXX for now avg_load is not computed and always 0 so we
8814 * select the 1st one.
8816 if (sgs->avg_load <= busiest->avg_load)
8820 case group_has_spare:
8822 * Select not overloaded group with lowest number of idle cpus
8823 * and highest number of running tasks. We could also compare
8824 * the spare capacity which is more stable but it can end up
8825 * that the group has less spare capacity but finally more idle
8826 * CPUs which means less opportunity to pull tasks.
8828 if (sgs->idle_cpus > busiest->idle_cpus)
8830 else if ((sgs->idle_cpus == busiest->idle_cpus) &&
8831 (sgs->sum_nr_running <= busiest->sum_nr_running))
8838 * Candidate sg has no more than one task per CPU and has higher
8839 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
8840 * throughput. Maximize throughput, power/energy consequences are not
8843 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
8844 (sgs->group_type <= group_fully_busy) &&
8845 (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu))))
8851 #ifdef CONFIG_NUMA_BALANCING
8852 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8854 if (sgs->sum_h_nr_running > sgs->nr_numa_running)
8856 if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
8861 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8863 if (rq->nr_running > rq->nr_numa_running)
8865 if (rq->nr_running > rq->nr_preferred_running)
8870 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8875 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8879 #endif /* CONFIG_NUMA_BALANCING */
8885 * task_running_on_cpu - return 1 if @p is running on @cpu.
8888 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
8890 /* Task has no contribution or is new */
8891 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
8894 if (task_on_rq_queued(p))
8901 * idle_cpu_without - would a given CPU be idle without p ?
8902 * @cpu: the processor on which idleness is tested.
8903 * @p: task which should be ignored.
8905 * Return: 1 if the CPU would be idle. 0 otherwise.
8907 static int idle_cpu_without(int cpu, struct task_struct *p)
8909 struct rq *rq = cpu_rq(cpu);
8911 if (rq->curr != rq->idle && rq->curr != p)
8915 * rq->nr_running can't be used but an updated version without the
8916 * impact of p on cpu must be used instead. The updated nr_running
8917 * be computed and tested before calling idle_cpu_without().
8921 if (rq->ttwu_pending)
8929 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
8930 * @sd: The sched_domain level to look for idlest group.
8931 * @group: sched_group whose statistics are to be updated.
8932 * @sgs: variable to hold the statistics for this group.
8933 * @p: The task for which we look for the idlest group/CPU.
8935 static inline void update_sg_wakeup_stats(struct sched_domain *sd,
8936 struct sched_group *group,
8937 struct sg_lb_stats *sgs,
8938 struct task_struct *p)
8942 memset(sgs, 0, sizeof(*sgs));
8944 for_each_cpu(i, sched_group_span(group)) {
8945 struct rq *rq = cpu_rq(i);
8948 sgs->group_load += cpu_load_without(rq, p);
8949 sgs->group_util += cpu_util_without(i, p);
8950 sgs->group_runnable += cpu_runnable_without(rq, p);
8951 local = task_running_on_cpu(i, p);
8952 sgs->sum_h_nr_running += rq->cfs.h_nr_running - local;
8954 nr_running = rq->nr_running - local;
8955 sgs->sum_nr_running += nr_running;
8958 * No need to call idle_cpu_without() if nr_running is not 0
8960 if (!nr_running && idle_cpu_without(i, p))
8965 /* Check if task fits in the group */
8966 if (sd->flags & SD_ASYM_CPUCAPACITY &&
8967 !task_fits_capacity(p, group->sgc->max_capacity)) {
8968 sgs->group_misfit_task_load = 1;
8971 sgs->group_capacity = group->sgc->capacity;
8973 sgs->group_weight = group->group_weight;
8975 sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
8978 * Computing avg_load makes sense only when group is fully busy or
8981 if (sgs->group_type == group_fully_busy ||
8982 sgs->group_type == group_overloaded)
8983 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
8984 sgs->group_capacity;
8987 static bool update_pick_idlest(struct sched_group *idlest,
8988 struct sg_lb_stats *idlest_sgs,
8989 struct sched_group *group,
8990 struct sg_lb_stats *sgs)
8992 if (sgs->group_type < idlest_sgs->group_type)
8995 if (sgs->group_type > idlest_sgs->group_type)
8999 * The candidate and the current idlest group are the same type of
9000 * group. Let check which one is the idlest according to the type.
9003 switch (sgs->group_type) {
9004 case group_overloaded:
9005 case group_fully_busy:
9006 /* Select the group with lowest avg_load. */
9007 if (idlest_sgs->avg_load <= sgs->avg_load)
9011 case group_imbalanced:
9012 case group_asym_packing:
9013 /* Those types are not used in the slow wakeup path */
9016 case group_misfit_task:
9017 /* Select group with the highest max capacity */
9018 if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
9022 case group_has_spare:
9023 /* Select group with most idle CPUs */
9024 if (idlest_sgs->idle_cpus > sgs->idle_cpus)
9027 /* Select group with lowest group_util */
9028 if (idlest_sgs->idle_cpus == sgs->idle_cpus &&
9029 idlest_sgs->group_util <= sgs->group_util)
9039 * Allow a NUMA imbalance if busy CPUs is less than 25% of the domain.
9040 * This is an approximation as the number of running tasks may not be
9041 * related to the number of busy CPUs due to sched_setaffinity.
9044 allow_numa_imbalance(unsigned int running, unsigned int weight)
9046 return (running < (weight >> 2));
9050 * find_idlest_group() finds and returns the least busy CPU group within the
9053 * Assumes p is allowed on at least one CPU in sd.
9055 static struct sched_group *
9056 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
9058 struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
9059 struct sg_lb_stats local_sgs, tmp_sgs;
9060 struct sg_lb_stats *sgs;
9061 unsigned long imbalance;
9062 struct sg_lb_stats idlest_sgs = {
9063 .avg_load = UINT_MAX,
9064 .group_type = group_overloaded,
9070 /* Skip over this group if it has no CPUs allowed */
9071 if (!cpumask_intersects(sched_group_span(group),
9075 /* Skip over this group if no cookie matched */
9076 if (!sched_group_cookie_match(cpu_rq(this_cpu), p, group))
9079 local_group = cpumask_test_cpu(this_cpu,
9080 sched_group_span(group));
9089 update_sg_wakeup_stats(sd, group, sgs, p);
9091 if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
9096 } while (group = group->next, group != sd->groups);
9099 /* There is no idlest group to push tasks to */
9103 /* The local group has been skipped because of CPU affinity */
9108 * If the local group is idler than the selected idlest group
9109 * don't try and push the task.
9111 if (local_sgs.group_type < idlest_sgs.group_type)
9115 * If the local group is busier than the selected idlest group
9116 * try and push the task.
9118 if (local_sgs.group_type > idlest_sgs.group_type)
9121 switch (local_sgs.group_type) {
9122 case group_overloaded:
9123 case group_fully_busy:
9125 /* Calculate allowed imbalance based on load */
9126 imbalance = scale_load_down(NICE_0_LOAD) *
9127 (sd->imbalance_pct-100) / 100;
9130 * When comparing groups across NUMA domains, it's possible for
9131 * the local domain to be very lightly loaded relative to the
9132 * remote domains but "imbalance" skews the comparison making
9133 * remote CPUs look much more favourable. When considering
9134 * cross-domain, add imbalance to the load on the remote node
9135 * and consider staying local.
9138 if ((sd->flags & SD_NUMA) &&
9139 ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
9143 * If the local group is less loaded than the selected
9144 * idlest group don't try and push any tasks.
9146 if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
9149 if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
9153 case group_imbalanced:
9154 case group_asym_packing:
9155 /* Those type are not used in the slow wakeup path */
9158 case group_misfit_task:
9159 /* Select group with the highest max capacity */
9160 if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
9164 case group_has_spare:
9165 if (sd->flags & SD_NUMA) {
9166 #ifdef CONFIG_NUMA_BALANCING
9169 * If there is spare capacity at NUMA, try to select
9170 * the preferred node
9172 if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
9175 idlest_cpu = cpumask_first(sched_group_span(idlest));
9176 if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
9180 * Otherwise, keep the task close to the wakeup source
9181 * and improve locality if the number of running tasks
9182 * would remain below threshold where an imbalance is
9183 * allowed. If there is a real need of migration,
9184 * periodic load balance will take care of it.
9186 if (allow_numa_imbalance(local_sgs.sum_nr_running + 1, local_sgs.group_weight))
9191 * Select group with highest number of idle CPUs. We could also
9192 * compare the utilization which is more stable but it can end
9193 * up that the group has less spare capacity but finally more
9194 * idle CPUs which means more opportunity to run task.
9196 if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
9205 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
9206 * @env: The load balancing environment.
9207 * @sds: variable to hold the statistics for this sched_domain.
