1 // SPDX-License-Identifier: GPL-2.0
3 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
5 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
7 * Interactivity improvements by Mike Galbraith
8 * (C) 2007 Mike Galbraith <efault@gmx.de>
10 * Various enhancements by Dmitry Adamushko.
11 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
13 * Group scheduling enhancements by Srivatsa Vaddagiri
14 * Copyright IBM Corporation, 2007
15 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
17 * Scaled math optimizations by Thomas Gleixner
18 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
20 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
21 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
25 #include <trace/events/sched.h>
28 * Targeted preemption latency for CPU-bound tasks:
30 * NOTE: this latency value is not the same as the concept of
31 * 'timeslice length' - timeslices in CFS are of variable length
32 * and have no persistent notion like in traditional, time-slice
33 * based scheduling concepts.
35 * (to see the precise effective timeslice length of your workload,
36 * run vmstat and monitor the context-switches (cs) field)
38 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
40 unsigned int sysctl_sched_latency = 6000000ULL;
41 static unsigned int normalized_sysctl_sched_latency = 6000000ULL;
44 * The initial- and re-scaling of tunables is configurable
48 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
49 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
50 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
52 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
54 enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
57 * Minimal preemption granularity for CPU-bound tasks:
59 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
61 unsigned int sysctl_sched_min_granularity = 750000ULL;
62 static unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
65 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
67 static unsigned int sched_nr_latency = 8;
70 * After fork, child runs first. If set to 0 (default) then
71 * parent will (try to) run first.
73 unsigned int sysctl_sched_child_runs_first __read_mostly;
76 * SCHED_OTHER wake-up granularity.
78 * This option delays the preemption effects of decoupled workloads
79 * and reduces their over-scheduling. Synchronous workloads will still
80 * have immediate wakeup/sleep latencies.
82 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
84 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
85 static unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
87 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
91 * For asym packing, by default the lower numbered CPU has higher priority.
93 int __weak arch_asym_cpu_priority(int cpu)
99 * The margin used when comparing utilization with CPU capacity:
100 * util * margin < capacity * 1024
104 static unsigned int capacity_margin = 1280;
107 #ifdef CONFIG_CFS_BANDWIDTH
109 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
110 * each time a cfs_rq requests quota.
112 * Note: in the case that the slice exceeds the runtime remaining (either due
113 * to consumption or the quota being specified to be smaller than the slice)
114 * we will always only issue the remaining available time.
116 * (default: 5 msec, units: microseconds)
118 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
121 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
127 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
133 static inline void update_load_set(struct load_weight *lw, unsigned long w)
140 * Increase the granularity value when there are more CPUs,
141 * because with more CPUs the 'effective latency' as visible
142 * to users decreases. But the relationship is not linear,
143 * so pick a second-best guess by going with the log2 of the
146 * This idea comes from the SD scheduler of Con Kolivas:
148 static unsigned int get_update_sysctl_factor(void)
150 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
153 switch (sysctl_sched_tunable_scaling) {
154 case SCHED_TUNABLESCALING_NONE:
157 case SCHED_TUNABLESCALING_LINEAR:
160 case SCHED_TUNABLESCALING_LOG:
162 factor = 1 + ilog2(cpus);
169 static void update_sysctl(void)
171 unsigned int factor = get_update_sysctl_factor();
173 #define SET_SYSCTL(name) \
174 (sysctl_##name = (factor) * normalized_sysctl_##name)
175 SET_SYSCTL(sched_min_granularity);
176 SET_SYSCTL(sched_latency);
177 SET_SYSCTL(sched_wakeup_granularity);
181 void sched_init_granularity(void)
186 #define WMULT_CONST (~0U)
187 #define WMULT_SHIFT 32
189 static void __update_inv_weight(struct load_weight *lw)
193 if (likely(lw->inv_weight))
196 w = scale_load_down(lw->weight);
198 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
200 else if (unlikely(!w))
201 lw->inv_weight = WMULT_CONST;
203 lw->inv_weight = WMULT_CONST / w;
207 * delta_exec * weight / lw.weight
209 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
211 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
212 * we're guaranteed shift stays positive because inv_weight is guaranteed to
213 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
215 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
216 * weight/lw.weight <= 1, and therefore our shift will also be positive.
218 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
220 u64 fact = scale_load_down(weight);
221 int shift = WMULT_SHIFT;
223 __update_inv_weight(lw);
225 if (unlikely(fact >> 32)) {
232 /* hint to use a 32x32->64 mul */
233 fact = (u64)(u32)fact * lw->inv_weight;
240 return mul_u64_u32_shr(delta_exec, fact, shift);
244 const struct sched_class fair_sched_class;
246 /**************************************************************
247 * CFS operations on generic schedulable entities:
250 #ifdef CONFIG_FAIR_GROUP_SCHED
251 static inline struct task_struct *task_of(struct sched_entity *se)
253 SCHED_WARN_ON(!entity_is_task(se));
254 return container_of(se, struct task_struct, se);
257 /* Walk up scheduling entities hierarchy */
258 #define for_each_sched_entity(se) \
259 for (; se; se = se->parent)
261 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
266 /* runqueue on which this entity is (to be) queued */
267 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
272 /* runqueue "owned" by this group */
273 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
278 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
283 if (cfs_rq && task_group_is_autogroup(cfs_rq->tg))
284 autogroup_path(cfs_rq->tg, path, len);
285 else if (cfs_rq && cfs_rq->tg->css.cgroup)
286 cgroup_path(cfs_rq->tg->css.cgroup, path, len);
288 strlcpy(path, "(null)", len);
291 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
293 struct rq *rq = rq_of(cfs_rq);
294 int cpu = cpu_of(rq);
297 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
302 * Ensure we either appear before our parent (if already
303 * enqueued) or force our parent to appear after us when it is
304 * enqueued. The fact that we always enqueue bottom-up
305 * reduces this to two cases and a special case for the root
306 * cfs_rq. Furthermore, it also means that we will always reset
307 * tmp_alone_branch either when the branch is connected
308 * to a tree or when we reach the top of the tree
310 if (cfs_rq->tg->parent &&
311 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
313 * If parent is already on the list, we add the child
314 * just before. Thanks to circular linked property of
315 * the list, this means to put the child at the tail
316 * of the list that starts by parent.
318 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
319 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
321 * The branch is now connected to its tree so we can
322 * reset tmp_alone_branch to the beginning of the
325 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
329 if (!cfs_rq->tg->parent) {
331 * cfs rq without parent should be put
332 * at the tail of the list.
334 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
335 &rq->leaf_cfs_rq_list);
337 * We have reach the top of a tree so we can reset
338 * tmp_alone_branch to the beginning of the list.
340 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
345 * The parent has not already been added so we want to
346 * make sure that it will be put after us.
347 * tmp_alone_branch points to the begin of the branch
348 * where we will add parent.
350 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
352 * update tmp_alone_branch to points to the new begin
355 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
359 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
361 if (cfs_rq->on_list) {
362 struct rq *rq = rq_of(cfs_rq);
365 * With cfs_rq being unthrottled/throttled during an enqueue,
366 * it can happen the tmp_alone_branch points the a leaf that
367 * we finally want to del. In this case, tmp_alone_branch moves
368 * to the prev element but it will point to rq->leaf_cfs_rq_list
369 * at the end of the enqueue.
