1 // SPDX-License-Identifier: GPL-2.0
3 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
5 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
7 * Interactivity improvements by Mike Galbraith
8 * (C) 2007 Mike Galbraith <efault@gmx.de>
10 * Various enhancements by Dmitry Adamushko.
11 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
13 * Group scheduling enhancements by Srivatsa Vaddagiri
14 * Copyright IBM Corporation, 2007
15 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
17 * Scaled math optimizations by Thomas Gleixner
18 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
20 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
21 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
25 #include <trace/events/sched.h>
28 * Targeted preemption latency for CPU-bound tasks:
30 * NOTE: this latency value is not the same as the concept of
31 * 'timeslice length' - timeslices in CFS are of variable length
32 * and have no persistent notion like in traditional, time-slice
33 * based scheduling concepts.
35 * (to see the precise effective timeslice length of your workload,
36 * run vmstat and monitor the context-switches (cs) field)
38 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
40 unsigned int sysctl_sched_latency = 6000000ULL;
41 static unsigned int normalized_sysctl_sched_latency = 6000000ULL;
44 * The initial- and re-scaling of tunables is configurable
48 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
49 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
50 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
52 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
54 enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
57 * Minimal preemption granularity for CPU-bound tasks:
59 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
61 unsigned int sysctl_sched_min_granularity = 750000ULL;
62 static unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
65 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
67 static unsigned int sched_nr_latency = 8;
70 * After fork, child runs first. If set to 0 (default) then
71 * parent will (try to) run first.
73 unsigned int sysctl_sched_child_runs_first __read_mostly;
76 * SCHED_OTHER wake-up granularity.
78 * This option delays the preemption effects of decoupled workloads
79 * and reduces their over-scheduling. Synchronous workloads will still
80 * have immediate wakeup/sleep latencies.
82 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
84 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
85 static unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
87 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
89 int sched_thermal_decay_shift;
90 static int __init setup_sched_thermal_decay_shift(char *str)
94 if (kstrtoint(str, 0, &_shift))
95 pr_warn("Unable to set scheduler thermal pressure decay shift parameter\n");
97 sched_thermal_decay_shift = clamp(_shift, 0, 10);
100 __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift);
104 * For asym packing, by default the lower numbered CPU has higher priority.
106 int __weak arch_asym_cpu_priority(int cpu)
112 * The margin used when comparing utilization with CPU capacity.
116 #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024)
120 #ifdef CONFIG_CFS_BANDWIDTH
122 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
123 * each time a cfs_rq requests quota.
125 * Note: in the case that the slice exceeds the runtime remaining (either due
126 * to consumption or the quota being specified to be smaller than the slice)
127 * we will always only issue the remaining available time.
129 * (default: 5 msec, units: microseconds)
131 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
134 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
140 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
146 static inline void update_load_set(struct load_weight *lw, unsigned long w)
153 * Increase the granularity value when there are more CPUs,
154 * because with more CPUs the 'effective latency' as visible
155 * to users decreases. But the relationship is not linear,
156 * so pick a second-best guess by going with the log2 of the
159 * This idea comes from the SD scheduler of Con Kolivas:
161 static unsigned int get_update_sysctl_factor(void)
163 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
166 switch (sysctl_sched_tunable_scaling) {
167 case SCHED_TUNABLESCALING_NONE:
170 case SCHED_TUNABLESCALING_LINEAR:
173 case SCHED_TUNABLESCALING_LOG:
175 factor = 1 + ilog2(cpus);
182 static void update_sysctl(void)
184 unsigned int factor = get_update_sysctl_factor();
186 #define SET_SYSCTL(name) \
187 (sysctl_##name = (factor) * normalized_sysctl_##name)
188 SET_SYSCTL(sched_min_granularity);
189 SET_SYSCTL(sched_latency);
190 SET_SYSCTL(sched_wakeup_granularity);
194 void sched_init_granularity(void)
199 #define WMULT_CONST (~0U)
200 #define WMULT_SHIFT 32
202 static void __update_inv_weight(struct load_weight *lw)
206 if (likely(lw->inv_weight))
209 w = scale_load_down(lw->weight);
211 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
213 else if (unlikely(!w))
214 lw->inv_weight = WMULT_CONST;
216 lw->inv_weight = WMULT_CONST / w;
220 * delta_exec * weight / lw.weight
222 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
224 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
225 * we're guaranteed shift stays positive because inv_weight is guaranteed to
226 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
228 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
229 * weight/lw.weight <= 1, and therefore our shift will also be positive.
231 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
233 u64 fact = scale_load_down(weight);
234 int shift = WMULT_SHIFT;
236 __update_inv_weight(lw);
238 if (unlikely(fact >> 32)) {
245 fact = mul_u32_u32(fact, lw->inv_weight);
252 return mul_u64_u32_shr(delta_exec, fact, shift);
256 const struct sched_class fair_sched_class;
258 /**************************************************************
259 * CFS operations on generic schedulable entities:
262 #ifdef CONFIG_FAIR_GROUP_SCHED
263 static inline struct task_struct *task_of(struct sched_entity *se)
265 SCHED_WARN_ON(!entity_is_task(se));
266 return container_of(se, struct task_struct, se);
269 /* Walk up scheduling entities hierarchy */
270 #define for_each_sched_entity(se) \
271 for (; se; se = se->parent)
273 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
278 /* runqueue on which this entity is (to be) queued */
279 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
284 /* runqueue "owned" by this group */
285 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
290 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
295 if (cfs_rq && task_group_is_autogroup(cfs_rq->tg))
296 autogroup_path(cfs_rq->tg, path, len);
297 else if (cfs_rq && cfs_rq->tg->css.cgroup)
298 cgroup_path(cfs_rq->tg->css.cgroup, path, len);
300 strlcpy(path, "(null)", len);
303 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
305 struct rq *rq = rq_of(cfs_rq);
306 int cpu = cpu_of(rq);
309 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
314 * Ensure we either appear before our parent (if already
315 * enqueued) or force our parent to appear after us when it is
316 * enqueued. The fact that we always enqueue bottom-up
317 * reduces this to two cases and a special case for the root
318 * cfs_rq. Furthermore, it also means that we will always reset
319 * tmp_alone_branch either when the branch is connected
320 * to a tree or when we reach the top of the tree
322 if (cfs_rq->tg->parent &&
323 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
325 * If parent is already on the list, we add the child
326 * just before. Thanks to circular linked property of
327 * the list, this means to put the child at the tail
328 * of the list that starts by parent.
330 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
331 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
333 * The branch is now connected to its tree so we can
334 * reset tmp_alone_branch to the beginning of the
337 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
341 if (!cfs_rq->tg->parent) {
343 * cfs rq without parent should be put
344 * at the tail of the list.
346 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
347 &rq->leaf_cfs_rq_list);
349 * We have reach the top of a tree so we can reset
350 * tmp_alone_branch to the beginning of the list.
352 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
357 * The parent has not already been added so we want to
358 * make sure that it will be put after us.
359 * tmp_alone_branch points to the begin of the branch
360 * where we will add parent.
362 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
364 * update tmp_alone_branch to points to the new begin
367 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
371 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
373 if (cfs_rq->on_list) {
374 struct rq *rq = rq_of(cfs_rq);
377 * With cfs_rq being unthrottled/throttled during an enqueue,
378 * it can happen the tmp_alone_branch points the a leaf that
379 * we finally want to del. In this case, tmp_alone_branch moves
380 * to the prev element but it will point to rq->leaf_cfs_rq_list
381 * at the end of the enqueue.
383 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
384 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
386 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
391 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
393 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
396 /* Iterate thr' all leaf cfs_rq's on a runqueue */
397 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
398 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
401 /* Do the two (enqueued) entities belong to the same group ? */
402 static inline struct cfs_rq *
403 is_same_group(struct sched_entity *se, struct sched_entity *pse)
405 if (se->cfs_rq == pse->cfs_rq)
411 static inline struct sched_entity *parent_entity(struct sched_entity *se)
417 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
419 int se_depth, pse_depth;
422 * preemption test can be made between sibling entities who are in the
423 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
424 * both tasks until we find their ancestors who are siblings of common
428 /* First walk up until both entities are at same depth */
429 se_depth = (*se)->depth;
430 pse_depth = (*pse)->depth;
432 while (se_depth > pse_depth) {
434 *se = parent_entity(*se);
437 while (pse_depth > se_depth) {
439 *pse = parent_entity(*pse);
442 while (!is_same_group(*se, *pse)) {
443 *se = parent_entity(*se);
444 *pse = parent_entity(*pse);
448 #else /* !CONFIG_FAIR_GROUP_SCHED */
450 static inline struct task_struct *task_of(struct sched_entity *se)
452 return container_of(se, struct task_struct, se);
455 #define for_each_sched_entity(se) \
456 for (; se; se = NULL)
458 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
460 return &task_rq(p)->cfs;
463 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
465 struct task_struct *p = task_of(se);
466 struct rq *rq = task_rq(p);
471 /* runqueue "owned" by this group */
472 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
477 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
480 strlcpy(path, "(null)", len);
483 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
488 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
492 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
496 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
497 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
499 static inline struct sched_entity *parent_entity(struct sched_entity *se)
505 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
509 #endif /* CONFIG_FAIR_GROUP_SCHED */
511 static __always_inline
512 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
514 /**************************************************************
515 * Scheduling class tree data structure manipulation methods:
518 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
520 s64 delta = (s64)(vruntime - max_vruntime);
522 max_vruntime = vruntime;
527 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
529 s64 delta = (s64)(vruntime - min_vruntime);
531 min_vruntime = vruntime;
536 static inline int entity_before(struct sched_entity *a,
537 struct sched_entity *b)
539 return (s64)(a->vruntime - b->vruntime) < 0;
542 static void update_min_vruntime(struct cfs_rq *cfs_rq)
544 struct sched_entity *curr = cfs_rq->curr;
545 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
547 u64 vruntime = cfs_rq->min_vruntime;
551 vruntime = curr->vruntime;
556 if (leftmost) { /* non-empty tree */
557 struct sched_entity *se;
558 se = rb_entry(leftmost, struct sched_entity, run_node);
561 vruntime = se->vruntime;
563 vruntime = min_vruntime(vruntime, se->vruntime);
566 /* ensure we never gain time by being placed backwards. */
567 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
570 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
575 * Enqueue an entity into the rb-tree:
577 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
579 struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node;
580 struct rb_node *parent = NULL;
581 struct sched_entity *entry;
582 bool leftmost = true;
585 * Find the right place in the rbtree:
589 entry = rb_entry(parent, struct sched_entity, run_node);
591 * We dont care about collisions. Nodes with
592 * the same key stay together.
594 if (entity_before(se, entry)) {
595 link = &parent->rb_left;
597 link = &parent->rb_right;
602 rb_link_node(&se->run_node, parent, link);
603 rb_insert_color_cached(&se->run_node,
604 &cfs_rq->tasks_timeline, leftmost);
607 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
609 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
612 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
614 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
619 return rb_entry(left, struct sched_entity, run_node);
622 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
624 struct rb_node *next = rb_next(&se->run_node);
629 return rb_entry(next, struct sched_entity, run_node);
632 #ifdef CONFIG_SCHED_DEBUG
633 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
635 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
640 return rb_entry(last, struct sched_entity, run_node);
643 /**************************************************************
644 * Scheduling class statistics methods:
647 int sched_proc_update_handler(struct ctl_table *table, int write,
648 void *buffer, size_t *lenp, loff_t *ppos)
650 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
651 unsigned int factor = get_update_sysctl_factor();
656 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
657 sysctl_sched_min_granularity);
659 #define WRT_SYSCTL(name) \
660 (normalized_sysctl_##name = sysctl_##name / (factor))
661 WRT_SYSCTL(sched_min_granularity);
662 WRT_SYSCTL(sched_latency);
663 WRT_SYSCTL(sched_wakeup_granularity);
673 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
675 if (unlikely(se->load.weight != NICE_0_LOAD))
676 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
682 * The idea is to set a period in which each task runs once.
684 * When there are too many tasks (sched_nr_latency) we have to stretch
685 * this period because otherwise the slices get too small.
687 * p = (nr <= nl) ? l : l*nr/nl
689 static u64 __sched_period(unsigned long nr_running)
691 if (unlikely(nr_running > sched_nr_latency))
692 return nr_running * sysctl_sched_min_granularity;
694 return sysctl_sched_latency;
698 * We calculate the wall-time slice from the period by taking a part
699 * proportional to the weight.
703 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
705 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
707 for_each_sched_entity(se) {
708 struct load_weight *load;
709 struct load_weight lw;
711 cfs_rq = cfs_rq_of(se);
712 load = &cfs_rq->load;
714 if (unlikely(!se->on_rq)) {
717 update_load_add(&lw, se->load.weight);
720 slice = __calc_delta(slice, se->load.weight, load);
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 = cpu_scale;
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, int force)
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 schedstat_set(curr->statistics.exec_max,
860 max(delta_exec, curr->statistics.exec_max));
862 curr->sum_exec_runtime += delta_exec;
863 schedstat_add(cfs_rq->exec_clock, delta_exec);
865 curr->vruntime += calc_delta_fair(delta_exec, curr);
866 update_min_vruntime(cfs_rq);
868 if (entity_is_task(curr)) {
869 struct task_struct *curtask = task_of(curr);
871 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
872 cgroup_account_cputime(curtask, delta_exec);
873 account_group_exec_runtime(curtask, delta_exec);
876 account_cfs_rq_runtime(cfs_rq, delta_exec);
879 static void update_curr_fair(struct rq *rq)
881 update_curr(cfs_rq_of(&rq->curr->se));
885 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
887 u64 wait_start, prev_wait_start;
889 if (!schedstat_enabled())
892 wait_start = rq_clock(rq_of(cfs_rq));
893 prev_wait_start = schedstat_val(se->statistics.wait_start);
895 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
896 likely(wait_start > prev_wait_start))
897 wait_start -= prev_wait_start;
899 __schedstat_set(se->statistics.wait_start, wait_start);
903 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
905 struct task_struct *p;
908 if (!schedstat_enabled())
911 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
913 if (entity_is_task(se)) {
915 if (task_on_rq_migrating(p)) {
917 * Preserve migrating task's wait time so wait_start
918 * time stamp can be adjusted to accumulate wait time
919 * prior to migration.
921 __schedstat_set(se->statistics.wait_start, delta);
924 trace_sched_stat_wait(p, delta);
927 __schedstat_set(se->statistics.wait_max,
928 max(schedstat_val(se->statistics.wait_max), delta));
929 __schedstat_inc(se->statistics.wait_count);
930 __schedstat_add(se->statistics.wait_sum, delta);
931 __schedstat_set(se->statistics.wait_start, 0);
935 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
937 struct task_struct *tsk = NULL;
938 u64 sleep_start, block_start;
940 if (!schedstat_enabled())
943 sleep_start = schedstat_val(se->statistics.sleep_start);
944 block_start = schedstat_val(se->statistics.block_start);
946 if (entity_is_task(se))
950 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
955 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
956 __schedstat_set(se->statistics.sleep_max, delta);
958 __schedstat_set(se->statistics.sleep_start, 0);
959 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
962 account_scheduler_latency(tsk, delta >> 10, 1);
963 trace_sched_stat_sleep(tsk, delta);
967 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
972 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
973 __schedstat_set(se->statistics.block_max, delta);
975 __schedstat_set(se->statistics.block_start, 0);
976 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
979 if (tsk->in_iowait) {
980 __schedstat_add(se->statistics.iowait_sum, delta);
981 __schedstat_inc(se->statistics.iowait_count);
982 trace_sched_stat_iowait(tsk, delta);
985 trace_sched_stat_blocked(tsk, delta);
988 * Blocking time is in units of nanosecs, so shift by
989 * 20 to get a milliseconds-range estimation of the
990 * amount of time that the task spent sleeping:
992 if (unlikely(prof_on == SLEEP_PROFILING)) {
993 profile_hits(SLEEP_PROFILING,
994 (void *)get_wchan(tsk),
997 account_scheduler_latency(tsk, delta >> 10, 0);
1003 * Task is being enqueued - update stats:
1006 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1008 if (!schedstat_enabled())
1012 * Are we enqueueing a waiting task? (for current tasks
1013 * a dequeue/enqueue event is a NOP)
1015 if (se != cfs_rq->curr)
1016 update_stats_wait_start(cfs_rq, se);
1018 if (flags & ENQUEUE_WAKEUP)
1019 update_stats_enqueue_sleeper(cfs_rq, se);
1023 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1026 if (!schedstat_enabled())
1030 * Mark the end of the wait period if dequeueing a
1033 if (se != cfs_rq->curr)
1034 update_stats_wait_end(cfs_rq, se);
1036 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1037 struct task_struct *tsk = task_of(se);
1039 if (tsk->state & TASK_INTERRUPTIBLE)
1040 __schedstat_set(se->statistics.sleep_start,
1041 rq_clock(rq_of(cfs_rq)));
1042 if (tsk->state & TASK_UNINTERRUPTIBLE)
1043 __schedstat_set(se->statistics.block_start,
1044 rq_clock(rq_of(cfs_rq)));
1049 * We are picking a new current task - update its stats:
1052 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1055 * We are starting a new run period:
1057 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1060 /**************************************************
1061 * Scheduling class queueing methods:
1064 #ifdef CONFIG_NUMA_BALANCING
1066 * Approximate time to scan a full NUMA task in ms. The task scan period is
1067 * calculated based on the tasks virtual memory size and
1068 * numa_balancing_scan_size.
1070 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1071 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1073 /* Portion of address space to scan in MB */
1074 unsigned int sysctl_numa_balancing_scan_size = 256;
1076 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1077 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1080 refcount_t refcount;
1082 spinlock_t lock; /* nr_tasks, tasks */
1087 struct rcu_head rcu;
1088 unsigned long total_faults;
1089 unsigned long max_faults_cpu;
1091 * Faults_cpu is used to decide whether memory should move
1092 * towards the CPU. As a consequence, these stats are weighted
1093 * more by CPU use than by memory faults.
1095 unsigned long *faults_cpu;
1096 unsigned long faults[0];
1100 * For functions that can be called in multiple contexts that permit reading
1101 * ->numa_group (see struct task_struct for locking rules).
1103 static struct numa_group *deref_task_numa_group(struct task_struct *p)
1105 return rcu_dereference_check(p->numa_group, p == current ||
1106 (lockdep_is_held(&task_rq(p)->lock) && !READ_ONCE(p->on_cpu)));
1109 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1111 return rcu_dereference_protected(p->numa_group, p == current);
1114 static inline unsigned long group_faults_priv(struct numa_group *ng);
1115 static inline unsigned long group_faults_shared(struct numa_group *ng);
1117 static unsigned int task_nr_scan_windows(struct task_struct *p)
1119 unsigned long rss = 0;
1120 unsigned long nr_scan_pages;
1123 * Calculations based on RSS as non-present and empty pages are skipped
1124 * by the PTE scanner and NUMA hinting faults should be trapped based
1127 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1128 rss = get_mm_rss(p->mm);
1130 rss = nr_scan_pages;
1132 rss = round_up(rss, nr_scan_pages);
1133 return rss / nr_scan_pages;
1136 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1137 #define MAX_SCAN_WINDOW 2560
1139 static unsigned int task_scan_min(struct task_struct *p)
1141 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1142 unsigned int scan, floor;
1143 unsigned int windows = 1;
1145 if (scan_size < MAX_SCAN_WINDOW)
1146 windows = MAX_SCAN_WINDOW / scan_size;
1147 floor = 1000 / windows;
1149 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1150 return max_t(unsigned int, floor, scan);
1153 static unsigned int task_scan_start(struct task_struct *p)
1155 unsigned long smin = task_scan_min(p);
1156 unsigned long period = smin;
1157 struct numa_group *ng;
1159 /* Scale the maximum scan period with the amount of shared memory. */
1161 ng = rcu_dereference(p->numa_group);
1163 unsigned long shared = group_faults_shared(ng);
1164 unsigned long private = group_faults_priv(ng);
1166 period *= refcount_read(&ng->refcount);
1167 period *= shared + 1;
1168 period /= private + shared + 1;
1172 return max(smin, period);
1175 static unsigned int task_scan_max(struct task_struct *p)
1177 unsigned long smin = task_scan_min(p);
1179 struct numa_group *ng;
1181 /* Watch for min being lower than max due to floor calculations */
1182 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1184 /* Scale the maximum scan period with the amount of shared memory. */
1185 ng = deref_curr_numa_group(p);
1187 unsigned long shared = group_faults_shared(ng);
1188 unsigned long private = group_faults_priv(ng);
1189 unsigned long period = smax;
1191 period *= refcount_read(&ng->refcount);
1192 period *= shared + 1;
1193 period /= private + shared + 1;
1195 smax = max(smax, period);
1198 return max(smin, smax);
1201 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1203 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1204 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1207 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1209 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1210 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1213 /* Shared or private faults. */
1214 #define NR_NUMA_HINT_FAULT_TYPES 2
1216 /* Memory and CPU locality */
1217 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1219 /* Averaged statistics, and temporary buffers. */
1220 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1222 pid_t task_numa_group_id(struct task_struct *p)
1224 struct numa_group *ng;
1228 ng = rcu_dereference(p->numa_group);
1237 * The averaged statistics, shared & private, memory & CPU,
1238 * occupy the first half of the array. The second half of the
1239 * array is for current counters, which are averaged into the
1240 * first set by task_numa_placement.
1242 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1244 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1247 static inline unsigned long task_faults(struct task_struct *p, int nid)
1249 if (!p->numa_faults)
1252 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1253 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1256 static inline unsigned long group_faults(struct task_struct *p, int nid)
1258 struct numa_group *ng = deref_task_numa_group(p);
1263 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1264 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1267 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1269 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1270 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1273 static inline unsigned long group_faults_priv(struct numa_group *ng)
1275 unsigned long faults = 0;
1278 for_each_online_node(node) {
1279 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1285 static inline unsigned long group_faults_shared(struct numa_group *ng)
1287 unsigned long faults = 0;
1290 for_each_online_node(node) {
1291 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1298 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1299 * considered part of a numa group's pseudo-interleaving set. Migrations
1300 * between these nodes are slowed down, to allow things to settle down.
1302 #define ACTIVE_NODE_FRACTION 3
1304 static bool numa_is_active_node(int nid, struct numa_group *ng)
1306 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1309 /* Handle placement on systems where not all nodes are directly connected. */
1310 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1311 int maxdist, bool task)
1313 unsigned long score = 0;
1317 * All nodes are directly connected, and the same distance
1318 * from each other. No need for fancy placement algorithms.
1320 if (sched_numa_topology_type == NUMA_DIRECT)
1324 * This code is called for each node, introducing N^2 complexity,
1325 * which should be ok given the number of nodes rarely exceeds 8.
1327 for_each_online_node(node) {
1328 unsigned long faults;
1329 int dist = node_distance(nid, node);
1332 * The furthest away nodes in the system are not interesting
1333 * for placement; nid was already counted.
1335 if (dist == sched_max_numa_distance || node == nid)
1339 * On systems with a backplane NUMA topology, compare groups
1340 * of nodes, and move tasks towards the group with the most
1341 * memory accesses. When comparing two nodes at distance
1342 * "hoplimit", only nodes closer by than "hoplimit" are part
1343 * of each group. Skip other nodes.
1345 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1349 /* Add up the faults from nearby nodes. */
1351 faults = task_faults(p, node);
1353 faults = group_faults(p, node);
1356 * On systems with a glueless mesh NUMA topology, there are
1357 * no fixed "groups of nodes". Instead, nodes that are not
1358 * directly connected bounce traffic through intermediate
1359 * nodes; a numa_group can occupy any set of nodes.
1360 * The further away a node is, the less the faults count.
1361 * This seems to result in good task placement.
1363 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1364 faults *= (sched_max_numa_distance - dist);
1365 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1375 * These return the fraction of accesses done by a particular task, or
1376 * task group, on a particular numa node. The group weight is given a
1377 * larger multiplier, in order to group tasks together that are almost
1378 * evenly spread out between numa nodes.
1380 static inline unsigned long task_weight(struct task_struct *p, int nid,
1383 unsigned long faults, total_faults;
1385 if (!p->numa_faults)
1388 total_faults = p->total_numa_faults;
1393 faults = task_faults(p, nid);
1394 faults += score_nearby_nodes(p, nid, dist, true);
1396 return 1000 * faults / total_faults;
1399 static inline unsigned long group_weight(struct task_struct *p, int nid,
1402 struct numa_group *ng = deref_task_numa_group(p);
1403 unsigned long faults, total_faults;
1408 total_faults = ng->total_faults;
1413 faults = group_faults(p, nid);
1414 faults += score_nearby_nodes(p, nid, dist, false);
1416 return 1000 * faults / total_faults;
1419 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1420 int src_nid, int dst_cpu)
1422 struct numa_group *ng = deref_curr_numa_group(p);
1423 int dst_nid = cpu_to_node(dst_cpu);
1424 int last_cpupid, this_cpupid;
1426 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1427 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1430 * Allow first faults or private faults to migrate immediately early in
1431 * the lifetime of a task. The magic number 4 is based on waiting for
1432 * two full passes of the "multi-stage node selection" test that is
1435 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1436 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1440 * Multi-stage node selection is used in conjunction with a periodic
1441 * migration fault to build a temporal task<->page relation. By using
1442 * a two-stage filter we remove short/unlikely relations.
1444 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1445 * a task's usage of a particular page (n_p) per total usage of this
1446 * page (n_t) (in a given time-span) to a probability.
1448 * Our periodic faults will sample this probability and getting the
1449 * same result twice in a row, given these samples are fully
1450 * independent, is then given by P(n)^2, provided our sample period
1451 * is sufficiently short compared to the usage pattern.
1453 * This quadric squishes small probabilities, making it less likely we
1454 * act on an unlikely task<->page relation.
1456 if (!cpupid_pid_unset(last_cpupid) &&
1457 cpupid_to_nid(last_cpupid) != dst_nid)
1460 /* Always allow migrate on private faults */
1461 if (cpupid_match_pid(p, last_cpupid))
1464 /* A shared fault, but p->numa_group has not been set up yet. */
1469 * Destination node is much more heavily used than the source
1470 * node? Allow migration.
1472 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1473 ACTIVE_NODE_FRACTION)
1477 * Distribute memory according to CPU & memory use on each node,
1478 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1480 * faults_cpu(dst) 3 faults_cpu(src)
1481 * --------------- * - > ---------------
1482 * faults_mem(dst) 4 faults_mem(src)
1484 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1485 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1489 * 'numa_type' describes the node at the moment of load balancing.
1492 /* The node has spare capacity that can be used to run more tasks. */
1495 * The node is fully used and the tasks don't compete for more CPU
1496 * cycles. Nevertheless, some tasks might wait before running.
1500 * The node is overloaded and can't provide expected CPU cycles to all
1506 /* Cached statistics for all CPUs within a node */
1510 /* Total compute capacity of CPUs on a node */
1511 unsigned long compute_capacity;
1512 unsigned int nr_running;
1513 unsigned int weight;
1514 enum numa_type node_type;
1518 static inline bool is_core_idle(int cpu)
1520 #ifdef CONFIG_SCHED_SMT
1523 for_each_cpu(sibling, cpu_smt_mask(cpu)) {
1535 struct task_numa_env {
1536 struct task_struct *p;
1538 int src_cpu, src_nid;
1539 int dst_cpu, dst_nid;
1541 struct numa_stats src_stats, dst_stats;
1546 struct task_struct *best_task;
1551 static unsigned long cpu_load(struct rq *rq);
1552 static unsigned long cpu_util(int cpu);
1553 static inline long adjust_numa_imbalance(int imbalance, int src_nr_running);
1556 numa_type numa_classify(unsigned int imbalance_pct,
1557 struct numa_stats *ns)
1559 if ((ns->nr_running > ns->weight) &&
1560 ((ns->compute_capacity * 100) < (ns->util * imbalance_pct)))
1561 return node_overloaded;
1563 if ((ns->nr_running < ns->weight) ||
1564 ((ns->compute_capacity * 100) > (ns->util * imbalance_pct)))
1565 return node_has_spare;
1567 return node_fully_busy;
1570 #ifdef CONFIG_SCHED_SMT
1571 /* Forward declarations of select_idle_sibling helpers */
1572 static inline bool test_idle_cores(int cpu, bool def);
1573 static inline int numa_idle_core(int idle_core, int cpu)
1575 if (!static_branch_likely(&sched_smt_present) ||
1576 idle_core >= 0 || !test_idle_cores(cpu, false))
1580 * Prefer cores instead of packing HT siblings
1581 * and triggering future load balancing.
1583 if (is_core_idle(cpu))
1589 static inline int numa_idle_core(int idle_core, int cpu)
1596 * Gather all necessary information to make NUMA balancing placement
1597 * decisions that are compatible with standard load balancer. This
1598 * borrows code and logic from update_sg_lb_stats but sharing a
1599 * common implementation is impractical.
1601 static void update_numa_stats(struct task_numa_env *env,
1602 struct numa_stats *ns, int nid,
1605 int cpu, idle_core = -1;
1607 memset(ns, 0, sizeof(*ns));
1611 for_each_cpu(cpu, cpumask_of_node(nid)) {
1612 struct rq *rq = cpu_rq(cpu);
1614 ns->load += cpu_load(rq);
1615 ns->util += cpu_util(cpu);
1616 ns->nr_running += rq->cfs.h_nr_running;
1617 ns->compute_capacity += capacity_of(cpu);
1619 if (find_idle && !rq->nr_running && idle_cpu(cpu)) {
1620 if (READ_ONCE(rq->numa_migrate_on) ||
1621 !cpumask_test_cpu(cpu, env->p->cpus_ptr))
1624 if (ns->idle_cpu == -1)
1627 idle_core = numa_idle_core(idle_core, cpu);
1632 ns->weight = cpumask_weight(cpumask_of_node(nid));
1634 ns->node_type = numa_classify(env->imbalance_pct, ns);
1637 ns->idle_cpu = idle_core;
1640 static void task_numa_assign(struct task_numa_env *env,
1641 struct task_struct *p, long imp)
1643 struct rq *rq = cpu_rq(env->dst_cpu);
1645 /* Check if run-queue part of active NUMA balance. */
1646 if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) {
1648 int start = env->dst_cpu;
1650 /* Find alternative idle CPU. */
1651 for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start) {
1652 if (cpu == env->best_cpu || !idle_cpu(cpu) ||
1653 !cpumask_test_cpu(cpu, env->p->cpus_ptr)) {
1658 rq = cpu_rq(env->dst_cpu);
1659 if (!xchg(&rq->numa_migrate_on, 1))
1663 /* Failed to find an alternative idle CPU */
1669 * Clear previous best_cpu/rq numa-migrate flag, since task now
1670 * found a better CPU to move/swap.
1672 if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) {
1673 rq = cpu_rq(env->best_cpu);
1674 WRITE_ONCE(rq->numa_migrate_on, 0);
1678 put_task_struct(env->best_task);
1683 env->best_imp = imp;
1684 env->best_cpu = env->dst_cpu;
1687 static bool load_too_imbalanced(long src_load, long dst_load,
1688 struct task_numa_env *env)
1691 long orig_src_load, orig_dst_load;
1692 long src_capacity, dst_capacity;
1695 * The load is corrected for the CPU capacity available on each node.