9210 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
9212 struct sched_domain *child = env->sd->child;
9213 struct sched_group *sg = env->sd->groups;
9214 struct sg_lb_stats *local = &sds->local_stat;
9215 struct sg_lb_stats tmp_sgs;
9219 struct sg_lb_stats *sgs = &tmp_sgs;
9222 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
9227 if (env->idle != CPU_NEWLY_IDLE ||
9228 time_after_eq(jiffies, sg->sgc->next_update))
9229 update_group_capacity(env->sd, env->dst_cpu);
9232 update_sg_lb_stats(env, sds, sg, sgs, &sg_status);
9238 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
9240 sds->busiest_stat = *sgs;
9244 /* Now, start updating sd_lb_stats */
9245 sds->total_load += sgs->group_load;
9246 sds->total_capacity += sgs->group_capacity;
9249 } while (sg != env->sd->groups);
9251 /* Tag domain that child domain prefers tasks go to siblings first */
9252 sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
9255 if (env->sd->flags & SD_NUMA)
9256 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
9258 if (!env->sd->parent) {
9259 struct root_domain *rd = env->dst_rq->rd;
9261 /* update overload indicator if we are at root domain */
9262 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
9264 /* Update over-utilization (tipping point, U >= 0) indicator */
9265 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
9266 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
9267 } else if (sg_status & SG_OVERUTILIZED) {
9268 struct root_domain *rd = env->dst_rq->rd;
9270 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
9271 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
9275 #define NUMA_IMBALANCE_MIN 2
9277 static inline long adjust_numa_imbalance(int imbalance,
9278 int dst_running, int dst_weight)
9280 if (!allow_numa_imbalance(dst_running, dst_weight))
9284 * Allow a small imbalance based on a simple pair of communicating
9285 * tasks that remain local when the destination is lightly loaded.
9287 if (imbalance <= NUMA_IMBALANCE_MIN)
9294 * calculate_imbalance - Calculate the amount of imbalance present within the
9295 * groups of a given sched_domain during load balance.
9296 * @env: load balance environment
9297 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
9299 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
9301 struct sg_lb_stats *local, *busiest;
9303 local = &sds->local_stat;
9304 busiest = &sds->busiest_stat;
9306 if (busiest->group_type == group_misfit_task) {
9307 /* Set imbalance to allow misfit tasks to be balanced. */
9308 env->migration_type = migrate_misfit;
9313 if (busiest->group_type == group_asym_packing) {
9315 * In case of asym capacity, we will try to migrate all load to
9316 * the preferred CPU.
9318 env->migration_type = migrate_task;
9319 env->imbalance = busiest->sum_h_nr_running;
9323 if (busiest->group_type == group_imbalanced) {
9325 * In the group_imb case we cannot rely on group-wide averages
9326 * to ensure CPU-load equilibrium, try to move any task to fix
9327 * the imbalance. The next load balance will take care of
9328 * balancing back the system.
9330 env->migration_type = migrate_task;
9336 * Try to use spare capacity of local group without overloading it or
9339 if (local->group_type == group_has_spare) {
9340 if ((busiest->group_type > group_fully_busy) &&
9341 !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) {
9343 * If busiest is overloaded, try to fill spare
9344 * capacity. This might end up creating spare capacity
9345 * in busiest or busiest still being overloaded but
9346 * there is no simple way to directly compute the
9347 * amount of load to migrate in order to balance the
9350 env->migration_type = migrate_util;
9351 env->imbalance = max(local->group_capacity, local->group_util) -
9355 * In some cases, the group's utilization is max or even
9356 * higher than capacity because of migrations but the
9357 * local CPU is (newly) idle. There is at least one
9358 * waiting task in this overloaded busiest group. Let's
9361 if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) {
9362 env->migration_type = migrate_task;
9369 if (busiest->group_weight == 1 || sds->prefer_sibling) {
9370 unsigned int nr_diff = busiest->sum_nr_running;
9372 * When prefer sibling, evenly spread running tasks on
9375 env->migration_type = migrate_task;
9376 lsub_positive(&nr_diff, local->sum_nr_running);
9377 env->imbalance = nr_diff >> 1;
9381 * If there is no overload, we just want to even the number of
9384 env->migration_type = migrate_task;
9385 env->imbalance = max_t(long, 0, (local->idle_cpus -
9386 busiest->idle_cpus) >> 1);
9389 /* Consider allowing a small imbalance between NUMA groups */
9390 if (env->sd->flags & SD_NUMA) {
9391 env->imbalance = adjust_numa_imbalance(env->imbalance,
9392 local->sum_nr_running + 1, local->group_weight);
9399 * Local is fully busy but has to take more load to relieve the
9402 if (local->group_type < group_overloaded) {
9404 * Local will become overloaded so the avg_load metrics are
9408 local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
9409 local->group_capacity;
9411 sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
9412 sds->total_capacity;
9414 * If the local group is more loaded than the selected
9415 * busiest group don't try to pull any tasks.
9417 if (local->avg_load >= busiest->avg_load) {
9424 * Both group are or will become overloaded and we're trying to get all
9425 * the CPUs to the average_load, so we don't want to push ourselves
9426 * above the average load, nor do we wish to reduce the max loaded CPU
9427 * below the average load. At the same time, we also don't want to
9428 * reduce the group load below the group capacity. Thus we look for
9429 * the minimum possible imbalance.
9431 env->migration_type = migrate_load;
9432 env->imbalance = min(
9433 (busiest->avg_load - sds->avg_load) * busiest->group_capacity,
9434 (sds->avg_load - local->avg_load) * local->group_capacity
9435 ) / SCHED_CAPACITY_SCALE;
9438 /******* find_busiest_group() helpers end here *********************/
9441 * Decision matrix according to the local and busiest group type:
9443 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
9444 * has_spare nr_idle balanced N/A N/A balanced balanced
9445 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
9446 * misfit_task force N/A N/A N/A force force
9447 * asym_packing force force N/A N/A force force
9448 * imbalanced force force N/A N/A force force
9449 * overloaded force force N/A N/A force avg_load
9451 * N/A : Not Applicable because already filtered while updating
9453 * balanced : The system is balanced for these 2 groups.
9454 * force : Calculate the imbalance as load migration is probably needed.
9455 * avg_load : Only if imbalance is significant enough.
9456 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
9457 * different in groups.
9461 * find_busiest_group - Returns the busiest group within the sched_domain
9462 * if there is an imbalance.
9463 * @env: The load balancing environment.
9465 * Also calculates the amount of runnable load which should be moved
9466 * to restore balance.
9468 * Return: - The busiest group if imbalance exists.
9470 static struct sched_group *find_busiest_group(struct lb_env *env)
9472 struct sg_lb_stats *local, *busiest;
9473 struct sd_lb_stats sds;
9475 init_sd_lb_stats(&sds);
9478 * Compute the various statistics relevant for load balancing at
9481 update_sd_lb_stats(env, &sds);
9483 if (sched_energy_enabled()) {
9484 struct root_domain *rd = env->dst_rq->rd;
9486 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
9490 local = &sds.local_stat;
9491 busiest = &sds.busiest_stat;
9493 /* There is no busy sibling group to pull tasks from */
9497 /* Misfit tasks should be dealt with regardless of the avg load */
9498 if (busiest->group_type == group_misfit_task)
9501 /* ASYM feature bypasses nice load balance check */
9502 if (busiest->group_type == group_asym_packing)
9506 * If the busiest group is imbalanced the below checks don't
9507 * work because they assume all things are equal, which typically
9508 * isn't true due to cpus_ptr constraints and the like.
9510 if (busiest->group_type == group_imbalanced)
9514 * If the local group is busier than the selected busiest group
9515 * don't try and pull any tasks.
9517 if (local->group_type > busiest->group_type)
9521 * When groups are overloaded, use the avg_load to ensure fairness
9524 if (local->group_type == group_overloaded) {
9526 * If the local group is more loaded than the selected
9527 * busiest group don't try to pull any tasks.
9529 if (local->avg_load >= busiest->avg_load)
9532 /* XXX broken for overlapping NUMA groups */
9533 sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
9537 * Don't pull any tasks if this group is already above the
9538 * domain average load.
9540 if (local->avg_load >= sds.avg_load)
9544 * If the busiest group is more loaded, use imbalance_pct to be
9547 if (100 * busiest->avg_load <=
9548 env->sd->imbalance_pct * local->avg_load)
9552 /* Try to move all excess tasks to child's sibling domain */
9553 if (sds.prefer_sibling && local->group_type == group_has_spare &&
9554 busiest->sum_nr_running > local->sum_nr_running + 1)
9557 if (busiest->group_type != group_overloaded) {
9558 if (env->idle == CPU_NOT_IDLE)
9560 * If the busiest group is not overloaded (and as a
9561 * result the local one too) but this CPU is already
9562 * busy, let another idle CPU try to pull task.
9566 if (busiest->group_weight > 1 &&
9567 local->idle_cpus <= (busiest->idle_cpus + 1))
9569 * If the busiest group is not overloaded
9570 * and there is no imbalance between this and busiest
9571 * group wrt idle CPUs, it is balanced. The imbalance
9572 * becomes significant if the diff is greater than 1
9573 * otherwise we might end up to just move the imbalance
9574 * on another group. Of course this applies only if
9575 * there is more than 1 CPU per group.