371 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
372 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
374 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
379 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
381 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
384 /* Iterate thr' all leaf cfs_rq's on a runqueue */
385 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
386 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
389 /* Do the two (enqueued) entities belong to the same group ? */
390 static inline struct cfs_rq *
391 is_same_group(struct sched_entity *se, struct sched_entity *pse)
393 if (se->cfs_rq == pse->cfs_rq)
399 static inline struct sched_entity *parent_entity(struct sched_entity *se)
405 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
407 int se_depth, pse_depth;
410 * preemption test can be made between sibling entities who are in the
411 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
412 * both tasks until we find their ancestors who are siblings of common
416 /* First walk up until both entities are at same depth */
417 se_depth = (*se)->depth;
418 pse_depth = (*pse)->depth;
420 while (se_depth > pse_depth) {
422 *se = parent_entity(*se);
425 while (pse_depth > se_depth) {
427 *pse = parent_entity(*pse);
430 while (!is_same_group(*se, *pse)) {
431 *se = parent_entity(*se);
432 *pse = parent_entity(*pse);
436 #else /* !CONFIG_FAIR_GROUP_SCHED */
438 static inline struct task_struct *task_of(struct sched_entity *se)
440 return container_of(se, struct task_struct, se);
443 #define for_each_sched_entity(se) \
444 for (; se; se = NULL)
446 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
448 return &task_rq(p)->cfs;
451 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
453 struct task_struct *p = task_of(se);
454 struct rq *rq = task_rq(p);
459 /* runqueue "owned" by this group */
460 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
465 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
468 strlcpy(path, "(null)", len);
471 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
476 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
480 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
484 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
485 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
487 static inline struct sched_entity *parent_entity(struct sched_entity *se)
493 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
497 #endif /* CONFIG_FAIR_GROUP_SCHED */
499 static __always_inline
500 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
502 /**************************************************************
503 * Scheduling class tree data structure manipulation methods:
506 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
508 s64 delta = (s64)(vruntime - max_vruntime);
510 max_vruntime = vruntime;
515 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
517 s64 delta = (s64)(vruntime - min_vruntime);
519 min_vruntime = vruntime;
524 static inline int entity_before(struct sched_entity *a,
525 struct sched_entity *b)
527 return (s64)(a->vruntime - b->vruntime) < 0;
530 static void update_min_vruntime(struct cfs_rq *cfs_rq)
532 struct sched_entity *curr = cfs_rq->curr;
533 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
535 u64 vruntime = cfs_rq->min_vruntime;
539 vruntime = curr->vruntime;
544 if (leftmost) { /* non-empty tree */
545 struct sched_entity *se;
546 se = rb_entry(leftmost, struct sched_entity, run_node);
549 vruntime = se->vruntime;
551 vruntime = min_vruntime(vruntime, se->vruntime);
554 /* ensure we never gain time by being placed backwards. */
555 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
558 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
563 * Enqueue an entity into the rb-tree:
565 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
567 struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node;
568 struct rb_node *parent = NULL;
569 struct sched_entity *entry;
570 bool leftmost = true;
573 * Find the right place in the rbtree:
577 entry = rb_entry(parent, struct sched_entity, run_node);
579 * We dont care about collisions. Nodes with
580 * the same key stay together.
582 if (entity_before(se, entry)) {
583 link = &parent->rb_left;
585 link = &parent->rb_right;
590 rb_link_node(&se->run_node, parent, link);
591 rb_insert_color_cached(&se->run_node,
592 &cfs_rq->tasks_timeline, leftmost);
595 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
597 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
600 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
602 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
607 return rb_entry(left, struct sched_entity, run_node);
610 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
612 struct rb_node *next = rb_next(&se->run_node);
617 return rb_entry(next, struct sched_entity, run_node);
620 #ifdef CONFIG_SCHED_DEBUG
621 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
623 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
628 return rb_entry(last, struct sched_entity, run_node);
631 /**************************************************************
632 * Scheduling class statistics methods:
635 int sched_proc_update_handler(struct ctl_table *table, int write,
636 void __user *buffer, size_t *lenp,
639 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
640 unsigned int factor = get_update_sysctl_factor();
645 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
646 sysctl_sched_min_granularity);
648 #define WRT_SYSCTL(name) \
649 (normalized_sysctl_##name = sysctl_##name / (factor))
650 WRT_SYSCTL(sched_min_granularity);
651 WRT_SYSCTL(sched_latency);
652 WRT_SYSCTL(sched_wakeup_granularity);
662 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
664 if (unlikely(se->load.weight != NICE_0_LOAD))
665 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
671 * The idea is to set a period in which each task runs once.
673 * When there are too many tasks (sched_nr_latency) we have to stretch
674 * this period because otherwise the slices get too small.
676 * p = (nr <= nl) ? l : l*nr/nl
678 static u64 __sched_period(unsigned long nr_running)
680 if (unlikely(nr_running > sched_nr_latency))
681 return nr_running * sysctl_sched_min_granularity;
683 return sysctl_sched_latency;
687 * We calculate the wall-time slice from the period by taking a part
688 * proportional to the weight.
692 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
694 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
696 for_each_sched_entity(se) {
697 struct load_weight *load;
698 struct load_weight lw;
700 cfs_rq = cfs_rq_of(se);
701 load = &cfs_rq->load;
703 if (unlikely(!se->on_rq)) {
706 update_load_add(&lw, se->load.weight);
709 slice = __calc_delta(slice, se->load.weight, load);
715 * We calculate the vruntime slice of a to-be-inserted task.
719 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
721 return calc_delta_fair(sched_slice(cfs_rq, se), se);
727 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
728 static unsigned long task_h_load(struct task_struct *p);
729 static unsigned long capacity_of(int cpu);
731 /* Give new sched_entity start runnable values to heavy its load in infant time */
732 void init_entity_runnable_average(struct sched_entity *se)
734 struct sched_avg *sa = &se->avg;
736 memset(sa, 0, sizeof(*sa));
739 * Tasks are initialized with full load to be seen as heavy tasks until
740 * they get a chance to stabilize to their real load level.
741 * Group entities are initialized with zero load to reflect the fact that
742 * nothing has been attached to the task group yet.
744 if (entity_is_task(se))
745 sa->runnable_load_avg = sa->load_avg = scale_load_down(se->load.weight);
747 se->runnable_weight = se->load.weight;
749 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
752 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
753 static void attach_entity_cfs_rq(struct sched_entity *se);
756 * With new tasks being created, their initial util_avgs are extrapolated
757 * based on the cfs_rq's current util_avg:
759 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
761 * However, in many cases, the above util_avg does not give a desired
762 * value. Moreover, the sum of the util_avgs may be divergent, such
763 * as when the series is a harmonic series.
765 * To solve this problem, we also cap the util_avg of successive tasks to
766 * only 1/2 of the left utilization budget:
768 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
770 * where n denotes the nth task and cpu_scale the CPU capacity.
772 * For example, for a CPU with 1024 of capacity, a simplest series from
773 * the beginning would be like:
775 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
776 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
778 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
779 * if util_avg > util_avg_cap.
781 void post_init_entity_util_avg(struct task_struct *p)
783 struct sched_entity *se = &p->se;
784 struct cfs_rq *cfs_rq = cfs_rq_of(se);
785 struct sched_avg *sa = &se->avg;
786 long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
787 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
790 if (cfs_rq->avg.util_avg != 0) {
791 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
792 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
794 if (sa->util_avg > cap)
801 if (p->sched_class != &fair_sched_class) {
803 * For !fair tasks do:
805 update_cfs_rq_load_avg(now, cfs_rq);
806 attach_entity_load_avg(cfs_rq, se, 0);
807 switched_from_fair(rq, p);
809 * such that the next switched_to_fair() has the
812 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
816 attach_entity_cfs_rq(se);
819 #else /* !CONFIG_SMP */
820 void init_entity_runnable_average(struct sched_entity *se)
823 void post_init_entity_util_avg(struct task_struct *p)
826 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
829 #endif /* CONFIG_SMP */
832 * Update the current task's runtime statistics.