1698 * ------------ vs ---------
1699 * src_capacity dst_capacity
1701 src_capacity = env->src_stats.compute_capacity;
1702 dst_capacity = env->dst_stats.compute_capacity;
1704 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1706 orig_src_load = env->src_stats.load;
1707 orig_dst_load = env->dst_stats.load;
1709 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1711 /* Would this change make things worse? */
1712 return (imb > old_imb);
1716 * Maximum NUMA importance can be 1998 (2*999);
1717 * SMALLIMP @ 30 would be close to 1998/64.
1718 * Used to deter task migration.
1723 * This checks if the overall compute and NUMA accesses of the system would
1724 * be improved if the source tasks was migrated to the target dst_cpu taking
1725 * into account that it might be best if task running on the dst_cpu should
1726 * be exchanged with the source task
1728 static bool task_numa_compare(struct task_numa_env *env,
1729 long taskimp, long groupimp, bool maymove)
1731 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
1732 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1733 long imp = p_ng ? groupimp : taskimp;
1734 struct task_struct *cur;
1735 long src_load, dst_load;
1736 int dist = env->dist;
1739 bool stopsearch = false;
1741 if (READ_ONCE(dst_rq->numa_migrate_on))
1745 cur = rcu_dereference(dst_rq->curr);
1746 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1750 * Because we have preemption enabled we can get migrated around and
1751 * end try selecting ourselves (current == env->p) as a swap candidate.
1753 if (cur == env->p) {
1759 if (maymove && moveimp >= env->best_imp)
1765 /* Skip this swap candidate if cannot move to the source cpu. */
1766 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
1770 * Skip this swap candidate if it is not moving to its preferred
1771 * node and the best task is.
1773 if (env->best_task &&
1774 env->best_task->numa_preferred_nid == env->src_nid &&
1775 cur->numa_preferred_nid != env->src_nid) {
1780 * "imp" is the fault differential for the source task between the
1781 * source and destination node. Calculate the total differential for
1782 * the source task and potential destination task. The more negative
1783 * the value is, the more remote accesses that would be expected to
1784 * be incurred if the tasks were swapped.
1786 * If dst and source tasks are in the same NUMA group, or not
1787 * in any group then look only at task weights.
1789 cur_ng = rcu_dereference(cur->numa_group);
1790 if (cur_ng == p_ng) {
1791 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1792 task_weight(cur, env->dst_nid, dist);
1794 * Add some hysteresis to prevent swapping the
1795 * tasks within a group over tiny differences.
1801 * Compare the group weights. If a task is all by itself
1802 * (not part of a group), use the task weight instead.
1805 imp += group_weight(cur, env->src_nid, dist) -
1806 group_weight(cur, env->dst_nid, dist);
1808 imp += task_weight(cur, env->src_nid, dist) -
1809 task_weight(cur, env->dst_nid, dist);
1812 /* Discourage picking a task already on its preferred node */
1813 if (cur->numa_preferred_nid == env->dst_nid)
1817 * Encourage picking a task that moves to its preferred node.
1818 * This potentially makes imp larger than it's maximum of
1819 * 1998 (see SMALLIMP and task_weight for why) but in this
1820 * case, it does not matter.
1822 if (cur->numa_preferred_nid == env->src_nid)
1825 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1832 * Prefer swapping with a task moving to its preferred node over a
1835 if (env->best_task && cur->numa_preferred_nid == env->src_nid &&
1836 env->best_task->numa_preferred_nid != env->src_nid) {
1841 * If the NUMA importance is less than SMALLIMP,
1842 * task migration might only result in ping pong
1843 * of tasks and also hurt performance due to cache
1846 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1850 * In the overloaded case, try and keep the load balanced.
1852 load = task_h_load(env->p) - task_h_load(cur);
1856 dst_load = env->dst_stats.load + load;
1857 src_load = env->src_stats.load - load;
1859 if (load_too_imbalanced(src_load, dst_load, env))
1863 /* Evaluate an idle CPU for a task numa move. */
1865 int cpu = env->dst_stats.idle_cpu;
1867 /* Nothing cached so current CPU went idle since the search. */
1872 * If the CPU is no longer truly idle and the previous best CPU
1873 * is, keep using it.
1875 if (!idle_cpu(cpu) && env->best_cpu >= 0 &&
1876 idle_cpu(env->best_cpu)) {
1877 cpu = env->best_cpu;
1883 task_numa_assign(env, cur, imp);
1886 * If a move to idle is allowed because there is capacity or load
1887 * balance improves then stop the search. While a better swap
1888 * candidate may exist, a search is not free.
1890 if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu))
1894 * If a swap candidate must be identified and the current best task
1895 * moves its preferred node then stop the search.
1897 if (!maymove && env->best_task &&
1898 env->best_task->numa_preferred_nid == env->src_nid) {
1907 static void task_numa_find_cpu(struct task_numa_env *env,
1908 long taskimp, long groupimp)
1910 bool maymove = false;
1914 * If dst node has spare capacity, then check if there is an
1915 * imbalance that would be overruled by the load balancer.
1917 if (env->dst_stats.node_type == node_has_spare) {
1918 unsigned int imbalance;
1919 int src_running, dst_running;
1922 * Would movement cause an imbalance? Note that if src has
1923 * more running tasks that the imbalance is ignored as the
1924 * move improves the imbalance from the perspective of the
1925 * CPU load balancer.
1927 src_running = env->src_stats.nr_running - 1;
1928 dst_running = env->dst_stats.nr_running + 1;
1929 imbalance = max(0, dst_running - src_running);
1930 imbalance = adjust_numa_imbalance(imbalance, src_running);
1932 /* Use idle CPU if there is no imbalance */
1935 if (env->dst_stats.idle_cpu >= 0) {
1936 env->dst_cpu = env->dst_stats.idle_cpu;
1937 task_numa_assign(env, NULL, 0);
1942 long src_load, dst_load, load;
1944 * If the improvement from just moving env->p direction is better
1945 * than swapping tasks around, check if a move is possible.
1947 load = task_h_load(env->p);
1948 dst_load = env->dst_stats.load + load;
1949 src_load = env->src_stats.load - load;
1950 maymove = !load_too_imbalanced(src_load, dst_load, env);
1953 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1954 /* Skip this CPU if the source task cannot migrate */
1955 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
1959 if (task_numa_compare(env, taskimp, groupimp, maymove))
1964 static int task_numa_migrate(struct task_struct *p)
1966 struct task_numa_env env = {
1969 .src_cpu = task_cpu(p),
1970 .src_nid = task_node(p),
1972 .imbalance_pct = 112,
1978 unsigned long taskweight, groupweight;
1979 struct sched_domain *sd;
1980 long taskimp, groupimp;
1981 struct numa_group *ng;
1986 * Pick the lowest SD_NUMA domain, as that would have the smallest
1987 * imbalance and would be the first to start moving tasks about.
1989 * And we want to avoid any moving of tasks about, as that would create
1990 * random movement of tasks -- counter the numa conditions we're trying
1994 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1996 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
2000 * Cpusets can break the scheduler domain tree into smaller
2001 * balance domains, some of which do not cross NUMA boundaries.
2002 * Tasks that are "trapped" in such domains cannot be migrated
2003 * elsewhere, so there is no point in (re)trying.
2005 if (unlikely(!sd)) {
2006 sched_setnuma(p, task_node(p));
2010 env.dst_nid = p->numa_preferred_nid;
2011 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
2012 taskweight = task_weight(p, env.src_nid, dist);
2013 groupweight = group_weight(p, env.src_nid, dist);
2014 update_numa_stats(&env, &env.src_stats, env.src_nid, false);
2015 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
2016 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
2017 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2019 /* Try to find a spot on the preferred nid. */
2020 task_numa_find_cpu(&env, taskimp, groupimp);
2023 * Look at other nodes in these cases:
2024 * - there is no space available on the preferred_nid
2025 * - the task is part of a numa_group that is interleaved across
2026 * multiple NUMA nodes; in order to better consolidate the group,
2027 * we need to check other locations.
2029 ng = deref_curr_numa_group(p);
2030 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
2031 for_each_online_node(nid) {
2032 if (nid == env.src_nid || nid == p->numa_preferred_nid)
2035 dist = node_distance(env.src_nid, env.dst_nid);
2036 if (sched_numa_topology_type == NUMA_BACKPLANE &&
2038 taskweight = task_weight(p, env.src_nid, dist);
2039 groupweight = group_weight(p, env.src_nid, dist);
2042 /* Only consider nodes where both task and groups benefit */
2043 taskimp = task_weight(p, nid, dist) - taskweight;
2044 groupimp = group_weight(p, nid, dist) - groupweight;
2045 if (taskimp < 0 && groupimp < 0)
2050 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2051 task_numa_find_cpu(&env, taskimp, groupimp);
2056 * If the task is part of a workload that spans multiple NUMA nodes,
2057 * and is migrating into one of the workload's active nodes, remember
2058 * this node as the task's preferred numa node, so the workload can
2060 * A task that migrated to a second choice node will be better off
2061 * trying for a better one later. Do not set the preferred node here.
2064 if (env.best_cpu == -1)
2067 nid = cpu_to_node(env.best_cpu);
2069 if (nid != p->numa_preferred_nid)
2070 sched_setnuma(p, nid);
2073 /* No better CPU than the current one was found. */
2074 if (env.best_cpu == -1) {
2075 trace_sched_stick_numa(p, env.src_cpu, NULL, -1);
2079 best_rq = cpu_rq(env.best_cpu);
2080 if (env.best_task == NULL) {
2081 ret = migrate_task_to(p, env.best_cpu);
2082 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2084 trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu);
2088 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
2089 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2092 trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu);
2093 put_task_struct(env.best_task);
2097 /* Attempt to migrate a task to a CPU on the preferred node. */
2098 static void numa_migrate_preferred(struct task_struct *p)
2100 unsigned long interval = HZ;
2102 /* This task has no NUMA fault statistics yet */
2103 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
2106 /* Periodically retry migrating the task to the preferred node */
2107 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
2108 p->numa_migrate_retry = jiffies + interval;
2110 /* Success if task is already running on preferred CPU */
2111 if (task_node(p) == p->numa_preferred_nid)
2114 /* Otherwise, try migrate to a CPU on the preferred node */
2115 task_numa_migrate(p);
2119 * Find out how many nodes on the workload is actively running on. Do this by
2120 * tracking the nodes from which NUMA hinting faults are triggered. This can
2121 * be different from the set of nodes where the workload's memory is currently
2124 static void numa_group_count_active_nodes(struct numa_group *numa_group)
2126 unsigned long faults, max_faults = 0;
2127 int nid, active_nodes = 0;
2129 for_each_online_node(nid) {
2130 faults = group_faults_cpu(numa_group, nid);
2131 if (faults > max_faults)
2132 max_faults = faults;
2135 for_each_online_node(nid) {
2136 faults = group_faults_cpu(numa_group, nid);
2137 if (faults * ACTIVE_NODE_FRACTION > max_faults)
2141 numa_group->max_faults_cpu = max_faults;
2142 numa_group->active_nodes = active_nodes;
2146 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
2147 * increments. The more local the fault statistics are, the higher the scan
2148 * period will be for the next scan window. If local/(local+remote) ratio is
2149 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
2150 * the scan period will decrease. Aim for 70% local accesses.
2152 #define NUMA_PERIOD_SLOTS 10
2153 #define NUMA_PERIOD_THRESHOLD 7
2156 * Increase the scan period (slow down scanning) if the majority of
2157 * our memory is already on our local node, or if the majority of
2158 * the page accesses are shared with other processes.
2159 * Otherwise, decrease the scan period.
2161 static void update_task_scan_period(struct task_struct *p,
2162 unsigned long shared, unsigned long private)
2164 unsigned int period_slot;
2165 int lr_ratio, ps_ratio;
2168 unsigned long remote = p->numa_faults_locality[0];
2169 unsigned long local = p->numa_faults_locality[1];
2172 * If there were no record hinting faults then either the task is
2173 * completely idle or all activity is areas that are not of interest
2174 * to automatic numa balancing. Related to that, if there were failed
2175 * migration then it implies we are migrating too quickly or the local
2176 * node is overloaded. In either case, scan slower
2178 if (local + shared == 0 || p->numa_faults_locality[2]) {
2179 p->numa_scan_period = min(p->numa_scan_period_max,
2180 p->numa_scan_period << 1);
2182 p->mm->numa_next_scan = jiffies +
2183 msecs_to_jiffies(p->numa_scan_period);
2189 * Prepare to scale scan period relative to the current period.
2190 * == NUMA_PERIOD_THRESHOLD scan period stays the same
2191 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2192 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2194 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
2195 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
2196 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
2198 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
2200 * Most memory accesses are local. There is no need to
2201 * do fast NUMA scanning, since memory is already local.
2203 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2206 diff = slot * period_slot;
2207 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2209 * Most memory accesses are shared with other tasks.
2210 * There is no point in continuing fast NUMA scanning,
2211 * since other tasks may just move the memory elsewhere.
2213 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2216 diff = slot * period_slot;
2219 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2220 * yet they are not on the local NUMA node. Speed up
2221 * NUMA scanning to get the memory moved over.
2223 int ratio = max(lr_ratio, ps_ratio);
2224 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2227 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2228 task_scan_min(p), task_scan_max(p));
2229 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2233 * Get the fraction of time the task has been running since the last
2234 * NUMA placement cycle. The scheduler keeps similar statistics, but
2235 * decays those on a 32ms period, which is orders of magnitude off
2236 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2237 * stats only if the task is so new there are no NUMA statistics yet.
2239 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2241 u64 runtime, delta, now;
2242 /* Use the start of this time slice to avoid calculations. */
2243 now = p->se.exec_start;
2244 runtime = p->se.sum_exec_runtime;
2246 if (p->last_task_numa_placement) {
2247 delta = runtime - p->last_sum_exec_runtime;
2248 *period = now - p->last_task_numa_placement;
2250 /* Avoid time going backwards, prevent potential divide error: */
2251 if (unlikely((s64)*period < 0))
2254 delta = p->se.avg.load_sum;
2255 *period = LOAD_AVG_MAX;
2258 p->last_sum_exec_runtime = runtime;
2259 p->last_task_numa_placement = now;
2265 * Determine the preferred nid for a task in a numa_group. This needs to
2266 * be done in a way that produces consistent results with group_weight,
2267 * otherwise workloads might not converge.
2269 static int preferred_group_nid(struct task_struct *p, int nid)
2274 /* Direct connections between all NUMA nodes. */
2275 if (sched_numa_topology_type == NUMA_DIRECT)
2279 * On a system with glueless mesh NUMA topology, group_weight
2280 * scores nodes according to the number of NUMA hinting faults on
2281 * both the node itself, and on nearby nodes.
2283 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2284 unsigned long score, max_score = 0;
2285 int node, max_node = nid;
2287 dist = sched_max_numa_distance;
2289 for_each_online_node(node) {
2290 score = group_weight(p, node, dist);
2291 if (score > max_score) {
2300 * Finding the preferred nid in a system with NUMA backplane
2301 * interconnect topology is more involved. The goal is to locate
2302 * tasks from numa_groups near each other in the system, and
2303 * untangle workloads from different sides of the system. This requires
2304 * searching down the hierarchy of node groups, recursively searching
2305 * inside the highest scoring group of nodes. The nodemask tricks
2306 * keep the complexity of the search down.
2308 nodes = node_online_map;
2309 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2310 unsigned long max_faults = 0;
2311 nodemask_t max_group = NODE_MASK_NONE;
2314 /* Are there nodes at this distance from each other? */
2315 if (!find_numa_distance(dist))
2318 for_each_node_mask(a, nodes) {
2319 unsigned long faults = 0;
2320 nodemask_t this_group;
2321 nodes_clear(this_group);
2323 /* Sum group's NUMA faults; includes a==b case. */
2324 for_each_node_mask(b, nodes) {
2325 if (node_distance(a, b) < dist) {
2326 faults += group_faults(p, b);
2327 node_set(b, this_group);
2328 node_clear(b, nodes);
2332 /* Remember the top group. */
2333 if (faults > max_faults) {
2334 max_faults = faults;
2335 max_group = this_group;
2337 * subtle: at the smallest distance there is
2338 * just one node left in each "group", the
2339 * winner is the preferred nid.
2344 /* Next round, evaluate the nodes within max_group. */
2352 static void task_numa_placement(struct task_struct *p)
2354 int seq, nid, max_nid = NUMA_NO_NODE;
2355 unsigned long max_faults = 0;
2356 unsigned long fault_types[2] = { 0, 0 };
2357 unsigned long total_faults;
2358 u64 runtime, period;
2359 spinlock_t *group_lock = NULL;
2360 struct numa_group *ng;
2363 * The p->mm->numa_scan_seq field gets updated without
2364 * exclusive access. Use READ_ONCE() here to ensure
2365 * that the field is read in a single access:
2367 seq = READ_ONCE(p->mm->numa_scan_seq);
2368 if (p->numa_scan_seq == seq)
2370 p->numa_scan_seq = seq;
2371 p->numa_scan_period_max = task_scan_max(p);
2373 total_faults = p->numa_faults_locality[0] +
2374 p->numa_faults_locality[1];
2375 runtime = numa_get_avg_runtime(p, &period);
2377 /* If the task is part of a group prevent parallel updates to group stats */
2378 ng = deref_curr_numa_group(p);
2380 group_lock = &ng->lock;
2381 spin_lock_irq(group_lock);
2384 /* Find the node with the highest number of faults */
2385 for_each_online_node(nid) {
2386 /* Keep track of the offsets in numa_faults array */
2387 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2388 unsigned long faults = 0, group_faults = 0;
2391 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2392 long diff, f_diff, f_weight;
2394 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2395 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2396 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2397 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2399 /* Decay existing window, copy faults since last scan */
2400 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2401 fault_types[priv] += p->numa_faults[membuf_idx];
2402 p->numa_faults[membuf_idx] = 0;
2405 * Normalize the faults_from, so all tasks in a group
2406 * count according to CPU use, instead of by the raw
2407 * number of faults. Tasks with little runtime have
2408 * little over-all impact on throughput, and thus their
2409 * faults are less important.
2411 f_weight = div64_u64(runtime << 16, period + 1);
2412 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2414 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2415 p->numa_faults[cpubuf_idx] = 0;
2417 p->numa_faults[mem_idx] += diff;
2418 p->numa_faults[cpu_idx] += f_diff;
2419 faults += p->numa_faults[mem_idx];
2420 p->total_numa_faults += diff;
2423 * safe because we can only change our own group
2425 * mem_idx represents the offset for a given
2426 * nid and priv in a specific region because it
2427 * is at the beginning of the numa_faults array.
2429 ng->faults[mem_idx] += diff;
2430 ng->faults_cpu[mem_idx] += f_diff;
2431 ng->total_faults += diff;
2432 group_faults += ng->faults[mem_idx];
2437 if (faults > max_faults) {
2438 max_faults = faults;
2441 } else if (group_faults > max_faults) {
2442 max_faults = group_faults;
2448 numa_group_count_active_nodes(ng);
2449 spin_unlock_irq(group_lock);
2450 max_nid = preferred_group_nid(p, max_nid);
2454 /* Set the new preferred node */
2455 if (max_nid != p->numa_preferred_nid)
2456 sched_setnuma(p, max_nid);
2459 update_task_scan_period(p, fault_types[0], fault_types[1]);
2462 static inline int get_numa_group(struct numa_group *grp)
2464 return refcount_inc_not_zero(&grp->refcount);
2467 static inline void put_numa_group(struct numa_group *grp)
2469 if (refcount_dec_and_test(&grp->refcount))
2470 kfree_rcu(grp, rcu);
2473 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2476 struct numa_group *grp, *my_grp;
2477 struct task_struct *tsk;
2479 int cpu = cpupid_to_cpu(cpupid);
2482 if (unlikely(!deref_curr_numa_group(p))) {
2483 unsigned int size = sizeof(struct numa_group) +
2484 4*nr_node_ids*sizeof(unsigned long);
2486 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2490 refcount_set(&grp->refcount, 1);
2491 grp->active_nodes = 1;
2492 grp->max_faults_cpu = 0;
2493 spin_lock_init(&grp->lock);
2495 /* Second half of the array tracks nids where faults happen */
2496 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2499 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2500 grp->faults[i] = p->numa_faults[i];
2502 grp->total_faults = p->total_numa_faults;
2505 rcu_assign_pointer(p->numa_group, grp);
2509 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2511 if (!cpupid_match_pid(tsk, cpupid))
2514 grp = rcu_dereference(tsk->numa_group);
2518 my_grp = deref_curr_numa_group(p);
2523 * Only join the other group if its bigger; if we're the bigger group,
2524 * the other task will join us.
2526 if (my_grp->nr_tasks > grp->nr_tasks)
2530 * Tie-break on the grp address.
2532 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2535 /* Always join threads in the same process. */
2536 if (tsk->mm == current->mm)
2539 /* Simple filter to avoid false positives due to PID collisions */
2540 if (flags & TNF_SHARED)
2543 /* Update priv based on whether false sharing was detected */
2546 if (join && !get_numa_group(grp))
2554 BUG_ON(irqs_disabled());
2555 double_lock_irq(&my_grp->lock, &grp->lock);
2557 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2558 my_grp->faults[i] -= p->numa_faults[i];
2559 grp->faults[i] += p->numa_faults[i];
2561 my_grp->total_faults -= p->total_numa_faults;
2562 grp->total_faults += p->total_numa_faults;
2567 spin_unlock(&my_grp->lock);
2568 spin_unlock_irq(&grp->lock);
2570 rcu_assign_pointer(p->numa_group, grp);
2572 put_numa_group(my_grp);
2581 * Get rid of NUMA staticstics associated with a task (either current or dead).
2582 * If @final is set, the task is dead and has reached refcount zero, so we can
2583 * safely free all relevant data structures. Otherwise, there might be
2584 * concurrent reads from places like load balancing and procfs, and we should
2585 * reset the data back to default state without freeing ->numa_faults.
2587 void task_numa_free(struct task_struct *p, bool final)
2589 /* safe: p either is current or is being freed by current */
2590 struct numa_group *grp = rcu_dereference_raw(p->numa_group);
2591 unsigned long *numa_faults = p->numa_faults;
2592 unsigned long flags;
2599 spin_lock_irqsave(&grp->lock, flags);
2600 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2601 grp->faults[i] -= p->numa_faults[i];
2602 grp->total_faults -= p->total_numa_faults;
2605 spin_unlock_irqrestore(&grp->lock, flags);
2606 RCU_INIT_POINTER(p->numa_group, NULL);
2607 put_numa_group(grp);
2611 p->numa_faults = NULL;
2614 p->total_numa_faults = 0;
2615 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2621 * Got a PROT_NONE fault for a page on @node.
2623 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2625 struct task_struct *p = current;
2626 bool migrated = flags & TNF_MIGRATED;
2627 int cpu_node = task_node(current);
2628 int local = !!(flags & TNF_FAULT_LOCAL);
2629 struct numa_group *ng;
2632 if (!static_branch_likely(&sched_numa_balancing))
2635 /* for example, ksmd faulting in a user's mm */
2639 /* Allocate buffer to track faults on a per-node basis */
2640 if (unlikely(!p->numa_faults)) {
2641 int size = sizeof(*p->numa_faults) *
2642 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2644 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2645 if (!p->numa_faults)
2648 p->total_numa_faults = 0;
2649 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2653 * First accesses are treated as private, otherwise consider accesses
2654 * to be private if the accessing pid has not changed
2656 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2659 priv = cpupid_match_pid(p, last_cpupid);
2660 if (!priv && !(flags & TNF_NO_GROUP))
2661 task_numa_group(p, last_cpupid, flags, &priv);
2665 * If a workload spans multiple NUMA nodes, a shared fault that
2666 * occurs wholly within the set of nodes that the workload is
2667 * actively using should be counted as local. This allows the
2668 * scan rate to slow down when a workload has settled down.
2670 ng = deref_curr_numa_group(p);
2671 if (!priv && !local && ng && ng->active_nodes > 1 &&
2672 numa_is_active_node(cpu_node, ng) &&
2673 numa_is_active_node(mem_node, ng))
2677 * Retry to migrate task to preferred node periodically, in case it
2678 * previously failed, or the scheduler moved us.
2680 if (time_after(jiffies, p->numa_migrate_retry)) {
2681 task_numa_placement(p);
2682 numa_migrate_preferred(p);
2686 p->numa_pages_migrated += pages;
2687 if (flags & TNF_MIGRATE_FAIL)
2688 p->numa_faults_locality[2] += pages;
2690 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2691 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2692 p->numa_faults_locality[local] += pages;
2695 static void reset_ptenuma_scan(struct task_struct *p)
2698 * We only did a read acquisition of the mmap sem, so
2699 * p->mm->numa_scan_seq is written to without exclusive access
2700 * and the update is not guaranteed to be atomic. That's not
2701 * much of an issue though, since this is just used for
2702 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2703 * expensive, to avoid any form of compiler optimizations:
2705 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2706 p->mm->numa_scan_offset = 0;
2710 * The expensive part of numa migration is done from task_work context.
2711 * Triggered from task_tick_numa().
2713 static void task_numa_work(struct callback_head *work)
2715 unsigned long migrate, next_scan, now = jiffies;
2716 struct task_struct *p = current;
2717 struct mm_struct *mm = p->mm;
2718 u64 runtime = p->se.sum_exec_runtime;
2719 struct vm_area_struct *vma;
2720 unsigned long start, end;
2721 unsigned long nr_pte_updates = 0;
2722 long pages, virtpages;
2724 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2728 * Who cares about NUMA placement when they're dying.
2730 * NOTE: make sure not to dereference p->mm before this check,
2731 * exit_task_work() happens _after_ exit_mm() so we could be called
2732 * without p->mm even though we still had it when we enqueued this
2735 if (p->flags & PF_EXITING)
2738 if (!mm->numa_next_scan) {
2739 mm->numa_next_scan = now +
2740 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2744 * Enforce maximal scan/migration frequency..
2746 migrate = mm->numa_next_scan;
2747 if (time_before(now, migrate))
2750 if (p->numa_scan_period == 0) {
2751 p->numa_scan_period_max = task_scan_max(p);
2752 p->numa_scan_period = task_scan_start(p);
2755 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2756 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2760 * Delay this task enough that another task of this mm will likely win
2761 * the next time around.
2763 p->node_stamp += 2 * TICK_NSEC;
2765 start = mm->numa_scan_offset;
2766 pages = sysctl_numa_balancing_scan_size;
2767 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2768 virtpages = pages * 8; /* Scan up to this much virtual space */
2773 if (!down_read_trylock(&mm->mmap_sem))
2775 vma = find_vma(mm, start);
2777 reset_ptenuma_scan(p);
2781 for (; vma; vma = vma->vm_next) {
2782 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2783 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2788 * Shared library pages mapped by multiple processes are not
2789 * migrated as it is expected they are cache replicated. Avoid
2790 * hinting faults in read-only file-backed mappings or the vdso
2791 * as migrating the pages will be of marginal benefit.
2794 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2798 * Skip inaccessible VMAs to avoid any confusion between
2799 * PROT_NONE and NUMA hinting ptes
2801 if (!vma_is_accessible(vma))
2805 start = max(start, vma->vm_start);
2806 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2807 end = min(end, vma->vm_end);
2808 nr_pte_updates = change_prot_numa(vma, start, end);
2811 * Try to scan sysctl_numa_balancing_size worth of
2812 * hpages that have at least one present PTE that
2813 * is not already pte-numa. If the VMA contains
2814 * areas that are unused or already full of prot_numa
2815 * PTEs, scan up to virtpages, to skip through those
2819 pages -= (end - start) >> PAGE_SHIFT;
2820 virtpages -= (end - start) >> PAGE_SHIFT;
2823 if (pages <= 0 || virtpages <= 0)
2827 } while (end != vma->vm_end);
2832 * It is possible to reach the end of the VMA list but the last few
2833 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2834 * would find the !migratable VMA on the next scan but not reset the
2835 * scanner to the start so check it now.
2838 mm->numa_scan_offset = start;
2840 reset_ptenuma_scan(p);
2841 up_read(&mm->mmap_sem);
2844 * Make sure tasks use at least 32x as much time to run other code
2845 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2846 * Usually update_task_scan_period slows down scanning enough; on an
2847 * overloaded system we need to limit overhead on a per task basis.
2849 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2850 u64 diff = p->se.sum_exec_runtime - runtime;
2851 p->node_stamp += 32 * diff;
2855 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
2858 struct mm_struct *mm = p->mm;
2861 mm_users = atomic_read(&mm->mm_users);
2862 if (mm_users == 1) {
2863 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2864 mm->numa_scan_seq = 0;
2868 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
2869 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2870 /* Protect against double add, see task_tick_numa and task_numa_work */
2871 p->numa_work.next = &p->numa_work;
2872 p->numa_faults = NULL;
2873 RCU_INIT_POINTER(p->numa_group, NULL);
2874 p->last_task_numa_placement = 0;
2875 p->last_sum_exec_runtime = 0;
2877 init_task_work(&p->numa_work, task_numa_work);
2879 /* New address space, reset the preferred nid */
2880 if (!(clone_flags & CLONE_VM)) {
2881 p->numa_preferred_nid = NUMA_NO_NODE;
2886 * New thread, keep existing numa_preferred_nid which should be copied
2887 * already by arch_dup_task_struct but stagger when scans start.
2892 delay = min_t(unsigned int, task_scan_max(current),
2893 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
2894 delay += 2 * TICK_NSEC;
2895 p->node_stamp = delay;
2900 * Drive the periodic memory faults..
2902 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2904 struct callback_head *work = &curr->numa_work;
2908 * We don't care about NUMA placement if we don't have memory.
2910 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2914 * Using runtime rather than walltime has the dual advantage that
2915 * we (mostly) drive the selection from busy threads and that the
2916 * task needs to have done some actual work before we bother with
2919 now = curr->se.sum_exec_runtime;
2920 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2922 if (now > curr->node_stamp + period) {
2923 if (!curr->node_stamp)
2924 curr->numa_scan_period = task_scan_start(curr);
2925 curr->node_stamp += period;
2927 if (!time_before(jiffies, curr->mm->numa_next_scan))
2928 task_work_add(curr, work, true);
2932 static void update_scan_period(struct task_struct *p, int new_cpu)
2934 int src_nid = cpu_to_node(task_cpu(p));
2935 int dst_nid = cpu_to_node(new_cpu);
2937 if (!static_branch_likely(&sched_numa_balancing))
2940 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2943 if (src_nid == dst_nid)
2947 * Allow resets if faults have been trapped before one scan
2948 * has completed. This is most likely due to a new task that
2949 * is pulled cross-node due to wakeups or load balancing.
2951 if (p->numa_scan_seq) {
2953 * Avoid scan adjustments if moving to the preferred
2954 * node or if the task was not previously running on
2955 * the preferred node.
2957 if (dst_nid == p->numa_preferred_nid ||
2958 (p->numa_preferred_nid != NUMA_NO_NODE &&
2959 src_nid != p->numa_preferred_nid))
2963 p->numa_scan_period = task_scan_start(p);
2967 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2971 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2975 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2979 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2983 #endif /* CONFIG_NUMA_BALANCING */
2986 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2988 update_load_add(&cfs_rq->load, se->load.weight);
2990 if (entity_is_task(se)) {
2991 struct rq *rq = rq_of(cfs_rq);
2993 account_numa_enqueue(rq, task_of(se));
2994 list_add(&se->group_node, &rq->cfs_tasks);
2997 cfs_rq->nr_running++;
3001 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3003 update_load_sub(&cfs_rq->load, se->load.weight);
3005 if (entity_is_task(se)) {
3006 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
3007 list_del_init(&se->group_node);
3010 cfs_rq->nr_running--;
3014 * Signed add and clamp on underflow.