9579 if (busiest->sum_h_nr_running == 1)
9581 * busiest doesn't have any tasks waiting to run
9587 /* Looks like there is an imbalance. Compute it */
9588 calculate_imbalance(env, &sds);
9589 return env->imbalance ? sds.busiest : NULL;
9597 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
9599 static struct rq *find_busiest_queue(struct lb_env *env,
9600 struct sched_group *group)
9602 struct rq *busiest = NULL, *rq;
9603 unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
9604 unsigned int busiest_nr = 0;
9607 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
9608 unsigned long capacity, load, util;
9609 unsigned int nr_running;
9613 rt = fbq_classify_rq(rq);
9616 * We classify groups/runqueues into three groups:
9617 * - regular: there are !numa tasks
9618 * - remote: there are numa tasks that run on the 'wrong' node
9619 * - all: there is no distinction
9621 * In order to avoid migrating ideally placed numa tasks,
9622 * ignore those when there's better options.
9624 * If we ignore the actual busiest queue to migrate another
9625 * task, the next balance pass can still reduce the busiest
9626 * queue by moving tasks around inside the node.
9628 * If we cannot move enough load due to this classification
9629 * the next pass will adjust the group classification and
9630 * allow migration of more tasks.
9632 * Both cases only affect the total convergence complexity.
9634 if (rt > env->fbq_type)
9637 nr_running = rq->cfs.h_nr_running;
9641 capacity = capacity_of(i);
9644 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
9645 * eventually lead to active_balancing high->low capacity.
9646 * Higher per-CPU capacity is considered better than balancing
9649 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
9650 !capacity_greater(capacity_of(env->dst_cpu), capacity) &&
9654 /* Make sure we only pull tasks from a CPU of lower priority */
9655 if ((env->sd->flags & SD_ASYM_PACKING) &&
9656 sched_asym_prefer(i, env->dst_cpu) &&
9660 switch (env->migration_type) {
9663 * When comparing with load imbalance, use cpu_load()
9664 * which is not scaled with the CPU capacity.
9666 load = cpu_load(rq);
9668 if (nr_running == 1 && load > env->imbalance &&
9669 !check_cpu_capacity(rq, env->sd))
9673 * For the load comparisons with the other CPUs,
9674 * consider the cpu_load() scaled with the CPU
9675 * capacity, so that the load can be moved away
9676 * from the CPU that is potentially running at a
9679 * Thus we're looking for max(load_i / capacity_i),
9680 * crosswise multiplication to rid ourselves of the
9681 * division works out to:
9682 * load_i * capacity_j > load_j * capacity_i;
9683 * where j is our previous maximum.
9685 if (load * busiest_capacity > busiest_load * capacity) {
9686 busiest_load = load;
9687 busiest_capacity = capacity;
9693 util = cpu_util_cfs(i);
9696 * Don't try to pull utilization from a CPU with one
9697 * running task. Whatever its utilization, we will fail
9700 if (nr_running <= 1)
9703 if (busiest_util < util) {
9704 busiest_util = util;
9710 if (busiest_nr < nr_running) {
9711 busiest_nr = nr_running;
9716 case migrate_misfit:
9718 * For ASYM_CPUCAPACITY domains with misfit tasks we
9719 * simply seek the "biggest" misfit task.
9721 if (rq->misfit_task_load > busiest_load) {
9722 busiest_load = rq->misfit_task_load;
9735 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
9736 * so long as it is large enough.
9738 #define MAX_PINNED_INTERVAL 512
9741 asym_active_balance(struct lb_env *env)
9744 * ASYM_PACKING needs to force migrate tasks from busy but
9745 * lower priority CPUs in order to pack all tasks in the
9746 * highest priority CPUs.
9748 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
9749 sched_asym_prefer(env->dst_cpu, env->src_cpu);
9753 imbalanced_active_balance(struct lb_env *env)
9755 struct sched_domain *sd = env->sd;
9758 * The imbalanced case includes the case of pinned tasks preventing a fair
9759 * distribution of the load on the system but also the even distribution of the
9760 * threads on a system with spare capacity
9762 if ((env->migration_type == migrate_task) &&
9763 (sd->nr_balance_failed > sd->cache_nice_tries+2))
9769 static int need_active_balance(struct lb_env *env)
9771 struct sched_domain *sd = env->sd;
9773 if (asym_active_balance(env))
9776 if (imbalanced_active_balance(env))
9780 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
9781 * It's worth migrating the task if the src_cpu's capacity is reduced
9782 * because of other sched_class or IRQs if more capacity stays
9783 * available on dst_cpu.
9785 if ((env->idle != CPU_NOT_IDLE) &&
9786 (env->src_rq->cfs.h_nr_running == 1)) {
9787 if ((check_cpu_capacity(env->src_rq, sd)) &&
9788 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
9792 if (env->migration_type == migrate_misfit)
9798 static int active_load_balance_cpu_stop(void *data);
9800 static int should_we_balance(struct lb_env *env)
9802 struct sched_group *sg = env->sd->groups;
9806 * Ensure the balancing environment is consistent; can happen
9807 * when the softirq triggers 'during' hotplug.
9809 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
9813 * In the newly idle case, we will allow all the CPUs
9814 * to do the newly idle load balance.
9816 if (env->idle == CPU_NEWLY_IDLE)
9819 /* Try to find first idle CPU */
9820 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
9824 /* Are we the first idle CPU? */
9825 return cpu == env->dst_cpu;
9828 /* Are we the first CPU of this group ? */
9829 return group_balance_cpu(sg) == env->dst_cpu;
9833 * Check this_cpu to ensure it is balanced within domain. Attempt to move
9834 * tasks if there is an imbalance.
9836 static int load_balance(int this_cpu, struct rq *this_rq,
9837 struct sched_domain *sd, enum cpu_idle_type idle,
9838 int *continue_balancing)
9840 int ld_moved, cur_ld_moved, active_balance = 0;
9841 struct sched_domain *sd_parent = sd->parent;
9842 struct sched_group *group;
9845 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
9847 struct lb_env env = {
9849 .dst_cpu = this_cpu,
9851 .dst_grpmask = sched_group_span(sd->groups),
9853 .loop_break = sched_nr_migrate_break,
9856 .tasks = LIST_HEAD_INIT(env.tasks),
9859 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
9861 schedstat_inc(sd->lb_count[idle]);
9864 if (!should_we_balance(&env)) {
9865 *continue_balancing = 0;
9869 group = find_busiest_group(&env);
9871 schedstat_inc(sd->lb_nobusyg[idle]);
9875 busiest = find_busiest_queue(&env, group);
9877 schedstat_inc(sd->lb_nobusyq[idle]);
9881 BUG_ON(busiest == env.dst_rq);
9883 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
9885 env.src_cpu = busiest->cpu;
9886 env.src_rq = busiest;
9889 /* Clear this flag as soon as we find a pullable task */
9890 env.flags |= LBF_ALL_PINNED;
9891 if (busiest->nr_running > 1) {
9893 * Attempt to move tasks. If find_busiest_group has found
9894 * an imbalance but busiest->nr_running <= 1, the group is
9895 * still unbalanced. ld_moved simply stays zero, so it is
9896 * correctly treated as an imbalance.
9898 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
9901 rq_lock_irqsave(busiest, &rf);
9902 update_rq_clock(busiest);
9905 * cur_ld_moved - load moved in current iteration
9906 * ld_moved - cumulative load moved across iterations
9908 cur_ld_moved = detach_tasks(&env);
9911 * We've detached some tasks from busiest_rq. Every
9912 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
9913 * unlock busiest->lock, and we are able to be sure
9914 * that nobody can manipulate the tasks in parallel.
9915 * See task_rq_lock() family for the details.
9918 rq_unlock(busiest, &rf);
9922 ld_moved += cur_ld_moved;
9925 local_irq_restore(rf.flags);
9927 if (env.flags & LBF_NEED_BREAK) {
9928 env.flags &= ~LBF_NEED_BREAK;
9933 * Revisit (affine) tasks on src_cpu that couldn't be moved to
9934 * us and move them to an alternate dst_cpu in our sched_group
9935 * where they can run. The upper limit on how many times we
9936 * iterate on same src_cpu is dependent on number of CPUs in our
9939 * This changes load balance semantics a bit on who can move
9940 * load to a given_cpu. In addition to the given_cpu itself
9941 * (or a ilb_cpu acting on its behalf where given_cpu is
9942 * nohz-idle), we now have balance_cpu in a position to move
9943 * load to given_cpu. In rare situations, this may cause
9944 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
9945 * _independently_ and at _same_ time to move some load to
9946 * given_cpu) causing excess load to be moved to given_cpu.
9947 * This however should not happen so much in practice and
9948 * moreover subsequent load balance cycles should correct the
9949 * excess load moved.
9951 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
9953 /* Prevent to re-select dst_cpu via env's CPUs */
9954 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
9956 env.dst_rq = cpu_rq(env.new_dst_cpu);
9957 env.dst_cpu = env.new_dst_cpu;
9958 env.flags &= ~LBF_DST_PINNED;
9960 env.loop_break = sched_nr_migrate_break;
9963 * Go back to "more_balance" rather than "redo" since we
9964 * need to continue with same src_cpu.
9970 * We failed to reach balance because of affinity.
9973 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9975 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
9976 *group_imbalance = 1;
9979 /* All tasks on this runqueue were pinned by CPU affinity */
9980 if (unlikely(env.flags & LBF_ALL_PINNED)) {
9981 __cpumask_clear_cpu(cpu_of(busiest), cpus);
9983 * Attempting to continue load balancing at the current
9984 * sched_domain level only makes sense if there are
9985 * active CPUs remaining as possible busiest CPUs to
9986 * pull load from which are not contained within the
9987 * destination group that is receiving any migrated
9990 if (!cpumask_subset(cpus, env.dst_grpmask)) {
9992 env.loop_break = sched_nr_migrate_break;
9995 goto out_all_pinned;
10000 schedstat_inc(sd->lb_failed[idle]);
10002 * Increment the failure counter only on periodic balance.