834 static void update_curr(struct cfs_rq *cfs_rq)
836 struct sched_entity *curr = cfs_rq->curr;
837 u64 now = rq_clock_task(rq_of(cfs_rq));
843 delta_exec = now - curr->exec_start;
844 if (unlikely((s64)delta_exec <= 0))
847 curr->exec_start = now;
849 schedstat_set(curr->statistics.exec_max,
850 max(delta_exec, curr->statistics.exec_max));
852 curr->sum_exec_runtime += delta_exec;
853 schedstat_add(cfs_rq->exec_clock, delta_exec);
855 curr->vruntime += calc_delta_fair(delta_exec, curr);
856 update_min_vruntime(cfs_rq);
858 if (entity_is_task(curr)) {
859 struct task_struct *curtask = task_of(curr);
861 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
862 cgroup_account_cputime(curtask, delta_exec);
863 account_group_exec_runtime(curtask, delta_exec);
866 account_cfs_rq_runtime(cfs_rq, delta_exec);
869 static void update_curr_fair(struct rq *rq)
871 update_curr(cfs_rq_of(&rq->curr->se));
875 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
877 u64 wait_start, prev_wait_start;
879 if (!schedstat_enabled())
882 wait_start = rq_clock(rq_of(cfs_rq));
883 prev_wait_start = schedstat_val(se->statistics.wait_start);
885 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
886 likely(wait_start > prev_wait_start))
887 wait_start -= prev_wait_start;
889 __schedstat_set(se->statistics.wait_start, wait_start);
893 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
895 struct task_struct *p;
898 if (!schedstat_enabled())
901 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
903 if (entity_is_task(se)) {
905 if (task_on_rq_migrating(p)) {
907 * Preserve migrating task's wait time so wait_start
908 * time stamp can be adjusted to accumulate wait time
909 * prior to migration.
911 __schedstat_set(se->statistics.wait_start, delta);
914 trace_sched_stat_wait(p, delta);
917 __schedstat_set(se->statistics.wait_max,
918 max(schedstat_val(se->statistics.wait_max), delta));
919 __schedstat_inc(se->statistics.wait_count);
920 __schedstat_add(se->statistics.wait_sum, delta);
921 __schedstat_set(se->statistics.wait_start, 0);
925 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
927 struct task_struct *tsk = NULL;
928 u64 sleep_start, block_start;
930 if (!schedstat_enabled())
933 sleep_start = schedstat_val(se->statistics.sleep_start);
934 block_start = schedstat_val(se->statistics.block_start);
936 if (entity_is_task(se))
940 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
945 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
946 __schedstat_set(se->statistics.sleep_max, delta);
948 __schedstat_set(se->statistics.sleep_start, 0);
949 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
952 account_scheduler_latency(tsk, delta >> 10, 1);
953 trace_sched_stat_sleep(tsk, delta);
957 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
962 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
963 __schedstat_set(se->statistics.block_max, delta);
965 __schedstat_set(se->statistics.block_start, 0);
966 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
969 if (tsk->in_iowait) {
970 __schedstat_add(se->statistics.iowait_sum, delta);
971 __schedstat_inc(se->statistics.iowait_count);
972 trace_sched_stat_iowait(tsk, delta);
975 trace_sched_stat_blocked(tsk, delta);
978 * Blocking time is in units of nanosecs, so shift by
979 * 20 to get a milliseconds-range estimation of the
980 * amount of time that the task spent sleeping:
982 if (unlikely(prof_on == SLEEP_PROFILING)) {
983 profile_hits(SLEEP_PROFILING,
984 (void *)get_wchan(tsk),
987 account_scheduler_latency(tsk, delta >> 10, 0);
993 * Task is being enqueued - update stats:
996 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
998 if (!schedstat_enabled())
1002 * Are we enqueueing a waiting task? (for current tasks
1003 * a dequeue/enqueue event is a NOP)
1005 if (se != cfs_rq->curr)
1006 update_stats_wait_start(cfs_rq, se);
1008 if (flags & ENQUEUE_WAKEUP)
1009 update_stats_enqueue_sleeper(cfs_rq, se);
1013 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1016 if (!schedstat_enabled())
1020 * Mark the end of the wait period if dequeueing a
1023 if (se != cfs_rq->curr)
1024 update_stats_wait_end(cfs_rq, se);
1026 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1027 struct task_struct *tsk = task_of(se);
1029 if (tsk->state & TASK_INTERRUPTIBLE)
1030 __schedstat_set(se->statistics.sleep_start,
1031 rq_clock(rq_of(cfs_rq)));
1032 if (tsk->state & TASK_UNINTERRUPTIBLE)
1033 __schedstat_set(se->statistics.block_start,
1034 rq_clock(rq_of(cfs_rq)));
1039 * We are picking a new current task - update its stats:
1042 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1045 * We are starting a new run period:
1047 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1050 /**************************************************
1051 * Scheduling class queueing methods:
1054 #ifdef CONFIG_NUMA_BALANCING
1056 * Approximate time to scan a full NUMA task in ms. The task scan period is
1057 * calculated based on the tasks virtual memory size and
1058 * numa_balancing_scan_size.
1060 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1061 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1063 /* Portion of address space to scan in MB */
1064 unsigned int sysctl_numa_balancing_scan_size = 256;
1066 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1067 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1070 refcount_t refcount;
1072 spinlock_t lock; /* nr_tasks, tasks */
1077 struct rcu_head rcu;
1078 unsigned long total_faults;
1079 unsigned long max_faults_cpu;
1081 * Faults_cpu is used to decide whether memory should move
1082 * towards the CPU. As a consequence, these stats are weighted
1083 * more by CPU use than by memory faults.
1085 unsigned long *faults_cpu;
1086 unsigned long faults[0];
1089 static inline unsigned long group_faults_priv(struct numa_group *ng);
1090 static inline unsigned long group_faults_shared(struct numa_group *ng);
1092 static unsigned int task_nr_scan_windows(struct task_struct *p)
1094 unsigned long rss = 0;
1095 unsigned long nr_scan_pages;
1098 * Calculations based on RSS as non-present and empty pages are skipped
1099 * by the PTE scanner and NUMA hinting faults should be trapped based
1102 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1103 rss = get_mm_rss(p->mm);
1105 rss = nr_scan_pages;
1107 rss = round_up(rss, nr_scan_pages);
1108 return rss / nr_scan_pages;
1111 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1112 #define MAX_SCAN_WINDOW 2560
1114 static unsigned int task_scan_min(struct task_struct *p)
1116 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1117 unsigned int scan, floor;
1118 unsigned int windows = 1;
1120 if (scan_size < MAX_SCAN_WINDOW)
1121 windows = MAX_SCAN_WINDOW / scan_size;
1122 floor = 1000 / windows;
1124 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1125 return max_t(unsigned int, floor, scan);
1128 static unsigned int task_scan_start(struct task_struct *p)
1130 unsigned long smin = task_scan_min(p);
1131 unsigned long period = smin;
1133 /* Scale the maximum scan period with the amount of shared memory. */
1134 if (p->numa_group) {
1135 struct numa_group *ng = p->numa_group;
1136 unsigned long shared = group_faults_shared(ng);
1137 unsigned long private = group_faults_priv(ng);
1139 period *= refcount_read(&ng->refcount);
1140 period *= shared + 1;
1141 period /= private + shared + 1;
1144 return max(smin, period);
1147 static unsigned int task_scan_max(struct task_struct *p)
1149 unsigned long smin = task_scan_min(p);
1152 /* Watch for min being lower than max due to floor calculations */
1153 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1155 /* Scale the maximum scan period with the amount of shared memory. */
1156 if (p->numa_group) {
1157 struct numa_group *ng = p->numa_group;
1158 unsigned long shared = group_faults_shared(ng);
1159 unsigned long private = group_faults_priv(ng);
1160 unsigned long period = smax;
1162 period *= refcount_read(&ng->refcount);
1163 period *= shared + 1;
1164 period /= private + shared + 1;
1166 smax = max(smax, period);
1169 return max(smin, smax);
1172 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
1175 struct mm_struct *mm = p->mm;
1178 mm_users = atomic_read(&mm->mm_users);
1179 if (mm_users == 1) {
1180 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1181 mm->numa_scan_seq = 0;
1185 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
1186 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1187 p->numa_work.next = &p->numa_work;
1188 p->numa_faults = NULL;
1189 p->numa_group = NULL;
1190 p->last_task_numa_placement = 0;
1191 p->last_sum_exec_runtime = 0;
1193 /* New address space, reset the preferred nid */
1194 if (!(clone_flags & CLONE_VM)) {
1195 p->numa_preferred_nid = NUMA_NO_NODE;
1200 * New thread, keep existing numa_preferred_nid which should be copied
1201 * already by arch_dup_task_struct but stagger when scans start.