3016 * Explicitly do a load-store to ensure the intermediate value never hits
3017 * memory. This allows lockless observations without ever seeing the negative
3020 #define add_positive(_ptr, _val) do { \
3021 typeof(_ptr) ptr = (_ptr); \
3022 typeof(_val) val = (_val); \
3023 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3027 if (val < 0 && res > var) \
3030 WRITE_ONCE(*ptr, res); \
3034 * Unsigned subtract and clamp on underflow.
3036 * Explicitly do a load-store to ensure the intermediate value never hits
3037 * memory. This allows lockless observations without ever seeing the negative
3040 #define sub_positive(_ptr, _val) do { \
3041 typeof(_ptr) ptr = (_ptr); \
3042 typeof(*ptr) val = (_val); \
3043 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3047 WRITE_ONCE(*ptr, res); \
3051 * Remove and clamp on negative, from a local variable.
3053 * A variant of sub_positive(), which does not use explicit load-store
3054 * and is thus optimized for local variable updates.
3056 #define lsub_positive(_ptr, _val) do { \
3057 typeof(_ptr) ptr = (_ptr); \
3058 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
3063 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3065 cfs_rq->avg.load_avg += se->avg.load_avg;
3066 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
3070 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3072 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3073 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
3077 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3079 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3082 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
3083 unsigned long weight)
3086 /* commit outstanding execution time */
3087 if (cfs_rq->curr == se)
3088 update_curr(cfs_rq);
3089 account_entity_dequeue(cfs_rq, se);
3091 dequeue_load_avg(cfs_rq, se);
3093 update_load_set(&se->load, weight);
3097 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
3099 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
3103 enqueue_load_avg(cfs_rq, se);
3105 account_entity_enqueue(cfs_rq, se);
3109 void reweight_task(struct task_struct *p, int prio)
3111 struct sched_entity *se = &p->se;
3112 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3113 struct load_weight *load = &se->load;
3114 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
3116 reweight_entity(cfs_rq, se, weight);
3117 load->inv_weight = sched_prio_to_wmult[prio];
3120 #ifdef CONFIG_FAIR_GROUP_SCHED
3123 * All this does is approximate the hierarchical proportion which includes that
3124 * global sum we all love to hate.
3126 * That is, the weight of a group entity, is the proportional share of the
3127 * group weight based on the group runqueue weights. That is:
3129 * tg->weight * grq->load.weight
3130 * ge->load.weight = ----------------------------- (1)
3131 * \Sum grq->load.weight
3133 * Now, because computing that sum is prohibitively expensive to compute (been
3134 * there, done that) we approximate it with this average stuff. The average
3135 * moves slower and therefore the approximation is cheaper and more stable.
3137 * So instead of the above, we substitute:
3139 * grq->load.weight -> grq->avg.load_avg (2)
3141 * which yields the following:
3143 * tg->weight * grq->avg.load_avg
3144 * ge->load.weight = ------------------------------ (3)
3147 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3149 * That is shares_avg, and it is right (given the approximation (2)).
3151 * The problem with it is that because the average is slow -- it was designed
3152 * to be exactly that of course -- this leads to transients in boundary
3153 * conditions. In specific, the case where the group was idle and we start the
3154 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3155 * yielding bad latency etc..
3157 * Now, in that special case (1) reduces to:
3159 * tg->weight * grq->load.weight
3160 * ge->load.weight = ----------------------------- = tg->weight (4)
3163 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3165 * So what we do is modify our approximation (3) to approach (4) in the (near)
3170 * tg->weight * grq->load.weight
3171 * --------------------------------------------------- (5)
3172 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3174 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3175 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3178 * tg->weight * grq->load.weight
3179 * ge->load.weight = ----------------------------- (6)
3184 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3185 * max(grq->load.weight, grq->avg.load_avg)
3187 * And that is shares_weight and is icky. In the (near) UP case it approaches
3188 * (4) while in the normal case it approaches (3). It consistently
3189 * overestimates the ge->load.weight and therefore:
3191 * \Sum ge->load.weight >= tg->weight
3195 static long calc_group_shares(struct cfs_rq *cfs_rq)
3197 long tg_weight, tg_shares, load, shares;
3198 struct task_group *tg = cfs_rq->tg;
3200 tg_shares = READ_ONCE(tg->shares);
3202 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3204 tg_weight = atomic_long_read(&tg->load_avg);
3206 /* Ensure tg_weight >= load */
3207 tg_weight -= cfs_rq->tg_load_avg_contrib;
3210 shares = (tg_shares * load);
3212 shares /= tg_weight;
3215 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3216 * of a group with small tg->shares value. It is a floor value which is
3217 * assigned as a minimum load.weight to the sched_entity representing
3218 * the group on a CPU.
3220 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3221 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3222 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3223 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3226 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3228 #endif /* CONFIG_SMP */
3230 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3233 * Recomputes the group entity based on the current state of its group
3236 static void update_cfs_group(struct sched_entity *se)
3238 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3244 if (throttled_hierarchy(gcfs_rq))
3248 shares = READ_ONCE(gcfs_rq->tg->shares);
3250 if (likely(se->load.weight == shares))
3253 shares = calc_group_shares(gcfs_rq);
3256 reweight_entity(cfs_rq_of(se), se, shares);
3259 #else /* CONFIG_FAIR_GROUP_SCHED */
3260 static inline void update_cfs_group(struct sched_entity *se)
3263 #endif /* CONFIG_FAIR_GROUP_SCHED */
3265 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3267 struct rq *rq = rq_of(cfs_rq);
3269 if (&rq->cfs == cfs_rq) {
3271 * There are a few boundary cases this might miss but it should
3272 * get called often enough that that should (hopefully) not be
3275 * It will not get called when we go idle, because the idle
3276 * thread is a different class (!fair), nor will the utilization
3277 * number include things like RT tasks.
3279 * As is, the util number is not freq-invariant (we'd have to
3280 * implement arch_scale_freq_capacity() for that).
3284 cpufreq_update_util(rq, flags);
3289 #ifdef CONFIG_FAIR_GROUP_SCHED
3291 * update_tg_load_avg - update the tg's load avg
3292 * @cfs_rq: the cfs_rq whose avg changed
3293 * @force: update regardless of how small the difference
3295 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3296 * However, because tg->load_avg is a global value there are performance
3299 * In order to avoid having to look at the other cfs_rq's, we use a
3300 * differential update where we store the last value we propagated. This in
3301 * turn allows skipping updates if the differential is 'small'.
3303 * Updating tg's load_avg is necessary before update_cfs_share().
3305 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3307 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3310 * No need to update load_avg for root_task_group as it is not used.
3312 if (cfs_rq->tg == &root_task_group)
3315 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3316 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3317 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3322 * Called within set_task_rq() right before setting a task's CPU. The
3323 * caller only guarantees p->pi_lock is held; no other assumptions,
3324 * including the state of rq->lock, should be made.
3326 void set_task_rq_fair(struct sched_entity *se,
3327 struct cfs_rq *prev, struct cfs_rq *next)
3329 u64 p_last_update_time;
3330 u64 n_last_update_time;
3332 if (!sched_feat(ATTACH_AGE_LOAD))
3336 * We are supposed to update the task to "current" time, then its up to
3337 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3338 * getting what current time is, so simply throw away the out-of-date
3339 * time. This will result in the wakee task is less decayed, but giving
3340 * the wakee more load sounds not bad.
3342 if (!(se->avg.last_update_time && prev))
3345 #ifndef CONFIG_64BIT
3347 u64 p_last_update_time_copy;
3348 u64 n_last_update_time_copy;
3351 p_last_update_time_copy = prev->load_last_update_time_copy;
3352 n_last_update_time_copy = next->load_last_update_time_copy;
3356 p_last_update_time = prev->avg.last_update_time;
3357 n_last_update_time = next->avg.last_update_time;
3359 } while (p_last_update_time != p_last_update_time_copy ||
3360 n_last_update_time != n_last_update_time_copy);
3363 p_last_update_time = prev->avg.last_update_time;
3364 n_last_update_time = next->avg.last_update_time;
3366 __update_load_avg_blocked_se(p_last_update_time, se);
3367 se->avg.last_update_time = n_last_update_time;
3372 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3373 * propagate its contribution. The key to this propagation is the invariant
3374 * that for each group:
3376 * ge->avg == grq->avg (1)
3378 * _IFF_ we look at the pure running and runnable sums. Because they
3379 * represent the very same entity, just at different points in the hierarchy.
3381 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
3382 * and simply copies the running/runnable sum over (but still wrong, because
3383 * the group entity and group rq do not have their PELT windows aligned).
3385 * However, update_tg_cfs_load() is more complex. So we have:
3387 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3389 * And since, like util, the runnable part should be directly transferable,
3390 * the following would _appear_ to be the straight forward approach:
3392 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3394 * And per (1) we have:
3396 * ge->avg.runnable_avg == grq->avg.runnable_avg
3400 * ge->load.weight * grq->avg.load_avg
3401 * ge->avg.load_avg = ----------------------------------- (4)
3404 * Except that is wrong!
3406 * Because while for entities historical weight is not important and we
3407 * really only care about our future and therefore can consider a pure
3408 * runnable sum, runqueues can NOT do this.
3410 * We specifically want runqueues to have a load_avg that includes
3411 * historical weights. Those represent the blocked load, the load we expect
3412 * to (shortly) return to us. This only works by keeping the weights as
3413 * integral part of the sum. We therefore cannot decompose as per (3).
3415 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3416 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3417 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3418 * runnable section of these tasks overlap (or not). If they were to perfectly
3419 * align the rq as a whole would be runnable 2/3 of the time. If however we
3420 * always have at least 1 runnable task, the rq as a whole is always runnable.
3422 * So we'll have to approximate.. :/
3424 * Given the constraint:
3426 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3428 * We can construct a rule that adds runnable to a rq by assuming minimal
3431 * On removal, we'll assume each task is equally runnable; which yields:
3433 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3435 * XXX: only do this for the part of runnable > running ?
3440 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3442 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3444 /* Nothing to update */
3449 * The relation between sum and avg is:
3451 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3453 * however, the PELT windows are not aligned between grq and gse.
3456 /* Set new sched_entity's utilization */
3457 se->avg.util_avg = gcfs_rq->avg.util_avg;
3458 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3460 /* Update parent cfs_rq utilization */
3461 add_positive(&cfs_rq->avg.util_avg, delta);
3462 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3466 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3468 long delta = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
3470 /* Nothing to update */
3475 * The relation between sum and avg is:
3477 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3479 * however, the PELT windows are not aligned between grq and gse.
3482 /* Set new sched_entity's runnable */
3483 se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
3484 se->avg.runnable_sum = se->avg.runnable_avg * LOAD_AVG_MAX;
3486 /* Update parent cfs_rq runnable */
3487 add_positive(&cfs_rq->avg.runnable_avg, delta);
3488 cfs_rq->avg.runnable_sum = cfs_rq->avg.runnable_avg * LOAD_AVG_MAX;
3492 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3494 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3495 unsigned long load_avg;
3502 gcfs_rq->prop_runnable_sum = 0;
3504 if (runnable_sum >= 0) {
3506 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3507 * the CPU is saturated running == runnable.
3509 runnable_sum += se->avg.load_sum;
3510 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3513 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3514 * assuming all tasks are equally runnable.
3516 if (scale_load_down(gcfs_rq->load.weight)) {
3517 load_sum = div_s64(gcfs_rq->avg.load_sum,
3518 scale_load_down(gcfs_rq->load.weight));
3521 /* But make sure to not inflate se's runnable */
3522 runnable_sum = min(se->avg.load_sum, load_sum);
3526 * runnable_sum can't be lower than running_sum
3527 * Rescale running sum to be in the same range as runnable sum
3528 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3529 * runnable_sum is in [0 : LOAD_AVG_MAX]
3531 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3532 runnable_sum = max(runnable_sum, running_sum);
3534 load_sum = (s64)se_weight(se) * runnable_sum;
3535 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3537 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3538 delta_avg = load_avg - se->avg.load_avg;
3540 se->avg.load_sum = runnable_sum;
3541 se->avg.load_avg = load_avg;
3542 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3543 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3546 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3548 cfs_rq->propagate = 1;
3549 cfs_rq->prop_runnable_sum += runnable_sum;
3552 /* Update task and its cfs_rq load average */
3553 static inline int propagate_entity_load_avg(struct sched_entity *se)
3555 struct cfs_rq *cfs_rq, *gcfs_rq;
3557 if (entity_is_task(se))
3560 gcfs_rq = group_cfs_rq(se);
3561 if (!gcfs_rq->propagate)
3564 gcfs_rq->propagate = 0;
3566 cfs_rq = cfs_rq_of(se);
3568 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3570 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3571 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3572 update_tg_cfs_load(cfs_rq, se, gcfs_rq);
3574 trace_pelt_cfs_tp(cfs_rq);
3575 trace_pelt_se_tp(se);
3581 * Check if we need to update the load and the utilization of a blocked
3584 static inline bool skip_blocked_update(struct sched_entity *se)
3586 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3589 * If sched_entity still have not zero load or utilization, we have to
3592 if (se->avg.load_avg || se->avg.util_avg)
3596 * If there is a pending propagation, we have to update the load and
3597 * the utilization of the sched_entity:
3599 if (gcfs_rq->propagate)
3603 * Otherwise, the load and the utilization of the sched_entity is
3604 * already zero and there is no pending propagation, so it will be a
3605 * waste of time to try to decay it:
3610 #else /* CONFIG_FAIR_GROUP_SCHED */
3612 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3614 static inline int propagate_entity_load_avg(struct sched_entity *se)
3619 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3621 #endif /* CONFIG_FAIR_GROUP_SCHED */
3624 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3625 * @now: current time, as per cfs_rq_clock_pelt()
3626 * @cfs_rq: cfs_rq to update
3628 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3629 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3630 * post_init_entity_util_avg().
3632 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3634 * Returns true if the load decayed or we removed load.
3636 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3637 * call update_tg_load_avg() when this function returns true.
3640 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3642 unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
3643 struct sched_avg *sa = &cfs_rq->avg;
3646 if (cfs_rq->removed.nr) {
3648 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3650 raw_spin_lock(&cfs_rq->removed.lock);
3651 swap(cfs_rq->removed.util_avg, removed_util);
3652 swap(cfs_rq->removed.load_avg, removed_load);
3653 swap(cfs_rq->removed.runnable_avg, removed_runnable);
3654 cfs_rq->removed.nr = 0;
3655 raw_spin_unlock(&cfs_rq->removed.lock);
3658 sub_positive(&sa->load_avg, r);
3659 sub_positive(&sa->load_sum, r * divider);
3662 sub_positive(&sa->util_avg, r);
3663 sub_positive(&sa->util_sum, r * divider);
3665 r = removed_runnable;
3666 sub_positive(&sa->runnable_avg, r);
3667 sub_positive(&sa->runnable_sum, r * divider);
3670 * removed_runnable is the unweighted version of removed_load so we
3671 * can use it to estimate removed_load_sum.
3673 add_tg_cfs_propagate(cfs_rq,
3674 -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
3679 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
3681 #ifndef CONFIG_64BIT
3683 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3690 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3691 * @cfs_rq: cfs_rq to attach to
3692 * @se: sched_entity to attach
3694 * Must call update_cfs_rq_load_avg() before this, since we rely on
3695 * cfs_rq->avg.last_update_time being current.
3697 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3699 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3702 * When we attach the @se to the @cfs_rq, we must align the decay
3703 * window because without that, really weird and wonderful things can
3708 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3709 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3712 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3713 * period_contrib. This isn't strictly correct, but since we're
3714 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3717 se->avg.util_sum = se->avg.util_avg * divider;
3719 se->avg.runnable_sum = se->avg.runnable_avg * divider;
3721 se->avg.load_sum = divider;
3722 if (se_weight(se)) {
3724 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3727 enqueue_load_avg(cfs_rq, se);
3728 cfs_rq->avg.util_avg += se->avg.util_avg;
3729 cfs_rq->avg.util_sum += se->avg.util_sum;
3730 cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
3731 cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
3733 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3735 cfs_rq_util_change(cfs_rq, 0);
3737 trace_pelt_cfs_tp(cfs_rq);
3741 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3742 * @cfs_rq: cfs_rq to detach from
3743 * @se: sched_entity to detach
3745 * Must call update_cfs_rq_load_avg() before this, since we rely on
3746 * cfs_rq->avg.last_update_time being current.
3748 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3750 dequeue_load_avg(cfs_rq, se);
3751 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3752 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3753 sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);
3754 sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum);
3756 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3758 cfs_rq_util_change(cfs_rq, 0);
3760 trace_pelt_cfs_tp(cfs_rq);
3764 * Optional action to be done while updating the load average
3766 #define UPDATE_TG 0x1
3767 #define SKIP_AGE_LOAD 0x2
3768 #define DO_ATTACH 0x4
3770 /* Update task and its cfs_rq load average */
3771 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3773 u64 now = cfs_rq_clock_pelt(cfs_rq);
3777 * Track task load average for carrying it to new CPU after migrated, and
3778 * track group sched_entity load average for task_h_load calc in migration
3780 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3781 __update_load_avg_se(now, cfs_rq, se);
3783 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3784 decayed |= propagate_entity_load_avg(se);
3786 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3789 * DO_ATTACH means we're here from enqueue_entity().
3790 * !last_update_time means we've passed through
3791 * migrate_task_rq_fair() indicating we migrated.
3793 * IOW we're enqueueing a task on a new CPU.
3795 attach_entity_load_avg(cfs_rq, se);
3796 update_tg_load_avg(cfs_rq, 0);
3798 } else if (decayed) {
3799 cfs_rq_util_change(cfs_rq, 0);
3801 if (flags & UPDATE_TG)
3802 update_tg_load_avg(cfs_rq, 0);
3806 #ifndef CONFIG_64BIT
3807 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3809 u64 last_update_time_copy;
3810 u64 last_update_time;
3813 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3815 last_update_time = cfs_rq->avg.last_update_time;
3816 } while (last_update_time != last_update_time_copy);
3818 return last_update_time;
3821 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3823 return cfs_rq->avg.last_update_time;
3828 * Synchronize entity load avg of dequeued entity without locking
3831 static void sync_entity_load_avg(struct sched_entity *se)
3833 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3834 u64 last_update_time;
3836 last_update_time = cfs_rq_last_update_time(cfs_rq);
3837 __update_load_avg_blocked_se(last_update_time, se);
3841 * Task first catches up with cfs_rq, and then subtract
3842 * itself from the cfs_rq (task must be off the queue now).
3844 static void remove_entity_load_avg(struct sched_entity *se)
3846 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3847 unsigned long flags;
3850 * tasks cannot exit without having gone through wake_up_new_task() ->
3851 * post_init_entity_util_avg() which will have added things to the
3852 * cfs_rq, so we can remove unconditionally.
3855 sync_entity_load_avg(se);
3857 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3858 ++cfs_rq->removed.nr;
3859 cfs_rq->removed.util_avg += se->avg.util_avg;
3860 cfs_rq->removed.load_avg += se->avg.load_avg;
3861 cfs_rq->removed.runnable_avg += se->avg.runnable_avg;
3862 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3865 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
3867 return cfs_rq->avg.runnable_avg;
3870 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3872 return cfs_rq->avg.load_avg;
3875 static inline unsigned long task_util(struct task_struct *p)
3877 return READ_ONCE(p->se.avg.util_avg);
3880 static inline unsigned long _task_util_est(struct task_struct *p)
3882 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3884 return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED);
3887 static inline unsigned long task_util_est(struct task_struct *p)
3889 return max(task_util(p), _task_util_est(p));
3892 #ifdef CONFIG_UCLAMP_TASK
3893 static inline unsigned long uclamp_task_util(struct task_struct *p)
3895 return clamp(task_util_est(p),
3896 uclamp_eff_value(p, UCLAMP_MIN),
3897 uclamp_eff_value(p, UCLAMP_MAX));
3900 static inline unsigned long uclamp_task_util(struct task_struct *p)
3902 return task_util_est(p);
3906 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3907 struct task_struct *p)
3909 unsigned int enqueued;
3911 if (!sched_feat(UTIL_EST))
3914 /* Update root cfs_rq's estimated utilization */
3915 enqueued = cfs_rq->avg.util_est.enqueued;
3916 enqueued += _task_util_est(p);
3917 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3921 * Check if a (signed) value is within a specified (unsigned) margin,
3922 * based on the observation that:
3924 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3926 * NOTE: this only works when value + maring < INT_MAX.
3928 static inline bool within_margin(int value, int margin)
3930 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3934 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3936 long last_ewma_diff;
3940 if (!sched_feat(UTIL_EST))
3943 /* Update root cfs_rq's estimated utilization */
3944 ue.enqueued = cfs_rq->avg.util_est.enqueued;
3945 ue.enqueued -= min_t(unsigned int, ue.enqueued, _task_util_est(p));
3946 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3949 * Skip update of task's estimated utilization when the task has not
3950 * yet completed an activation, e.g. being migrated.
3956 * If the PELT values haven't changed since enqueue time,
3957 * skip the util_est update.
3959 ue = p->se.avg.util_est;
3960 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3964 * Reset EWMA on utilization increases, the moving average is used only
3965 * to smooth utilization decreases.
3967 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3968 if (sched_feat(UTIL_EST_FASTUP)) {
3969 if (ue.ewma < ue.enqueued) {
3970 ue.ewma = ue.enqueued;
3976 * Skip update of task's estimated utilization when its EWMA is
3977 * already ~1% close to its last activation value.
3979 last_ewma_diff = ue.enqueued - ue.ewma;
3980 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3984 * To avoid overestimation of actual task utilization, skip updates if
3985 * we cannot grant there is idle time in this CPU.
3987 cpu = cpu_of(rq_of(cfs_rq));
3988 if (task_util(p) > capacity_orig_of(cpu))
3992 * Update Task's estimated utilization
3994 * When *p completes an activation we can consolidate another sample
3995 * of the task size. This is done by storing the current PELT value
3996 * as ue.enqueued and by using this value to update the Exponential
3997 * Weighted Moving Average (EWMA):
3999 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4000 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4001 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4002 * = w * ( last_ewma_diff ) + ewma(t-1)
4003 * = w * (last_ewma_diff + ewma(t-1) / w)
4005 * Where 'w' is the weight of new samples, which is configured to be
4006 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4008 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
4009 ue.ewma += last_ewma_diff;
4010 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
4012 WRITE_ONCE(p->se.avg.util_est, ue);
4015 static inline int task_fits_capacity(struct task_struct *p, long capacity)
4017 return fits_capacity(uclamp_task_util(p), capacity);
4020 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
4022 if (!static_branch_unlikely(&sched_asym_cpucapacity))
4026 rq->misfit_task_load = 0;
4030 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
4031 rq->misfit_task_load = 0;
4035 rq->misfit_task_load = task_h_load(p);
4038 #else /* CONFIG_SMP */
4040 #define UPDATE_TG 0x0
4041 #define SKIP_AGE_LOAD 0x0
4042 #define DO_ATTACH 0x0
4044 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
4046 cfs_rq_util_change(cfs_rq, 0);
4049 static inline void remove_entity_load_avg(struct sched_entity *se) {}
4052 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4054 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4056 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
4062 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4065 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
4067 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
4069 #endif /* CONFIG_SMP */
4071 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
4073 #ifdef CONFIG_SCHED_DEBUG
4074 s64 d = se->vruntime - cfs_rq->min_vruntime;
4079 if (d > 3*sysctl_sched_latency)
4080 schedstat_inc(cfs_rq->nr_spread_over);
4085 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
4087 u64 vruntime = cfs_rq->min_vruntime;
4090 * The 'current' period is already promised to the current tasks,
4091 * however the extra weight of the new task will slow them down a
4092 * little, place the new task so that it fits in the slot that
4093 * stays open at the end.
4095 if (initial && sched_feat(START_DEBIT))
4096 vruntime += sched_vslice(cfs_rq, se);
4098 /* sleeps up to a single latency don't count. */
4100 unsigned long thresh = sysctl_sched_latency;
4103 * Halve their sleep time's effect, to allow
4104 * for a gentler effect of sleepers:
4106 if (sched_feat(GENTLE_FAIR_SLEEPERS))
4112 /* ensure we never gain time by being placed backwards. */
4113 se->vruntime = max_vruntime(se->vruntime, vruntime);
4116 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
4118 static inline void check_schedstat_required(void)
4120 #ifdef CONFIG_SCHEDSTATS
4121 if (schedstat_enabled())
4124 /* Force schedstat enabled if a dependent tracepoint is active */
4125 if (trace_sched_stat_wait_enabled() ||
4126 trace_sched_stat_sleep_enabled() ||
4127 trace_sched_stat_iowait_enabled() ||
4128 trace_sched_stat_blocked_enabled() ||
4129 trace_sched_stat_runtime_enabled()) {
4130 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
4131 "stat_blocked and stat_runtime require the "
4132 "kernel parameter schedstats=enable or "
4133 "kernel.sched_schedstats=1\n");
4138 static inline bool cfs_bandwidth_used(void);
4145 * update_min_vruntime()
4146 * vruntime -= min_vruntime
4150 * update_min_vruntime()
4151 * vruntime += min_vruntime
4153 * this way the vruntime transition between RQs is done when both
4154 * min_vruntime are up-to-date.
4158 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
4159 * vruntime -= min_vruntime
4163 * update_min_vruntime()
4164 * vruntime += min_vruntime
4166 * this way we don't have the most up-to-date min_vruntime on the originating
4167 * CPU and an up-to-date min_vruntime on the destination CPU.
4171 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4173 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
4174 bool curr = cfs_rq->curr == se;
4177 * If we're the current task, we must renormalise before calling
4181 se->vruntime += cfs_rq->min_vruntime;
4183 update_curr(cfs_rq);
4186 * Otherwise, renormalise after, such that we're placed at the current
4187 * moment in time, instead of some random moment in the past. Being
4188 * placed in the past could significantly boost this task to the
4189 * fairness detriment of existing tasks.
4191 if (renorm && !curr)
4192 se->vruntime += cfs_rq->min_vruntime;
4195 * When enqueuing a sched_entity, we must:
4196 * - Update loads to have both entity and cfs_rq synced with now.
4197 * - Add its load to cfs_rq->runnable_avg
4198 * - For group_entity, update its weight to reflect the new share of
4200 * - Add its new weight to cfs_rq->load.weight
4202 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
4203 se_update_runnable(se);
4204 update_cfs_group(se);
4205 account_entity_enqueue(cfs_rq, se);
4207 if (flags & ENQUEUE_WAKEUP)
4208 place_entity(cfs_rq, se, 0);
4210 check_schedstat_required();
4211 update_stats_enqueue(cfs_rq, se, flags);
4212 check_spread(cfs_rq, se);
4214 __enqueue_entity(cfs_rq, se);
4218 * When bandwidth control is enabled, cfs might have been removed
4219 * because of a parent been throttled but cfs->nr_running > 1. Try to
4220 * add it unconditionnally.
4222 if (cfs_rq->nr_running == 1 || cfs_bandwidth_used())
4223 list_add_leaf_cfs_rq(cfs_rq);
4225 if (cfs_rq->nr_running == 1)
4226 check_enqueue_throttle(cfs_rq);
4229 static void __clear_buddies_last(struct sched_entity *se)
4231 for_each_sched_entity(se) {
4232 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4233 if (cfs_rq->last != se)
4236 cfs_rq->last = NULL;
4240 static void __clear_buddies_next(struct sched_entity *se)
4242 for_each_sched_entity(se) {
4243 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4244 if (cfs_rq->next != se)
4247 cfs_rq->next = NULL;
4251 static void __clear_buddies_skip(struct sched_entity *se)
4253 for_each_sched_entity(se) {
4254 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4255 if (cfs_rq->skip != se)
4258 cfs_rq->skip = NULL;
4262 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4264 if (cfs_rq->last == se)
4265 __clear_buddies_last(se);
4267 if (cfs_rq->next == se)
4268 __clear_buddies_next(se);
4270 if (cfs_rq->skip == se)
4271 __clear_buddies_skip(se);
4274 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4277 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4280 * Update run-time statistics of the 'current'.
4282 update_curr(cfs_rq);
4285 * When dequeuing a sched_entity, we must:
4286 * - Update loads to have both entity and cfs_rq synced with now.
4287 * - Subtract its load from the cfs_rq->runnable_avg.
4288 * - Subtract its previous weight from cfs_rq->load.weight.
4289 * - For group entity, update its weight to reflect the new share
4290 * of its group cfs_rq.
4292 update_load_avg(cfs_rq, se, UPDATE_TG);
4293 se_update_runnable(se);
4295 update_stats_dequeue(cfs_rq, se, flags);
4297 clear_buddies(cfs_rq, se);
4299 if (se != cfs_rq->curr)
4300 __dequeue_entity(cfs_rq, se);
4302 account_entity_dequeue(cfs_rq, se);
4305 * Normalize after update_curr(); which will also have moved
4306 * min_vruntime if @se is the one holding it back. But before doing
4307 * update_min_vruntime() again, which will discount @se's position and
4308 * can move min_vruntime forward still more.
4310 if (!(flags & DEQUEUE_SLEEP))
4311 se->vruntime -= cfs_rq->min_vruntime;
4313 /* return excess runtime on last dequeue */
4314 return_cfs_rq_runtime(cfs_rq);
4316 update_cfs_group(se);
4319 * Now advance min_vruntime if @se was the entity holding it back,
4320 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4321 * put back on, and if we advance min_vruntime, we'll be placed back
4322 * further than we started -- ie. we'll be penalized.
4324 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4325 update_min_vruntime(cfs_rq);
4329 * Preempt the current task with a newly woken task if needed:
4332 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4334 unsigned long ideal_runtime, delta_exec;
4335 struct sched_entity *se;
4338 ideal_runtime = sched_slice(cfs_rq, curr);
4339 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4340 if (delta_exec > ideal_runtime) {
4341 resched_curr(rq_of(cfs_rq));
4343 * The current task ran long enough, ensure it doesn't get
4344 * re-elected due to buddy favours.
4346 clear_buddies(cfs_rq, curr);
4351 * Ensure that a task that missed wakeup preemption by a
4352 * narrow margin doesn't have to wait for a full slice.
4353 * This also mitigates buddy induced latencies under load.