10003 * We do not want newidle balance, which can be very
10004 * frequent, pollute the failure counter causing
10005 * excessive cache_hot migrations and active balances.
10007 if (idle != CPU_NEWLY_IDLE)
10008 sd->nr_balance_failed++;
10010 if (need_active_balance(&env)) {
10011 unsigned long flags;
10013 raw_spin_rq_lock_irqsave(busiest, flags);
10016 * Don't kick the active_load_balance_cpu_stop,
10017 * if the curr task on busiest CPU can't be
10018 * moved to this_cpu:
10020 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
10021 raw_spin_rq_unlock_irqrestore(busiest, flags);
10022 goto out_one_pinned;
10025 /* Record that we found at least one task that could run on this_cpu */
10026 env.flags &= ~LBF_ALL_PINNED;
10029 * ->active_balance synchronizes accesses to
10030 * ->active_balance_work. Once set, it's cleared
10031 * only after active load balance is finished.
10033 if (!busiest->active_balance) {
10034 busiest->active_balance = 1;
10035 busiest->push_cpu = this_cpu;
10036 active_balance = 1;
10038 raw_spin_rq_unlock_irqrestore(busiest, flags);
10040 if (active_balance) {
10041 stop_one_cpu_nowait(cpu_of(busiest),
10042 active_load_balance_cpu_stop, busiest,
10043 &busiest->active_balance_work);
10047 sd->nr_balance_failed = 0;
10050 if (likely(!active_balance) || need_active_balance(&env)) {
10051 /* We were unbalanced, so reset the balancing interval */
10052 sd->balance_interval = sd->min_interval;
10059 * We reach balance although we may have faced some affinity
10060 * constraints. Clear the imbalance flag only if other tasks got
10061 * a chance to move and fix the imbalance.
10063 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
10064 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
10066 if (*group_imbalance)
10067 *group_imbalance = 0;
10072 * We reach balance because all tasks are pinned at this level so
10073 * we can't migrate them. Let the imbalance flag set so parent level
10074 * can try to migrate them.
10076 schedstat_inc(sd->lb_balanced[idle]);
10078 sd->nr_balance_failed = 0;
10084 * newidle_balance() disregards balance intervals, so we could
10085 * repeatedly reach this code, which would lead to balance_interval
10086 * skyrocketing in a short amount of time. Skip the balance_interval
10087 * increase logic to avoid that.
10089 if (env.idle == CPU_NEWLY_IDLE)
10092 /* tune up the balancing interval */
10093 if ((env.flags & LBF_ALL_PINNED &&
10094 sd->balance_interval < MAX_PINNED_INTERVAL) ||
10095 sd->balance_interval < sd->max_interval)
10096 sd->balance_interval *= 2;
10101 static inline unsigned long
10102 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
10104 unsigned long interval = sd->balance_interval;
10107 interval *= sd->busy_factor;
10109 /* scale ms to jiffies */
10110 interval = msecs_to_jiffies(interval);
10113 * Reduce likelihood of busy balancing at higher domains racing with
10114 * balancing at lower domains by preventing their balancing periods
10115 * from being multiples of each other.
10120 interval = clamp(interval, 1UL, max_load_balance_interval);
10126 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
10128 unsigned long interval, next;
10130 /* used by idle balance, so cpu_busy = 0 */
10131 interval = get_sd_balance_interval(sd, 0);
10132 next = sd->last_balance + interval;
10134 if (time_after(*next_balance, next))
10135 *next_balance = next;
10139 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
10140 * running tasks off the busiest CPU onto idle CPUs. It requires at
10141 * least 1 task to be running on each physical CPU where possible, and
10142 * avoids physical / logical imbalances.
10144 static int active_load_balance_cpu_stop(void *data)
10146 struct rq *busiest_rq = data;
10147 int busiest_cpu = cpu_of(busiest_rq);
10148 int target_cpu = busiest_rq->push_cpu;
10149 struct rq *target_rq = cpu_rq(target_cpu);
10150 struct sched_domain *sd;
10151 struct task_struct *p = NULL;
10152 struct rq_flags rf;
10154 rq_lock_irq(busiest_rq, &rf);
10156 * Between queueing the stop-work and running it is a hole in which
10157 * CPUs can become inactive. We should not move tasks from or to
10160 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
10163 /* Make sure the requested CPU hasn't gone down in the meantime: */
10164 if (unlikely(busiest_cpu != smp_processor_id() ||
10165 !busiest_rq->active_balance))
10168 /* Is there any task to move? */
10169 if (busiest_rq->nr_running <= 1)
10173 * This condition is "impossible", if it occurs
10174 * we need to fix it. Originally reported by
10175 * Bjorn Helgaas on a 128-CPU setup.
10177 BUG_ON(busiest_rq == target_rq);
10179 /* Search for an sd spanning us and the target CPU. */
10181 for_each_domain(target_cpu, sd) {
10182 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
10187 struct lb_env env = {
10189 .dst_cpu = target_cpu,
10190 .dst_rq = target_rq,
10191 .src_cpu = busiest_rq->cpu,
10192 .src_rq = busiest_rq,
10194 .flags = LBF_ACTIVE_LB,
10197 schedstat_inc(sd->alb_count);
10198 update_rq_clock(busiest_rq);
10200 p = detach_one_task(&env);
10202 schedstat_inc(sd->alb_pushed);
10203 /* Active balancing done, reset the failure counter. */
10204 sd->nr_balance_failed = 0;
10206 schedstat_inc(sd->alb_failed);
10211 busiest_rq->active_balance = 0;
10212 rq_unlock(busiest_rq, &rf);
10215 attach_one_task(target_rq, p);
10217 local_irq_enable();
10222 static DEFINE_SPINLOCK(balancing);
10225 * Scale the max load_balance interval with the number of CPUs in the system.
10226 * This trades load-balance latency on larger machines for less cross talk.
10228 void update_max_interval(void)
10230 max_load_balance_interval = HZ*num_online_cpus()/10;
10233 static inline bool update_newidle_cost(struct sched_domain *sd, u64 cost)
10235 if (cost > sd->max_newidle_lb_cost) {
10237 * Track max cost of a domain to make sure to not delay the
10238 * next wakeup on the CPU.
10240 sd->max_newidle_lb_cost = cost;
10241 sd->last_decay_max_lb_cost = jiffies;
10242 } else if (time_after(jiffies, sd->last_decay_max_lb_cost + HZ)) {
10244 * Decay the newidle max times by ~1% per second to ensure that
10245 * it is not outdated and the current max cost is actually
10248 sd->max_newidle_lb_cost = (sd->max_newidle_lb_cost * 253) / 256;
10249 sd->last_decay_max_lb_cost = jiffies;
10258 * It checks each scheduling domain to see if it is due to be balanced,
10259 * and initiates a balancing operation if so.
10261 * Balancing parameters are set up in init_sched_domains.
10263 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
10265 int continue_balancing = 1;
10267 int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
10268 unsigned long interval;
10269 struct sched_domain *sd;
10270 /* Earliest time when we have to do rebalance again */
10271 unsigned long next_balance = jiffies + 60*HZ;
10272 int update_next_balance = 0;
10273 int need_serialize, need_decay = 0;
10277 for_each_domain(cpu, sd) {
10279 * Decay the newidle max times here because this is a regular
10280 * visit to all the domains.
10282 need_decay = update_newidle_cost(sd, 0);
10283 max_cost += sd->max_newidle_lb_cost;
10286 * Stop the load balance at this level. There is another
10287 * CPU in our sched group which is doing load balancing more
10290 if (!continue_balancing) {
10296 interval = get_sd_balance_interval(sd, busy);
10298 need_serialize = sd->flags & SD_SERIALIZE;
10299 if (need_serialize) {
10300 if (!spin_trylock(&balancing))
10304 if (time_after_eq(jiffies, sd->last_balance + interval)) {
10305 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
10307 * The LBF_DST_PINNED logic could have changed
10308 * env->dst_cpu, so we can't know our idle
10309 * state even if we migrated tasks. Update it.
10311 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
10312 busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
10314 sd->last_balance = jiffies;
10315 interval = get_sd_balance_interval(sd, busy);
10317 if (need_serialize)
10318 spin_unlock(&balancing);
10320 if (time_after(next_balance, sd->last_balance + interval)) {
10321 next_balance = sd->last_balance + interval;
10322 update_next_balance = 1;
10327 * Ensure the rq-wide value also decays but keep it at a
10328 * reasonable floor to avoid funnies with rq->avg_idle.
10330 rq->max_idle_balance_cost =
10331 max((u64)sysctl_sched_migration_cost, max_cost);
10336 * next_balance will be updated only when there is a need.
10337 * When the cpu is attached to null domain for ex, it will not be
10340 if (likely(update_next_balance))
10341 rq->next_balance = next_balance;
10345 static inline int on_null_domain(struct rq *rq)
10347 return unlikely(!rcu_dereference_sched(rq->sd));
10350 #ifdef CONFIG_NO_HZ_COMMON
10352 * idle load balancing details
10353 * - When one of the busy CPUs notice that there may be an idle rebalancing
10354 * needed, they will kick the idle load balancer, which then does idle
10355 * load balancing for all the idle CPUs.