1206 delay = min_t(unsigned int, task_scan_max(current),
1207 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
1208 delay += 2 * TICK_NSEC;
1209 p->node_stamp = delay;
1213 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1215 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1216 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1219 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1221 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1222 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1225 /* Shared or private faults. */
1226 #define NR_NUMA_HINT_FAULT_TYPES 2
1228 /* Memory and CPU locality */
1229 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1231 /* Averaged statistics, and temporary buffers. */
1232 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1234 pid_t task_numa_group_id(struct task_struct *p)
1236 return p->numa_group ? p->numa_group->gid : 0;
1240 * The averaged statistics, shared & private, memory & CPU,
1241 * occupy the first half of the array. The second half of the
1242 * array is for current counters, which are averaged into the
1243 * first set by task_numa_placement.
1245 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1247 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1250 static inline unsigned long task_faults(struct task_struct *p, int nid)
1252 if (!p->numa_faults)
1255 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1256 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1259 static inline unsigned long group_faults(struct task_struct *p, int nid)
1264 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1265 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1268 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1270 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1271 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1274 static inline unsigned long group_faults_priv(struct numa_group *ng)
1276 unsigned long faults = 0;
1279 for_each_online_node(node) {
1280 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1286 static inline unsigned long group_faults_shared(struct numa_group *ng)
1288 unsigned long faults = 0;
1291 for_each_online_node(node) {
1292 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1299 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1300 * considered part of a numa group's pseudo-interleaving set. Migrations
1301 * between these nodes are slowed down, to allow things to settle down.
1303 #define ACTIVE_NODE_FRACTION 3
1305 static bool numa_is_active_node(int nid, struct numa_group *ng)
1307 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1310 /* Handle placement on systems where not all nodes are directly connected. */
1311 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1312 int maxdist, bool task)
1314 unsigned long score = 0;
1318 * All nodes are directly connected, and the same distance
1319 * from each other. No need for fancy placement algorithms.
1321 if (sched_numa_topology_type == NUMA_DIRECT)
1325 * This code is called for each node, introducing N^2 complexity,
1326 * which should be ok given the number of nodes rarely exceeds 8.
1328 for_each_online_node(node) {
1329 unsigned long faults;
1330 int dist = node_distance(nid, node);
1333 * The furthest away nodes in the system are not interesting
1334 * for placement; nid was already counted.
1336 if (dist == sched_max_numa_distance || node == nid)
1340 * On systems with a backplane NUMA topology, compare groups
1341 * of nodes, and move tasks towards the group with the most
1342 * memory accesses. When comparing two nodes at distance
1343 * "hoplimit", only nodes closer by than "hoplimit" are part
1344 * of each group. Skip other nodes.
1346 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1350 /* Add up the faults from nearby nodes. */
1352 faults = task_faults(p, node);
1354 faults = group_faults(p, node);
1357 * On systems with a glueless mesh NUMA topology, there are
1358 * no fixed "groups of nodes". Instead, nodes that are not
1359 * directly connected bounce traffic through intermediate
1360 * nodes; a numa_group can occupy any set of nodes.
1361 * The further away a node is, the less the faults count.
1362 * This seems to result in good task placement.
1364 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1365 faults *= (sched_max_numa_distance - dist);
1366 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1376 * These return the fraction of accesses done by a particular task, or
1377 * task group, on a particular numa node. The group weight is given a
1378 * larger multiplier, in order to group tasks together that are almost
1379 * evenly spread out between numa nodes.
1381 static inline unsigned long task_weight(struct task_struct *p, int nid,
1384 unsigned long faults, total_faults;
1386 if (!p->numa_faults)
1389 total_faults = p->total_numa_faults;
1394 faults = task_faults(p, nid);
1395 faults += score_nearby_nodes(p, nid, dist, true);
1397 return 1000 * faults / total_faults;
1400 static inline unsigned long group_weight(struct task_struct *p, int nid,
1403 unsigned long faults, total_faults;
1408 total_faults = p->numa_group->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 = p->numa_group;
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;
1488 static unsigned long weighted_cpuload(struct rq *rq);
1490 /* Cached statistics for all CPUs within a node */
1494 /* Total compute capacity of CPUs on a node */
1495 unsigned long compute_capacity;
1499 * XXX borrowed from update_sg_lb_stats
1501 static void update_numa_stats(struct numa_stats *ns, int nid)
1505 memset(ns, 0, sizeof(*ns));
1506 for_each_cpu(cpu, cpumask_of_node(nid)) {
1507 struct rq *rq = cpu_rq(cpu);
1509 ns->load += weighted_cpuload(rq);
1510 ns->compute_capacity += capacity_of(cpu);
1515 struct task_numa_env {
1516 struct task_struct *p;
1518 int src_cpu, src_nid;
1519 int dst_cpu, dst_nid;
1521 struct numa_stats src_stats, dst_stats;
1526 struct task_struct *best_task;
1531 static void task_numa_assign(struct task_numa_env *env,
1532 struct task_struct *p, long imp)
1534 struct rq *rq = cpu_rq(env->dst_cpu);
1536 /* Bail out if run-queue part of active NUMA balance. */
1537 if (xchg(&rq->numa_migrate_on, 1))
1541 * Clear previous best_cpu/rq numa-migrate flag, since task now
1542 * found a better CPU to move/swap.
1544 if (env->best_cpu != -1) {
1545 rq = cpu_rq(env->best_cpu);
1546 WRITE_ONCE(rq->numa_migrate_on, 0);
1550 put_task_struct(env->best_task);
1555 env->best_imp = imp;
1556 env->best_cpu = env->dst_cpu;
1559 static bool load_too_imbalanced(long src_load, long dst_load,
1560 struct task_numa_env *env)
1563 long orig_src_load, orig_dst_load;
1564 long src_capacity, dst_capacity;
1567 * The load is corrected for the CPU capacity available on each node.
1570 * ------------ vs ---------
1571 * src_capacity dst_capacity
1573 src_capacity = env->src_stats.compute_capacity;
1574 dst_capacity = env->dst_stats.compute_capacity;
1576 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1578 orig_src_load = env->src_stats.load;
1579 orig_dst_load = env->dst_stats.load;
1581 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1583 /* Would this change make things worse? */
1584 return (imb > old_imb);
1588 * Maximum NUMA importance can be 1998 (2*999);
1589 * SMALLIMP @ 30 would be close to 1998/64.
1590 * Used to deter task migration.
1595 * This checks if the overall compute and NUMA accesses of the system would
1596 * be improved if the source tasks was migrated to the target dst_cpu taking
1597 * into account that it might be best if task running on the dst_cpu should
1598 * be exchanged with the source task
1600 static void task_numa_compare(struct task_numa_env *env,
1601 long taskimp, long groupimp, bool maymove)
1603 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1604 struct task_struct *cur;
1605 long src_load, dst_load;
1607 long imp = env->p->numa_group ? groupimp : taskimp;
1609 int dist = env->dist;
1611 if (READ_ONCE(dst_rq->numa_migrate_on))
1615 cur = task_rcu_dereference(&dst_rq->curr);
1616 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1620 * Because we have preemption enabled we can get migrated around and
1621 * end try selecting ourselves (current == env->p) as a swap candidate.
1627 if (maymove && moveimp >= env->best_imp)
1634 * "imp" is the fault differential for the source task between the
1635 * source and destination node. Calculate the total differential for
1636 * the source task and potential destination task. The more negative
1637 * the value is, the more remote accesses that would be expected to
1638 * be incurred if the tasks were swapped.