4355 if (delta_exec < sysctl_sched_min_granularity)
4358 se = __pick_first_entity(cfs_rq);
4359 delta = curr->vruntime - se->vruntime;
4364 if (delta > ideal_runtime)
4365 resched_curr(rq_of(cfs_rq));
4369 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4371 /* 'current' is not kept within the tree. */
4374 * Any task has to be enqueued before it get to execute on
4375 * a CPU. So account for the time it spent waiting on the
4378 update_stats_wait_end(cfs_rq, se);
4379 __dequeue_entity(cfs_rq, se);
4380 update_load_avg(cfs_rq, se, UPDATE_TG);
4383 update_stats_curr_start(cfs_rq, se);
4387 * Track our maximum slice length, if the CPU's load is at
4388 * least twice that of our own weight (i.e. dont track it
4389 * when there are only lesser-weight tasks around):
4391 if (schedstat_enabled() &&
4392 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
4393 schedstat_set(se->statistics.slice_max,
4394 max((u64)schedstat_val(se->statistics.slice_max),
4395 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4398 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4402 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4405 * Pick the next process, keeping these things in mind, in this order:
4406 * 1) keep things fair between processes/task groups
4407 * 2) pick the "next" process, since someone really wants that to run
4408 * 3) pick the "last" process, for cache locality
4409 * 4) do not run the "skip" process, if something else is available
4411 static struct sched_entity *
4412 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4414 struct sched_entity *left = __pick_first_entity(cfs_rq);
4415 struct sched_entity *se;
4418 * If curr is set we have to see if its left of the leftmost entity
4419 * still in the tree, provided there was anything in the tree at all.
4421 if (!left || (curr && entity_before(curr, left)))
4424 se = left; /* ideally we run the leftmost entity */
4427 * Avoid running the skip buddy, if running something else can
4428 * be done without getting too unfair.
4430 if (cfs_rq->skip == se) {
4431 struct sched_entity *second;
4434 second = __pick_first_entity(cfs_rq);
4436 second = __pick_next_entity(se);
4437 if (!second || (curr && entity_before(curr, second)))
4441 if (second && wakeup_preempt_entity(second, left) < 1)
4446 * Prefer last buddy, try to return the CPU to a preempted task.
4448 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4452 * Someone really wants this to run. If it's not unfair, run it.
4454 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4457 clear_buddies(cfs_rq, se);
4462 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4464 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4467 * If still on the runqueue then deactivate_task()
4468 * was not called and update_curr() has to be done:
4471 update_curr(cfs_rq);
4473 /* throttle cfs_rqs exceeding runtime */
4474 check_cfs_rq_runtime(cfs_rq);
4476 check_spread(cfs_rq, prev);
4479 update_stats_wait_start(cfs_rq, prev);
4480 /* Put 'current' back into the tree. */
4481 __enqueue_entity(cfs_rq, prev);
4482 /* in !on_rq case, update occurred at dequeue */
4483 update_load_avg(cfs_rq, prev, 0);
4485 cfs_rq->curr = NULL;
4489 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4492 * Update run-time statistics of the 'current'.
4494 update_curr(cfs_rq);
4497 * Ensure that runnable average is periodically updated.
4499 update_load_avg(cfs_rq, curr, UPDATE_TG);
4500 update_cfs_group(curr);
4502 #ifdef CONFIG_SCHED_HRTICK
4504 * queued ticks are scheduled to match the slice, so don't bother
4505 * validating it and just reschedule.
4508 resched_curr(rq_of(cfs_rq));
4512 * don't let the period tick interfere with the hrtick preemption
4514 if (!sched_feat(DOUBLE_TICK) &&
4515 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4519 if (cfs_rq->nr_running > 1)
4520 check_preempt_tick(cfs_rq, curr);
4524 /**************************************************
4525 * CFS bandwidth control machinery
4528 #ifdef CONFIG_CFS_BANDWIDTH
4530 #ifdef CONFIG_JUMP_LABEL
4531 static struct static_key __cfs_bandwidth_used;
4533 static inline bool cfs_bandwidth_used(void)
4535 return static_key_false(&__cfs_bandwidth_used);
4538 void cfs_bandwidth_usage_inc(void)
4540 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4543 void cfs_bandwidth_usage_dec(void)
4545 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4547 #else /* CONFIG_JUMP_LABEL */
4548 static bool cfs_bandwidth_used(void)
4553 void cfs_bandwidth_usage_inc(void) {}
4554 void cfs_bandwidth_usage_dec(void) {}
4555 #endif /* CONFIG_JUMP_LABEL */
4558 * default period for cfs group bandwidth.
4559 * default: 0.1s, units: nanoseconds
4561 static inline u64 default_cfs_period(void)
4563 return 100000000ULL;
4566 static inline u64 sched_cfs_bandwidth_slice(void)
4568 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4572 * Replenish runtime according to assigned quota. We use sched_clock_cpu
4573 * directly instead of rq->clock to avoid adding additional synchronization
4576 * requires cfs_b->lock
4578 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4580 if (cfs_b->quota != RUNTIME_INF)
4581 cfs_b->runtime = cfs_b->quota;
4584 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4586 return &tg->cfs_bandwidth;
4589 /* returns 0 on failure to allocate runtime */
4590 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4592 struct task_group *tg = cfs_rq->tg;
4593 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4594 u64 amount = 0, min_amount;
4596 /* note: this is a positive sum as runtime_remaining <= 0 */
4597 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4599 raw_spin_lock(&cfs_b->lock);
4600 if (cfs_b->quota == RUNTIME_INF)
4601 amount = min_amount;
4603 start_cfs_bandwidth(cfs_b);
4605 if (cfs_b->runtime > 0) {
4606 amount = min(cfs_b->runtime, min_amount);
4607 cfs_b->runtime -= amount;
4611 raw_spin_unlock(&cfs_b->lock);
4613 cfs_rq->runtime_remaining += amount;
4615 return cfs_rq->runtime_remaining > 0;
4618 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4620 /* dock delta_exec before expiring quota (as it could span periods) */
4621 cfs_rq->runtime_remaining -= delta_exec;
4623 if (likely(cfs_rq->runtime_remaining > 0))
4626 if (cfs_rq->throttled)
4629 * if we're unable to extend our runtime we resched so that the active
4630 * hierarchy can be throttled
4632 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4633 resched_curr(rq_of(cfs_rq));
4636 static __always_inline
4637 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4639 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4642 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4645 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4647 return cfs_bandwidth_used() && cfs_rq->throttled;
4650 /* check whether cfs_rq, or any parent, is throttled */
4651 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4653 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4657 * Ensure that neither of the group entities corresponding to src_cpu or
4658 * dest_cpu are members of a throttled hierarchy when performing group
4659 * load-balance operations.
4661 static inline int throttled_lb_pair(struct task_group *tg,
4662 int src_cpu, int dest_cpu)
4664 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4666 src_cfs_rq = tg->cfs_rq[src_cpu];
4667 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4669 return throttled_hierarchy(src_cfs_rq) ||
4670 throttled_hierarchy(dest_cfs_rq);
4673 static int tg_unthrottle_up(struct task_group *tg, void *data)
4675 struct rq *rq = data;
4676 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4678 cfs_rq->throttle_count--;
4679 if (!cfs_rq->throttle_count) {
4680 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4681 cfs_rq->throttled_clock_task;
4683 /* Add cfs_rq with already running entity in the list */
4684 if (cfs_rq->nr_running >= 1)
4685 list_add_leaf_cfs_rq(cfs_rq);
4691 static int tg_throttle_down(struct task_group *tg, void *data)
4693 struct rq *rq = data;
4694 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4696 /* group is entering throttled state, stop time */
4697 if (!cfs_rq->throttle_count) {
4698 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4699 list_del_leaf_cfs_rq(cfs_rq);
4701 cfs_rq->throttle_count++;
4706 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4708 struct rq *rq = rq_of(cfs_rq);
4709 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4710 struct sched_entity *se;
4711 long task_delta, idle_task_delta, dequeue = 1;
4714 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4716 /* freeze hierarchy runnable averages while throttled */
4718 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4721 task_delta = cfs_rq->h_nr_running;
4722 idle_task_delta = cfs_rq->idle_h_nr_running;
4723 for_each_sched_entity(se) {
4724 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4725 /* throttled entity or throttle-on-deactivate */
4730 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4732 update_load_avg(qcfs_rq, se, 0);
4733 se_update_runnable(se);
4736 qcfs_rq->h_nr_running -= task_delta;
4737 qcfs_rq->idle_h_nr_running -= idle_task_delta;
4739 if (qcfs_rq->load.weight)
4744 sub_nr_running(rq, task_delta);
4746 cfs_rq->throttled = 1;
4747 cfs_rq->throttled_clock = rq_clock(rq);
4748 raw_spin_lock(&cfs_b->lock);
4749 empty = list_empty(&cfs_b->throttled_cfs_rq);
4752 * Add to the _head_ of the list, so that an already-started
4753 * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
4754 * not running add to the tail so that later runqueues don't get starved.
4756 if (cfs_b->distribute_running)
4757 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4759 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4762 * If we're the first throttled task, make sure the bandwidth
4766 start_cfs_bandwidth(cfs_b);
4768 raw_spin_unlock(&cfs_b->lock);
4771 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4773 struct rq *rq = rq_of(cfs_rq);
4774 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4775 struct sched_entity *se;
4777 long task_delta, idle_task_delta;
4779 se = cfs_rq->tg->se[cpu_of(rq)];
4781 cfs_rq->throttled = 0;
4783 update_rq_clock(rq);
4785 raw_spin_lock(&cfs_b->lock);
4786 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4787 list_del_rcu(&cfs_rq->throttled_list);
4788 raw_spin_unlock(&cfs_b->lock);
4790 /* update hierarchical throttle state */
4791 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4793 if (!cfs_rq->load.weight)
4796 task_delta = cfs_rq->h_nr_running;
4797 idle_task_delta = cfs_rq->idle_h_nr_running;
4798 for_each_sched_entity(se) {
4802 cfs_rq = cfs_rq_of(se);
4804 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4806 update_load_avg(cfs_rq, se, 0);
4807 se_update_runnable(se);
4810 cfs_rq->h_nr_running += task_delta;
4811 cfs_rq->idle_h_nr_running += idle_task_delta;
4813 if (cfs_rq_throttled(cfs_rq))
4818 add_nr_running(rq, task_delta);
4821 * The cfs_rq_throttled() breaks in the above iteration can result in
4822 * incomplete leaf list maintenance, resulting in triggering the
4825 for_each_sched_entity(se) {
4826 cfs_rq = cfs_rq_of(se);
4828 list_add_leaf_cfs_rq(cfs_rq);
4831 assert_list_leaf_cfs_rq(rq);
4833 /* Determine whether we need to wake up potentially idle CPU: */
4834 if (rq->curr == rq->idle && rq->cfs.nr_running)
4838 static void distribute_cfs_runtime(struct cfs_bandwidth *cfs_b)
4840 struct cfs_rq *cfs_rq;
4841 u64 runtime, remaining = 1;
4844 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4846 struct rq *rq = rq_of(cfs_rq);
4849 rq_lock_irqsave(rq, &rf);
4850 if (!cfs_rq_throttled(cfs_rq))
4853 /* By the above check, this should never be true */
4854 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
4856 raw_spin_lock(&cfs_b->lock);
4857 runtime = -cfs_rq->runtime_remaining + 1;
4858 if (runtime > cfs_b->runtime)
4859 runtime = cfs_b->runtime;
4860 cfs_b->runtime -= runtime;
4861 remaining = cfs_b->runtime;
4862 raw_spin_unlock(&cfs_b->lock);
4864 cfs_rq->runtime_remaining += runtime;
4866 /* we check whether we're throttled above */
4867 if (cfs_rq->runtime_remaining > 0)
4868 unthrottle_cfs_rq(cfs_rq);
4871 rq_unlock_irqrestore(rq, &rf);
4880 * Responsible for refilling a task_group's bandwidth and unthrottling its
4881 * cfs_rqs as appropriate. If there has been no activity within the last
4882 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4883 * used to track this state.
4885 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
4889 /* no need to continue the timer with no bandwidth constraint */
4890 if (cfs_b->quota == RUNTIME_INF)
4891 goto out_deactivate;
4893 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4894 cfs_b->nr_periods += overrun;
4897 * idle depends on !throttled (for the case of a large deficit), and if
4898 * we're going inactive then everything else can be deferred
4900 if (cfs_b->idle && !throttled)
4901 goto out_deactivate;
4903 __refill_cfs_bandwidth_runtime(cfs_b);
4906 /* mark as potentially idle for the upcoming period */
4911 /* account preceding periods in which throttling occurred */
4912 cfs_b->nr_throttled += overrun;
4915 * This check is repeated as we release cfs_b->lock while we unthrottle.
4917 while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
4918 cfs_b->distribute_running = 1;
4919 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4920 /* we can't nest cfs_b->lock while distributing bandwidth */
4921 distribute_cfs_runtime(cfs_b);
4922 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4924 cfs_b->distribute_running = 0;
4925 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4929 * While we are ensured activity in the period following an
4930 * unthrottle, this also covers the case in which the new bandwidth is
4931 * insufficient to cover the existing bandwidth deficit. (Forcing the
4932 * timer to remain active while there are any throttled entities.)
4942 /* a cfs_rq won't donate quota below this amount */
4943 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4944 /* minimum remaining period time to redistribute slack quota */
4945 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4946 /* how long we wait to gather additional slack before distributing */
4947 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4950 * Are we near the end of the current quota period?
4952 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4953 * hrtimer base being cleared by hrtimer_start. In the case of
4954 * migrate_hrtimers, base is never cleared, so we are fine.
4956 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4958 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4961 /* if the call-back is running a quota refresh is already occurring */
4962 if (hrtimer_callback_running(refresh_timer))
4965 /* is a quota refresh about to occur? */
4966 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4967 if (remaining < min_expire)
4973 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4975 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4977 /* if there's a quota refresh soon don't bother with slack */
4978 if (runtime_refresh_within(cfs_b, min_left))
4981 /* don't push forwards an existing deferred unthrottle */
4982 if (cfs_b->slack_started)
4984 cfs_b->slack_started = true;
4986 hrtimer_start(&cfs_b->slack_timer,
4987 ns_to_ktime(cfs_bandwidth_slack_period),
4991 /* we know any runtime found here is valid as update_curr() precedes return */
4992 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4994 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4995 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4997 if (slack_runtime <= 0)
5000 raw_spin_lock(&cfs_b->lock);
5001 if (cfs_b->quota != RUNTIME_INF) {
5002 cfs_b->runtime += slack_runtime;
5004 /* we are under rq->lock, defer unthrottling using a timer */
5005 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
5006 !list_empty(&cfs_b->throttled_cfs_rq))
5007 start_cfs_slack_bandwidth(cfs_b);
5009 raw_spin_unlock(&cfs_b->lock);
5011 /* even if it's not valid for return we don't want to try again */
5012 cfs_rq->runtime_remaining -= slack_runtime;
5015 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5017 if (!cfs_bandwidth_used())
5020 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
5023 __return_cfs_rq_runtime(cfs_rq);
5027 * This is done with a timer (instead of inline with bandwidth return) since
5028 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
5030 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
5032 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
5033 unsigned long flags;
5035 /* confirm we're still not at a refresh boundary */
5036 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5037 cfs_b->slack_started = false;
5038 if (cfs_b->distribute_running) {
5039 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5043 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
5044 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5048 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
5049 runtime = cfs_b->runtime;
5052 cfs_b->distribute_running = 1;
5054 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5059 distribute_cfs_runtime(cfs_b);
5061 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5062 cfs_b->distribute_running = 0;
5063 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5067 * When a group wakes up we want to make sure that its quota is not already
5068 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
5069 * runtime as update_curr() throttling can not not trigger until it's on-rq.
5071 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
5073 if (!cfs_bandwidth_used())
5076 /* an active group must be handled by the update_curr()->put() path */
5077 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
5080 /* ensure the group is not already throttled */
5081 if (cfs_rq_throttled(cfs_rq))
5084 /* update runtime allocation */
5085 account_cfs_rq_runtime(cfs_rq, 0);
5086 if (cfs_rq->runtime_remaining <= 0)
5087 throttle_cfs_rq(cfs_rq);
5090 static void sync_throttle(struct task_group *tg, int cpu)
5092 struct cfs_rq *pcfs_rq, *cfs_rq;
5094 if (!cfs_bandwidth_used())
5100 cfs_rq = tg->cfs_rq[cpu];
5101 pcfs_rq = tg->parent->cfs_rq[cpu];
5103 cfs_rq->throttle_count = pcfs_rq->throttle_count;
5104 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
5107 /* conditionally throttle active cfs_rq's from put_prev_entity() */
5108 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5110 if (!cfs_bandwidth_used())
5113 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
5117 * it's possible for a throttled entity to be forced into a running
5118 * state (e.g. set_curr_task), in this case we're finished.
5120 if (cfs_rq_throttled(cfs_rq))
5123 throttle_cfs_rq(cfs_rq);
5127 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
5129 struct cfs_bandwidth *cfs_b =
5130 container_of(timer, struct cfs_bandwidth, slack_timer);
5132 do_sched_cfs_slack_timer(cfs_b);
5134 return HRTIMER_NORESTART;
5137 extern const u64 max_cfs_quota_period;
5139 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
5141 struct cfs_bandwidth *cfs_b =
5142 container_of(timer, struct cfs_bandwidth, period_timer);
5143 unsigned long flags;
5148 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5150 overrun = hrtimer_forward_now(timer, cfs_b->period);
5155 u64 new, old = ktime_to_ns(cfs_b->period);
5158 * Grow period by a factor of 2 to avoid losing precision.
5159 * Precision loss in the quota/period ratio can cause __cfs_schedulable
5163 if (new < max_cfs_quota_period) {
5164 cfs_b->period = ns_to_ktime(new);
5167 pr_warn_ratelimited(
5168 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5170 div_u64(new, NSEC_PER_USEC),
5171 div_u64(cfs_b->quota, NSEC_PER_USEC));
5173 pr_warn_ratelimited(
5174 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5176 div_u64(old, NSEC_PER_USEC),
5177 div_u64(cfs_b->quota, NSEC_PER_USEC));
5180 /* reset count so we don't come right back in here */
5184 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
5187 cfs_b->period_active = 0;
5188 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5190 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
5193 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5195 raw_spin_lock_init(&cfs_b->lock);
5197 cfs_b->quota = RUNTIME_INF;
5198 cfs_b->period = ns_to_ktime(default_cfs_period());
5200 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
5201 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
5202 cfs_b->period_timer.function = sched_cfs_period_timer;
5203 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5204 cfs_b->slack_timer.function = sched_cfs_slack_timer;
5205 cfs_b->distribute_running = 0;
5206 cfs_b->slack_started = false;
5209 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5211 cfs_rq->runtime_enabled = 0;
5212 INIT_LIST_HEAD(&cfs_rq->throttled_list);
5215 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5217 lockdep_assert_held(&cfs_b->lock);
5219 if (cfs_b->period_active)
5222 cfs_b->period_active = 1;
5223 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5224 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5227 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5229 /* init_cfs_bandwidth() was not called */
5230 if (!cfs_b->throttled_cfs_rq.next)
5233 hrtimer_cancel(&cfs_b->period_timer);
5234 hrtimer_cancel(&cfs_b->slack_timer);
5238 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5240 * The race is harmless, since modifying bandwidth settings of unhooked group
5241 * bits doesn't do much.
5244 /* cpu online calback */
5245 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5247 struct task_group *tg;
5249 lockdep_assert_held(&rq->lock);
5252 list_for_each_entry_rcu(tg, &task_groups, list) {
5253 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5254 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5256 raw_spin_lock(&cfs_b->lock);
5257 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5258 raw_spin_unlock(&cfs_b->lock);
5263 /* cpu offline callback */
5264 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5266 struct task_group *tg;
5268 lockdep_assert_held(&rq->lock);
5271 list_for_each_entry_rcu(tg, &task_groups, list) {
5272 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5274 if (!cfs_rq->runtime_enabled)
5278 * clock_task is not advancing so we just need to make sure
5279 * there's some valid quota amount
5281 cfs_rq->runtime_remaining = 1;
5283 * Offline rq is schedulable till CPU is completely disabled
5284 * in take_cpu_down(), so we prevent new cfs throttling here.
5286 cfs_rq->runtime_enabled = 0;
5288 if (cfs_rq_throttled(cfs_rq))
5289 unthrottle_cfs_rq(cfs_rq);
5294 #else /* CONFIG_CFS_BANDWIDTH */
5296 static inline bool cfs_bandwidth_used(void)
5301 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5302 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5303 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5304 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5305 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5307 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5312 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5317 static inline int throttled_lb_pair(struct task_group *tg,
5318 int src_cpu, int dest_cpu)
5323 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5325 #ifdef CONFIG_FAIR_GROUP_SCHED
5326 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5329 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5333 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5334 static inline void update_runtime_enabled(struct rq *rq) {}
5335 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5337 #endif /* CONFIG_CFS_BANDWIDTH */
5339 /**************************************************
5340 * CFS operations on tasks:
5343 #ifdef CONFIG_SCHED_HRTICK
5344 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5346 struct sched_entity *se = &p->se;
5347 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5349 SCHED_WARN_ON(task_rq(p) != rq);
5351 if (rq->cfs.h_nr_running > 1) {
5352 u64 slice = sched_slice(cfs_rq, se);
5353 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5354 s64 delta = slice - ran;
5361 hrtick_start(rq, delta);
5366 * called from enqueue/dequeue and updates the hrtick when the
5367 * current task is from our class and nr_running is low enough
5370 static void hrtick_update(struct rq *rq)
5372 struct task_struct *curr = rq->curr;
5374 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5377 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5378 hrtick_start_fair(rq, curr);
5380 #else /* !CONFIG_SCHED_HRTICK */
5382 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5386 static inline void hrtick_update(struct rq *rq)
5392 static inline unsigned long cpu_util(int cpu);
5394 static inline bool cpu_overutilized(int cpu)
5396 return !fits_capacity(cpu_util(cpu), capacity_of(cpu));
5399 static inline void update_overutilized_status(struct rq *rq)
5401 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
5402 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5403 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
5407 static inline void update_overutilized_status(struct rq *rq) { }
5410 /* Runqueue only has SCHED_IDLE tasks enqueued */
5411 static int sched_idle_rq(struct rq *rq)
5413 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
5418 static int sched_idle_cpu(int cpu)
5420 return sched_idle_rq(cpu_rq(cpu));
5425 * The enqueue_task method is called before nr_running is
5426 * increased. Here we update the fair scheduling stats and
5427 * then put the task into the rbtree:
5430 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5432 struct cfs_rq *cfs_rq;
5433 struct sched_entity *se = &p->se;
5434 int idle_h_nr_running = task_has_idle_policy(p);
5437 * The code below (indirectly) updates schedutil which looks at
5438 * the cfs_rq utilization to select a frequency.
5439 * Let's add the task's estimated utilization to the cfs_rq's
5440 * estimated utilization, before we update schedutil.
5442 util_est_enqueue(&rq->cfs, p);
5445 * If in_iowait is set, the code below may not trigger any cpufreq
5446 * utilization updates, so do it here explicitly with the IOWAIT flag
5450 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5452 for_each_sched_entity(se) {
5455 cfs_rq = cfs_rq_of(se);
5456 enqueue_entity(cfs_rq, se, flags);
5458 cfs_rq->h_nr_running++;
5459 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5461 /* end evaluation on encountering a throttled cfs_rq */
5462 if (cfs_rq_throttled(cfs_rq))
5463 goto enqueue_throttle;
5465 flags = ENQUEUE_WAKEUP;
5468 for_each_sched_entity(se) {
5469 cfs_rq = cfs_rq_of(se);
5471 update_load_avg(cfs_rq, se, UPDATE_TG);
5472 se_update_runnable(se);
5473 update_cfs_group(se);
5475 cfs_rq->h_nr_running++;
5476 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5478 /* end evaluation on encountering a throttled cfs_rq */
5479 if (cfs_rq_throttled(cfs_rq))
5480 goto enqueue_throttle;
5485 add_nr_running(rq, 1);
5487 * Since new tasks are assigned an initial util_avg equal to
5488 * half of the spare capacity of their CPU, tiny tasks have the
5489 * ability to cross the overutilized threshold, which will
5490 * result in the load balancer ruining all the task placement
5491 * done by EAS. As a way to mitigate that effect, do not account
5492 * for the first enqueue operation of new tasks during the
5493 * overutilized flag detection.
5495 * A better way of solving this problem would be to wait for
5496 * the PELT signals of tasks to converge before taking them
5497 * into account, but that is not straightforward to implement,
5498 * and the following generally works well enough in practice.
5500 if (flags & ENQUEUE_WAKEUP)
5501 update_overutilized_status(rq);
5505 if (cfs_bandwidth_used()) {
5507 * When bandwidth control is enabled; the cfs_rq_throttled()
5508 * breaks in the above iteration can result in incomplete
5509 * leaf list maintenance, resulting in triggering the assertion
5512 for_each_sched_entity(se) {
5513 cfs_rq = cfs_rq_of(se);
5515 if (list_add_leaf_cfs_rq(cfs_rq))
5520 assert_list_leaf_cfs_rq(rq);
5525 static void set_next_buddy(struct sched_entity *se);
5528 * The dequeue_task method is called before nr_running is
5529 * decreased. We remove the task from the rbtree and
5530 * update the fair scheduling stats:
5532 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5534 struct cfs_rq *cfs_rq;
5535 struct sched_entity *se = &p->se;
5536 int task_sleep = flags & DEQUEUE_SLEEP;
5537 int idle_h_nr_running = task_has_idle_policy(p);
5538 bool was_sched_idle = sched_idle_rq(rq);
5540 for_each_sched_entity(se) {
5541 cfs_rq = cfs_rq_of(se);
5542 dequeue_entity(cfs_rq, se, flags);
5544 cfs_rq->h_nr_running--;
5545 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5547 /* end evaluation on encountering a throttled cfs_rq */
5548 if (cfs_rq_throttled(cfs_rq))
5549 goto dequeue_throttle;
5551 /* Don't dequeue parent if it has other entities besides us */
5552 if (cfs_rq->load.weight) {
5553 /* Avoid re-evaluating load for this entity: */
5554 se = parent_entity(se);
5556 * Bias pick_next to pick a task from this cfs_rq, as
5557 * p is sleeping when it is within its sched_slice.
5559 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5563 flags |= DEQUEUE_SLEEP;
5566 for_each_sched_entity(se) {
5567 cfs_rq = cfs_rq_of(se);
5569 update_load_avg(cfs_rq, se, UPDATE_TG);
5570 se_update_runnable(se);
5571 update_cfs_group(se);
5573 cfs_rq->h_nr_running--;
5574 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5576 /* end evaluation on encountering a throttled cfs_rq */
5577 if (cfs_rq_throttled(cfs_rq))
5578 goto dequeue_throttle;
5584 sub_nr_running(rq, 1);
5586 /* balance early to pull high priority tasks */
5587 if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
5588 rq->next_balance = jiffies;
5590 util_est_dequeue(&rq->cfs, p, task_sleep);
5596 /* Working cpumask for: load_balance, load_balance_newidle. */
5597 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5598 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5600 #ifdef CONFIG_NO_HZ_COMMON
5603 cpumask_var_t idle_cpus_mask;
5605 int has_blocked; /* Idle CPUS has blocked load */
5606 unsigned long next_balance; /* in jiffy units */
5607 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5608 } nohz ____cacheline_aligned;
5610 #endif /* CONFIG_NO_HZ_COMMON */
5612 static unsigned long cpu_load(struct rq *rq)
5614 return cfs_rq_load_avg(&rq->cfs);
5618 * cpu_load_without - compute CPU load without any contributions from *p
5619 * @cpu: the CPU which load is requested
5620 * @p: the task which load should be discounted
5622 * The load of a CPU is defined by the load of tasks currently enqueued on that
5623 * CPU as well as tasks which are currently sleeping after an execution on that
5626 * This method returns the load of the specified CPU by discounting the load of
5627 * the specified task, whenever the task is currently contributing to the CPU
5630 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
5632 struct cfs_rq *cfs_rq;
5635 /* Task has no contribution or is new */
5636 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
5637 return cpu_load(rq);
5640 load = READ_ONCE(cfs_rq->avg.load_avg);
5642 /* Discount task's util from CPU's util */
5643 lsub_positive(&load, task_h_load(p));
5648 static unsigned long cpu_runnable(struct rq *rq)
5650 return cfs_rq_runnable_avg(&rq->cfs);
5653 static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
5655 struct cfs_rq *cfs_rq;
5656 unsigned int runnable;
5658 /* Task has no contribution or is new */
5659 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
5660 return cpu_runnable(rq);
5663 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
5665 /* Discount task's runnable from CPU's runnable */
5666 lsub_positive(&runnable, p->se.avg.runnable_avg);
5671 static unsigned long capacity_of(int cpu)
5673 return cpu_rq(cpu)->cpu_capacity;
5676 static void record_wakee(struct task_struct *p)
5679 * Only decay a single time; tasks that have less then 1 wakeup per
5680 * jiffy will not have built up many flips.
5682 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5683 current->wakee_flips >>= 1;
5684 current->wakee_flip_decay_ts = jiffies;
5687 if (current->last_wakee != p) {
5688 current->last_wakee = p;
5689 current->wakee_flips++;
5694 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5696 * A waker of many should wake a different task than the one last awakened
5697 * at a frequency roughly N times higher than one of its wakees.
5699 * In order to determine whether we should let the load spread vs consolidating
5700 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5701 * partner, and a factor of lls_size higher frequency in the other.
5703 * With both conditions met, we can be relatively sure that the relationship is
5704 * non-monogamous, with partner count exceeding socket size.
5706 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5707 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5710 static int wake_wide(struct task_struct *p)
5712 unsigned int master = current->wakee_flips;
5713 unsigned int slave = p->wakee_flips;
5714 int factor = this_cpu_read(sd_llc_size);
5717 swap(master, slave);
5718 if (slave < factor || master < slave * factor)
5724 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5725 * soonest. For the purpose of speed we only consider the waking and previous
5728 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5729 * cache-affine and is (or will be) idle.
5731 * wake_affine_weight() - considers the weight to reflect the average
5732 * scheduling latency of the CPUs. This seems to work
5733 * for the overloaded case.
5736 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5739 * If this_cpu is idle, it implies the wakeup is from interrupt
5740 * context. Only allow the move if cache is shared. Otherwise an
5741 * interrupt intensive workload could force all tasks onto one
5742 * node depending on the IO topology or IRQ affinity settings.
5744 * If the prev_cpu is idle and cache affine then avoid a migration.
5745 * There is no guarantee that the cache hot data from an interrupt
5746 * is more important than cache hot data on the prev_cpu and from
5747 * a cpufreq perspective, it's better to have higher utilisation
5750 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5751 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5753 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5756 return nr_cpumask_bits;
5760 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5761 int this_cpu, int prev_cpu, int sync)
5763 s64 this_eff_load, prev_eff_load;
5764 unsigned long task_load;
5766 this_eff_load = cpu_load(cpu_rq(this_cpu));
5769 unsigned long current_load = task_h_load(current);
5771 if (current_load > this_eff_load)
5774 this_eff_load -= current_load;
5777 task_load = task_h_load(p);
5779 this_eff_load += task_load;
5780 if (sched_feat(WA_BIAS))
5781 this_eff_load *= 100;
5782 this_eff_load *= capacity_of(prev_cpu);
5784 prev_eff_load = cpu_load(cpu_rq(prev_cpu));
5785 prev_eff_load -= task_load;
5786 if (sched_feat(WA_BIAS))
5787 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5788 prev_eff_load *= capacity_of(this_cpu);
5791 * If sync, adjust the weight of prev_eff_load such that if
5792 * prev_eff == this_eff that select_idle_sibling() will consider
5793 * stacking the wakee on top of the waker if no other CPU is
5799 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5802 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5803 int this_cpu, int prev_cpu, int sync)
5805 int target = nr_cpumask_bits;
5807 if (sched_feat(WA_IDLE))
5808 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5810 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5811 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5813 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5814 if (target == nr_cpumask_bits)
5817 schedstat_inc(sd->ttwu_move_affine);
5818 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5822 static struct sched_group *
5823 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5824 int this_cpu, int sd_flag);
5827 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5830 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5832 unsigned long load, min_load = ULONG_MAX;
5833 unsigned int min_exit_latency = UINT_MAX;
5834 u64 latest_idle_timestamp = 0;
5835 int least_loaded_cpu = this_cpu;
5836 int shallowest_idle_cpu = -1;
5839 /* Check if we have any choice: */
5840 if (group->group_weight == 1)
5841 return cpumask_first(sched_group_span(group));
5843 /* Traverse only the allowed CPUs */
5844 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
5845 if (sched_idle_cpu(i))
5848 if (available_idle_cpu(i)) {
5849 struct rq *rq = cpu_rq(i);
5850 struct cpuidle_state *idle = idle_get_state(rq);
5851 if (idle && idle->exit_latency < min_exit_latency) {
5853 * We give priority to a CPU whose idle state
5854 * has the smallest exit latency irrespective
5855 * of any idle timestamp.