10356 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
10360 static inline int find_new_ilb(void)
10363 const struct cpumask *hk_mask;
10365 hk_mask = housekeeping_cpumask(HK_FLAG_MISC);
10367 for_each_cpu_and(ilb, nohz.idle_cpus_mask, hk_mask) {
10369 if (ilb == smp_processor_id())
10380 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
10381 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
10383 static void kick_ilb(unsigned int flags)
10388 * Increase nohz.next_balance only when if full ilb is triggered but
10389 * not if we only update stats.
10391 if (flags & NOHZ_BALANCE_KICK)
10392 nohz.next_balance = jiffies+1;
10394 ilb_cpu = find_new_ilb();
10396 if (ilb_cpu >= nr_cpu_ids)
10400 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
10401 * the first flag owns it; cleared by nohz_csd_func().
10403 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
10404 if (flags & NOHZ_KICK_MASK)
10408 * This way we generate an IPI on the target CPU which
10409 * is idle. And the softirq performing nohz idle load balance
10410 * will be run before returning from the IPI.
10412 smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd);
10416 * Current decision point for kicking the idle load balancer in the presence
10417 * of idle CPUs in the system.
10419 static void nohz_balancer_kick(struct rq *rq)
10421 unsigned long now = jiffies;
10422 struct sched_domain_shared *sds;
10423 struct sched_domain *sd;
10424 int nr_busy, i, cpu = rq->cpu;
10425 unsigned int flags = 0;
10427 if (unlikely(rq->idle_balance))
10431 * We may be recently in ticked or tickless idle mode. At the first
10432 * busy tick after returning from idle, we will update the busy stats.
10434 nohz_balance_exit_idle(rq);
10437 * None are in tickless mode and hence no need for NOHZ idle load
10440 if (likely(!atomic_read(&nohz.nr_cpus)))
10443 if (READ_ONCE(nohz.has_blocked) &&
10444 time_after(now, READ_ONCE(nohz.next_blocked)))
10445 flags = NOHZ_STATS_KICK;
10447 if (time_before(now, nohz.next_balance))
10450 if (rq->nr_running >= 2) {
10451 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
10457 sd = rcu_dereference(rq->sd);
10460 * If there's a CFS task and the current CPU has reduced
10461 * capacity; kick the ILB to see if there's a better CPU to run
10464 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
10465 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
10470 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
10473 * When ASYM_PACKING; see if there's a more preferred CPU
10474 * currently idle; in which case, kick the ILB to move tasks
10477 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
10478 if (sched_asym_prefer(i, cpu)) {
10479 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
10485 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
10488 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
10489 * to run the misfit task on.
10491 if (check_misfit_status(rq, sd)) {
10492 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
10497 * For asymmetric systems, we do not want to nicely balance
10498 * cache use, instead we want to embrace asymmetry and only
10499 * ensure tasks have enough CPU capacity.
10501 * Skip the LLC logic because it's not relevant in that case.
10506 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
10509 * If there is an imbalance between LLC domains (IOW we could
10510 * increase the overall cache use), we need some less-loaded LLC
10511 * domain to pull some load. Likewise, we may need to spread
10512 * load within the current LLC domain (e.g. packed SMT cores but
10513 * other CPUs are idle). We can't really know from here how busy
10514 * the others are - so just get a nohz balance going if it looks
10515 * like this LLC domain has tasks we could move.
10517 nr_busy = atomic_read(&sds->nr_busy_cpus);
10519 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
10526 if (READ_ONCE(nohz.needs_update))
10527 flags |= NOHZ_NEXT_KICK;
10533 static void set_cpu_sd_state_busy(int cpu)
10535 struct sched_domain *sd;
10538 sd = rcu_dereference(per_cpu(sd_llc, cpu));
10540 if (!sd || !sd->nohz_idle)
10544 atomic_inc(&sd->shared->nr_busy_cpus);
10549 void nohz_balance_exit_idle(struct rq *rq)
10551 SCHED_WARN_ON(rq != this_rq());
10553 if (likely(!rq->nohz_tick_stopped))
10556 rq->nohz_tick_stopped = 0;
10557 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
10558 atomic_dec(&nohz.nr_cpus);
10560 set_cpu_sd_state_busy(rq->cpu);
10563 static void set_cpu_sd_state_idle(int cpu)
10565 struct sched_domain *sd;
10568 sd = rcu_dereference(per_cpu(sd_llc, cpu));
10570 if (!sd || sd->nohz_idle)
10574 atomic_dec(&sd->shared->nr_busy_cpus);
10580 * This routine will record that the CPU is going idle with tick stopped.
10581 * This info will be used in performing idle load balancing in the future.
10583 void nohz_balance_enter_idle(int cpu)
10585 struct rq *rq = cpu_rq(cpu);
10587 SCHED_WARN_ON(cpu != smp_processor_id());
10589 /* If this CPU is going down, then nothing needs to be done: */
10590 if (!cpu_active(cpu))
10593 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
10594 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
10598 * Can be set safely without rq->lock held
10599 * If a clear happens, it will have evaluated last additions because
10600 * rq->lock is held during the check and the clear
10602 rq->has_blocked_load = 1;
10605 * The tick is still stopped but load could have been added in the
10606 * meantime. We set the nohz.has_blocked flag to trig a check of the
10607 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
10608 * of nohz.has_blocked can only happen after checking the new load
10610 if (rq->nohz_tick_stopped)
10613 /* If we're a completely isolated CPU, we don't play: */
10614 if (on_null_domain(rq))
10617 rq->nohz_tick_stopped = 1;
10619 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
10620 atomic_inc(&nohz.nr_cpus);
10623 * Ensures that if nohz_idle_balance() fails to observe our
10624 * @idle_cpus_mask store, it must observe the @has_blocked
10625 * and @needs_update stores.
10627 smp_mb__after_atomic();
10629 set_cpu_sd_state_idle(cpu);
10631 WRITE_ONCE(nohz.needs_update, 1);
10634 * Each time a cpu enter idle, we assume that it has blocked load and
10635 * enable the periodic update of the load of idle cpus
10637 WRITE_ONCE(nohz.has_blocked, 1);
10640 static bool update_nohz_stats(struct rq *rq)
10642 unsigned int cpu = rq->cpu;
10644 if (!rq->has_blocked_load)
10647 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
10650 if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick)))
10653 update_blocked_averages(cpu);
10655 return rq->has_blocked_load;
10659 * Internal function that runs load balance for all idle cpus. The load balance
10660 * can be a simple update of blocked load or a complete load balance with
10661 * tasks movement depending of flags.
10663 static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
10664 enum cpu_idle_type idle)
10666 /* Earliest time when we have to do rebalance again */
10667 unsigned long now = jiffies;
10668 unsigned long next_balance = now + 60*HZ;
10669 bool has_blocked_load = false;
10670 int update_next_balance = 0;
10671 int this_cpu = this_rq->cpu;
10675 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
10678 * We assume there will be no idle load after this update and clear
10679 * the has_blocked flag. If a cpu enters idle in the mean time, it will
10680 * set the has_blocked flag and trigger another update of idle load.
10681 * Because a cpu that becomes idle, is added to idle_cpus_mask before
10682 * setting the flag, we are sure to not clear the state and not
10683 * check the load of an idle cpu.
10685 * Same applies to idle_cpus_mask vs needs_update.
10687 if (flags & NOHZ_STATS_KICK)
10688 WRITE_ONCE(nohz.has_blocked, 0);
10689 if (flags & NOHZ_NEXT_KICK)
10690 WRITE_ONCE(nohz.needs_update, 0);
10693 * Ensures that if we miss the CPU, we must see the has_blocked
10694 * store from nohz_balance_enter_idle().
10699 * Start with the next CPU after this_cpu so we will end with this_cpu and let a
10700 * chance for other idle cpu to pull load.
10702 for_each_cpu_wrap(balance_cpu, nohz.idle_cpus_mask, this_cpu+1) {
10703 if (!idle_cpu(balance_cpu))
10707 * If this CPU gets work to do, stop the load balancing
10708 * work being done for other CPUs. Next load
10709 * balancing owner will pick it up.
10711 if (need_resched()) {
10712 if (flags & NOHZ_STATS_KICK)
10713 has_blocked_load = true;
10714 if (flags & NOHZ_NEXT_KICK)
10715 WRITE_ONCE(nohz.needs_update, 1);
10719 rq = cpu_rq(balance_cpu);
10721 if (flags & NOHZ_STATS_KICK)
10722 has_blocked_load |= update_nohz_stats(rq);
10725 * If time for next balance is due,
10728 if (time_after_eq(jiffies, rq->next_balance)) {
10729 struct rq_flags rf;
10731 rq_lock_irqsave(rq, &rf);
10732 update_rq_clock(rq);
10733 rq_unlock_irqrestore(rq, &rf);
10735 if (flags & NOHZ_BALANCE_KICK)
10736 rebalance_domains(rq, CPU_IDLE);
10739 if (time_after(next_balance, rq->next_balance)) {
10740 next_balance = rq->next_balance;
10741 update_next_balance = 1;
10746 * next_balance will be updated only when there is a need.