1640 /* Skip this swap candidate if cannot move to the source cpu */
1641 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
1645 * If dst and source tasks are in the same NUMA group, or not
1646 * in any group then look only at task weights.
1648 if (cur->numa_group == env->p->numa_group) {
1649 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1650 task_weight(cur, env->dst_nid, dist);
1652 * Add some hysteresis to prevent swapping the
1653 * tasks within a group over tiny differences.
1655 if (cur->numa_group)
1659 * Compare the group weights. If a task is all by itself
1660 * (not part of a group), use the task weight instead.
1662 if (cur->numa_group && env->p->numa_group)
1663 imp += group_weight(cur, env->src_nid, dist) -
1664 group_weight(cur, env->dst_nid, dist);
1666 imp += task_weight(cur, env->src_nid, dist) -
1667 task_weight(cur, env->dst_nid, dist);
1670 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1677 * If the NUMA importance is less than SMALLIMP,
1678 * task migration might only result in ping pong
1679 * of tasks and also hurt performance due to cache
1682 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1686 * In the overloaded case, try and keep the load balanced.
1688 load = task_h_load(env->p) - task_h_load(cur);
1692 dst_load = env->dst_stats.load + load;
1693 src_load = env->src_stats.load - load;
1695 if (load_too_imbalanced(src_load, dst_load, env))
1700 * One idle CPU per node is evaluated for a task numa move.
1701 * Call select_idle_sibling to maybe find a better one.
1705 * select_idle_siblings() uses an per-CPU cpumask that
1706 * can be used from IRQ context.
1708 local_irq_disable();
1709 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1714 task_numa_assign(env, cur, imp);
1719 static void task_numa_find_cpu(struct task_numa_env *env,
1720 long taskimp, long groupimp)
1722 long src_load, dst_load, load;
1723 bool maymove = false;
1726 load = task_h_load(env->p);
1727 dst_load = env->dst_stats.load + load;
1728 src_load = env->src_stats.load - load;
1731 * If the improvement from just moving env->p direction is better
1732 * than swapping tasks around, check if a move is possible.
1734 maymove = !load_too_imbalanced(src_load, dst_load, env);
1736 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1737 /* Skip this CPU if the source task cannot migrate */
1738 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
1742 task_numa_compare(env, taskimp, groupimp, maymove);
1746 static int task_numa_migrate(struct task_struct *p)
1748 struct task_numa_env env = {
1751 .src_cpu = task_cpu(p),
1752 .src_nid = task_node(p),
1754 .imbalance_pct = 112,
1760 struct sched_domain *sd;
1762 unsigned long taskweight, groupweight;
1764 long taskimp, groupimp;
1767 * Pick the lowest SD_NUMA domain, as that would have the smallest
1768 * imbalance and would be the first to start moving tasks about.
1770 * And we want to avoid any moving of tasks about, as that would create
1771 * random movement of tasks -- counter the numa conditions we're trying
1775 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1777 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1781 * Cpusets can break the scheduler domain tree into smaller
1782 * balance domains, some of which do not cross NUMA boundaries.
1783 * Tasks that are "trapped" in such domains cannot be migrated
1784 * elsewhere, so there is no point in (re)trying.
1786 if (unlikely(!sd)) {
1787 sched_setnuma(p, task_node(p));
1791 env.dst_nid = p->numa_preferred_nid;
1792 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1793 taskweight = task_weight(p, env.src_nid, dist);
1794 groupweight = group_weight(p, env.src_nid, dist);
1795 update_numa_stats(&env.src_stats, env.src_nid);
1796 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1797 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1798 update_numa_stats(&env.dst_stats, env.dst_nid);
1800 /* Try to find a spot on the preferred nid. */
1801 task_numa_find_cpu(&env, taskimp, groupimp);
1804 * Look at other nodes in these cases:
1805 * - there is no space available on the preferred_nid
1806 * - the task is part of a numa_group that is interleaved across
1807 * multiple NUMA nodes; in order to better consolidate the group,
1808 * we need to check other locations.
1810 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1811 for_each_online_node(nid) {
1812 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1815 dist = node_distance(env.src_nid, env.dst_nid);
1816 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1818 taskweight = task_weight(p, env.src_nid, dist);
1819 groupweight = group_weight(p, env.src_nid, dist);
1822 /* Only consider nodes where both task and groups benefit */
1823 taskimp = task_weight(p, nid, dist) - taskweight;
1824 groupimp = group_weight(p, nid, dist) - groupweight;
1825 if (taskimp < 0 && groupimp < 0)
1830 update_numa_stats(&env.dst_stats, env.dst_nid);
1831 task_numa_find_cpu(&env, taskimp, groupimp);
1836 * If the task is part of a workload that spans multiple NUMA nodes,
1837 * and is migrating into one of the workload's active nodes, remember
1838 * this node as the task's preferred numa node, so the workload can
1840 * A task that migrated to a second choice node will be better off
1841 * trying for a better one later. Do not set the preferred node here.
1843 if (p->numa_group) {
1844 if (env.best_cpu == -1)
1847 nid = cpu_to_node(env.best_cpu);
1849 if (nid != p->numa_preferred_nid)
1850 sched_setnuma(p, nid);
1853 /* No better CPU than the current one was found. */
1854 if (env.best_cpu == -1)
1857 best_rq = cpu_rq(env.best_cpu);
1858 if (env.best_task == NULL) {
1859 ret = migrate_task_to(p, env.best_cpu);
1860 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1862 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1866 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
1867 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1870 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1871 put_task_struct(env.best_task);
1875 /* Attempt to migrate a task to a CPU on the preferred node. */
1876 static void numa_migrate_preferred(struct task_struct *p)
1878 unsigned long interval = HZ;
1880 /* This task has no NUMA fault statistics yet */
1881 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
1884 /* Periodically retry migrating the task to the preferred node */
1885 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1886 p->numa_migrate_retry = jiffies + interval;
1888 /* Success if task is already running on preferred CPU */
1889 if (task_node(p) == p->numa_preferred_nid)
1892 /* Otherwise, try migrate to a CPU on the preferred node */
1893 task_numa_migrate(p);
1897 * Find out how many nodes on the workload is actively running on. Do this by
1898 * tracking the nodes from which NUMA hinting faults are triggered. This can
1899 * be different from the set of nodes where the workload's memory is currently
1902 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1904 unsigned long faults, max_faults = 0;
1905 int nid, active_nodes = 0;
1907 for_each_online_node(nid) {
1908 faults = group_faults_cpu(numa_group, nid);
1909 if (faults > max_faults)
1910 max_faults = faults;
1913 for_each_online_node(nid) {
1914 faults = group_faults_cpu(numa_group, nid);
1915 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1919 numa_group->max_faults_cpu = max_faults;
1920 numa_group->active_nodes = active_nodes;
1924 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1925 * increments. The more local the fault statistics are, the higher the scan
1926 * period will be for the next scan window. If local/(local+remote) ratio is
1927 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1928 * the scan period will decrease. Aim for 70% local accesses.
1930 #define NUMA_PERIOD_SLOTS 10
1931 #define NUMA_PERIOD_THRESHOLD 7
1934 * Increase the scan period (slow down scanning) if the majority of
1935 * our memory is already on our local node, or if the majority of
1936 * the page accesses are shared with other processes.
1937 * Otherwise, decrease the scan period.
1939 static void update_task_scan_period(struct task_struct *p,
1940 unsigned long shared, unsigned long private)
1942 unsigned int period_slot;
1943 int lr_ratio, ps_ratio;
1946 unsigned long remote = p->numa_faults_locality[0];
1947 unsigned long local = p->numa_faults_locality[1];
1950 * If there were no record hinting faults then either the task is
1951 * completely idle or all activity is areas that are not of interest
1952 * to automatic numa balancing. Related to that, if there were failed
1953 * migration then it implies we are migrating too quickly or the local
1954 * node is overloaded. In either case, scan slower
1956 if (local + shared == 0 || p->numa_faults_locality[2]) {
1957 p->numa_scan_period = min(p->numa_scan_period_max,
1958 p->numa_scan_period << 1);
1960 p->mm->numa_next_scan = jiffies +
1961 msecs_to_jiffies(p->numa_scan_period);
1967 * Prepare to scale scan period relative to the current period.