5857 min_exit_latency = idle->exit_latency;
5858 latest_idle_timestamp = rq->idle_stamp;
5859 shallowest_idle_cpu = i;
5860 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5861 rq->idle_stamp > latest_idle_timestamp) {
5863 * If equal or no active idle state, then
5864 * the most recently idled CPU might have
5867 latest_idle_timestamp = rq->idle_stamp;
5868 shallowest_idle_cpu = i;
5870 } else if (shallowest_idle_cpu == -1) {
5871 load = cpu_load(cpu_rq(i));
5872 if (load < min_load) {
5874 least_loaded_cpu = i;
5879 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5882 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5883 int cpu, int prev_cpu, int sd_flag)
5887 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
5891 * We need task's util for cpu_util_without, sync it up to
5892 * prev_cpu's last_update_time.
5894 if (!(sd_flag & SD_BALANCE_FORK))
5895 sync_entity_load_avg(&p->se);
5898 struct sched_group *group;
5899 struct sched_domain *tmp;
5902 if (!(sd->flags & sd_flag)) {
5907 group = find_idlest_group(sd, p, cpu, sd_flag);
5913 new_cpu = find_idlest_group_cpu(group, p, cpu);
5914 if (new_cpu == cpu) {
5915 /* Now try balancing at a lower domain level of 'cpu': */
5920 /* Now try balancing at a lower domain level of 'new_cpu': */
5922 weight = sd->span_weight;
5924 for_each_domain(cpu, tmp) {
5925 if (weight <= tmp->span_weight)
5927 if (tmp->flags & sd_flag)
5935 #ifdef CONFIG_SCHED_SMT
5936 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5937 EXPORT_SYMBOL_GPL(sched_smt_present);
5939 static inline void set_idle_cores(int cpu, int val)
5941 struct sched_domain_shared *sds;
5943 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5945 WRITE_ONCE(sds->has_idle_cores, val);
5948 static inline bool test_idle_cores(int cpu, bool def)
5950 struct sched_domain_shared *sds;
5952 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5954 return READ_ONCE(sds->has_idle_cores);
5960 * Scans the local SMT mask to see if the entire core is idle, and records this
5961 * information in sd_llc_shared->has_idle_cores.
5963 * Since SMT siblings share all cache levels, inspecting this limited remote
5964 * state should be fairly cheap.
5966 void __update_idle_core(struct rq *rq)
5968 int core = cpu_of(rq);
5972 if (test_idle_cores(core, true))
5975 for_each_cpu(cpu, cpu_smt_mask(core)) {
5979 if (!available_idle_cpu(cpu))
5983 set_idle_cores(core, 1);
5989 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5990 * there are no idle cores left in the system; tracked through
5991 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5993 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5995 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5998 if (!static_branch_likely(&sched_smt_present))
6001 if (!test_idle_cores(target, false))
6004 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6006 for_each_cpu_wrap(core, cpus, target) {
6009 for_each_cpu(cpu, cpu_smt_mask(core)) {
6010 if (!available_idle_cpu(cpu)) {
6015 cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
6022 * Failed to find an idle core; stop looking for one.
6024 set_idle_cores(target, 0);
6030 * Scan the local SMT mask for idle CPUs.
6032 static int select_idle_smt(struct task_struct *p, int target)
6036 if (!static_branch_likely(&sched_smt_present))
6039 for_each_cpu(cpu, cpu_smt_mask(target)) {
6040 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6042 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
6049 #else /* CONFIG_SCHED_SMT */
6051 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6056 static inline int select_idle_smt(struct task_struct *p, int target)
6061 #endif /* CONFIG_SCHED_SMT */
6064 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6065 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6066 * average idle time for this rq (as found in rq->avg_idle).
6068 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
6070 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6071 struct sched_domain *this_sd;
6072 u64 avg_cost, avg_idle;
6074 int this = smp_processor_id();
6075 int cpu, nr = INT_MAX;
6077 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6082 * Due to large variance we need a large fuzz factor; hackbench in
6083 * particularly is sensitive here.
6085 avg_idle = this_rq()->avg_idle / 512;
6086 avg_cost = this_sd->avg_scan_cost + 1;
6088 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
6091 if (sched_feat(SIS_PROP)) {
6092 u64 span_avg = sd->span_weight * avg_idle;
6093 if (span_avg > 4*avg_cost)
6094 nr = div_u64(span_avg, avg_cost);
6099 time = cpu_clock(this);
6101 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6103 for_each_cpu_wrap(cpu, cpus, target) {
6106 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
6110 time = cpu_clock(this) - time;
6111 update_avg(&this_sd->avg_scan_cost, time);
6117 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
6118 * the task fits. If no CPU is big enough, but there are idle ones, try to
6119 * maximize capacity.
6122 select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
6124 unsigned long best_cap = 0;
6125 int cpu, best_cpu = -1;
6126 struct cpumask *cpus;
6128 sync_entity_load_avg(&p->se);
6130 cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6131 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6133 for_each_cpu_wrap(cpu, cpus, target) {
6134 unsigned long cpu_cap = capacity_of(cpu);
6136 if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
6138 if (task_fits_capacity(p, cpu_cap))
6141 if (cpu_cap > best_cap) {
6151 * Try and locate an idle core/thread in the LLC cache domain.
6153 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6155 struct sched_domain *sd;
6156 int i, recent_used_cpu;
6159 * For asymmetric CPU capacity systems, our domain of interest is
6160 * sd_asym_cpucapacity rather than sd_llc.
6162 if (static_branch_unlikely(&sched_asym_cpucapacity)) {
6163 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
6165 * On an asymmetric CPU capacity system where an exclusive
6166 * cpuset defines a symmetric island (i.e. one unique
6167 * capacity_orig value through the cpuset), the key will be set
6168 * but the CPUs within that cpuset will not have a domain with
6169 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
6175 i = select_idle_capacity(p, sd, target);
6176 return ((unsigned)i < nr_cpumask_bits) ? i : target;
6180 if (available_idle_cpu(target) || sched_idle_cpu(target))
6184 * If the previous CPU is cache affine and idle, don't be stupid:
6186 if (prev != target && cpus_share_cache(prev, target) &&
6187 (available_idle_cpu(prev) || sched_idle_cpu(prev)))
6191 * Allow a per-cpu kthread to stack with the wakee if the
6192 * kworker thread and the tasks previous CPUs are the same.
6193 * The assumption is that the wakee queued work for the
6194 * per-cpu kthread that is now complete and the wakeup is
6195 * essentially a sync wakeup. An obvious example of this
6196 * pattern is IO completions.
6198 if (is_per_cpu_kthread(current) &&
6199 prev == smp_processor_id() &&
6200 this_rq()->nr_running <= 1) {
6204 /* Check a recently used CPU as a potential idle candidate: */
6205 recent_used_cpu = p->recent_used_cpu;
6206 if (recent_used_cpu != prev &&
6207 recent_used_cpu != target &&
6208 cpus_share_cache(recent_used_cpu, target) &&
6209 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
6210 cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr)) {
6212 * Replace recent_used_cpu with prev as it is a potential
6213 * candidate for the next wake:
6215 p->recent_used_cpu = prev;
6216 return recent_used_cpu;
6219 sd = rcu_dereference(per_cpu(sd_llc, target));
6223 i = select_idle_core(p, sd, target);
6224 if ((unsigned)i < nr_cpumask_bits)
6227 i = select_idle_cpu(p, sd, target);
6228 if ((unsigned)i < nr_cpumask_bits)
6231 i = select_idle_smt(p, target);
6232 if ((unsigned)i < nr_cpumask_bits)
6239 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6240 * @cpu: the CPU to get the utilization of
6242 * The unit of the return value must be the one of capacity so we can compare
6243 * the utilization with the capacity of the CPU that is available for CFS task
6244 * (ie cpu_capacity).
6246 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6247 * recent utilization of currently non-runnable tasks on a CPU. It represents
6248 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6249 * capacity_orig is the cpu_capacity available at the highest frequency
6250 * (arch_scale_freq_capacity()).
6251 * The utilization of a CPU converges towards a sum equal to or less than the
6252 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6253 * the running time on this CPU scaled by capacity_curr.
6255 * The estimated utilization of a CPU is defined to be the maximum between its
6256 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6257 * currently RUNNABLE on that CPU.
6258 * This allows to properly represent the expected utilization of a CPU which
6259 * has just got a big task running since a long sleep period. At the same time
6260 * however it preserves the benefits of the "blocked utilization" in
6261 * describing the potential for other tasks waking up on the same CPU.
6263 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6264 * higher than capacity_orig because of unfortunate rounding in
6265 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6266 * the average stabilizes with the new running time. We need to check that the
6267 * utilization stays within the range of [0..capacity_orig] and cap it if
6268 * necessary. Without utilization capping, a group could be seen as overloaded
6269 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6270 * available capacity. We allow utilization to overshoot capacity_curr (but not
6271 * capacity_orig) as it useful for predicting the capacity required after task
6272 * migrations (scheduler-driven DVFS).
6274 * Return: the (estimated) utilization for the specified CPU
6276 static inline unsigned long cpu_util(int cpu)
6278 struct cfs_rq *cfs_rq;
6281 cfs_rq = &cpu_rq(cpu)->cfs;
6282 util = READ_ONCE(cfs_rq->avg.util_avg);
6284 if (sched_feat(UTIL_EST))
6285 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6287 return min_t(unsigned long, util, capacity_orig_of(cpu));
6291 * cpu_util_without: compute cpu utilization without any contributions from *p
6292 * @cpu: the CPU which utilization is requested
6293 * @p: the task which utilization should be discounted
6295 * The utilization of a CPU is defined by the utilization of tasks currently
6296 * enqueued on that CPU as well as tasks which are currently sleeping after an
6297 * execution on that CPU.
6299 * This method returns the utilization of the specified CPU by discounting the
6300 * utilization of the specified task, whenever the task is currently
6301 * contributing to the CPU utilization.
6303 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6305 struct cfs_rq *cfs_rq;
6308 /* Task has no contribution or is new */
6309 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6310 return cpu_util(cpu);
6312 cfs_rq = &cpu_rq(cpu)->cfs;
6313 util = READ_ONCE(cfs_rq->avg.util_avg);
6315 /* Discount task's util from CPU's util */
6316 lsub_positive(&util, task_util(p));
6321 * a) if *p is the only task sleeping on this CPU, then:
6322 * cpu_util (== task_util) > util_est (== 0)
6323 * and thus we return:
6324 * cpu_util_without = (cpu_util - task_util) = 0
6326 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6328 * cpu_util >= task_util
6329 * cpu_util > util_est (== 0)
6330 * and thus we discount *p's blocked utilization to return:
6331 * cpu_util_without = (cpu_util - task_util) >= 0
6333 * c) if other tasks are RUNNABLE on that CPU and
6334 * util_est > cpu_util
6335 * then we use util_est since it returns a more restrictive
6336 * estimation of the spare capacity on that CPU, by just
6337 * considering the expected utilization of tasks already
6338 * runnable on that CPU.
6340 * Cases a) and b) are covered by the above code, while case c) is
6341 * covered by the following code when estimated utilization is
6344 if (sched_feat(UTIL_EST)) {
6345 unsigned int estimated =
6346 READ_ONCE(cfs_rq->avg.util_est.enqueued);
6349 * Despite the following checks we still have a small window
6350 * for a possible race, when an execl's select_task_rq_fair()
6351 * races with LB's detach_task():
6354 * p->on_rq = TASK_ON_RQ_MIGRATING;
6355 * ---------------------------------- A
6356 * deactivate_task() \
6357 * dequeue_task() + RaceTime
6358 * util_est_dequeue() /
6359 * ---------------------------------- B
6361 * The additional check on "current == p" it's required to
6362 * properly fix the execl regression and it helps in further
6363 * reducing the chances for the above race.
6365 if (unlikely(task_on_rq_queued(p) || current == p))
6366 lsub_positive(&estimated, _task_util_est(p));
6368 util = max(util, estimated);
6372 * Utilization (estimated) can exceed the CPU capacity, thus let's
6373 * clamp to the maximum CPU capacity to ensure consistency with
6374 * the cpu_util call.
6376 return min_t(unsigned long, util, capacity_orig_of(cpu));
6380 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6383 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6385 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6386 unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
6389 * If @p migrates from @cpu to another, remove its contribution. Or,
6390 * if @p migrates from another CPU to @cpu, add its contribution. In
6391 * the other cases, @cpu is not impacted by the migration, so the
6392 * util_avg should already be correct.
6394 if (task_cpu(p) == cpu && dst_cpu != cpu)
6395 sub_positive(&util, task_util(p));
6396 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6397 util += task_util(p);
6399 if (sched_feat(UTIL_EST)) {
6400 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6403 * During wake-up, the task isn't enqueued yet and doesn't
6404 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6405 * so just add it (if needed) to "simulate" what will be
6406 * cpu_util() after the task has been enqueued.
6409 util_est += _task_util_est(p);
6411 util = max(util, util_est);
6414 return min(util, capacity_orig_of(cpu));
6418 * compute_energy(): Estimates the energy that @pd would consume if @p was
6419 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6420 * landscape of @pd's CPUs after the task migration, and uses the Energy Model
6421 * to compute what would be the energy if we decided to actually migrate that
6425 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
6427 struct cpumask *pd_mask = perf_domain_span(pd);
6428 unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask));
6429 unsigned long max_util = 0, sum_util = 0;
6433 * The capacity state of CPUs of the current rd can be driven by CPUs
6434 * of another rd if they belong to the same pd. So, account for the
6435 * utilization of these CPUs too by masking pd with cpu_online_mask
6436 * instead of the rd span.
6438 * If an entire pd is outside of the current rd, it will not appear in
6439 * its pd list and will not be accounted by compute_energy().
6441 for_each_cpu_and(cpu, pd_mask, cpu_online_mask) {
6442 unsigned long cpu_util, util_cfs = cpu_util_next(cpu, p, dst_cpu);
6443 struct task_struct *tsk = cpu == dst_cpu ? p : NULL;
6446 * Busy time computation: utilization clamping is not
6447 * required since the ratio (sum_util / cpu_capacity)
6448 * is already enough to scale the EM reported power
6449 * consumption at the (eventually clamped) cpu_capacity.
6451 sum_util += schedutil_cpu_util(cpu, util_cfs, cpu_cap,
6455 * Performance domain frequency: utilization clamping
6456 * must be considered since it affects the selection
6457 * of the performance domain frequency.
6458 * NOTE: in case RT tasks are running, by default the
6459 * FREQUENCY_UTIL's utilization can be max OPP.
6461 cpu_util = schedutil_cpu_util(cpu, util_cfs, cpu_cap,
6462 FREQUENCY_UTIL, tsk);
6463 max_util = max(max_util, cpu_util);
6466 return em_pd_energy(pd->em_pd, max_util, sum_util);
6470 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6471 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6472 * spare capacity in each performance domain and uses it as a potential
6473 * candidate to execute the task. Then, it uses the Energy Model to figure
6474 * out which of the CPU candidates is the most energy-efficient.
6476 * The rationale for this heuristic is as follows. In a performance domain,
6477 * all the most energy efficient CPU candidates (according to the Energy
6478 * Model) are those for which we'll request a low frequency. When there are
6479 * several CPUs for which the frequency request will be the same, we don't
6480 * have enough data to break the tie between them, because the Energy Model
6481 * only includes active power costs. With this model, if we assume that
6482 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6483 * the maximum spare capacity in a performance domain is guaranteed to be among
6484 * the best candidates of the performance domain.
6486 * In practice, it could be preferable from an energy standpoint to pack
6487 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6488 * but that could also hurt our chances to go cluster idle, and we have no
6489 * ways to tell with the current Energy Model if this is actually a good
6490 * idea or not. So, find_energy_efficient_cpu() basically favors
6491 * cluster-packing, and spreading inside a cluster. That should at least be
6492 * a good thing for latency, and this is consistent with the idea that most
6493 * of the energy savings of EAS come from the asymmetry of the system, and
6494 * not so much from breaking the tie between identical CPUs. That's also the
6495 * reason why EAS is enabled in the topology code only for systems where
6496 * SD_ASYM_CPUCAPACITY is set.
6498 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6499 * they don't have any useful utilization data yet and it's not possible to
6500 * forecast their impact on energy consumption. Consequently, they will be
6501 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6502 * to be energy-inefficient in some use-cases. The alternative would be to
6503 * bias new tasks towards specific types of CPUs first, or to try to infer
6504 * their util_avg from the parent task, but those heuristics could hurt
6505 * other use-cases too. So, until someone finds a better way to solve this,
6506 * let's keep things simple by re-using the existing slow path.
6508 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6510 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
6511 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6512 unsigned long cpu_cap, util, base_energy = 0;
6513 int cpu, best_energy_cpu = prev_cpu;
6514 struct sched_domain *sd;
6515 struct perf_domain *pd;
6518 pd = rcu_dereference(rd->pd);
6519 if (!pd || READ_ONCE(rd->overutilized))
6523 * Energy-aware wake-up happens on the lowest sched_domain starting
6524 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6526 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6527 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6532 sync_entity_load_avg(&p->se);
6533 if (!task_util_est(p))
6536 for (; pd; pd = pd->next) {
6537 unsigned long cur_delta, spare_cap, max_spare_cap = 0;
6538 unsigned long base_energy_pd;
6539 int max_spare_cap_cpu = -1;
6541 /* Compute the 'base' energy of the pd, without @p */
6542 base_energy_pd = compute_energy(p, -1, pd);
6543 base_energy += base_energy_pd;
6545 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
6546 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6549 util = cpu_util_next(cpu, p, cpu);
6550 cpu_cap = capacity_of(cpu);
6551 spare_cap = cpu_cap - util;
6554 * Skip CPUs that cannot satisfy the capacity request.
6555 * IOW, placing the task there would make the CPU
6556 * overutilized. Take uclamp into account to see how
6557 * much capacity we can get out of the CPU; this is
6558 * aligned with schedutil_cpu_util().
6560 util = uclamp_rq_util_with(cpu_rq(cpu), util, p);
6561 if (!fits_capacity(util, cpu_cap))
6564 /* Always use prev_cpu as a candidate. */
6565 if (cpu == prev_cpu) {
6566 prev_delta = compute_energy(p, prev_cpu, pd);
6567 prev_delta -= base_energy_pd;
6568 best_delta = min(best_delta, prev_delta);
6572 * Find the CPU with the maximum spare capacity in
6573 * the performance domain
6575 if (spare_cap > max_spare_cap) {
6576 max_spare_cap = spare_cap;
6577 max_spare_cap_cpu = cpu;
6581 /* Evaluate the energy impact of using this CPU. */
6582 if (max_spare_cap_cpu >= 0 && max_spare_cap_cpu != prev_cpu) {
6583 cur_delta = compute_energy(p, max_spare_cap_cpu, pd);
6584 cur_delta -= base_energy_pd;
6585 if (cur_delta < best_delta) {
6586 best_delta = cur_delta;
6587 best_energy_cpu = max_spare_cap_cpu;
6595 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6596 * least 6% of the energy used by prev_cpu.
6598 if (prev_delta == ULONG_MAX)
6599 return best_energy_cpu;
6601 if ((prev_delta - best_delta) > ((prev_delta + base_energy) >> 4))
6602 return best_energy_cpu;
6613 * select_task_rq_fair: Select target runqueue for the waking task in domains
6614 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6615 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6617 * Balances load by selecting the idlest CPU in the idlest group, or under
6618 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6620 * Returns the target CPU number.
6622 * preempt must be disabled.
6625 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6627 struct sched_domain *tmp, *sd = NULL;
6628 int cpu = smp_processor_id();
6629 int new_cpu = prev_cpu;
6630 int want_affine = 0;
6631 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6633 if (sd_flag & SD_BALANCE_WAKE) {
6636 if (sched_energy_enabled()) {
6637 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
6643 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
6647 for_each_domain(cpu, tmp) {
6648 if (!(tmp->flags & SD_LOAD_BALANCE))
6652 * If both 'cpu' and 'prev_cpu' are part of this domain,
6653 * cpu is a valid SD_WAKE_AFFINE target.
6655 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6656 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6657 if (cpu != prev_cpu)
6658 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6660 sd = NULL; /* Prefer wake_affine over balance flags */
6664 if (tmp->flags & sd_flag)
6666 else if (!want_affine)
6672 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6673 } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6676 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6679 current->recent_used_cpu = cpu;
6686 static void detach_entity_cfs_rq(struct sched_entity *se);
6689 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6690 * cfs_rq_of(p) references at time of call are still valid and identify the
6691 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6693 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6696 * As blocked tasks retain absolute vruntime the migration needs to
6697 * deal with this by subtracting the old and adding the new
6698 * min_vruntime -- the latter is done by enqueue_entity() when placing
6699 * the task on the new runqueue.
6701 if (p->state == TASK_WAKING) {
6702 struct sched_entity *se = &p->se;
6703 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6706 #ifndef CONFIG_64BIT
6707 u64 min_vruntime_copy;
6710 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6712 min_vruntime = cfs_rq->min_vruntime;
6713 } while (min_vruntime != min_vruntime_copy);
6715 min_vruntime = cfs_rq->min_vruntime;
6718 se->vruntime -= min_vruntime;
6721 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6723 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6724 * rq->lock and can modify state directly.
6726 lockdep_assert_held(&task_rq(p)->lock);
6727 detach_entity_cfs_rq(&p->se);
6731 * We are supposed to update the task to "current" time, then
6732 * its up to date and ready to go to new CPU/cfs_rq. But we
6733 * have difficulty in getting what current time is, so simply
6734 * throw away the out-of-date time. This will result in the
6735 * wakee task is less decayed, but giving the wakee more load
6738 remove_entity_load_avg(&p->se);
6741 /* Tell new CPU we are migrated */
6742 p->se.avg.last_update_time = 0;
6744 /* We have migrated, no longer consider this task hot */
6745 p->se.exec_start = 0;
6747 update_scan_period(p, new_cpu);
6750 static void task_dead_fair(struct task_struct *p)
6752 remove_entity_load_avg(&p->se);
6756 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6761 return newidle_balance(rq, rf) != 0;
6763 #endif /* CONFIG_SMP */
6765 static unsigned long wakeup_gran(struct sched_entity *se)
6767 unsigned long gran = sysctl_sched_wakeup_granularity;
6770 * Since its curr running now, convert the gran from real-time
6771 * to virtual-time in his units.
6773 * By using 'se' instead of 'curr' we penalize light tasks, so
6774 * they get preempted easier. That is, if 'se' < 'curr' then
6775 * the resulting gran will be larger, therefore penalizing the
6776 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6777 * be smaller, again penalizing the lighter task.
6779 * This is especially important for buddies when the leftmost
6780 * task is higher priority than the buddy.
6782 return calc_delta_fair(gran, se);
6786 * Should 'se' preempt 'curr'.
6800 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6802 s64 gran, vdiff = curr->vruntime - se->vruntime;
6807 gran = wakeup_gran(se);
6814 static void set_last_buddy(struct sched_entity *se)
6816 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6819 for_each_sched_entity(se) {
6820 if (SCHED_WARN_ON(!se->on_rq))
6822 cfs_rq_of(se)->last = se;
6826 static void set_next_buddy(struct sched_entity *se)
6828 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6831 for_each_sched_entity(se) {
6832 if (SCHED_WARN_ON(!se->on_rq))
6834 cfs_rq_of(se)->next = se;
6838 static void set_skip_buddy(struct sched_entity *se)
6840 for_each_sched_entity(se)
6841 cfs_rq_of(se)->skip = se;
6845 * Preempt the current task with a newly woken task if needed:
6847 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6849 struct task_struct *curr = rq->curr;
6850 struct sched_entity *se = &curr->se, *pse = &p->se;
6851 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6852 int scale = cfs_rq->nr_running >= sched_nr_latency;
6853 int next_buddy_marked = 0;
6855 if (unlikely(se == pse))
6859 * This is possible from callers such as attach_tasks(), in which we
6860 * unconditionally check_prempt_curr() after an enqueue (which may have
6861 * lead to a throttle). This both saves work and prevents false
6862 * next-buddy nomination below.
6864 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6867 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6868 set_next_buddy(pse);
6869 next_buddy_marked = 1;
6873 * We can come here with TIF_NEED_RESCHED already set from new task
6876 * Note: this also catches the edge-case of curr being in a throttled
6877 * group (e.g. via set_curr_task), since update_curr() (in the
6878 * enqueue of curr) will have resulted in resched being set. This
6879 * prevents us from potentially nominating it as a false LAST_BUDDY
6882 if (test_tsk_need_resched(curr))
6885 /* Idle tasks are by definition preempted by non-idle tasks. */
6886 if (unlikely(task_has_idle_policy(curr)) &&
6887 likely(!task_has_idle_policy(p)))
6891 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6892 * is driven by the tick):
6894 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6897 find_matching_se(&se, &pse);
6898 update_curr(cfs_rq_of(se));
6900 if (wakeup_preempt_entity(se, pse) == 1) {
6902 * Bias pick_next to pick the sched entity that is
6903 * triggering this preemption.
6905 if (!next_buddy_marked)
6906 set_next_buddy(pse);
6915 * Only set the backward buddy when the current task is still
6916 * on the rq. This can happen when a wakeup gets interleaved
6917 * with schedule on the ->pre_schedule() or idle_balance()
6918 * point, either of which can * drop the rq lock.
6920 * Also, during early boot the idle thread is in the fair class,
6921 * for obvious reasons its a bad idea to schedule back to it.
6923 if (unlikely(!se->on_rq || curr == rq->idle))
6926 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6930 struct task_struct *
6931 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6933 struct cfs_rq *cfs_rq = &rq->cfs;
6934 struct sched_entity *se;
6935 struct task_struct *p;
6939 if (!sched_fair_runnable(rq))
6942 #ifdef CONFIG_FAIR_GROUP_SCHED
6943 if (!prev || prev->sched_class != &fair_sched_class)
6947 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6948 * likely that a next task is from the same cgroup as the current.
6950 * Therefore attempt to avoid putting and setting the entire cgroup
6951 * hierarchy, only change the part that actually changes.
6955 struct sched_entity *curr = cfs_rq->curr;
6958 * Since we got here without doing put_prev_entity() we also
6959 * have to consider cfs_rq->curr. If it is still a runnable
6960 * entity, update_curr() will update its vruntime, otherwise
6961 * forget we've ever seen it.
6965 update_curr(cfs_rq);
6970 * This call to check_cfs_rq_runtime() will do the
6971 * throttle and dequeue its entity in the parent(s).
6972 * Therefore the nr_running test will indeed
6975 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6978 if (!cfs_rq->nr_running)
6985 se = pick_next_entity(cfs_rq, curr);
6986 cfs_rq = group_cfs_rq(se);
6992 * Since we haven't yet done put_prev_entity and if the selected task
6993 * is a different task than we started out with, try and touch the
6994 * least amount of cfs_rqs.
6997 struct sched_entity *pse = &prev->se;
6999 while (!(cfs_rq = is_same_group(se, pse))) {
7000 int se_depth = se->depth;
7001 int pse_depth = pse->depth;
7003 if (se_depth <= pse_depth) {
7004 put_prev_entity(cfs_rq_of(pse), pse);
7005 pse = parent_entity(pse);
7007 if (se_depth >= pse_depth) {
7008 set_next_entity(cfs_rq_of(se), se);
7009 se = parent_entity(se);
7013 put_prev_entity(cfs_rq, pse);
7014 set_next_entity(cfs_rq, se);
7021 put_prev_task(rq, prev);
7024 se = pick_next_entity(cfs_rq, NULL);
7025 set_next_entity(cfs_rq, se);
7026 cfs_rq = group_cfs_rq(se);
7031 done: __maybe_unused;
7034 * Move the next running task to the front of
7035 * the list, so our cfs_tasks list becomes MRU
7038 list_move(&p->se.group_node, &rq->cfs_tasks);
7041 if (hrtick_enabled(rq))
7042 hrtick_start_fair(rq, p);
7044 update_misfit_status(p, rq);
7052 new_tasks = newidle_balance(rq, rf);
7055 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
7056 * possible for any higher priority task to appear. In that case we
7057 * must re-start the pick_next_entity() loop.
7066 * rq is about to be idle, check if we need to update the
7067 * lost_idle_time of clock_pelt
7069 update_idle_rq_clock_pelt(rq);
7074 static struct task_struct *__pick_next_task_fair(struct rq *rq)
7076 return pick_next_task_fair(rq, NULL, NULL);
7080 * Account for a descheduled task:
7082 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
7084 struct sched_entity *se = &prev->se;
7085 struct cfs_rq *cfs_rq;
7087 for_each_sched_entity(se) {
7088 cfs_rq = cfs_rq_of(se);
7089 put_prev_entity(cfs_rq, se);
7094 * sched_yield() is very simple
7096 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7098 static void yield_task_fair(struct rq *rq)
7100 struct task_struct *curr = rq->curr;
7101 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7102 struct sched_entity *se = &curr->se;
7105 * Are we the only task in the tree?
7107 if (unlikely(rq->nr_running == 1))
7110 clear_buddies(cfs_rq, se);
7112 if (curr->policy != SCHED_BATCH) {
7113 update_rq_clock(rq);
7115 * Update run-time statistics of the 'current'.
7117 update_curr(cfs_rq);
7119 * Tell update_rq_clock() that we've just updated,
7120 * so we don't do microscopic update in schedule()
7121 * and double the fastpath cost.
7123 rq_clock_skip_update(rq);
7129 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
7131 struct sched_entity *se = &p->se;
7133 /* throttled hierarchies are not runnable */
7134 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
7137 /* Tell the scheduler that we'd really like pse to run next. */
7140 yield_task_fair(rq);
7146 /**************************************************
7147 * Fair scheduling class load-balancing methods.
7151 * The purpose of load-balancing is to achieve the same basic fairness the
7152 * per-CPU scheduler provides, namely provide a proportional amount of compute
7153 * time to each task. This is expressed in the following equation:
7155 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7157 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7158 * W_i,0 is defined as:
7160 * W_i,0 = \Sum_j w_i,j (2)
7162 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7163 * is derived from the nice value as per sched_prio_to_weight[].
7165 * The weight average is an exponential decay average of the instantaneous
7168 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7170 * C_i is the compute capacity of CPU i, typically it is the
7171 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7172 * can also include other factors [XXX].