10747 * When the CPU is attached to null domain for ex, it will not be
10750 if (likely(update_next_balance))
10751 nohz.next_balance = next_balance;
10753 if (flags & NOHZ_STATS_KICK)
10754 WRITE_ONCE(nohz.next_blocked,
10755 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
10758 /* There is still blocked load, enable periodic update */
10759 if (has_blocked_load)
10760 WRITE_ONCE(nohz.has_blocked, 1);
10764 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
10765 * rebalancing for all the cpus for whom scheduler ticks are stopped.
10767 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
10769 unsigned int flags = this_rq->nohz_idle_balance;
10774 this_rq->nohz_idle_balance = 0;
10776 if (idle != CPU_IDLE)
10779 _nohz_idle_balance(this_rq, flags, idle);
10785 * Check if we need to run the ILB for updating blocked load before entering
10788 void nohz_run_idle_balance(int cpu)
10790 unsigned int flags;
10792 flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu));
10795 * Update the blocked load only if no SCHED_SOFTIRQ is about to happen
10796 * (ie NOHZ_STATS_KICK set) and will do the same.
10798 if ((flags == NOHZ_NEWILB_KICK) && !need_resched())
10799 _nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK, CPU_IDLE);
10802 static void nohz_newidle_balance(struct rq *this_rq)
10804 int this_cpu = this_rq->cpu;
10807 * This CPU doesn't want to be disturbed by scheduler
10810 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
10813 /* Will wake up very soon. No time for doing anything else*/
10814 if (this_rq->avg_idle < sysctl_sched_migration_cost)
10817 /* Don't need to update blocked load of idle CPUs*/
10818 if (!READ_ONCE(nohz.has_blocked) ||
10819 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
10823 * Set the need to trigger ILB in order to update blocked load
10824 * before entering idle state.
10826 atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu));
10829 #else /* !CONFIG_NO_HZ_COMMON */
10830 static inline void nohz_balancer_kick(struct rq *rq) { }
10832 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
10837 static inline void nohz_newidle_balance(struct rq *this_rq) { }
10838 #endif /* CONFIG_NO_HZ_COMMON */
10841 * newidle_balance is called by schedule() if this_cpu is about to become
10842 * idle. Attempts to pull tasks from other CPUs.
10845 * < 0 - we released the lock and there are !fair tasks present
10846 * 0 - failed, no new tasks
10847 * > 0 - success, new (fair) tasks present
10849 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
10851 unsigned long next_balance = jiffies + HZ;
10852 int this_cpu = this_rq->cpu;
10853 u64 t0, t1, curr_cost = 0;
10854 struct sched_domain *sd;
10855 int pulled_task = 0;
10857 update_misfit_status(NULL, this_rq);
10860 * There is a task waiting to run. No need to search for one.
10861 * Return 0; the task will be enqueued when switching to idle.
10863 if (this_rq->ttwu_pending)
10867 * We must set idle_stamp _before_ calling idle_balance(), such that we
10868 * measure the duration of idle_balance() as idle time.
10870 this_rq->idle_stamp = rq_clock(this_rq);
10873 * Do not pull tasks towards !active CPUs...
10875 if (!cpu_active(this_cpu))
10879 * This is OK, because current is on_cpu, which avoids it being picked
10880 * for load-balance and preemption/IRQs are still disabled avoiding
10881 * further scheduler activity on it and we're being very careful to
10882 * re-start the picking loop.
10884 rq_unpin_lock(this_rq, rf);
10887 sd = rcu_dereference_check_sched_domain(this_rq->sd);
10889 if (!READ_ONCE(this_rq->rd->overload) ||
10890 (sd && this_rq->avg_idle < sd->max_newidle_lb_cost)) {
10893 update_next_balance(sd, &next_balance);
10900 raw_spin_rq_unlock(this_rq);
10902 t0 = sched_clock_cpu(this_cpu);
10903 update_blocked_averages(this_cpu);
10906 for_each_domain(this_cpu, sd) {
10907 int continue_balancing = 1;
10910 update_next_balance(sd, &next_balance);
10912 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
10915 if (sd->flags & SD_BALANCE_NEWIDLE) {
10917 pulled_task = load_balance(this_cpu, this_rq,
10918 sd, CPU_NEWLY_IDLE,
10919 &continue_balancing);
10921 t1 = sched_clock_cpu(this_cpu);
10922 domain_cost = t1 - t0;
10923 update_newidle_cost(sd, domain_cost);
10925 curr_cost += domain_cost;
10930 * Stop searching for tasks to pull if there are
10931 * now runnable tasks on this rq.
10933 if (pulled_task || this_rq->nr_running > 0 ||
10934 this_rq->ttwu_pending)
10939 raw_spin_rq_lock(this_rq);
10941 if (curr_cost > this_rq->max_idle_balance_cost)
10942 this_rq->max_idle_balance_cost = curr_cost;
10945 * While browsing the domains, we released the rq lock, a task could
10946 * have been enqueued in the meantime. Since we're not going idle,
10947 * pretend we pulled a task.
10949 if (this_rq->cfs.h_nr_running && !pulled_task)
10952 /* Is there a task of a high priority class? */
10953 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
10957 /* Move the next balance forward */
10958 if (time_after(this_rq->next_balance, next_balance))
10959 this_rq->next_balance = next_balance;
10962 this_rq->idle_stamp = 0;
10964 nohz_newidle_balance(this_rq);
10966 rq_repin_lock(this_rq, rf);
10968 return pulled_task;
10972 * run_rebalance_domains is triggered when needed from the scheduler tick.
10973 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
10975 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
10977 struct rq *this_rq = this_rq();
10978 enum cpu_idle_type idle = this_rq->idle_balance ?
10979 CPU_IDLE : CPU_NOT_IDLE;
10982 * If this CPU has a pending nohz_balance_kick, then do the
10983 * balancing on behalf of the other idle CPUs whose ticks are
10984 * stopped. Do nohz_idle_balance *before* rebalance_domains to
10985 * give the idle CPUs a chance to load balance. Else we may
10986 * load balance only within the local sched_domain hierarchy
10987 * and abort nohz_idle_balance altogether if we pull some load.
10989 if (nohz_idle_balance(this_rq, idle))
10992 /* normal load balance */
10993 update_blocked_averages(this_rq->cpu);
10994 rebalance_domains(this_rq, idle);
10998 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
11000 void trigger_load_balance(struct rq *rq)
11003 * Don't need to rebalance while attached to NULL domain or
11004 * runqueue CPU is not active
11006 if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq))))
11009 if (time_after_eq(jiffies, rq->next_balance))
11010 raise_softirq(SCHED_SOFTIRQ);
11012 nohz_balancer_kick(rq);
11015 static void rq_online_fair(struct rq *rq)
11019 update_runtime_enabled(rq);
11022 static void rq_offline_fair(struct rq *rq)
11026 /* Ensure any throttled groups are reachable by pick_next_task */
11027 unthrottle_offline_cfs_rqs(rq);
11030 #endif /* CONFIG_SMP */
11032 #ifdef CONFIG_SCHED_CORE
11034 __entity_slice_used(struct sched_entity *se, int min_nr_tasks)
11036 u64 slice = sched_slice(cfs_rq_of(se), se);
11037 u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime;
11039 return (rtime * min_nr_tasks > slice);
11042 #define MIN_NR_TASKS_DURING_FORCEIDLE 2
11043 static inline void task_tick_core(struct rq *rq, struct task_struct *curr)
11045 if (!sched_core_enabled(rq))
11049 * If runqueue has only one task which used up its slice and
11050 * if the sibling is forced idle, then trigger schedule to
11051 * give forced idle task a chance.
11053 * sched_slice() considers only this active rq and it gets the
11054 * whole slice. But during force idle, we have siblings acting
11055 * like a single runqueue and hence we need to consider runnable
11056 * tasks on this CPU and the forced idle CPU. Ideally, we should
11057 * go through the forced idle rq, but that would be a perf hit.
11058 * We can assume that the forced idle CPU has at least
11059 * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check
11060 * if we need to give up the CPU.
11062 if (rq->core->core_forceidle_count && rq->cfs.nr_running == 1 &&
11063 __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE))
11068 * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed.
11070 static void se_fi_update(struct sched_entity *se, unsigned int fi_seq, bool forceidle)
11072 for_each_sched_entity(se) {
11073 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11076 if (cfs_rq->forceidle_seq == fi_seq)
11078 cfs_rq->forceidle_seq = fi_seq;
11081 cfs_rq->min_vruntime_fi = cfs_rq->min_vruntime;
11085 void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi)
11087 struct sched_entity *se = &p->se;
11089 if (p->sched_class != &fair_sched_class)
11092 se_fi_update(se, rq->core->core_forceidle_seq, in_fi);
11095 bool cfs_prio_less(struct task_struct *a, struct task_struct *b, bool in_fi)
11097 struct rq *rq = task_rq(a);
11098 struct sched_entity *sea = &a->se;
11099 struct sched_entity *seb = &b->se;
11100 struct cfs_rq *cfs_rqa;
11101 struct cfs_rq *cfs_rqb;
11104 SCHED_WARN_ON(task_rq(b)->core != rq->core);
11106 #ifdef CONFIG_FAIR_GROUP_SCHED
11108 * Find an se in the hierarchy for tasks a and b, such that the se's
11109 * are immediate siblings.