1968 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1969 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1970 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1972 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1973 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1974 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1976 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1978 * Most memory accesses are local. There is no need to
1979 * do fast NUMA scanning, since memory is already local.
1981 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
1984 diff = slot * period_slot;
1985 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
1987 * Most memory accesses are shared with other tasks.
1988 * There is no point in continuing fast NUMA scanning,
1989 * since other tasks may just move the memory elsewhere.
1991 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
1994 diff = slot * period_slot;
1997 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
1998 * yet they are not on the local NUMA node. Speed up
1999 * NUMA scanning to get the memory moved over.
2001 int ratio = max(lr_ratio, ps_ratio);
2002 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2005 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2006 task_scan_min(p), task_scan_max(p));
2007 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2011 * Get the fraction of time the task has been running since the last
2012 * NUMA placement cycle. The scheduler keeps similar statistics, but
2013 * decays those on a 32ms period, which is orders of magnitude off
2014 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2015 * stats only if the task is so new there are no NUMA statistics yet.
2017 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2019 u64 runtime, delta, now;
2020 /* Use the start of this time slice to avoid calculations. */
2021 now = p->se.exec_start;
2022 runtime = p->se.sum_exec_runtime;
2024 if (p->last_task_numa_placement) {
2025 delta = runtime - p->last_sum_exec_runtime;
2026 *period = now - p->last_task_numa_placement;
2028 /* Avoid time going backwards, prevent potential divide error: */
2029 if (unlikely((s64)*period < 0))
2032 delta = p->se.avg.load_sum;
2033 *period = LOAD_AVG_MAX;
2036 p->last_sum_exec_runtime = runtime;
2037 p->last_task_numa_placement = now;
2043 * Determine the preferred nid for a task in a numa_group. This needs to
2044 * be done in a way that produces consistent results with group_weight,
2045 * otherwise workloads might not converge.
2047 static int preferred_group_nid(struct task_struct *p, int nid)
2052 /* Direct connections between all NUMA nodes. */
2053 if (sched_numa_topology_type == NUMA_DIRECT)
2057 * On a system with glueless mesh NUMA topology, group_weight
2058 * scores nodes according to the number of NUMA hinting faults on
2059 * both the node itself, and on nearby nodes.
2061 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2062 unsigned long score, max_score = 0;
2063 int node, max_node = nid;
2065 dist = sched_max_numa_distance;
2067 for_each_online_node(node) {
2068 score = group_weight(p, node, dist);
2069 if (score > max_score) {
2078 * Finding the preferred nid in a system with NUMA backplane
2079 * interconnect topology is more involved. The goal is to locate
2080 * tasks from numa_groups near each other in the system, and
2081 * untangle workloads from different sides of the system. This requires
2082 * searching down the hierarchy of node groups, recursively searching
2083 * inside the highest scoring group of nodes. The nodemask tricks
2084 * keep the complexity of the search down.
2086 nodes = node_online_map;
2087 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2088 unsigned long max_faults = 0;
2089 nodemask_t max_group = NODE_MASK_NONE;
2092 /* Are there nodes at this distance from each other? */
2093 if (!find_numa_distance(dist))
2096 for_each_node_mask(a, nodes) {
2097 unsigned long faults = 0;
2098 nodemask_t this_group;
2099 nodes_clear(this_group);
2101 /* Sum group's NUMA faults; includes a==b case. */
2102 for_each_node_mask(b, nodes) {
2103 if (node_distance(a, b) < dist) {
2104 faults += group_faults(p, b);
2105 node_set(b, this_group);
2106 node_clear(b, nodes);
2110 /* Remember the top group. */
2111 if (faults > max_faults) {
2112 max_faults = faults;
2113 max_group = this_group;
2115 * subtle: at the smallest distance there is
2116 * just one node left in each "group", the
2117 * winner is the preferred nid.
2122 /* Next round, evaluate the nodes within max_group. */
2130 static void task_numa_placement(struct task_struct *p)
2132 int seq, nid, max_nid = NUMA_NO_NODE;
2133 unsigned long max_faults = 0;
2134 unsigned long fault_types[2] = { 0, 0 };
2135 unsigned long total_faults;
2136 u64 runtime, period;
2137 spinlock_t *group_lock = NULL;
2140 * The p->mm->numa_scan_seq field gets updated without
2141 * exclusive access. Use READ_ONCE() here to ensure
2142 * that the field is read in a single access:
2144 seq = READ_ONCE(p->mm->numa_scan_seq);
2145 if (p->numa_scan_seq == seq)
2147 p->numa_scan_seq = seq;
2148 p->numa_scan_period_max = task_scan_max(p);
2150 total_faults = p->numa_faults_locality[0] +
2151 p->numa_faults_locality[1];
2152 runtime = numa_get_avg_runtime(p, &period);
2154 /* If the task is part of a group prevent parallel updates to group stats */
2155 if (p->numa_group) {
2156 group_lock = &p->numa_group->lock;
2157 spin_lock_irq(group_lock);
2160 /* Find the node with the highest number of faults */
2161 for_each_online_node(nid) {
2162 /* Keep track of the offsets in numa_faults array */
2163 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2164 unsigned long faults = 0, group_faults = 0;
2167 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2168 long diff, f_diff, f_weight;
2170 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2171 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2172 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2173 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2175 /* Decay existing window, copy faults since last scan */
2176 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2177 fault_types[priv] += p->numa_faults[membuf_idx];
2178 p->numa_faults[membuf_idx] = 0;
2181 * Normalize the faults_from, so all tasks in a group
2182 * count according to CPU use, instead of by the raw
2183 * number of faults. Tasks with little runtime have
2184 * little over-all impact on throughput, and thus their
2185 * faults are less important.
2187 f_weight = div64_u64(runtime << 16, period + 1);
2188 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2190 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2191 p->numa_faults[cpubuf_idx] = 0;
2193 p->numa_faults[mem_idx] += diff;
2194 p->numa_faults[cpu_idx] += f_diff;
2195 faults += p->numa_faults[mem_idx];
2196 p->total_numa_faults += diff;
2197 if (p->numa_group) {
2199 * safe because we can only change our own group
2201 * mem_idx represents the offset for a given
2202 * nid and priv in a specific region because it
2203 * is at the beginning of the numa_faults array.
2205 p->numa_group->faults[mem_idx] += diff;
2206 p->numa_group->faults_cpu[mem_idx] += f_diff;
2207 p->numa_group->total_faults += diff;
2208 group_faults += p->numa_group->faults[mem_idx];
2212 if (!p->numa_group) {
2213 if (faults > max_faults) {
2214 max_faults = faults;
2217 } else if (group_faults > max_faults) {
2218 max_faults = group_faults;
2223 if (p->numa_group) {
2224 numa_group_count_active_nodes(p->numa_group);
2225 spin_unlock_irq(group_lock);
2226 max_nid = preferred_group_nid(p, max_nid);
2230 /* Set the new preferred node */
2231 if (max_nid != p->numa_preferred_nid)
2232 sched_setnuma(p, max_nid);
2235 update_task_scan_period(p, fault_types[0], fault_types[1]);
2238 static inline int get_numa_group(struct numa_group *grp)
2240 return refcount_inc_not_zero(&grp->refcount);
2243 static inline void put_numa_group(struct numa_group *grp)
2245 if (refcount_dec_and_test(&grp->refcount))
2246 kfree_rcu(grp, rcu);
2249 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2252 struct numa_group *grp, *my_grp;
2253 struct task_struct *tsk;
2255 int cpu = cpupid_to_cpu(cpupid);
2258 if (unlikely(!p->numa_group)) {
2259 unsigned int size = sizeof(struct numa_group) +
2260 4*nr_node_ids*sizeof(unsigned long);
2262 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2266 refcount_set(&grp->refcount, 1);
2267 grp->active_nodes = 1;
2268 grp->max_faults_cpu = 0;
2269 spin_lock_init(&grp->lock);
2271 /* Second half of the array tracks nids where faults happen */
2272 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2275 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2276 grp->faults[i] = p->numa_faults[i];
2278 grp->total_faults = p->total_numa_faults;
2281 rcu_assign_pointer(p->numa_group, grp);
2285 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2287 if (!cpupid_match_pid(tsk, cpupid))
2290 grp = rcu_dereference(tsk->numa_group);
2294 my_grp = p->numa_group;
2299 * Only join the other group if its bigger; if we're the bigger group,
2300 * the other task will join us.