7174 * To achieve this balance we define a measure of imbalance which follows
7175 * directly from (1):
7177 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7179 * We them move tasks around to minimize the imbalance. In the continuous
7180 * function space it is obvious this converges, in the discrete case we get
7181 * a few fun cases generally called infeasible weight scenarios.
7184 * - infeasible weights;
7185 * - local vs global optima in the discrete case. ]
7190 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7191 * for all i,j solution, we create a tree of CPUs that follows the hardware
7192 * topology where each level pairs two lower groups (or better). This results
7193 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7194 * tree to only the first of the previous level and we decrease the frequency
7195 * of load-balance at each level inv. proportional to the number of CPUs in
7201 * \Sum { --- * --- * 2^i } = O(n) (5)
7203 * `- size of each group
7204 * | | `- number of CPUs doing load-balance
7206 * `- sum over all levels
7208 * Coupled with a limit on how many tasks we can migrate every balance pass,
7209 * this makes (5) the runtime complexity of the balancer.
7211 * An important property here is that each CPU is still (indirectly) connected
7212 * to every other CPU in at most O(log n) steps:
7214 * The adjacency matrix of the resulting graph is given by:
7217 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7220 * And you'll find that:
7222 * A^(log_2 n)_i,j != 0 for all i,j (7)
7224 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7225 * The task movement gives a factor of O(m), giving a convergence complexity
7228 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7233 * In order to avoid CPUs going idle while there's still work to do, new idle
7234 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7235 * tree itself instead of relying on other CPUs to bring it work.
7237 * This adds some complexity to both (5) and (8) but it reduces the total idle
7245 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7248 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7253 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7255 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7257 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7260 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7261 * rewrite all of this once again.]
7264 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7266 enum fbq_type { regular, remote, all };
7269 * 'group_type' describes the group of CPUs at the moment of load balancing.
7271 * The enum is ordered by pulling priority, with the group with lowest priority
7272 * first so the group_type can simply be compared when selecting the busiest
7273 * group. See update_sd_pick_busiest().
7276 /* The group has spare capacity that can be used to run more tasks. */
7277 group_has_spare = 0,
7279 * The group is fully used and the tasks don't compete for more CPU
7280 * cycles. Nevertheless, some tasks might wait before running.
7284 * SD_ASYM_CPUCAPACITY only: One task doesn't fit with CPU's capacity
7285 * and must be migrated to a more powerful CPU.
7289 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
7290 * and the task should be migrated to it instead of running on the
7295 * The tasks' affinity constraints previously prevented the scheduler
7296 * from balancing the load across the system.
7300 * The CPU is overloaded and can't provide expected CPU cycles to all
7306 enum migration_type {
7313 #define LBF_ALL_PINNED 0x01
7314 #define LBF_NEED_BREAK 0x02
7315 #define LBF_DST_PINNED 0x04
7316 #define LBF_SOME_PINNED 0x08
7317 #define LBF_NOHZ_STATS 0x10
7318 #define LBF_NOHZ_AGAIN 0x20
7321 struct sched_domain *sd;
7329 struct cpumask *dst_grpmask;
7331 enum cpu_idle_type idle;
7333 /* The set of CPUs under consideration for load-balancing */
7334 struct cpumask *cpus;
7339 unsigned int loop_break;
7340 unsigned int loop_max;
7342 enum fbq_type fbq_type;
7343 enum migration_type migration_type;
7344 struct list_head tasks;
7348 * Is this task likely cache-hot:
7350 static int task_hot(struct task_struct *p, struct lb_env *env)
7354 lockdep_assert_held(&env->src_rq->lock);
7356 if (p->sched_class != &fair_sched_class)
7359 if (unlikely(task_has_idle_policy(p)))
7363 * Buddy candidates are cache hot:
7365 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7366 (&p->se == cfs_rq_of(&p->se)->next ||
7367 &p->se == cfs_rq_of(&p->se)->last))
7370 if (sysctl_sched_migration_cost == -1)
7372 if (sysctl_sched_migration_cost == 0)
7375 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7377 return delta < (s64)sysctl_sched_migration_cost;
7380 #ifdef CONFIG_NUMA_BALANCING
7382 * Returns 1, if task migration degrades locality
7383 * Returns 0, if task migration improves locality i.e migration preferred.
7384 * Returns -1, if task migration is not affected by locality.
7386 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7388 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7389 unsigned long src_weight, dst_weight;
7390 int src_nid, dst_nid, dist;
7392 if (!static_branch_likely(&sched_numa_balancing))
7395 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7398 src_nid = cpu_to_node(env->src_cpu);
7399 dst_nid = cpu_to_node(env->dst_cpu);
7401 if (src_nid == dst_nid)
7404 /* Migrating away from the preferred node is always bad. */
7405 if (src_nid == p->numa_preferred_nid) {
7406 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7412 /* Encourage migration to the preferred node. */
7413 if (dst_nid == p->numa_preferred_nid)
7416 /* Leaving a core idle is often worse than degrading locality. */
7417 if (env->idle == CPU_IDLE)
7420 dist = node_distance(src_nid, dst_nid);
7422 src_weight = group_weight(p, src_nid, dist);
7423 dst_weight = group_weight(p, dst_nid, dist);
7425 src_weight = task_weight(p, src_nid, dist);
7426 dst_weight = task_weight(p, dst_nid, dist);
7429 return dst_weight < src_weight;
7433 static inline int migrate_degrades_locality(struct task_struct *p,
7441 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7444 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7448 lockdep_assert_held(&env->src_rq->lock);
7451 * We do not migrate tasks that are:
7452 * 1) throttled_lb_pair, or
7453 * 2) cannot be migrated to this CPU due to cpus_ptr, or
7454 * 3) running (obviously), or
7455 * 4) are cache-hot on their current CPU.
7457 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7460 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
7463 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7465 env->flags |= LBF_SOME_PINNED;
7468 * Remember if this task can be migrated to any other CPU in
7469 * our sched_group. We may want to revisit it if we couldn't
7470 * meet load balance goals by pulling other tasks on src_cpu.
7472 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7473 * already computed one in current iteration.
7475 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7478 /* Prevent to re-select dst_cpu via env's CPUs: */
7479 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7480 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
7481 env->flags |= LBF_DST_PINNED;
7482 env->new_dst_cpu = cpu;
7490 /* Record that we found atleast one task that could run on dst_cpu */
7491 env->flags &= ~LBF_ALL_PINNED;
7493 if (task_running(env->src_rq, p)) {
7494 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7499 * Aggressive migration if:
7500 * 1) destination numa is preferred
7501 * 2) task is cache cold, or
7502 * 3) too many balance attempts have failed.
7504 tsk_cache_hot = migrate_degrades_locality(p, env);
7505 if (tsk_cache_hot == -1)
7506 tsk_cache_hot = task_hot(p, env);
7508 if (tsk_cache_hot <= 0 ||
7509 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7510 if (tsk_cache_hot == 1) {
7511 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7512 schedstat_inc(p->se.statistics.nr_forced_migrations);
7517 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7522 * detach_task() -- detach the task for the migration specified in env
7524 static void detach_task(struct task_struct *p, struct lb_env *env)
7526 lockdep_assert_held(&env->src_rq->lock);
7528 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7529 set_task_cpu(p, env->dst_cpu);
7533 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7534 * part of active balancing operations within "domain".
7536 * Returns a task if successful and NULL otherwise.
7538 static struct task_struct *detach_one_task(struct lb_env *env)
7540 struct task_struct *p;
7542 lockdep_assert_held(&env->src_rq->lock);
7544 list_for_each_entry_reverse(p,
7545 &env->src_rq->cfs_tasks, se.group_node) {
7546 if (!can_migrate_task(p, env))
7549 detach_task(p, env);
7552 * Right now, this is only the second place where
7553 * lb_gained[env->idle] is updated (other is detach_tasks)
7554 * so we can safely collect stats here rather than
7555 * inside detach_tasks().
7557 schedstat_inc(env->sd->lb_gained[env->idle]);
7563 static const unsigned int sched_nr_migrate_break = 32;
7566 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
7567 * busiest_rq, as part of a balancing operation within domain "sd".
7569 * Returns number of detached tasks if successful and 0 otherwise.
7571 static int detach_tasks(struct lb_env *env)
7573 struct list_head *tasks = &env->src_rq->cfs_tasks;
7574 unsigned long util, load;
7575 struct task_struct *p;
7578 lockdep_assert_held(&env->src_rq->lock);
7580 if (env->imbalance <= 0)
7583 while (!list_empty(tasks)) {
7585 * We don't want to steal all, otherwise we may be treated likewise,
7586 * which could at worst lead to a livelock crash.
7588 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7591 p = list_last_entry(tasks, struct task_struct, se.group_node);
7594 /* We've more or less seen every task there is, call it quits */
7595 if (env->loop > env->loop_max)
7598 /* take a breather every nr_migrate tasks */
7599 if (env->loop > env->loop_break) {
7600 env->loop_break += sched_nr_migrate_break;
7601 env->flags |= LBF_NEED_BREAK;
7605 if (!can_migrate_task(p, env))
7608 switch (env->migration_type) {
7610 load = task_h_load(p);
7612 if (sched_feat(LB_MIN) &&
7613 load < 16 && !env->sd->nr_balance_failed)
7617 * Make sure that we don't migrate too much load.
7618 * Nevertheless, let relax the constraint if
7619 * scheduler fails to find a good waiting task to
7622 if (load/2 > env->imbalance &&
7623 env->sd->nr_balance_failed <= env->sd->cache_nice_tries)
7626 env->imbalance -= load;
7630 util = task_util_est(p);
7632 if (util > env->imbalance)
7635 env->imbalance -= util;
7642 case migrate_misfit:
7643 /* This is not a misfit task */
7644 if (task_fits_capacity(p, capacity_of(env->src_cpu)))
7651 detach_task(p, env);
7652 list_add(&p->se.group_node, &env->tasks);
7656 #ifdef CONFIG_PREEMPTION
7658 * NEWIDLE balancing is a source of latency, so preemptible
7659 * kernels will stop after the first task is detached to minimize
7660 * the critical section.
7662 if (env->idle == CPU_NEWLY_IDLE)
7667 * We only want to steal up to the prescribed amount of
7670 if (env->imbalance <= 0)
7675 list_move(&p->se.group_node, tasks);
7679 * Right now, this is one of only two places we collect this stat
7680 * so we can safely collect detach_one_task() stats here rather
7681 * than inside detach_one_task().
7683 schedstat_add(env->sd->lb_gained[env->idle], detached);
7689 * attach_task() -- attach the task detached by detach_task() to its new rq.
7691 static void attach_task(struct rq *rq, struct task_struct *p)
7693 lockdep_assert_held(&rq->lock);
7695 BUG_ON(task_rq(p) != rq);
7696 activate_task(rq, p, ENQUEUE_NOCLOCK);
7697 check_preempt_curr(rq, p, 0);
7701 * attach_one_task() -- attaches the task returned from detach_one_task() to
7704 static void attach_one_task(struct rq *rq, struct task_struct *p)
7709 update_rq_clock(rq);
7715 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7718 static void attach_tasks(struct lb_env *env)
7720 struct list_head *tasks = &env->tasks;
7721 struct task_struct *p;
7724 rq_lock(env->dst_rq, &rf);
7725 update_rq_clock(env->dst_rq);
7727 while (!list_empty(tasks)) {
7728 p = list_first_entry(tasks, struct task_struct, se.group_node);
7729 list_del_init(&p->se.group_node);
7731 attach_task(env->dst_rq, p);
7734 rq_unlock(env->dst_rq, &rf);
7737 #ifdef CONFIG_NO_HZ_COMMON
7738 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7740 if (cfs_rq->avg.load_avg)
7743 if (cfs_rq->avg.util_avg)
7749 static inline bool others_have_blocked(struct rq *rq)
7751 if (READ_ONCE(rq->avg_rt.util_avg))
7754 if (READ_ONCE(rq->avg_dl.util_avg))
7757 if (thermal_load_avg(rq))
7760 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7761 if (READ_ONCE(rq->avg_irq.util_avg))
7768 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
7770 rq->last_blocked_load_update_tick = jiffies;
7773 rq->has_blocked_load = 0;
7776 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
7777 static inline bool others_have_blocked(struct rq *rq) { return false; }
7778 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
7781 static bool __update_blocked_others(struct rq *rq, bool *done)
7783 const struct sched_class *curr_class;
7784 u64 now = rq_clock_pelt(rq);
7785 unsigned long thermal_pressure;
7789 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
7790 * DL and IRQ signals have been updated before updating CFS.
7792 curr_class = rq->curr->sched_class;
7794 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
7796 decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
7797 update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
7798 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) |
7799 update_irq_load_avg(rq, 0);
7801 if (others_have_blocked(rq))
7807 #ifdef CONFIG_FAIR_GROUP_SCHED
7809 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7811 if (cfs_rq->load.weight)
7814 if (cfs_rq->avg.load_sum)
7817 if (cfs_rq->avg.util_sum)
7820 if (cfs_rq->avg.runnable_sum)
7826 static bool __update_blocked_fair(struct rq *rq, bool *done)
7828 struct cfs_rq *cfs_rq, *pos;
7829 bool decayed = false;
7830 int cpu = cpu_of(rq);
7833 * Iterates the task_group tree in a bottom up fashion, see
7834 * list_add_leaf_cfs_rq() for details.
7836 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7837 struct sched_entity *se;
7839 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
7840 update_tg_load_avg(cfs_rq, 0);
7842 if (cfs_rq == &rq->cfs)
7846 /* Propagate pending load changes to the parent, if any: */
7847 se = cfs_rq->tg->se[cpu];
7848 if (se && !skip_blocked_update(se))
7849 update_load_avg(cfs_rq_of(se), se, 0);
7852 * There can be a lot of idle CPU cgroups. Don't let fully
7853 * decayed cfs_rqs linger on the list.
7855 if (cfs_rq_is_decayed(cfs_rq))
7856 list_del_leaf_cfs_rq(cfs_rq);
7858 /* Don't need periodic decay once load/util_avg are null */
7859 if (cfs_rq_has_blocked(cfs_rq))
7867 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7868 * This needs to be done in a top-down fashion because the load of a child
7869 * group is a fraction of its parents load.
7871 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7873 struct rq *rq = rq_of(cfs_rq);
7874 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7875 unsigned long now = jiffies;
7878 if (cfs_rq->last_h_load_update == now)
7881 WRITE_ONCE(cfs_rq->h_load_next, NULL);
7882 for_each_sched_entity(se) {
7883 cfs_rq = cfs_rq_of(se);
7884 WRITE_ONCE(cfs_rq->h_load_next, se);
7885 if (cfs_rq->last_h_load_update == now)
7890 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7891 cfs_rq->last_h_load_update = now;
7894 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
7895 load = cfs_rq->h_load;
7896 load = div64_ul(load * se->avg.load_avg,
7897 cfs_rq_load_avg(cfs_rq) + 1);
7898 cfs_rq = group_cfs_rq(se);
7899 cfs_rq->h_load = load;
7900 cfs_rq->last_h_load_update = now;
7904 static unsigned long task_h_load(struct task_struct *p)
7906 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7908 update_cfs_rq_h_load(cfs_rq);
7909 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7910 cfs_rq_load_avg(cfs_rq) + 1);
7913 static bool __update_blocked_fair(struct rq *rq, bool *done)
7915 struct cfs_rq *cfs_rq = &rq->cfs;
7918 decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
7919 if (cfs_rq_has_blocked(cfs_rq))
7925 static unsigned long task_h_load(struct task_struct *p)
7927 return p->se.avg.load_avg;
7931 static void update_blocked_averages(int cpu)
7933 bool decayed = false, done = true;
7934 struct rq *rq = cpu_rq(cpu);
7937 rq_lock_irqsave(rq, &rf);
7938 update_rq_clock(rq);
7940 decayed |= __update_blocked_others(rq, &done);
7941 decayed |= __update_blocked_fair(rq, &done);
7943 update_blocked_load_status(rq, !done);
7945 cpufreq_update_util(rq, 0);
7946 rq_unlock_irqrestore(rq, &rf);
7949 /********** Helpers for find_busiest_group ************************/
7952 * sg_lb_stats - stats of a sched_group required for load_balancing
7954 struct sg_lb_stats {
7955 unsigned long avg_load; /*Avg load across the CPUs of the group */
7956 unsigned long group_load; /* Total load over the CPUs of the group */
7957 unsigned long group_capacity;
7958 unsigned long group_util; /* Total utilization over the CPUs of the group */
7959 unsigned long group_runnable; /* Total runnable time over the CPUs of the group */
7960 unsigned int sum_nr_running; /* Nr of tasks running in the group */
7961 unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
7962 unsigned int idle_cpus;
7963 unsigned int group_weight;
7964 enum group_type group_type;
7965 unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
7966 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
7967 #ifdef CONFIG_NUMA_BALANCING
7968 unsigned int nr_numa_running;
7969 unsigned int nr_preferred_running;
7974 * sd_lb_stats - Structure to store the statistics of a sched_domain
7975 * during load balancing.
7977 struct sd_lb_stats {
7978 struct sched_group *busiest; /* Busiest group in this sd */
7979 struct sched_group *local; /* Local group in this sd */
7980 unsigned long total_load; /* Total load of all groups in sd */
7981 unsigned long total_capacity; /* Total capacity of all groups in sd */
7982 unsigned long avg_load; /* Average load across all groups in sd */
7983 unsigned int prefer_sibling; /* tasks should go to sibling first */
7985 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7986 struct sg_lb_stats local_stat; /* Statistics of the local group */
7989 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7992 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7993 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7994 * We must however set busiest_stat::group_type and
7995 * busiest_stat::idle_cpus to the worst busiest group because
7996 * update_sd_pick_busiest() reads these before assignment.
7998 *sds = (struct sd_lb_stats){
8002 .total_capacity = 0UL,
8004 .idle_cpus = UINT_MAX,
8005 .group_type = group_has_spare,
8010 static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu)
8012 struct rq *rq = cpu_rq(cpu);
8013 unsigned long max = arch_scale_cpu_capacity(cpu);
8014 unsigned long used, free;
8017 irq = cpu_util_irq(rq);
8019 if (unlikely(irq >= max))
8023 * avg_rt.util_avg and avg_dl.util_avg track binary signals
8024 * (running and not running) with weights 0 and 1024 respectively.
8025 * avg_thermal.load_avg tracks thermal pressure and the weighted
8026 * average uses the actual delta max capacity(load).
8028 used = READ_ONCE(rq->avg_rt.util_avg);
8029 used += READ_ONCE(rq->avg_dl.util_avg);
8030 used += thermal_load_avg(rq);
8032 if (unlikely(used >= max))
8037 return scale_irq_capacity(free, irq, max);
8040 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
8042 unsigned long capacity = scale_rt_capacity(sd, cpu);
8043 struct sched_group *sdg = sd->groups;
8045 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
8050 cpu_rq(cpu)->cpu_capacity = capacity;
8051 sdg->sgc->capacity = capacity;
8052 sdg->sgc->min_capacity = capacity;
8053 sdg->sgc->max_capacity = capacity;
8056 void update_group_capacity(struct sched_domain *sd, int cpu)
8058 struct sched_domain *child = sd->child;
8059 struct sched_group *group, *sdg = sd->groups;
8060 unsigned long capacity, min_capacity, max_capacity;
8061 unsigned long interval;
8063 interval = msecs_to_jiffies(sd->balance_interval);
8064 interval = clamp(interval, 1UL, max_load_balance_interval);
8065 sdg->sgc->next_update = jiffies + interval;
8068 update_cpu_capacity(sd, cpu);
8073 min_capacity = ULONG_MAX;
8076 if (child->flags & SD_OVERLAP) {
8078 * SD_OVERLAP domains cannot assume that child groups
8079 * span the current group.
8082 for_each_cpu(cpu, sched_group_span(sdg)) {
8083 unsigned long cpu_cap = capacity_of(cpu);
8085 capacity += cpu_cap;
8086 min_capacity = min(cpu_cap, min_capacity);
8087 max_capacity = max(cpu_cap, max_capacity);
8091 * !SD_OVERLAP domains can assume that child groups
8092 * span the current group.
8095 group = child->groups;
8097 struct sched_group_capacity *sgc = group->sgc;
8099 capacity += sgc->capacity;
8100 min_capacity = min(sgc->min_capacity, min_capacity);
8101 max_capacity = max(sgc->max_capacity, max_capacity);
8102 group = group->next;
8103 } while (group != child->groups);
8106 sdg->sgc->capacity = capacity;
8107 sdg->sgc->min_capacity = min_capacity;
8108 sdg->sgc->max_capacity = max_capacity;
8112 * Check whether the capacity of the rq has been noticeably reduced by side
8113 * activity. The imbalance_pct is used for the threshold.
8114 * Return true is the capacity is reduced
8117 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
8119 return ((rq->cpu_capacity * sd->imbalance_pct) <
8120 (rq->cpu_capacity_orig * 100));
8124 * Check whether a rq has a misfit task and if it looks like we can actually
8125 * help that task: we can migrate the task to a CPU of higher capacity, or
8126 * the task's current CPU is heavily pressured.
8128 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
8130 return rq->misfit_task_load &&
8131 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
8132 check_cpu_capacity(rq, sd));
8136 * Group imbalance indicates (and tries to solve) the problem where balancing
8137 * groups is inadequate due to ->cpus_ptr constraints.
8139 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
8140 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
8143 * { 0 1 2 3 } { 4 5 6 7 }
8146 * If we were to balance group-wise we'd place two tasks in the first group and
8147 * two tasks in the second group. Clearly this is undesired as it will overload
8148 * cpu 3 and leave one of the CPUs in the second group unused.
8150 * The current solution to this issue is detecting the skew in the first group
8151 * by noticing the lower domain failed to reach balance and had difficulty
8152 * moving tasks due to affinity constraints.
8154 * When this is so detected; this group becomes a candidate for busiest; see
8155 * update_sd_pick_busiest(). And calculate_imbalance() and
8156 * find_busiest_group() avoid some of the usual balance conditions to allow it
8157 * to create an effective group imbalance.
8159 * This is a somewhat tricky proposition since the next run might not find the
8160 * group imbalance and decide the groups need to be balanced again. A most
8161 * subtle and fragile situation.
8164 static inline int sg_imbalanced(struct sched_group *group)
8166 return group->sgc->imbalance;
8170 * group_has_capacity returns true if the group has spare capacity that could
8171 * be used by some tasks.
8172 * We consider that a group has spare capacity if the * number of task is
8173 * smaller than the number of CPUs or if the utilization is lower than the
8174 * available capacity for CFS tasks.
8175 * For the latter, we use a threshold to stabilize the state, to take into
8176 * account the variance of the tasks' load and to return true if the available
8177 * capacity in meaningful for the load balancer.
8178 * As an example, an available capacity of 1% can appear but it doesn't make
8179 * any benefit for the load balance.
8182 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
8184 if (sgs->sum_nr_running < sgs->group_weight)
8187 if ((sgs->group_capacity * imbalance_pct) <
8188 (sgs->group_runnable * 100))
8191 if ((sgs->group_capacity * 100) >
8192 (sgs->group_util * imbalance_pct))
8199 * group_is_overloaded returns true if the group has more tasks than it can
8201 * group_is_overloaded is not equals to !group_has_capacity because a group
8202 * with the exact right number of tasks, has no more spare capacity but is not
8203 * overloaded so both group_has_capacity and group_is_overloaded return
8207 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
8209 if (sgs->sum_nr_running <= sgs->group_weight)
8212 if ((sgs->group_capacity * 100) <
8213 (sgs->group_util * imbalance_pct))
8216 if ((sgs->group_capacity * imbalance_pct) <
8217 (sgs->group_runnable * 100))
8224 * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller
8225 * per-CPU capacity than sched_group ref.
8228 group_smaller_min_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
8230 return fits_capacity(sg->sgc->min_capacity, ref->sgc->min_capacity);
8234 * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller
8235 * per-CPU capacity_orig than sched_group ref.
8238 group_smaller_max_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
8240 return fits_capacity(sg->sgc->max_capacity, ref->sgc->max_capacity);
8244 group_type group_classify(unsigned int imbalance_pct,
8245 struct sched_group *group,
8246 struct sg_lb_stats *sgs)
8248 if (group_is_overloaded(imbalance_pct, sgs))
8249 return group_overloaded;
8251 if (sg_imbalanced(group))
8252 return group_imbalanced;
8254 if (sgs->group_asym_packing)
8255 return group_asym_packing;
8257 if (sgs->group_misfit_task_load)
8258 return group_misfit_task;
8260 if (!group_has_capacity(imbalance_pct, sgs))
8261 return group_fully_busy;
8263 return group_has_spare;
8266 static bool update_nohz_stats(struct rq *rq, bool force)
8268 #ifdef CONFIG_NO_HZ_COMMON
8269 unsigned int cpu = rq->cpu;
8271 if (!rq->has_blocked_load)
8274 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
8277 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
8280 update_blocked_averages(cpu);
8282 return rq->has_blocked_load;
8289 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8290 * @env: The load balancing environment.
8291 * @group: sched_group whose statistics are to be updated.
8292 * @sgs: variable to hold the statistics for this group.
8293 * @sg_status: Holds flag indicating the status of the sched_group
8295 static inline void update_sg_lb_stats(struct lb_env *env,
8296 struct sched_group *group,
8297 struct sg_lb_stats *sgs,
8300 int i, nr_running, local_group;
8302 memset(sgs, 0, sizeof(*sgs));
8304 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(group));
8306 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8307 struct rq *rq = cpu_rq(i);
8309 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
8310 env->flags |= LBF_NOHZ_AGAIN;
8312 sgs->group_load += cpu_load(rq);
8313 sgs->group_util += cpu_util(i);
8314 sgs->group_runnable += cpu_runnable(rq);
8315 sgs->sum_h_nr_running += rq->cfs.h_nr_running;
8317 nr_running = rq->nr_running;
8318 sgs->sum_nr_running += nr_running;
8321 *sg_status |= SG_OVERLOAD;
8323 if (cpu_overutilized(i))
8324 *sg_status |= SG_OVERUTILIZED;
8326 #ifdef CONFIG_NUMA_BALANCING
8327 sgs->nr_numa_running += rq->nr_numa_running;
8328 sgs->nr_preferred_running += rq->nr_preferred_running;
8331 * No need to call idle_cpu() if nr_running is not 0
8333 if (!nr_running && idle_cpu(i)) {
8335 /* Idle cpu can't have misfit task */
8342 /* Check for a misfit task on the cpu */
8343 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8344 sgs->group_misfit_task_load < rq->misfit_task_load) {
8345 sgs->group_misfit_task_load = rq->misfit_task_load;
8346 *sg_status |= SG_OVERLOAD;
8350 /* Check if dst CPU is idle and preferred to this group */
8351 if (env->sd->flags & SD_ASYM_PACKING &&
8352 env->idle != CPU_NOT_IDLE &&
8353 sgs->sum_h_nr_running &&
8354 sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu)) {
8355 sgs->group_asym_packing = 1;
8358 sgs->group_capacity = group->sgc->capacity;
8360 sgs->group_weight = group->group_weight;
8362 sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
8364 /* Computing avg_load makes sense only when group is overloaded */
8365 if (sgs->group_type == group_overloaded)
8366 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
8367 sgs->group_capacity;
8371 * update_sd_pick_busiest - return 1 on busiest group
8372 * @env: The load balancing environment.
8373 * @sds: sched_domain statistics
8374 * @sg: sched_group candidate to be checked for being the busiest
8375 * @sgs: sched_group statistics
8377 * Determine if @sg is a busier group than the previously selected
8380 * Return: %true if @sg is a busier group than the previously selected
8381 * busiest group. %false otherwise.
8383 static bool update_sd_pick_busiest(struct lb_env *env,
8384 struct sd_lb_stats *sds,
8385 struct sched_group *sg,
8386 struct sg_lb_stats *sgs)
8388 struct sg_lb_stats *busiest = &sds->busiest_stat;
8390 /* Make sure that there is at least one task to pull */
8391 if (!sgs->sum_h_nr_running)
8395 * Don't try to pull misfit tasks we can't help.
8396 * We can use max_capacity here as reduction in capacity on some
8397 * CPUs in the group should either be possible to resolve
8398 * internally or be covered by avg_load imbalance (eventually).
8400 if (sgs->group_type == group_misfit_task &&
8401 (!group_smaller_max_cpu_capacity(sg, sds->local) ||
8402 sds->local_stat.group_type != group_has_spare))
8405 if (sgs->group_type > busiest->group_type)
8408 if (sgs->group_type < busiest->group_type)
8412 * The candidate and the current busiest group are the same type of
8413 * group. Let check which one is the busiest according to the type.
8416 switch (sgs->group_type) {
8417 case group_overloaded:
8418 /* Select the overloaded group with highest avg_load. */
8419 if (sgs->avg_load <= busiest->avg_load)
8423 case group_imbalanced:
8425 * Select the 1st imbalanced group as we don't have any way to
8426 * choose one more than another.
8430 case group_asym_packing:
8431 /* Prefer to move from lowest priority CPU's work */
8432 if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu))
8436 case group_misfit_task:
8438 * If we have more than one misfit sg go with the biggest
8441 if (sgs->group_misfit_task_load < busiest->group_misfit_task_load)
8445 case group_fully_busy:
8447 * Select the fully busy group with highest avg_load. In
8448 * theory, there is no need to pull task from such kind of
8449 * group because tasks have all compute capacity that they need
8450 * but we can still improve the overall throughput by reducing
8451 * contention when accessing shared HW resources.
8453 * XXX for now avg_load is not computed and always 0 so we
8454 * select the 1st one.
8456 if (sgs->avg_load <= busiest->avg_load)
8460 case group_has_spare:
8462 * Select not overloaded group with lowest number of idle cpus
8463 * and highest number of running tasks. We could also compare
8464 * the spare capacity which is more stable but it can end up
8465 * that the group has less spare capacity but finally more idle
8466 * CPUs which means less opportunity to pull tasks.
8468 if (sgs->idle_cpus > busiest->idle_cpus)
8470 else if ((sgs->idle_cpus == busiest->idle_cpus) &&
8471 (sgs->sum_nr_running <= busiest->sum_nr_running))
8478 * Candidate sg has no more than one task per CPU and has higher
8479 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
8480 * throughput. Maximize throughput, power/energy consequences are not
8483 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
8484 (sgs->group_type <= group_fully_busy) &&
8485 (group_smaller_min_cpu_capacity(sds->local, sg)))
8491 #ifdef CONFIG_NUMA_BALANCING
8492 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8494 if (sgs->sum_h_nr_running > sgs->nr_numa_running)
8496 if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
8501 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8503 if (rq->nr_running > rq->nr_numa_running)
8505 if (rq->nr_running > rq->nr_preferred_running)
8510 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8515 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8519 #endif /* CONFIG_NUMA_BALANCING */
8525 * task_running_on_cpu - return 1 if @p is running on @cpu.
8528 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
8530 /* Task has no contribution or is new */
8531 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
8534 if (task_on_rq_queued(p))
8541 * idle_cpu_without - would a given CPU be idle without p ?
8542 * @cpu: the processor on which idleness is tested.
8543 * @p: task which should be ignored.
8545 * Return: 1 if the CPU would be idle. 0 otherwise.