11111 while (sea->cfs_rq->tg != seb->cfs_rq->tg) {
11112 int sea_depth = sea->depth;
11113 int seb_depth = seb->depth;
11115 if (sea_depth >= seb_depth)
11116 sea = parent_entity(sea);
11117 if (sea_depth <= seb_depth)
11118 seb = parent_entity(seb);
11121 se_fi_update(sea, rq->core->core_forceidle_seq, in_fi);
11122 se_fi_update(seb, rq->core->core_forceidle_seq, in_fi);
11124 cfs_rqa = sea->cfs_rq;
11125 cfs_rqb = seb->cfs_rq;
11127 cfs_rqa = &task_rq(a)->cfs;
11128 cfs_rqb = &task_rq(b)->cfs;
11132 * Find delta after normalizing se's vruntime with its cfs_rq's
11133 * min_vruntime_fi, which would have been updated in prior calls
11134 * to se_fi_update().
11136 delta = (s64)(sea->vruntime - seb->vruntime) +
11137 (s64)(cfs_rqb->min_vruntime_fi - cfs_rqa->min_vruntime_fi);
11142 static inline void task_tick_core(struct rq *rq, struct task_struct *curr) {}
11146 * scheduler tick hitting a task of our scheduling class.
11148 * NOTE: This function can be called remotely by the tick offload that
11149 * goes along full dynticks. Therefore no local assumption can be made
11150 * and everything must be accessed through the @rq and @curr passed in
11153 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
11155 struct cfs_rq *cfs_rq;
11156 struct sched_entity *se = &curr->se;
11158 for_each_sched_entity(se) {
11159 cfs_rq = cfs_rq_of(se);
11160 entity_tick(cfs_rq, se, queued);
11163 if (static_branch_unlikely(&sched_numa_balancing))
11164 task_tick_numa(rq, curr);
11166 update_misfit_status(curr, rq);
11167 update_overutilized_status(task_rq(curr));
11169 task_tick_core(rq, curr);
11173 * called on fork with the child task as argument from the parent's context
11174 * - child not yet on the tasklist
11175 * - preemption disabled
11177 static void task_fork_fair(struct task_struct *p)
11179 struct cfs_rq *cfs_rq;
11180 struct sched_entity *se = &p->se, *curr;
11181 struct rq *rq = this_rq();
11182 struct rq_flags rf;
11185 update_rq_clock(rq);
11187 cfs_rq = task_cfs_rq(current);
11188 curr = cfs_rq->curr;
11190 update_curr(cfs_rq);
11191 se->vruntime = curr->vruntime;
11193 place_entity(cfs_rq, se, 1);
11195 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
11197 * Upon rescheduling, sched_class::put_prev_task() will place
11198 * 'current' within the tree based on its new key value.
11200 swap(curr->vruntime, se->vruntime);
11204 se->vruntime -= cfs_rq->min_vruntime;
11205 rq_unlock(rq, &rf);
11209 * Priority of the task has changed. Check to see if we preempt
11210 * the current task.
11213 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
11215 if (!task_on_rq_queued(p))
11218 if (rq->cfs.nr_running == 1)
11222 * Reschedule if we are currently running on this runqueue and
11223 * our priority decreased, or if we are not currently running on
11224 * this runqueue and our priority is higher than the current's
11226 if (task_current(rq, p)) {
11227 if (p->prio > oldprio)
11230 check_preempt_curr(rq, p, 0);
11233 static inline bool vruntime_normalized(struct task_struct *p)
11235 struct sched_entity *se = &p->se;
11238 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
11239 * the dequeue_entity(.flags=0) will already have normalized the
11246 * When !on_rq, vruntime of the task has usually NOT been normalized.
11247 * But there are some cases where it has already been normalized:
11249 * - A forked child which is waiting for being woken up by
11250 * wake_up_new_task().
11251 * - A task which has been woken up by try_to_wake_up() and
11252 * waiting for actually being woken up by sched_ttwu_pending().
11254 if (!se->sum_exec_runtime ||
11255 (READ_ONCE(p->__state) == TASK_WAKING && p->sched_remote_wakeup))
11261 #ifdef CONFIG_FAIR_GROUP_SCHED
11263 * Propagate the changes of the sched_entity across the tg tree to make it
11264 * visible to the root
11266 static void propagate_entity_cfs_rq(struct sched_entity *se)
11268 struct cfs_rq *cfs_rq;
11270 list_add_leaf_cfs_rq(cfs_rq_of(se));
11272 /* Start to propagate at parent */
11275 for_each_sched_entity(se) {
11276 cfs_rq = cfs_rq_of(se);
11278 if (!cfs_rq_throttled(cfs_rq)){
11279 update_load_avg(cfs_rq, se, UPDATE_TG);
11280 list_add_leaf_cfs_rq(cfs_rq);
11284 if (list_add_leaf_cfs_rq(cfs_rq))
11289 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
11292 static void detach_entity_cfs_rq(struct sched_entity *se)
11294 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11296 /* Catch up with the cfs_rq and remove our load when we leave */
11297 update_load_avg(cfs_rq, se, 0);
11298 detach_entity_load_avg(cfs_rq, se);
11299 update_tg_load_avg(cfs_rq);
11300 propagate_entity_cfs_rq(se);
11303 static void attach_entity_cfs_rq(struct sched_entity *se)
11305 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11307 #ifdef CONFIG_FAIR_GROUP_SCHED
11309 * Since the real-depth could have been changed (only FAIR
11310 * class maintain depth value), reset depth properly.
11312 se->depth = se->parent ? se->parent->depth + 1 : 0;
11315 /* Synchronize entity with its cfs_rq */
11316 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
11317 attach_entity_load_avg(cfs_rq, se);
11318 update_tg_load_avg(cfs_rq);
11319 propagate_entity_cfs_rq(se);
11322 static void detach_task_cfs_rq(struct task_struct *p)
11324 struct sched_entity *se = &p->se;
11325 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11327 if (!vruntime_normalized(p)) {
11329 * Fix up our vruntime so that the current sleep doesn't
11330 * cause 'unlimited' sleep bonus.
11332 place_entity(cfs_rq, se, 0);
11333 se->vruntime -= cfs_rq->min_vruntime;
11336 detach_entity_cfs_rq(se);
11339 static void attach_task_cfs_rq(struct task_struct *p)
11341 struct sched_entity *se = &p->se;
11342 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11344 attach_entity_cfs_rq(se);
11346 if (!vruntime_normalized(p))
11347 se->vruntime += cfs_rq->min_vruntime;
11350 static void switched_from_fair(struct rq *rq, struct task_struct *p)
11352 detach_task_cfs_rq(p);
11355 static void switched_to_fair(struct rq *rq, struct task_struct *p)
11357 attach_task_cfs_rq(p);
11359 if (task_on_rq_queued(p)) {
11361 * We were most likely switched from sched_rt, so
11362 * kick off the schedule if running, otherwise just see
11363 * if we can still preempt the current task.
11365 if (task_current(rq, p))
11368 check_preempt_curr(rq, p, 0);
11372 /* Account for a task changing its policy or group.
11374 * This routine is mostly called to set cfs_rq->curr field when a task
11375 * migrates between groups/classes.
11377 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
11379 struct sched_entity *se = &p->se;
11382 if (task_on_rq_queued(p)) {
11384 * Move the next running task to the front of the list, so our
11385 * cfs_tasks list becomes MRU one.
11387 list_move(&se->group_node, &rq->cfs_tasks);
11391 for_each_sched_entity(se) {
11392 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11394 set_next_entity(cfs_rq, se);
11395 /* ensure bandwidth has been allocated on our new cfs_rq */
11396 account_cfs_rq_runtime(cfs_rq, 0);
11400 void init_cfs_rq(struct cfs_rq *cfs_rq)
11402 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
11403 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
11404 #ifndef CONFIG_64BIT
11405 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
11408 raw_spin_lock_init(&cfs_rq->removed.lock);
11412 #ifdef CONFIG_FAIR_GROUP_SCHED
11413 static void task_set_group_fair(struct task_struct *p)
11415 struct sched_entity *se = &p->se;
11417 set_task_rq(p, task_cpu(p));
11418 se->depth = se->parent ? se->parent->depth + 1 : 0;
11421 static void task_move_group_fair(struct task_struct *p)
11423 detach_task_cfs_rq(p);
11424 set_task_rq(p, task_cpu(p));
11427 /* Tell se's cfs_rq has been changed -- migrated */
11428 p->se.avg.last_update_time = 0;
11430 attach_task_cfs_rq(p);
11433 static void task_change_group_fair(struct task_struct *p, int type)
11436 case TASK_SET_GROUP:
11437 task_set_group_fair(p);
11440 case TASK_MOVE_GROUP:
11441 task_move_group_fair(p);
11446 void free_fair_sched_group(struct task_group *tg)
11450 for_each_possible_cpu(i) {
11452 kfree(tg->cfs_rq[i]);
11461 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
11463 struct sched_entity *se;
11464 struct cfs_rq *cfs_rq;
11467 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
11470 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
11474 tg->shares = NICE_0_LOAD;
11476 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
11478 for_each_possible_cpu(i) {
11479 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
11480 GFP_KERNEL, cpu_to_node(i));
11484 se = kzalloc_node(sizeof(struct sched_entity_stats),
11485 GFP_KERNEL, cpu_to_node(i));
11489 init_cfs_rq(cfs_rq);
11490 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
11491 init_entity_runnable_average(se);
11502 void online_fair_sched_group(struct task_group *tg)
11504 struct sched_entity *se;
11505 struct rq_flags rf;
11509 for_each_possible_cpu(i) {
11512 rq_lock_irq(rq, &rf);
11513 update_rq_clock(rq);
11514 attach_entity_cfs_rq(se);
11515 sync_throttle(tg, i);
11516 rq_unlock_irq(rq, &rf);
11520 void unregister_fair_sched_group(struct task_group *tg)
11522 unsigned long flags;
11526 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
11528 for_each_possible_cpu(cpu) {
11530 remove_entity_load_avg(tg->se[cpu]);
11533 * Only empty task groups can be destroyed; so we can speculatively
11534 * check on_list without danger of it being re-added.