2302 if (my_grp->nr_tasks > grp->nr_tasks)
2306 * Tie-break on the grp address.
2308 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2311 /* Always join threads in the same process. */
2312 if (tsk->mm == current->mm)
2315 /* Simple filter to avoid false positives due to PID collisions */
2316 if (flags & TNF_SHARED)
2319 /* Update priv based on whether false sharing was detected */
2322 if (join && !get_numa_group(grp))
2330 BUG_ON(irqs_disabled());
2331 double_lock_irq(&my_grp->lock, &grp->lock);
2333 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2334 my_grp->faults[i] -= p->numa_faults[i];
2335 grp->faults[i] += p->numa_faults[i];
2337 my_grp->total_faults -= p->total_numa_faults;
2338 grp->total_faults += p->total_numa_faults;
2343 spin_unlock(&my_grp->lock);
2344 spin_unlock_irq(&grp->lock);
2346 rcu_assign_pointer(p->numa_group, grp);
2348 put_numa_group(my_grp);
2356 void task_numa_free(struct task_struct *p)
2358 struct numa_group *grp = p->numa_group;
2359 void *numa_faults = p->numa_faults;
2360 unsigned long flags;
2364 spin_lock_irqsave(&grp->lock, flags);
2365 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2366 grp->faults[i] -= p->numa_faults[i];
2367 grp->total_faults -= p->total_numa_faults;
2370 spin_unlock_irqrestore(&grp->lock, flags);
2371 RCU_INIT_POINTER(p->numa_group, NULL);
2372 put_numa_group(grp);
2375 p->numa_faults = NULL;
2380 * Got a PROT_NONE fault for a page on @node.
2382 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2384 struct task_struct *p = current;
2385 bool migrated = flags & TNF_MIGRATED;
2386 int cpu_node = task_node(current);
2387 int local = !!(flags & TNF_FAULT_LOCAL);
2388 struct numa_group *ng;
2391 if (!static_branch_likely(&sched_numa_balancing))
2394 /* for example, ksmd faulting in a user's mm */
2398 /* Allocate buffer to track faults on a per-node basis */
2399 if (unlikely(!p->numa_faults)) {
2400 int size = sizeof(*p->numa_faults) *
2401 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2403 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2404 if (!p->numa_faults)
2407 p->total_numa_faults = 0;
2408 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2412 * First accesses are treated as private, otherwise consider accesses
2413 * to be private if the accessing pid has not changed
2415 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2418 priv = cpupid_match_pid(p, last_cpupid);
2419 if (!priv && !(flags & TNF_NO_GROUP))
2420 task_numa_group(p, last_cpupid, flags, &priv);
2424 * If a workload spans multiple NUMA nodes, a shared fault that
2425 * occurs wholly within the set of nodes that the workload is
2426 * actively using should be counted as local. This allows the
2427 * scan rate to slow down when a workload has settled down.
2430 if (!priv && !local && ng && ng->active_nodes > 1 &&
2431 numa_is_active_node(cpu_node, ng) &&
2432 numa_is_active_node(mem_node, ng))
2436 * Retry to migrate task to preferred node periodically, in case it
2437 * previously failed, or the scheduler moved us.
2439 if (time_after(jiffies, p->numa_migrate_retry)) {
2440 task_numa_placement(p);
2441 numa_migrate_preferred(p);
2445 p->numa_pages_migrated += pages;
2446 if (flags & TNF_MIGRATE_FAIL)
2447 p->numa_faults_locality[2] += pages;
2449 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2450 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2451 p->numa_faults_locality[local] += pages;
2454 static void reset_ptenuma_scan(struct task_struct *p)
2457 * We only did a read acquisition of the mmap sem, so
2458 * p->mm->numa_scan_seq is written to without exclusive access
2459 * and the update is not guaranteed to be atomic. That's not
2460 * much of an issue though, since this is just used for
2461 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2462 * expensive, to avoid any form of compiler optimizations:
2464 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2465 p->mm->numa_scan_offset = 0;
2469 * The expensive part of numa migration is done from task_work context.
2470 * Triggered from task_tick_numa().
2472 void task_numa_work(struct callback_head *work)
2474 unsigned long migrate, next_scan, now = jiffies;
2475 struct task_struct *p = current;
2476 struct mm_struct *mm = p->mm;
2477 u64 runtime = p->se.sum_exec_runtime;
2478 struct vm_area_struct *vma;
2479 unsigned long start, end;
2480 unsigned long nr_pte_updates = 0;
2481 long pages, virtpages;
2483 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2485 work->next = work; /* protect against double add */
2487 * Who cares about NUMA placement when they're dying.
2489 * NOTE: make sure not to dereference p->mm before this check,
2490 * exit_task_work() happens _after_ exit_mm() so we could be called
2491 * without p->mm even though we still had it when we enqueued this
2494 if (p->flags & PF_EXITING)
2497 if (!mm->numa_next_scan) {
2498 mm->numa_next_scan = now +
2499 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2503 * Enforce maximal scan/migration frequency..
2505 migrate = mm->numa_next_scan;
2506 if (time_before(now, migrate))
2509 if (p->numa_scan_period == 0) {
2510 p->numa_scan_period_max = task_scan_max(p);
2511 p->numa_scan_period = task_scan_start(p);
2514 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2515 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2519 * Delay this task enough that another task of this mm will likely win
2520 * the next time around.
2522 p->node_stamp += 2 * TICK_NSEC;
2524 start = mm->numa_scan_offset;
2525 pages = sysctl_numa_balancing_scan_size;
2526 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2527 virtpages = pages * 8; /* Scan up to this much virtual space */
2532 if (!down_read_trylock(&mm->mmap_sem))
2534 vma = find_vma(mm, start);
2536 reset_ptenuma_scan(p);
2540 for (; vma; vma = vma->vm_next) {
2541 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2542 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2547 * Shared library pages mapped by multiple processes are not
2548 * migrated as it is expected they are cache replicated. Avoid
2549 * hinting faults in read-only file-backed mappings or the vdso
2550 * as migrating the pages will be of marginal benefit.
2553 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2557 * Skip inaccessible VMAs to avoid any confusion between
2558 * PROT_NONE and NUMA hinting ptes
2560 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2564 start = max(start, vma->vm_start);
2565 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2566 end = min(end, vma->vm_end);
2567 nr_pte_updates = change_prot_numa(vma, start, end);
2570 * Try to scan sysctl_numa_balancing_size worth of
2571 * hpages that have at least one present PTE that
2572 * is not already pte-numa. If the VMA contains
2573 * areas that are unused or already full of prot_numa
2574 * PTEs, scan up to virtpages, to skip through those
2578 pages -= (end - start) >> PAGE_SHIFT;
2579 virtpages -= (end - start) >> PAGE_SHIFT;
2582 if (pages <= 0 || virtpages <= 0)
2586 } while (end != vma->vm_end);
2591 * It is possible to reach the end of the VMA list but the last few
2592 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2593 * would find the !migratable VMA on the next scan but not reset the
2594 * scanner to the start so check it now.
2597 mm->numa_scan_offset = start;
2599 reset_ptenuma_scan(p);
2600 up_read(&mm->mmap_sem);
2603 * Make sure tasks use at least 32x as much time to run other code
2604 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2605 * Usually update_task_scan_period slows down scanning enough; on an
2606 * overloaded system we need to limit overhead on a per task basis.