8547 static int idle_cpu_without(int cpu, struct task_struct *p)
8549 struct rq *rq = cpu_rq(cpu);
8551 if (rq->curr != rq->idle && rq->curr != p)
8555 * rq->nr_running can't be used but an updated version without the
8556 * impact of p on cpu must be used instead. The updated nr_running
8557 * be computed and tested before calling idle_cpu_without().
8561 if (!llist_empty(&rq->wake_list))
8569 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
8570 * @sd: The sched_domain level to look for idlest group.
8571 * @group: sched_group whose statistics are to be updated.
8572 * @sgs: variable to hold the statistics for this group.
8573 * @p: The task for which we look for the idlest group/CPU.
8575 static inline void update_sg_wakeup_stats(struct sched_domain *sd,
8576 struct sched_group *group,
8577 struct sg_lb_stats *sgs,
8578 struct task_struct *p)
8582 memset(sgs, 0, sizeof(*sgs));
8584 for_each_cpu(i, sched_group_span(group)) {
8585 struct rq *rq = cpu_rq(i);
8588 sgs->group_load += cpu_load_without(rq, p);
8589 sgs->group_util += cpu_util_without(i, p);
8590 sgs->group_runnable += cpu_runnable_without(rq, p);
8591 local = task_running_on_cpu(i, p);
8592 sgs->sum_h_nr_running += rq->cfs.h_nr_running - local;
8594 nr_running = rq->nr_running - local;
8595 sgs->sum_nr_running += nr_running;
8598 * No need to call idle_cpu_without() if nr_running is not 0
8600 if (!nr_running && idle_cpu_without(i, p))
8605 /* Check if task fits in the group */
8606 if (sd->flags & SD_ASYM_CPUCAPACITY &&
8607 !task_fits_capacity(p, group->sgc->max_capacity)) {
8608 sgs->group_misfit_task_load = 1;
8611 sgs->group_capacity = group->sgc->capacity;
8613 sgs->group_weight = group->group_weight;
8615 sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
8618 * Computing avg_load makes sense only when group is fully busy or
8621 if (sgs->group_type == group_fully_busy ||
8622 sgs->group_type == group_overloaded)
8623 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
8624 sgs->group_capacity;
8627 static bool update_pick_idlest(struct sched_group *idlest,
8628 struct sg_lb_stats *idlest_sgs,
8629 struct sched_group *group,
8630 struct sg_lb_stats *sgs)
8632 if (sgs->group_type < idlest_sgs->group_type)
8635 if (sgs->group_type > idlest_sgs->group_type)
8639 * The candidate and the current idlest group are the same type of
8640 * group. Let check which one is the idlest according to the type.
8643 switch (sgs->group_type) {
8644 case group_overloaded:
8645 case group_fully_busy:
8646 /* Select the group with lowest avg_load. */
8647 if (idlest_sgs->avg_load <= sgs->avg_load)
8651 case group_imbalanced:
8652 case group_asym_packing:
8653 /* Those types are not used in the slow wakeup path */
8656 case group_misfit_task:
8657 /* Select group with the highest max capacity */
8658 if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
8662 case group_has_spare:
8663 /* Select group with most idle CPUs */
8664 if (idlest_sgs->idle_cpus >= sgs->idle_cpus)
8673 * find_idlest_group() finds and returns the least busy CPU group within the
8676 * Assumes p is allowed on at least one CPU in sd.
8678 static struct sched_group *
8679 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
8680 int this_cpu, int sd_flag)
8682 struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
8683 struct sg_lb_stats local_sgs, tmp_sgs;
8684 struct sg_lb_stats *sgs;
8685 unsigned long imbalance;
8686 struct sg_lb_stats idlest_sgs = {
8687 .avg_load = UINT_MAX,
8688 .group_type = group_overloaded,
8691 imbalance = scale_load_down(NICE_0_LOAD) *
8692 (sd->imbalance_pct-100) / 100;
8697 /* Skip over this group if it has no CPUs allowed */
8698 if (!cpumask_intersects(sched_group_span(group),
8702 local_group = cpumask_test_cpu(this_cpu,
8703 sched_group_span(group));
8712 update_sg_wakeup_stats(sd, group, sgs, p);
8714 if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
8719 } while (group = group->next, group != sd->groups);
8722 /* There is no idlest group to push tasks to */
8726 /* The local group has been skipped because of CPU affinity */
8731 * If the local group is idler than the selected idlest group
8732 * don't try and push the task.
8734 if (local_sgs.group_type < idlest_sgs.group_type)
8738 * If the local group is busier than the selected idlest group
8739 * try and push the task.
8741 if (local_sgs.group_type > idlest_sgs.group_type)
8744 switch (local_sgs.group_type) {
8745 case group_overloaded:
8746 case group_fully_busy:
8748 * When comparing groups across NUMA domains, it's possible for
8749 * the local domain to be very lightly loaded relative to the
8750 * remote domains but "imbalance" skews the comparison making
8751 * remote CPUs look much more favourable. When considering
8752 * cross-domain, add imbalance to the load on the remote node
8753 * and consider staying local.
8756 if ((sd->flags & SD_NUMA) &&
8757 ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
8761 * If the local group is less loaded than the selected
8762 * idlest group don't try and push any tasks.
8764 if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
8767 if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
8771 case group_imbalanced:
8772 case group_asym_packing:
8773 /* Those type are not used in the slow wakeup path */
8776 case group_misfit_task:
8777 /* Select group with the highest max capacity */
8778 if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
8782 case group_has_spare:
8783 if (sd->flags & SD_NUMA) {
8784 #ifdef CONFIG_NUMA_BALANCING
8787 * If there is spare capacity at NUMA, try to select
8788 * the preferred node
8790 if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
8793 idlest_cpu = cpumask_first(sched_group_span(idlest));
8794 if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
8798 * Otherwise, keep the task on this node to stay close
8799 * its wakeup source and improve locality. If there is
8800 * a real need of migration, periodic load balance will
8803 if (local_sgs.idle_cpus)
8808 * Select group with highest number of idle CPUs. We could also
8809 * compare the utilization which is more stable but it can end
8810 * up that the group has less spare capacity but finally more
8811 * idle CPUs which means more opportunity to run task.
8813 if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
8822 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8823 * @env: The load balancing environment.
8824 * @sds: variable to hold the statistics for this sched_domain.
8827 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8829 struct sched_domain *child = env->sd->child;
8830 struct sched_group *sg = env->sd->groups;
8831 struct sg_lb_stats *local = &sds->local_stat;
8832 struct sg_lb_stats tmp_sgs;
8835 #ifdef CONFIG_NO_HZ_COMMON
8836 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
8837 env->flags |= LBF_NOHZ_STATS;
8841 struct sg_lb_stats *sgs = &tmp_sgs;
8844 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
8849 if (env->idle != CPU_NEWLY_IDLE ||
8850 time_after_eq(jiffies, sg->sgc->next_update))
8851 update_group_capacity(env->sd, env->dst_cpu);
8854 update_sg_lb_stats(env, sg, sgs, &sg_status);
8860 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8862 sds->busiest_stat = *sgs;
8866 /* Now, start updating sd_lb_stats */
8867 sds->total_load += sgs->group_load;
8868 sds->total_capacity += sgs->group_capacity;
8871 } while (sg != env->sd->groups);
8873 /* Tag domain that child domain prefers tasks go to siblings first */
8874 sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
8876 #ifdef CONFIG_NO_HZ_COMMON
8877 if ((env->flags & LBF_NOHZ_AGAIN) &&
8878 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8880 WRITE_ONCE(nohz.next_blocked,
8881 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8885 if (env->sd->flags & SD_NUMA)
8886 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8888 if (!env->sd->parent) {
8889 struct root_domain *rd = env->dst_rq->rd;
8891 /* update overload indicator if we are at root domain */
8892 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
8894 /* Update over-utilization (tipping point, U >= 0) indicator */
8895 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
8896 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
8897 } else if (sg_status & SG_OVERUTILIZED) {
8898 struct root_domain *rd = env->dst_rq->rd;
8900 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
8901 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
8905 static inline long adjust_numa_imbalance(int imbalance, int src_nr_running)
8907 unsigned int imbalance_min;
8910 * Allow a small imbalance based on a simple pair of communicating
8911 * tasks that remain local when the source domain is almost idle.
8914 if (src_nr_running <= imbalance_min)
8921 * calculate_imbalance - Calculate the amount of imbalance present within the
8922 * groups of a given sched_domain during load balance.
8923 * @env: load balance environment
8924 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8926 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8928 struct sg_lb_stats *local, *busiest;
8930 local = &sds->local_stat;
8931 busiest = &sds->busiest_stat;
8933 if (busiest->group_type == group_misfit_task) {
8934 /* Set imbalance to allow misfit tasks to be balanced. */
8935 env->migration_type = migrate_misfit;
8940 if (busiest->group_type == group_asym_packing) {
8942 * In case of asym capacity, we will try to migrate all load to
8943 * the preferred CPU.
8945 env->migration_type = migrate_task;
8946 env->imbalance = busiest->sum_h_nr_running;
8950 if (busiest->group_type == group_imbalanced) {
8952 * In the group_imb case we cannot rely on group-wide averages
8953 * to ensure CPU-load equilibrium, try to move any task to fix
8954 * the imbalance. The next load balance will take care of
8955 * balancing back the system.
8957 env->migration_type = migrate_task;
8963 * Try to use spare capacity of local group without overloading it or
8966 if (local->group_type == group_has_spare) {
8967 if (busiest->group_type > group_fully_busy) {
8969 * If busiest is overloaded, try to fill spare
8970 * capacity. This might end up creating spare capacity
8971 * in busiest or busiest still being overloaded but
8972 * there is no simple way to directly compute the
8973 * amount of load to migrate in order to balance the
8976 env->migration_type = migrate_util;
8977 env->imbalance = max(local->group_capacity, local->group_util) -
8981 * In some cases, the group's utilization is max or even
8982 * higher than capacity because of migrations but the
8983 * local CPU is (newly) idle. There is at least one
8984 * waiting task in this overloaded busiest group. Let's
8987 if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) {
8988 env->migration_type = migrate_task;
8995 if (busiest->group_weight == 1 || sds->prefer_sibling) {
8996 unsigned int nr_diff = busiest->sum_nr_running;
8998 * When prefer sibling, evenly spread running tasks on
9001 env->migration_type = migrate_task;
9002 lsub_positive(&nr_diff, local->sum_nr_running);
9003 env->imbalance = nr_diff >> 1;
9007 * If there is no overload, we just want to even the number of
9010 env->migration_type = migrate_task;
9011 env->imbalance = max_t(long, 0, (local->idle_cpus -
9012 busiest->idle_cpus) >> 1);
9015 /* Consider allowing a small imbalance between NUMA groups */
9016 if (env->sd->flags & SD_NUMA)
9017 env->imbalance = adjust_numa_imbalance(env->imbalance,
9018 busiest->sum_nr_running);
9024 * Local is fully busy but has to take more load to relieve the
9027 if (local->group_type < group_overloaded) {
9029 * Local will become overloaded so the avg_load metrics are
9033 local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
9034 local->group_capacity;
9036 sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
9037 sds->total_capacity;
9039 * If the local group is more loaded than the selected
9040 * busiest group don't try to pull any tasks.
9042 if (local->avg_load >= busiest->avg_load) {
9049 * Both group are or will become overloaded and we're trying to get all
9050 * the CPUs to the average_load, so we don't want to push ourselves
9051 * above the average load, nor do we wish to reduce the max loaded CPU
9052 * below the average load. At the same time, we also don't want to
9053 * reduce the group load below the group capacity. Thus we look for
9054 * the minimum possible imbalance.
9056 env->migration_type = migrate_load;
9057 env->imbalance = min(
9058 (busiest->avg_load - sds->avg_load) * busiest->group_capacity,
9059 (sds->avg_load - local->avg_load) * local->group_capacity
9060 ) / SCHED_CAPACITY_SCALE;
9063 /******* find_busiest_group() helpers end here *********************/
9066 * Decision matrix according to the local and busiest group type:
9068 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
9069 * has_spare nr_idle balanced N/A N/A balanced balanced
9070 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
9071 * misfit_task force N/A N/A N/A force force
9072 * asym_packing force force N/A N/A force force
9073 * imbalanced force force N/A N/A force force
9074 * overloaded force force N/A N/A force avg_load
9076 * N/A : Not Applicable because already filtered while updating
9078 * balanced : The system is balanced for these 2 groups.
9079 * force : Calculate the imbalance as load migration is probably needed.
9080 * avg_load : Only if imbalance is significant enough.
9081 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
9082 * different in groups.
9086 * find_busiest_group - Returns the busiest group within the sched_domain
9087 * if there is an imbalance.
9089 * Also calculates the amount of runnable load which should be moved
9090 * to restore balance.
9092 * @env: The load balancing environment.
9094 * Return: - The busiest group if imbalance exists.
9096 static struct sched_group *find_busiest_group(struct lb_env *env)
9098 struct sg_lb_stats *local, *busiest;
9099 struct sd_lb_stats sds;
9101 init_sd_lb_stats(&sds);
9104 * Compute the various statistics relevant for load balancing at
9107 update_sd_lb_stats(env, &sds);
9109 if (sched_energy_enabled()) {
9110 struct root_domain *rd = env->dst_rq->rd;
9112 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
9116 local = &sds.local_stat;
9117 busiest = &sds.busiest_stat;
9119 /* There is no busy sibling group to pull tasks from */
9123 /* Misfit tasks should be dealt with regardless of the avg load */
9124 if (busiest->group_type == group_misfit_task)
9127 /* ASYM feature bypasses nice load balance check */
9128 if (busiest->group_type == group_asym_packing)
9132 * If the busiest group is imbalanced the below checks don't
9133 * work because they assume all things are equal, which typically
9134 * isn't true due to cpus_ptr constraints and the like.
9136 if (busiest->group_type == group_imbalanced)
9140 * If the local group is busier than the selected busiest group
9141 * don't try and pull any tasks.
9143 if (local->group_type > busiest->group_type)
9147 * When groups are overloaded, use the avg_load to ensure fairness
9150 if (local->group_type == group_overloaded) {
9152 * If the local group is more loaded than the selected
9153 * busiest group don't try to pull any tasks.
9155 if (local->avg_load >= busiest->avg_load)
9158 /* XXX broken for overlapping NUMA groups */
9159 sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
9163 * Don't pull any tasks if this group is already above the
9164 * domain average load.
9166 if (local->avg_load >= sds.avg_load)
9170 * If the busiest group is more loaded, use imbalance_pct to be
9173 if (100 * busiest->avg_load <=
9174 env->sd->imbalance_pct * local->avg_load)
9178 /* Try to move all excess tasks to child's sibling domain */
9179 if (sds.prefer_sibling && local->group_type == group_has_spare &&
9180 busiest->sum_nr_running > local->sum_nr_running + 1)
9183 if (busiest->group_type != group_overloaded) {
9184 if (env->idle == CPU_NOT_IDLE)
9186 * If the busiest group is not overloaded (and as a
9187 * result the local one too) but this CPU is already
9188 * busy, let another idle CPU try to pull task.
9192 if (busiest->group_weight > 1 &&
9193 local->idle_cpus <= (busiest->idle_cpus + 1))
9195 * If the busiest group is not overloaded
9196 * and there is no imbalance between this and busiest
9197 * group wrt idle CPUs, it is balanced. The imbalance
9198 * becomes significant if the diff is greater than 1
9199 * otherwise we might end up to just move the imbalance
9200 * on another group. Of course this applies only if
9201 * there is more than 1 CPU per group.
9205 if (busiest->sum_h_nr_running == 1)
9207 * busiest doesn't have any tasks waiting to run
9213 /* Looks like there is an imbalance. Compute it */
9214 calculate_imbalance(env, &sds);
9215 return env->imbalance ? sds.busiest : NULL;
9223 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
9225 static struct rq *find_busiest_queue(struct lb_env *env,
9226 struct sched_group *group)
9228 struct rq *busiest = NULL, *rq;
9229 unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
9230 unsigned int busiest_nr = 0;
9233 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
9234 unsigned long capacity, load, util;
9235 unsigned int nr_running;
9239 rt = fbq_classify_rq(rq);
9242 * We classify groups/runqueues into three groups:
9243 * - regular: there are !numa tasks
9244 * - remote: there are numa tasks that run on the 'wrong' node
9245 * - all: there is no distinction
9247 * In order to avoid migrating ideally placed numa tasks,
9248 * ignore those when there's better options.
9250 * If we ignore the actual busiest queue to migrate another
9251 * task, the next balance pass can still reduce the busiest
9252 * queue by moving tasks around inside the node.
9254 * If we cannot move enough load due to this classification
9255 * the next pass will adjust the group classification and
9256 * allow migration of more tasks.
9258 * Both cases only affect the total convergence complexity.
9260 if (rt > env->fbq_type)
9263 capacity = capacity_of(i);
9264 nr_running = rq->cfs.h_nr_running;
9267 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
9268 * eventually lead to active_balancing high->low capacity.
9269 * Higher per-CPU capacity is considered better than balancing
9272 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
9273 capacity_of(env->dst_cpu) < capacity &&
9277 switch (env->migration_type) {
9280 * When comparing with load imbalance, use cpu_load()
9281 * which is not scaled with the CPU capacity.
9283 load = cpu_load(rq);
9285 if (nr_running == 1 && load > env->imbalance &&
9286 !check_cpu_capacity(rq, env->sd))
9290 * For the load comparisons with the other CPUs,
9291 * consider the cpu_load() scaled with the CPU
9292 * capacity, so that the load can be moved away
9293 * from the CPU that is potentially running at a
9296 * Thus we're looking for max(load_i / capacity_i),
9297 * crosswise multiplication to rid ourselves of the
9298 * division works out to:
9299 * load_i * capacity_j > load_j * capacity_i;
9300 * where j is our previous maximum.
9302 if (load * busiest_capacity > busiest_load * capacity) {
9303 busiest_load = load;
9304 busiest_capacity = capacity;
9310 util = cpu_util(cpu_of(rq));
9313 * Don't try to pull utilization from a CPU with one
9314 * running task. Whatever its utilization, we will fail
9317 if (nr_running <= 1)
9320 if (busiest_util < util) {
9321 busiest_util = util;
9327 if (busiest_nr < nr_running) {
9328 busiest_nr = nr_running;
9333 case migrate_misfit:
9335 * For ASYM_CPUCAPACITY domains with misfit tasks we
9336 * simply seek the "biggest" misfit task.
9338 if (rq->misfit_task_load > busiest_load) {
9339 busiest_load = rq->misfit_task_load;
9352 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
9353 * so long as it is large enough.
9355 #define MAX_PINNED_INTERVAL 512
9358 asym_active_balance(struct lb_env *env)
9361 * ASYM_PACKING needs to force migrate tasks from busy but
9362 * lower priority CPUs in order to pack all tasks in the
9363 * highest priority CPUs.
9365 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
9366 sched_asym_prefer(env->dst_cpu, env->src_cpu);
9370 voluntary_active_balance(struct lb_env *env)
9372 struct sched_domain *sd = env->sd;
9374 if (asym_active_balance(env))
9378 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
9379 * It's worth migrating the task if the src_cpu's capacity is reduced
9380 * because of other sched_class or IRQs if more capacity stays
9381 * available on dst_cpu.
9383 if ((env->idle != CPU_NOT_IDLE) &&
9384 (env->src_rq->cfs.h_nr_running == 1)) {
9385 if ((check_cpu_capacity(env->src_rq, sd)) &&
9386 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
9390 if (env->migration_type == migrate_misfit)
9396 static int need_active_balance(struct lb_env *env)
9398 struct sched_domain *sd = env->sd;
9400 if (voluntary_active_balance(env))
9403 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
9406 static int active_load_balance_cpu_stop(void *data);
9408 static int should_we_balance(struct lb_env *env)
9410 struct sched_group *sg = env->sd->groups;
9411 int cpu, balance_cpu = -1;
9414 * Ensure the balancing environment is consistent; can happen
9415 * when the softirq triggers 'during' hotplug.
9417 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
9421 * In the newly idle case, we will allow all the CPUs
9422 * to do the newly idle load balance.
9424 if (env->idle == CPU_NEWLY_IDLE)
9427 /* Try to find first idle CPU */
9428 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
9436 if (balance_cpu == -1)
9437 balance_cpu = group_balance_cpu(sg);
9440 * First idle CPU or the first CPU(busiest) in this sched group
9441 * is eligible for doing load balancing at this and above domains.
9443 return balance_cpu == env->dst_cpu;
9447 * Check this_cpu to ensure it is balanced within domain. Attempt to move
9448 * tasks if there is an imbalance.
9450 static int load_balance(int this_cpu, struct rq *this_rq,
9451 struct sched_domain *sd, enum cpu_idle_type idle,
9452 int *continue_balancing)
9454 int ld_moved, cur_ld_moved, active_balance = 0;
9455 struct sched_domain *sd_parent = sd->parent;
9456 struct sched_group *group;
9459 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
9461 struct lb_env env = {
9463 .dst_cpu = this_cpu,
9465 .dst_grpmask = sched_group_span(sd->groups),
9467 .loop_break = sched_nr_migrate_break,
9470 .tasks = LIST_HEAD_INIT(env.tasks),
9473 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
9475 schedstat_inc(sd->lb_count[idle]);
9478 if (!should_we_balance(&env)) {
9479 *continue_balancing = 0;
9483 group = find_busiest_group(&env);
9485 schedstat_inc(sd->lb_nobusyg[idle]);
9489 busiest = find_busiest_queue(&env, group);
9491 schedstat_inc(sd->lb_nobusyq[idle]);
9495 BUG_ON(busiest == env.dst_rq);
9497 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
9499 env.src_cpu = busiest->cpu;
9500 env.src_rq = busiest;
9503 if (busiest->nr_running > 1) {
9505 * Attempt to move tasks. If find_busiest_group has found
9506 * an imbalance but busiest->nr_running <= 1, the group is
9507 * still unbalanced. ld_moved simply stays zero, so it is
9508 * correctly treated as an imbalance.
9510 env.flags |= LBF_ALL_PINNED;
9511 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
9514 rq_lock_irqsave(busiest, &rf);
9515 update_rq_clock(busiest);
9518 * cur_ld_moved - load moved in current iteration
9519 * ld_moved - cumulative load moved across iterations
9521 cur_ld_moved = detach_tasks(&env);
9524 * We've detached some tasks from busiest_rq. Every
9525 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
9526 * unlock busiest->lock, and we are able to be sure
9527 * that nobody can manipulate the tasks in parallel.
9528 * See task_rq_lock() family for the details.
9531 rq_unlock(busiest, &rf);
9535 ld_moved += cur_ld_moved;
9538 local_irq_restore(rf.flags);
9540 if (env.flags & LBF_NEED_BREAK) {
9541 env.flags &= ~LBF_NEED_BREAK;
9546 * Revisit (affine) tasks on src_cpu that couldn't be moved to
9547 * us and move them to an alternate dst_cpu in our sched_group
9548 * where they can run. The upper limit on how many times we
9549 * iterate on same src_cpu is dependent on number of CPUs in our
9552 * This changes load balance semantics a bit on who can move
9553 * load to a given_cpu. In addition to the given_cpu itself
9554 * (or a ilb_cpu acting on its behalf where given_cpu is
9555 * nohz-idle), we now have balance_cpu in a position to move
9556 * load to given_cpu. In rare situations, this may cause
9557 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
9558 * _independently_ and at _same_ time to move some load to
9559 * given_cpu) causing exceess load to be moved to given_cpu.
9560 * This however should not happen so much in practice and
9561 * moreover subsequent load balance cycles should correct the
9562 * excess load moved.
9564 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
9566 /* Prevent to re-select dst_cpu via env's CPUs */
9567 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
9569 env.dst_rq = cpu_rq(env.new_dst_cpu);
9570 env.dst_cpu = env.new_dst_cpu;
9571 env.flags &= ~LBF_DST_PINNED;
9573 env.loop_break = sched_nr_migrate_break;
9576 * Go back to "more_balance" rather than "redo" since we
9577 * need to continue with same src_cpu.
9583 * We failed to reach balance because of affinity.
9586 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9588 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
9589 *group_imbalance = 1;
9592 /* All tasks on this runqueue were pinned by CPU affinity */
9593 if (unlikely(env.flags & LBF_ALL_PINNED)) {
9594 __cpumask_clear_cpu(cpu_of(busiest), cpus);
9596 * Attempting to continue load balancing at the current
9597 * sched_domain level only makes sense if there are
9598 * active CPUs remaining as possible busiest CPUs to
9599 * pull load from which are not contained within the
9600 * destination group that is receiving any migrated
9603 if (!cpumask_subset(cpus, env.dst_grpmask)) {
9605 env.loop_break = sched_nr_migrate_break;
9608 goto out_all_pinned;
9613 schedstat_inc(sd->lb_failed[idle]);
9615 * Increment the failure counter only on periodic balance.
9616 * We do not want newidle balance, which can be very
9617 * frequent, pollute the failure counter causing
9618 * excessive cache_hot migrations and active balances.
9620 if (idle != CPU_NEWLY_IDLE)
9621 sd->nr_balance_failed++;
9623 if (need_active_balance(&env)) {
9624 unsigned long flags;
9626 raw_spin_lock_irqsave(&busiest->lock, flags);
9629 * Don't kick the active_load_balance_cpu_stop,
9630 * if the curr task on busiest CPU can't be
9631 * moved to this_cpu:
9633 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
9634 raw_spin_unlock_irqrestore(&busiest->lock,
9636 env.flags |= LBF_ALL_PINNED;
9637 goto out_one_pinned;
9641 * ->active_balance synchronizes accesses to
9642 * ->active_balance_work. Once set, it's cleared
9643 * only after active load balance is finished.
9645 if (!busiest->active_balance) {
9646 busiest->active_balance = 1;
9647 busiest->push_cpu = this_cpu;
9650 raw_spin_unlock_irqrestore(&busiest->lock, flags);
9652 if (active_balance) {
9653 stop_one_cpu_nowait(cpu_of(busiest),
9654 active_load_balance_cpu_stop, busiest,
9655 &busiest->active_balance_work);
9658 /* We've kicked active balancing, force task migration. */
9659 sd->nr_balance_failed = sd->cache_nice_tries+1;
9662 sd->nr_balance_failed = 0;
9664 if (likely(!active_balance) || voluntary_active_balance(&env)) {
9665 /* We were unbalanced, so reset the balancing interval */
9666 sd->balance_interval = sd->min_interval;
9669 * If we've begun active balancing, start to back off. This
9670 * case may not be covered by the all_pinned logic if there
9671 * is only 1 task on the busy runqueue (because we don't call
9674 if (sd->balance_interval < sd->max_interval)
9675 sd->balance_interval *= 2;
9682 * We reach balance although we may have faced some affinity
9683 * constraints. Clear the imbalance flag only if other tasks got
9684 * a chance to move and fix the imbalance.
9686 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
9687 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9689 if (*group_imbalance)
9690 *group_imbalance = 0;
9695 * We reach balance because all tasks are pinned at this level so
9696 * we can't migrate them. Let the imbalance flag set so parent level
9697 * can try to migrate them.
9699 schedstat_inc(sd->lb_balanced[idle]);
9701 sd->nr_balance_failed = 0;
9707 * newidle_balance() disregards balance intervals, so we could
9708 * repeatedly reach this code, which would lead to balance_interval
9709 * skyrocketting in a short amount of time. Skip the balance_interval
9710 * increase logic to avoid that.
9712 if (env.idle == CPU_NEWLY_IDLE)
9715 /* tune up the balancing interval */
9716 if ((env.flags & LBF_ALL_PINNED &&
9717 sd->balance_interval < MAX_PINNED_INTERVAL) ||
9718 sd->balance_interval < sd->max_interval)
9719 sd->balance_interval *= 2;
9724 static inline unsigned long
9725 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
9727 unsigned long interval = sd->balance_interval;
9730 interval *= sd->busy_factor;
9732 /* scale ms to jiffies */
9733 interval = msecs_to_jiffies(interval);
9734 interval = clamp(interval, 1UL, max_load_balance_interval);
9740 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
9742 unsigned long interval, next;
9744 /* used by idle balance, so cpu_busy = 0 */
9745 interval = get_sd_balance_interval(sd, 0);
9746 next = sd->last_balance + interval;
9748 if (time_after(*next_balance, next))
9749 *next_balance = next;
9753 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9754 * running tasks off the busiest CPU onto idle CPUs. It requires at
9755 * least 1 task to be running on each physical CPU where possible, and
9756 * avoids physical / logical imbalances.
9758 static int active_load_balance_cpu_stop(void *data)
9760 struct rq *busiest_rq = data;
9761 int busiest_cpu = cpu_of(busiest_rq);
9762 int target_cpu = busiest_rq->push_cpu;
9763 struct rq *target_rq = cpu_rq(target_cpu);
9764 struct sched_domain *sd;
9765 struct task_struct *p = NULL;
9768 rq_lock_irq(busiest_rq, &rf);
9770 * Between queueing the stop-work and running it is a hole in which
9771 * CPUs can become inactive. We should not move tasks from or to
9774 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
9777 /* Make sure the requested CPU hasn't gone down in the meantime: */
9778 if (unlikely(busiest_cpu != smp_processor_id() ||
9779 !busiest_rq->active_balance))
9782 /* Is there any task to move? */
9783 if (busiest_rq->nr_running <= 1)
9787 * This condition is "impossible", if it occurs
9788 * we need to fix it. Originally reported by
9789 * Bjorn Helgaas on a 128-CPU setup.
9791 BUG_ON(busiest_rq == target_rq);
9793 /* Search for an sd spanning us and the target CPU. */
9795 for_each_domain(target_cpu, sd) {
9796 if ((sd->flags & SD_LOAD_BALANCE) &&
9797 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9802 struct lb_env env = {
9804 .dst_cpu = target_cpu,
9805 .dst_rq = target_rq,
9806 .src_cpu = busiest_rq->cpu,
9807 .src_rq = busiest_rq,
9810 * can_migrate_task() doesn't need to compute new_dst_cpu
9811 * for active balancing. Since we have CPU_IDLE, but no
9812 * @dst_grpmask we need to make that test go away with lying
9815 .flags = LBF_DST_PINNED,
9818 schedstat_inc(sd->alb_count);
9819 update_rq_clock(busiest_rq);
9821 p = detach_one_task(&env);
9823 schedstat_inc(sd->alb_pushed);
9824 /* Active balancing done, reset the failure counter. */
9825 sd->nr_balance_failed = 0;
9827 schedstat_inc(sd->alb_failed);
9832 busiest_rq->active_balance = 0;
9833 rq_unlock(busiest_rq, &rf);
9836 attach_one_task(target_rq, p);
9843 static DEFINE_SPINLOCK(balancing);
9846 * Scale the max load_balance interval with the number of CPUs in the system.
9847 * This trades load-balance latency on larger machines for less cross talk.
9849 void update_max_interval(void)
9851 max_load_balance_interval = HZ*num_online_cpus()/10;
9855 * It checks each scheduling domain to see if it is due to be balanced,
9856 * and initiates a balancing operation if so.
9858 * Balancing parameters are set up in init_sched_domains.
9860 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9862 int continue_balancing = 1;
9864 int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
9865 unsigned long interval;
9866 struct sched_domain *sd;
9867 /* Earliest time when we have to do rebalance again */
9868 unsigned long next_balance = jiffies + 60*HZ;
9869 int update_next_balance = 0;
9870 int need_serialize, need_decay = 0;
9874 for_each_domain(cpu, sd) {
9876 * Decay the newidle max times here because this is a regular
9877 * visit to all the domains. Decay ~1% per second.