11536 if (!tg->cfs_rq[cpu]->on_list)
11541 raw_spin_rq_lock_irqsave(rq, flags);
11542 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
11543 raw_spin_rq_unlock_irqrestore(rq, flags);
11547 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
11548 struct sched_entity *se, int cpu,
11549 struct sched_entity *parent)
11551 struct rq *rq = cpu_rq(cpu);
11555 init_cfs_rq_runtime(cfs_rq);
11557 tg->cfs_rq[cpu] = cfs_rq;
11560 /* se could be NULL for root_task_group */
11565 se->cfs_rq = &rq->cfs;
11568 se->cfs_rq = parent->my_q;
11569 se->depth = parent->depth + 1;
11573 /* guarantee group entities always have weight */
11574 update_load_set(&se->load, NICE_0_LOAD);
11575 se->parent = parent;
11578 static DEFINE_MUTEX(shares_mutex);
11580 static int __sched_group_set_shares(struct task_group *tg, unsigned long shares)
11584 lockdep_assert_held(&shares_mutex);
11587 * We can't change the weight of the root cgroup.
11592 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
11594 if (tg->shares == shares)
11597 tg->shares = shares;
11598 for_each_possible_cpu(i) {
11599 struct rq *rq = cpu_rq(i);
11600 struct sched_entity *se = tg->se[i];
11601 struct rq_flags rf;
11603 /* Propagate contribution to hierarchy */
11604 rq_lock_irqsave(rq, &rf);
11605 update_rq_clock(rq);
11606 for_each_sched_entity(se) {
11607 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
11608 update_cfs_group(se);
11610 rq_unlock_irqrestore(rq, &rf);
11616 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
11620 mutex_lock(&shares_mutex);
11621 if (tg_is_idle(tg))
11624 ret = __sched_group_set_shares(tg, shares);
11625 mutex_unlock(&shares_mutex);
11630 int sched_group_set_idle(struct task_group *tg, long idle)
11634 if (tg == &root_task_group)
11637 if (idle < 0 || idle > 1)
11640 mutex_lock(&shares_mutex);
11642 if (tg->idle == idle) {
11643 mutex_unlock(&shares_mutex);
11649 for_each_possible_cpu(i) {
11650 struct rq *rq = cpu_rq(i);
11651 struct sched_entity *se = tg->se[i];
11652 struct cfs_rq *parent_cfs_rq, *grp_cfs_rq = tg->cfs_rq[i];
11653 bool was_idle = cfs_rq_is_idle(grp_cfs_rq);
11654 long idle_task_delta;
11655 struct rq_flags rf;
11657 rq_lock_irqsave(rq, &rf);
11659 grp_cfs_rq->idle = idle;
11660 if (WARN_ON_ONCE(was_idle == cfs_rq_is_idle(grp_cfs_rq)))
11664 parent_cfs_rq = cfs_rq_of(se);
11665 if (cfs_rq_is_idle(grp_cfs_rq))
11666 parent_cfs_rq->idle_nr_running++;
11668 parent_cfs_rq->idle_nr_running--;
11671 idle_task_delta = grp_cfs_rq->h_nr_running -
11672 grp_cfs_rq->idle_h_nr_running;
11673 if (!cfs_rq_is_idle(grp_cfs_rq))
11674 idle_task_delta *= -1;
11676 for_each_sched_entity(se) {
11677 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11682 cfs_rq->idle_h_nr_running += idle_task_delta;
11684 /* Already accounted at parent level and above. */
11685 if (cfs_rq_is_idle(cfs_rq))
11690 rq_unlock_irqrestore(rq, &rf);
11693 /* Idle groups have minimum weight. */
11694 if (tg_is_idle(tg))
11695 __sched_group_set_shares(tg, scale_load(WEIGHT_IDLEPRIO));
11697 __sched_group_set_shares(tg, NICE_0_LOAD);
11699 mutex_unlock(&shares_mutex);
11703 #else /* CONFIG_FAIR_GROUP_SCHED */
11705 void free_fair_sched_group(struct task_group *tg) { }
11707 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
11712 void online_fair_sched_group(struct task_group *tg) { }
11714 void unregister_fair_sched_group(struct task_group *tg) { }
11716 #endif /* CONFIG_FAIR_GROUP_SCHED */
11719 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
11721 struct sched_entity *se = &task->se;
11722 unsigned int rr_interval = 0;
11725 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
11728 if (rq->cfs.load.weight)
11729 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
11731 return rr_interval;
11735 * All the scheduling class methods:
11737 DEFINE_SCHED_CLASS(fair) = {
11739 .enqueue_task = enqueue_task_fair,
11740 .dequeue_task = dequeue_task_fair,
11741 .yield_task = yield_task_fair,
11742 .yield_to_task = yield_to_task_fair,
11744 .check_preempt_curr = check_preempt_wakeup,
11746 .pick_next_task = __pick_next_task_fair,
11747 .put_prev_task = put_prev_task_fair,
11748 .set_next_task = set_next_task_fair,
11751 .balance = balance_fair,
11752 .pick_task = pick_task_fair,
11753 .select_task_rq = select_task_rq_fair,
11754 .migrate_task_rq = migrate_task_rq_fair,
11756 .rq_online = rq_online_fair,
11757 .rq_offline = rq_offline_fair,
11759 .task_dead = task_dead_fair,
11760 .set_cpus_allowed = set_cpus_allowed_common,
11763 .task_tick = task_tick_fair,
11764 .task_fork = task_fork_fair,
11766 .prio_changed = prio_changed_fair,
11767 .switched_from = switched_from_fair,
11768 .switched_to = switched_to_fair,
11770 .get_rr_interval = get_rr_interval_fair,
11772 .update_curr = update_curr_fair,
11774 #ifdef CONFIG_FAIR_GROUP_SCHED
11775 .task_change_group = task_change_group_fair,
11778 #ifdef CONFIG_UCLAMP_TASK
11779 .uclamp_enabled = 1,
11783 #ifdef CONFIG_SCHED_DEBUG
11784 void print_cfs_stats(struct seq_file *m, int cpu)
11786 struct cfs_rq *cfs_rq, *pos;
11789 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
11790 print_cfs_rq(m, cpu, cfs_rq);
11794 #ifdef CONFIG_NUMA_BALANCING
11795 void show_numa_stats(struct task_struct *p, struct seq_file *m)
11798 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
11799 struct numa_group *ng;
11802 ng = rcu_dereference(p->numa_group);
11803 for_each_online_node(node) {
11804 if (p->numa_faults) {
11805 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
11806 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
11809 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
11810 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
11812 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
11816 #endif /* CONFIG_NUMA_BALANCING */
11817 #endif /* CONFIG_SCHED_DEBUG */
11819 __init void init_sched_fair_class(void)
11822 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
11824 #ifdef CONFIG_NO_HZ_COMMON
11825 nohz.next_balance = jiffies;
11826 nohz.next_blocked = jiffies;
11827 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
11834 * Helper functions to facilitate extracting info from tracepoints.
11837 const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq)
11840 return cfs_rq ? &cfs_rq->avg : NULL;
11845 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg);
11847 char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len)
11851 strlcpy(str, "(null)", len);
11856 cfs_rq_tg_path(cfs_rq, str, len);
11859 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path);
11861 int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq)
11863 return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1;
11865 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu);
11867 const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq)
11870 return rq ? &rq->avg_rt : NULL;
11875 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt);
11877 const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq)
11880 return rq ? &rq->avg_dl : NULL;
11885 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl);
11887 const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq)
11889 #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ)
11890 return rq ? &rq->avg_irq : NULL;
11895 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq);
11897 int sched_trace_rq_cpu(struct rq *rq)
11899 return rq ? cpu_of(rq) : -1;
11901 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu);
11903 int sched_trace_rq_cpu_capacity(struct rq *rq)
11909 SCHED_CAPACITY_SCALE
11913 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu_capacity);
11915 const struct cpumask *sched_trace_rd_span(struct root_domain *rd)
11918 return rd ? rd->span : NULL;
11923 EXPORT_SYMBOL_GPL(sched_trace_rd_span);
11925 int sched_trace_rq_nr_running(struct rq *rq)
11927 return rq ? rq->nr_running : -1;
11929 EXPORT_SYMBOL_GPL(sched_trace_rq_nr_running);