9879 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9880 sd->max_newidle_lb_cost =
9881 (sd->max_newidle_lb_cost * 253) / 256;
9882 sd->next_decay_max_lb_cost = jiffies + HZ;
9885 max_cost += sd->max_newidle_lb_cost;
9887 if (!(sd->flags & SD_LOAD_BALANCE))
9891 * Stop the load balance at this level. There is another
9892 * CPU in our sched group which is doing load balancing more
9895 if (!continue_balancing) {
9901 interval = get_sd_balance_interval(sd, busy);
9903 need_serialize = sd->flags & SD_SERIALIZE;
9904 if (need_serialize) {
9905 if (!spin_trylock(&balancing))
9909 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9910 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9912 * The LBF_DST_PINNED logic could have changed
9913 * env->dst_cpu, so we can't know our idle
9914 * state even if we migrated tasks. Update it.
9916 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9917 busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
9919 sd->last_balance = jiffies;
9920 interval = get_sd_balance_interval(sd, busy);
9923 spin_unlock(&balancing);
9925 if (time_after(next_balance, sd->last_balance + interval)) {
9926 next_balance = sd->last_balance + interval;
9927 update_next_balance = 1;
9932 * Ensure the rq-wide value also decays but keep it at a
9933 * reasonable floor to avoid funnies with rq->avg_idle.
9935 rq->max_idle_balance_cost =
9936 max((u64)sysctl_sched_migration_cost, max_cost);
9941 * next_balance will be updated only when there is a need.
9942 * When the cpu is attached to null domain for ex, it will not be
9945 if (likely(update_next_balance)) {
9946 rq->next_balance = next_balance;
9948 #ifdef CONFIG_NO_HZ_COMMON
9950 * If this CPU has been elected to perform the nohz idle
9951 * balance. Other idle CPUs have already rebalanced with
9952 * nohz_idle_balance() and nohz.next_balance has been
9953 * updated accordingly. This CPU is now running the idle load
9954 * balance for itself and we need to update the
9955 * nohz.next_balance accordingly.
9957 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9958 nohz.next_balance = rq->next_balance;
9963 static inline int on_null_domain(struct rq *rq)
9965 return unlikely(!rcu_dereference_sched(rq->sd));
9968 #ifdef CONFIG_NO_HZ_COMMON
9970 * idle load balancing details
9971 * - When one of the busy CPUs notice that there may be an idle rebalancing
9972 * needed, they will kick the idle load balancer, which then does idle
9973 * load balancing for all the idle CPUs.
9974 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
9978 static inline int find_new_ilb(void)
9982 for_each_cpu_and(ilb, nohz.idle_cpus_mask,
9983 housekeeping_cpumask(HK_FLAG_MISC)) {
9992 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
9993 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
9995 static void kick_ilb(unsigned int flags)
9999 nohz.next_balance++;
10001 ilb_cpu = find_new_ilb();
10003 if (ilb_cpu >= nr_cpu_ids)
10006 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
10007 if (flags & NOHZ_KICK_MASK)
10011 * Use smp_send_reschedule() instead of resched_cpu().
10012 * This way we generate a sched IPI on the target CPU which
10013 * is idle. And the softirq performing nohz idle load balance
10014 * will be run before returning from the IPI.
10016 smp_send_reschedule(ilb_cpu);
10020 * Current decision point for kicking the idle load balancer in the presence
10021 * of idle CPUs in the system.
10023 static void nohz_balancer_kick(struct rq *rq)
10025 unsigned long now = jiffies;
10026 struct sched_domain_shared *sds;
10027 struct sched_domain *sd;
10028 int nr_busy, i, cpu = rq->cpu;
10029 unsigned int flags = 0;
10031 if (unlikely(rq->idle_balance))
10035 * We may be recently in ticked or tickless idle mode. At the first
10036 * busy tick after returning from idle, we will update the busy stats.
10038 nohz_balance_exit_idle(rq);
10041 * None are in tickless mode and hence no need for NOHZ idle load
10044 if (likely(!atomic_read(&nohz.nr_cpus)))
10047 if (READ_ONCE(nohz.has_blocked) &&
10048 time_after(now, READ_ONCE(nohz.next_blocked)))
10049 flags = NOHZ_STATS_KICK;
10051 if (time_before(now, nohz.next_balance))
10054 if (rq->nr_running >= 2) {
10055 flags = NOHZ_KICK_MASK;
10061 sd = rcu_dereference(rq->sd);
10064 * If there's a CFS task and the current CPU has reduced
10065 * capacity; kick the ILB to see if there's a better CPU to run
10068 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
10069 flags = NOHZ_KICK_MASK;
10074 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
10077 * When ASYM_PACKING; see if there's a more preferred CPU
10078 * currently idle; in which case, kick the ILB to move tasks
10081 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
10082 if (sched_asym_prefer(i, cpu)) {
10083 flags = NOHZ_KICK_MASK;
10089 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
10092 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
10093 * to run the misfit task on.
10095 if (check_misfit_status(rq, sd)) {
10096 flags = NOHZ_KICK_MASK;
10101 * For asymmetric systems, we do not want to nicely balance
10102 * cache use, instead we want to embrace asymmetry and only
10103 * ensure tasks have enough CPU capacity.
10105 * Skip the LLC logic because it's not relevant in that case.
10110 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
10113 * If there is an imbalance between LLC domains (IOW we could
10114 * increase the overall cache use), we need some less-loaded LLC
10115 * domain to pull some load. Likewise, we may need to spread
10116 * load within the current LLC domain (e.g. packed SMT cores but
10117 * other CPUs are idle). We can't really know from here how busy
10118 * the others are - so just get a nohz balance going if it looks
10119 * like this LLC domain has tasks we could move.
10121 nr_busy = atomic_read(&sds->nr_busy_cpus);
10123 flags = NOHZ_KICK_MASK;
10134 static void set_cpu_sd_state_busy(int cpu)
10136 struct sched_domain *sd;
10139 sd = rcu_dereference(per_cpu(sd_llc, cpu));
10141 if (!sd || !sd->nohz_idle)
10145 atomic_inc(&sd->shared->nr_busy_cpus);
10150 void nohz_balance_exit_idle(struct rq *rq)
10152 SCHED_WARN_ON(rq != this_rq());
10154 if (likely(!rq->nohz_tick_stopped))
10157 rq->nohz_tick_stopped = 0;
10158 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
10159 atomic_dec(&nohz.nr_cpus);
10161 set_cpu_sd_state_busy(rq->cpu);
10164 static void set_cpu_sd_state_idle(int cpu)
10166 struct sched_domain *sd;
10169 sd = rcu_dereference(per_cpu(sd_llc, cpu));
10171 if (!sd || sd->nohz_idle)
10175 atomic_dec(&sd->shared->nr_busy_cpus);
10181 * This routine will record that the CPU is going idle with tick stopped.
10182 * This info will be used in performing idle load balancing in the future.
10184 void nohz_balance_enter_idle(int cpu)
10186 struct rq *rq = cpu_rq(cpu);
10188 SCHED_WARN_ON(cpu != smp_processor_id());
10190 /* If this CPU is going down, then nothing needs to be done: */
10191 if (!cpu_active(cpu))
10194 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
10195 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
10199 * Can be set safely without rq->lock held
10200 * If a clear happens, it will have evaluated last additions because
10201 * rq->lock is held during the check and the clear
10203 rq->has_blocked_load = 1;
10206 * The tick is still stopped but load could have been added in the
10207 * meantime. We set the nohz.has_blocked flag to trig a check of the
10208 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
10209 * of nohz.has_blocked can only happen after checking the new load
10211 if (rq->nohz_tick_stopped)
10214 /* If we're a completely isolated CPU, we don't play: */
10215 if (on_null_domain(rq))
10218 rq->nohz_tick_stopped = 1;
10220 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
10221 atomic_inc(&nohz.nr_cpus);
10224 * Ensures that if nohz_idle_balance() fails to observe our
10225 * @idle_cpus_mask store, it must observe the @has_blocked
10228 smp_mb__after_atomic();
10230 set_cpu_sd_state_idle(cpu);
10234 * Each time a cpu enter idle, we assume that it has blocked load and
10235 * enable the periodic update of the load of idle cpus
10237 WRITE_ONCE(nohz.has_blocked, 1);
10241 * Internal function that runs load balance for all idle cpus. The load balance
10242 * can be a simple update of blocked load or a complete load balance with
10243 * tasks movement depending of flags.
10244 * The function returns false if the loop has stopped before running
10245 * through all idle CPUs.
10247 static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
10248 enum cpu_idle_type idle)
10250 /* Earliest time when we have to do rebalance again */
10251 unsigned long now = jiffies;
10252 unsigned long next_balance = now + 60*HZ;
10253 bool has_blocked_load = false;
10254 int update_next_balance = 0;
10255 int this_cpu = this_rq->cpu;
10260 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
10263 * We assume there will be no idle load after this update and clear
10264 * the has_blocked flag. If a cpu enters idle in the mean time, it will
10265 * set the has_blocked flag and trig another update of idle load.
10266 * Because a cpu that becomes idle, is added to idle_cpus_mask before
10267 * setting the flag, we are sure to not clear the state and not
10268 * check the load of an idle cpu.
10270 WRITE_ONCE(nohz.has_blocked, 0);
10273 * Ensures that if we miss the CPU, we must see the has_blocked
10274 * store from nohz_balance_enter_idle().
10278 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
10279 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
10283 * If this CPU gets work to do, stop the load balancing
10284 * work being done for other CPUs. Next load
10285 * balancing owner will pick it up.
10287 if (need_resched()) {
10288 has_blocked_load = true;
10292 rq = cpu_rq(balance_cpu);
10294 has_blocked_load |= update_nohz_stats(rq, true);
10297 * If time for next balance is due,
10300 if (time_after_eq(jiffies, rq->next_balance)) {
10301 struct rq_flags rf;
10303 rq_lock_irqsave(rq, &rf);
10304 update_rq_clock(rq);
10305 rq_unlock_irqrestore(rq, &rf);
10307 if (flags & NOHZ_BALANCE_KICK)
10308 rebalance_domains(rq, CPU_IDLE);
10311 if (time_after(next_balance, rq->next_balance)) {
10312 next_balance = rq->next_balance;
10313 update_next_balance = 1;
10317 /* Newly idle CPU doesn't need an update */
10318 if (idle != CPU_NEWLY_IDLE) {
10319 update_blocked_averages(this_cpu);
10320 has_blocked_load |= this_rq->has_blocked_load;
10323 if (flags & NOHZ_BALANCE_KICK)
10324 rebalance_domains(this_rq, CPU_IDLE);
10326 WRITE_ONCE(nohz.next_blocked,
10327 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
10329 /* The full idle balance loop has been done */
10333 /* There is still blocked load, enable periodic update */
10334 if (has_blocked_load)
10335 WRITE_ONCE(nohz.has_blocked, 1);
10338 * next_balance will be updated only when there is a need.
10339 * When the CPU is attached to null domain for ex, it will not be
10342 if (likely(update_next_balance))
10343 nohz.next_balance = next_balance;
10349 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
10350 * rebalancing for all the cpus for whom scheduler ticks are stopped.
10352 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
10354 int this_cpu = this_rq->cpu;
10355 unsigned int flags;
10357 if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
10360 if (idle != CPU_IDLE) {
10361 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
10365 /* could be _relaxed() */
10366 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
10367 if (!(flags & NOHZ_KICK_MASK))
10370 _nohz_idle_balance(this_rq, flags, idle);
10375 static void nohz_newidle_balance(struct rq *this_rq)
10377 int this_cpu = this_rq->cpu;
10380 * This CPU doesn't want to be disturbed by scheduler
10383 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
10386 /* Will wake up very soon. No time for doing anything else*/
10387 if (this_rq->avg_idle < sysctl_sched_migration_cost)
10390 /* Don't need to update blocked load of idle CPUs*/
10391 if (!READ_ONCE(nohz.has_blocked) ||
10392 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
10395 raw_spin_unlock(&this_rq->lock);
10397 * This CPU is going to be idle and blocked load of idle CPUs
10398 * need to be updated. Run the ilb locally as it is a good
10399 * candidate for ilb instead of waking up another idle CPU.
10400 * Kick an normal ilb if we failed to do the update.
10402 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
10403 kick_ilb(NOHZ_STATS_KICK);
10404 raw_spin_lock(&this_rq->lock);
10407 #else /* !CONFIG_NO_HZ_COMMON */
10408 static inline void nohz_balancer_kick(struct rq *rq) { }
10410 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
10415 static inline void nohz_newidle_balance(struct rq *this_rq) { }
10416 #endif /* CONFIG_NO_HZ_COMMON */
10419 * idle_balance is called by schedule() if this_cpu is about to become
10420 * idle. Attempts to pull tasks from other CPUs.
10423 * < 0 - we released the lock and there are !fair tasks present
10424 * 0 - failed, no new tasks
10425 * > 0 - success, new (fair) tasks present
10427 int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
10429 unsigned long next_balance = jiffies + HZ;
10430 int this_cpu = this_rq->cpu;
10431 struct sched_domain *sd;
10432 int pulled_task = 0;
10435 update_misfit_status(NULL, this_rq);
10437 * We must set idle_stamp _before_ calling idle_balance(), such that we
10438 * measure the duration of idle_balance() as idle time.
10440 this_rq->idle_stamp = rq_clock(this_rq);
10443 * Do not pull tasks towards !active CPUs...
10445 if (!cpu_active(this_cpu))
10449 * This is OK, because current is on_cpu, which avoids it being picked
10450 * for load-balance and preemption/IRQs are still disabled avoiding
10451 * further scheduler activity on it and we're being very careful to
10452 * re-start the picking loop.
10454 rq_unpin_lock(this_rq, rf);
10456 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
10457 !READ_ONCE(this_rq->rd->overload)) {
10460 sd = rcu_dereference_check_sched_domain(this_rq->sd);
10462 update_next_balance(sd, &next_balance);
10465 nohz_newidle_balance(this_rq);
10470 raw_spin_unlock(&this_rq->lock);
10472 update_blocked_averages(this_cpu);
10474 for_each_domain(this_cpu, sd) {
10475 int continue_balancing = 1;
10476 u64 t0, domain_cost;
10478 if (!(sd->flags & SD_LOAD_BALANCE))
10481 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
10482 update_next_balance(sd, &next_balance);
10486 if (sd->flags & SD_BALANCE_NEWIDLE) {
10487 t0 = sched_clock_cpu(this_cpu);
10489 pulled_task = load_balance(this_cpu, this_rq,
10490 sd, CPU_NEWLY_IDLE,
10491 &continue_balancing);
10493 domain_cost = sched_clock_cpu(this_cpu) - t0;
10494 if (domain_cost > sd->max_newidle_lb_cost)
10495 sd->max_newidle_lb_cost = domain_cost;
10497 curr_cost += domain_cost;
10500 update_next_balance(sd, &next_balance);
10503 * Stop searching for tasks to pull if there are
10504 * now runnable tasks on this rq.
10506 if (pulled_task || this_rq->nr_running > 0)
10511 raw_spin_lock(&this_rq->lock);
10513 if (curr_cost > this_rq->max_idle_balance_cost)
10514 this_rq->max_idle_balance_cost = curr_cost;
10518 * While browsing the domains, we released the rq lock, a task could
10519 * have been enqueued in the meantime. Since we're not going idle,
10520 * pretend we pulled a task.
10522 if (this_rq->cfs.h_nr_running && !pulled_task)
10525 /* Move the next balance forward */
10526 if (time_after(this_rq->next_balance, next_balance))
10527 this_rq->next_balance = next_balance;
10529 /* Is there a task of a high priority class? */
10530 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
10534 this_rq->idle_stamp = 0;
10536 rq_repin_lock(this_rq, rf);
10538 return pulled_task;
10542 * run_rebalance_domains is triggered when needed from the scheduler tick.
10543 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
10545 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
10547 struct rq *this_rq = this_rq();
10548 enum cpu_idle_type idle = this_rq->idle_balance ?
10549 CPU_IDLE : CPU_NOT_IDLE;
10552 * If this CPU has a pending nohz_balance_kick, then do the
10553 * balancing on behalf of the other idle CPUs whose ticks are
10554 * stopped. Do nohz_idle_balance *before* rebalance_domains to
10555 * give the idle CPUs a chance to load balance. Else we may
10556 * load balance only within the local sched_domain hierarchy
10557 * and abort nohz_idle_balance altogether if we pull some load.
10559 if (nohz_idle_balance(this_rq, idle))
10562 /* normal load balance */
10563 update_blocked_averages(this_rq->cpu);
10564 rebalance_domains(this_rq, idle);
10568 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
10570 void trigger_load_balance(struct rq *rq)
10572 /* Don't need to rebalance while attached to NULL domain */
10573 if (unlikely(on_null_domain(rq)))
10576 if (time_after_eq(jiffies, rq->next_balance))
10577 raise_softirq(SCHED_SOFTIRQ);
10579 nohz_balancer_kick(rq);
10582 static void rq_online_fair(struct rq *rq)
10586 update_runtime_enabled(rq);
10589 static void rq_offline_fair(struct rq *rq)
10593 /* Ensure any throttled groups are reachable by pick_next_task */
10594 unthrottle_offline_cfs_rqs(rq);
10597 #endif /* CONFIG_SMP */
10600 * scheduler tick hitting a task of our scheduling class.
10602 * NOTE: This function can be called remotely by the tick offload that
10603 * goes along full dynticks. Therefore no local assumption can be made
10604 * and everything must be accessed through the @rq and @curr passed in
10607 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
10609 struct cfs_rq *cfs_rq;
10610 struct sched_entity *se = &curr->se;
10612 for_each_sched_entity(se) {
10613 cfs_rq = cfs_rq_of(se);
10614 entity_tick(cfs_rq, se, queued);
10617 if (static_branch_unlikely(&sched_numa_balancing))
10618 task_tick_numa(rq, curr);
10620 update_misfit_status(curr, rq);
10621 update_overutilized_status(task_rq(curr));
10625 * called on fork with the child task as argument from the parent's context
10626 * - child not yet on the tasklist
10627 * - preemption disabled
10629 static void task_fork_fair(struct task_struct *p)
10631 struct cfs_rq *cfs_rq;
10632 struct sched_entity *se = &p->se, *curr;
10633 struct rq *rq = this_rq();
10634 struct rq_flags rf;
10637 update_rq_clock(rq);
10639 cfs_rq = task_cfs_rq(current);
10640 curr = cfs_rq->curr;
10642 update_curr(cfs_rq);
10643 se->vruntime = curr->vruntime;
10645 place_entity(cfs_rq, se, 1);
10647 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
10649 * Upon rescheduling, sched_class::put_prev_task() will place
10650 * 'current' within the tree based on its new key value.
10652 swap(curr->vruntime, se->vruntime);
10656 se->vruntime -= cfs_rq->min_vruntime;
10657 rq_unlock(rq, &rf);
10661 * Priority of the task has changed. Check to see if we preempt
10662 * the current task.
10665 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
10667 if (!task_on_rq_queued(p))
10670 if (rq->cfs.nr_running == 1)
10674 * Reschedule if we are currently running on this runqueue and
10675 * our priority decreased, or if we are not currently running on
10676 * this runqueue and our priority is higher than the current's
10678 if (rq->curr == p) {
10679 if (p->prio > oldprio)
10682 check_preempt_curr(rq, p, 0);
10685 static inline bool vruntime_normalized(struct task_struct *p)
10687 struct sched_entity *se = &p->se;
10690 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
10691 * the dequeue_entity(.flags=0) will already have normalized the
10698 * When !on_rq, vruntime of the task has usually NOT been normalized.
10699 * But there are some cases where it has already been normalized:
10701 * - A forked child which is waiting for being woken up by
10702 * wake_up_new_task().
10703 * - A task which has been woken up by try_to_wake_up() and
10704 * waiting for actually being woken up by sched_ttwu_pending().
10706 if (!se->sum_exec_runtime ||
10707 (p->state == TASK_WAKING && p->sched_remote_wakeup))
10713 #ifdef CONFIG_FAIR_GROUP_SCHED
10715 * Propagate the changes of the sched_entity across the tg tree to make it
10716 * visible to the root
10718 static void propagate_entity_cfs_rq(struct sched_entity *se)
10720 struct cfs_rq *cfs_rq;
10722 /* Start to propagate at parent */
10725 for_each_sched_entity(se) {
10726 cfs_rq = cfs_rq_of(se);
10728 if (cfs_rq_throttled(cfs_rq))
10731 update_load_avg(cfs_rq, se, UPDATE_TG);
10735 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
10738 static void detach_entity_cfs_rq(struct sched_entity *se)
10740 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10742 /* Catch up with the cfs_rq and remove our load when we leave */
10743 update_load_avg(cfs_rq, se, 0);
10744 detach_entity_load_avg(cfs_rq, se);
10745 update_tg_load_avg(cfs_rq, false);
10746 propagate_entity_cfs_rq(se);
10749 static void attach_entity_cfs_rq(struct sched_entity *se)
10751 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10753 #ifdef CONFIG_FAIR_GROUP_SCHED
10755 * Since the real-depth could have been changed (only FAIR
10756 * class maintain depth value), reset depth properly.
10758 se->depth = se->parent ? se->parent->depth + 1 : 0;
10761 /* Synchronize entity with its cfs_rq */
10762 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
10763 attach_entity_load_avg(cfs_rq, se);
10764 update_tg_load_avg(cfs_rq, false);
10765 propagate_entity_cfs_rq(se);
10768 static void detach_task_cfs_rq(struct task_struct *p)
10770 struct sched_entity *se = &p->se;
10771 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10773 if (!vruntime_normalized(p)) {
10775 * Fix up our vruntime so that the current sleep doesn't
10776 * cause 'unlimited' sleep bonus.
10778 place_entity(cfs_rq, se, 0);
10779 se->vruntime -= cfs_rq->min_vruntime;
10782 detach_entity_cfs_rq(se);
10785 static void attach_task_cfs_rq(struct task_struct *p)
10787 struct sched_entity *se = &p->se;
10788 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10790 attach_entity_cfs_rq(se);
10792 if (!vruntime_normalized(p))
10793 se->vruntime += cfs_rq->min_vruntime;
10796 static void switched_from_fair(struct rq *rq, struct task_struct *p)
10798 detach_task_cfs_rq(p);
10801 static void switched_to_fair(struct rq *rq, struct task_struct *p)
10803 attach_task_cfs_rq(p);
10805 if (task_on_rq_queued(p)) {
10807 * We were most likely switched from sched_rt, so
10808 * kick off the schedule if running, otherwise just see
10809 * if we can still preempt the current task.
10814 check_preempt_curr(rq, p, 0);
10818 /* Account for a task changing its policy or group.
10820 * This routine is mostly called to set cfs_rq->curr field when a task
10821 * migrates between groups/classes.
10823 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
10825 struct sched_entity *se = &p->se;
10828 if (task_on_rq_queued(p)) {
10830 * Move the next running task to the front of the list, so our
10831 * cfs_tasks list becomes MRU one.
10833 list_move(&se->group_node, &rq->cfs_tasks);
10837 for_each_sched_entity(se) {
10838 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10840 set_next_entity(cfs_rq, se);
10841 /* ensure bandwidth has been allocated on our new cfs_rq */
10842 account_cfs_rq_runtime(cfs_rq, 0);
10846 void init_cfs_rq(struct cfs_rq *cfs_rq)
10848 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
10849 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
10850 #ifndef CONFIG_64BIT
10851 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
10854 raw_spin_lock_init(&cfs_rq->removed.lock);
10858 #ifdef CONFIG_FAIR_GROUP_SCHED
10859 static void task_set_group_fair(struct task_struct *p)
10861 struct sched_entity *se = &p->se;
10863 set_task_rq(p, task_cpu(p));
10864 se->depth = se->parent ? se->parent->depth + 1 : 0;
10867 static void task_move_group_fair(struct task_struct *p)
10869 detach_task_cfs_rq(p);
10870 set_task_rq(p, task_cpu(p));
10873 /* Tell se's cfs_rq has been changed -- migrated */
10874 p->se.avg.last_update_time = 0;
10876 attach_task_cfs_rq(p);
10879 static void task_change_group_fair(struct task_struct *p, int type)
10882 case TASK_SET_GROUP:
10883 task_set_group_fair(p);
10886 case TASK_MOVE_GROUP:
10887 task_move_group_fair(p);
10892 void free_fair_sched_group(struct task_group *tg)
10896 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
10898 for_each_possible_cpu(i) {
10900 kfree(tg->cfs_rq[i]);
10909 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10911 struct sched_entity *se;
10912 struct cfs_rq *cfs_rq;
10915 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
10918 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
10922 tg->shares = NICE_0_LOAD;
10924 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
10926 for_each_possible_cpu(i) {
10927 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
10928 GFP_KERNEL, cpu_to_node(i));
10932 se = kzalloc_node(sizeof(struct sched_entity),
10933 GFP_KERNEL, cpu_to_node(i));
10937 init_cfs_rq(cfs_rq);
10938 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
10939 init_entity_runnable_average(se);
10950 void online_fair_sched_group(struct task_group *tg)
10952 struct sched_entity *se;
10953 struct rq_flags rf;
10957 for_each_possible_cpu(i) {
10960 rq_lock_irq(rq, &rf);
10961 update_rq_clock(rq);
10962 attach_entity_cfs_rq(se);
10963 sync_throttle(tg, i);
10964 rq_unlock_irq(rq, &rf);
10968 void unregister_fair_sched_group(struct task_group *tg)
10970 unsigned long flags;
10974 for_each_possible_cpu(cpu) {
10976 remove_entity_load_avg(tg->se[cpu]);
10979 * Only empty task groups can be destroyed; so we can speculatively
10980 * check on_list without danger of it being re-added.
10982 if (!tg->cfs_rq[cpu]->on_list)
10987 raw_spin_lock_irqsave(&rq->lock, flags);
10988 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
10989 raw_spin_unlock_irqrestore(&rq->lock, flags);
10993 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
10994 struct sched_entity *se, int cpu,
10995 struct sched_entity *parent)
10997 struct rq *rq = cpu_rq(cpu);
11001 init_cfs_rq_runtime(cfs_rq);
11003 tg->cfs_rq[cpu] = cfs_rq;
11006 /* se could be NULL for root_task_group */
11011 se->cfs_rq = &rq->cfs;
11014 se->cfs_rq = parent->my_q;
11015 se->depth = parent->depth + 1;
11019 /* guarantee group entities always have weight */
11020 update_load_set(&se->load, NICE_0_LOAD);
11021 se->parent = parent;
11024 static DEFINE_MUTEX(shares_mutex);
11026 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
11031 * We can't change the weight of the root cgroup.
11036 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
11038 mutex_lock(&shares_mutex);
11039 if (tg->shares == shares)
11042 tg->shares = shares;
11043 for_each_possible_cpu(i) {
11044 struct rq *rq = cpu_rq(i);
11045 struct sched_entity *se = tg->se[i];
11046 struct rq_flags rf;
11048 /* Propagate contribution to hierarchy */
11049 rq_lock_irqsave(rq, &rf);
11050 update_rq_clock(rq);
11051 for_each_sched_entity(se) {
11052 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
11053 update_cfs_group(se);
11055 rq_unlock_irqrestore(rq, &rf);
11059 mutex_unlock(&shares_mutex);
11062 #else /* CONFIG_FAIR_GROUP_SCHED */
11064 void free_fair_sched_group(struct task_group *tg) { }
11066 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
11071 void online_fair_sched_group(struct task_group *tg) { }
11073 void unregister_fair_sched_group(struct task_group *tg) { }
11075 #endif /* CONFIG_FAIR_GROUP_SCHED */
11078 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
11080 struct sched_entity *se = &task->se;
11081 unsigned int rr_interval = 0;
11084 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
11087 if (rq->cfs.load.weight)
11088 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
11090 return rr_interval;
11094 * All the scheduling class methods:
11096 const struct sched_class fair_sched_class = {
11097 .next = &idle_sched_class,
11098 .enqueue_task = enqueue_task_fair,
11099 .dequeue_task = dequeue_task_fair,
11100 .yield_task = yield_task_fair,
11101 .yield_to_task = yield_to_task_fair,
11103 .check_preempt_curr = check_preempt_wakeup,
11105 .pick_next_task = __pick_next_task_fair,
11106 .put_prev_task = put_prev_task_fair,
11107 .set_next_task = set_next_task_fair,
11110 .balance = balance_fair,
11111 .select_task_rq = select_task_rq_fair,
11112 .migrate_task_rq = migrate_task_rq_fair,
11114 .rq_online = rq_online_fair,
11115 .rq_offline = rq_offline_fair,
11117 .task_dead = task_dead_fair,
11118 .set_cpus_allowed = set_cpus_allowed_common,
11121 .task_tick = task_tick_fair,
11122 .task_fork = task_fork_fair,
11124 .prio_changed = prio_changed_fair,
11125 .switched_from = switched_from_fair,
11126 .switched_to = switched_to_fair,
11128 .get_rr_interval = get_rr_interval_fair,
11130 .update_curr = update_curr_fair,
11132 #ifdef CONFIG_FAIR_GROUP_SCHED
11133 .task_change_group = task_change_group_fair,
11136 #ifdef CONFIG_UCLAMP_TASK
11137 .uclamp_enabled = 1,
11141 #ifdef CONFIG_SCHED_DEBUG
11142 void print_cfs_stats(struct seq_file *m, int cpu)
11144 struct cfs_rq *cfs_rq, *pos;
11147 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
11148 print_cfs_rq(m, cpu, cfs_rq);
11152 #ifdef CONFIG_NUMA_BALANCING
11153 void show_numa_stats(struct task_struct *p, struct seq_file *m)
11156 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
11157 struct numa_group *ng;
11160 ng = rcu_dereference(p->numa_group);
11161 for_each_online_node(node) {
11162 if (p->numa_faults) {
11163 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
11164 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
11167 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
11168 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
11170 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
11174 #endif /* CONFIG_NUMA_BALANCING */
11175 #endif /* CONFIG_SCHED_DEBUG */
11177 __init void init_sched_fair_class(void)
11180 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
11182 #ifdef CONFIG_NO_HZ_COMMON
11183 nohz.next_balance = jiffies;
11184 nohz.next_blocked = jiffies;
11185 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
11192 * Helper functions to facilitate extracting info from tracepoints.
11195 const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq)
11198 return cfs_rq ? &cfs_rq->avg : NULL;
11203 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg);
11205 char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len)
11209 strlcpy(str, "(null)", len);
11214 cfs_rq_tg_path(cfs_rq, str, len);
11217 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path);
11219 int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq)
11221 return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1;
11223 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu);
11225 const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq)
11228 return rq ? &rq->avg_rt : NULL;
11233 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt);
11235 const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq)
11238 return rq ? &rq->avg_dl : NULL;
11243 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl);
11245 const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq)
11247 #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ)
11248 return rq ? &rq->avg_irq : NULL;
11253 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq);
11255 int sched_trace_rq_cpu(struct rq *rq)
11257 return rq ? cpu_of(rq) : -1;
11259 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu);
11261 const struct cpumask *sched_trace_rd_span(struct root_domain *rd)
11264 return rd ? rd->span : NULL;
11269 EXPORT_SYMBOL_GPL(sched_trace_rd_span);