2608 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2609 u64 diff = p->se.sum_exec_runtime - runtime;
2610 p->node_stamp += 32 * diff;
2615 * Drive the periodic memory faults..
2617 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2619 struct callback_head *work = &curr->numa_work;
2623 * We don't care about NUMA placement if we don't have memory.
2625 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2629 * Using runtime rather than walltime has the dual advantage that
2630 * we (mostly) drive the selection from busy threads and that the
2631 * task needs to have done some actual work before we bother with
2634 now = curr->se.sum_exec_runtime;
2635 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2637 if (now > curr->node_stamp + period) {
2638 if (!curr->node_stamp)
2639 curr->numa_scan_period = task_scan_start(curr);
2640 curr->node_stamp += period;
2642 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2643 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2644 task_work_add(curr, work, true);
2649 static void update_scan_period(struct task_struct *p, int new_cpu)
2651 int src_nid = cpu_to_node(task_cpu(p));
2652 int dst_nid = cpu_to_node(new_cpu);
2654 if (!static_branch_likely(&sched_numa_balancing))
2657 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2660 if (src_nid == dst_nid)
2664 * Allow resets if faults have been trapped before one scan
2665 * has completed. This is most likely due to a new task that
2666 * is pulled cross-node due to wakeups or load balancing.
2668 if (p->numa_scan_seq) {
2670 * Avoid scan adjustments if moving to the preferred
2671 * node or if the task was not previously running on
2672 * the preferred node.
2674 if (dst_nid == p->numa_preferred_nid ||
2675 (p->numa_preferred_nid != NUMA_NO_NODE &&
2676 src_nid != p->numa_preferred_nid))
2680 p->numa_scan_period = task_scan_start(p);
2684 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2688 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2692 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2696 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2700 #endif /* CONFIG_NUMA_BALANCING */
2703 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2705 update_load_add(&cfs_rq->load, se->load.weight);
2707 if (entity_is_task(se)) {
2708 struct rq *rq = rq_of(cfs_rq);
2710 account_numa_enqueue(rq, task_of(se));
2711 list_add(&se->group_node, &rq->cfs_tasks);
2714 cfs_rq->nr_running++;
2718 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2720 update_load_sub(&cfs_rq->load, se->load.weight);
2722 if (entity_is_task(se)) {
2723 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2724 list_del_init(&se->group_node);
2727 cfs_rq->nr_running--;
2731 * Signed add and clamp on underflow.
2733 * Explicitly do a load-store to ensure the intermediate value never hits
2734 * memory. This allows lockless observations without ever seeing the negative
2737 #define add_positive(_ptr, _val) do { \
2738 typeof(_ptr) ptr = (_ptr); \
2739 typeof(_val) val = (_val); \
2740 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2744 if (val < 0 && res > var) \
2747 WRITE_ONCE(*ptr, res); \
2751 * Unsigned subtract and clamp on underflow.
2753 * Explicitly do a load-store to ensure the intermediate value never hits
2754 * memory. This allows lockless observations without ever seeing the negative
2757 #define sub_positive(_ptr, _val) do { \
2758 typeof(_ptr) ptr = (_ptr); \
2759 typeof(*ptr) val = (_val); \
2760 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2764 WRITE_ONCE(*ptr, res); \
2768 * Remove and clamp on negative, from a local variable.
2770 * A variant of sub_positive(), which does not use explicit load-store
2771 * and is thus optimized for local variable updates.
2773 #define lsub_positive(_ptr, _val) do { \
2774 typeof(_ptr) ptr = (_ptr); \
2775 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
2780 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2782 cfs_rq->runnable_weight += se->runnable_weight;
2784 cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2785 cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2789 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2791 cfs_rq->runnable_weight -= se->runnable_weight;
2793 sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2794 sub_positive(&cfs_rq->avg.runnable_load_sum,
2795 se_runnable(se) * se->avg.runnable_load_sum);
2799 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2801 cfs_rq->avg.load_avg += se->avg.load_avg;
2802 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2806 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2808 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2809 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2813 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2815 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2817 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2819 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2822 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2823 unsigned long weight, unsigned long runnable)
2826 /* commit outstanding execution time */
2827 if (cfs_rq->curr == se)
2828 update_curr(cfs_rq);
2829 account_entity_dequeue(cfs_rq, se);
2830 dequeue_runnable_load_avg(cfs_rq, se);
2832 dequeue_load_avg(cfs_rq, se);
2834 se->runnable_weight = runnable;
2835 update_load_set(&se->load, weight);
2839 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2841 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2842 se->avg.runnable_load_avg =
2843 div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2847 enqueue_load_avg(cfs_rq, se);
2849 account_entity_enqueue(cfs_rq, se);
2850 enqueue_runnable_load_avg(cfs_rq, se);
2854 void reweight_task(struct task_struct *p, int prio)
2856 struct sched_entity *se = &p->se;
2857 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2858 struct load_weight *load = &se->load;
2859 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2861 reweight_entity(cfs_rq, se, weight, weight);
2862 load->inv_weight = sched_prio_to_wmult[prio];
2865 #ifdef CONFIG_FAIR_GROUP_SCHED
2868 * All this does is approximate the hierarchical proportion which includes that
2869 * global sum we all love to hate.
2871 * That is, the weight of a group entity, is the proportional share of the
2872 * group weight based on the group runqueue weights. That is:
2874 * tg->weight * grq->load.weight
2875 * ge->load.weight = ----------------------------- (1)
2876 * \Sum grq->load.weight
2878 * Now, because computing that sum is prohibitively expensive to compute (been
2879 * there, done that) we approximate it with this average stuff. The average
2880 * moves slower and therefore the approximation is cheaper and more stable.
2882 * So instead of the above, we substitute:
2884 * grq->load.weight -> grq->avg.load_avg (2)
2886 * which yields the following:
2888 * tg->weight * grq->avg.load_avg
2889 * ge->load.weight = ------------------------------ (3)
2892 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2894 * That is shares_avg, and it is right (given the approximation (2)).
2896 * The problem with it is that because the average is slow -- it was designed
2897 * to be exactly that of course -- this leads to transients in boundary
2898 * conditions. In specific, the case where the group was idle and we start the
2899 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2900 * yielding bad latency etc..
2902 * Now, in that special case (1) reduces to:
2904 * tg->weight * grq->load.weight
2905 * ge->load.weight = ----------------------------- = tg->weight (4)
2908 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2910 * So what we do is modify our approximation (3) to approach (4) in the (near)
2915 * tg->weight * grq->load.weight
2916 * --------------------------------------------------- (5)
2917 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2919 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2920 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2923 * tg->weight * grq->load.weight
2924 * ge->load.weight = ----------------------------- (6)
2929 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2930 * max(grq->load.weight, grq->avg.load_avg)
2932 * And that is shares_weight and is icky. In the (near) UP case it approaches
2933 * (4) while in the normal case it approaches (3). It consistently
2934 * overestimates the ge->load.weight and therefore:
2936 * \Sum ge->load.weight >= tg->weight
2940 static long calc_group_shares(struct cfs_rq *cfs_rq)
2942 long tg_weight, tg_shares, load, shares;
2943 struct task_group *tg = cfs_rq->tg;
2945 tg_shares = READ_ONCE(tg->shares);
2947 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
2949 tg_weight = atomic_long_read(&tg->load_avg);
2951 /* Ensure tg_weight >= load */
2952 tg_weight -= cfs_rq->tg_load_avg_contrib;
2955 shares = (tg_shares * load);
2957 shares /= tg_weight;
2960 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2961 * of a group with small tg->shares value. It is a floor value which is
2962 * assigned as a minimum load.weight to the sched_entity representing
2963 * the group on a CPU.
2965 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2966 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2967 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2968 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2971 return clamp_t(long, shares, MIN_SHARES, tg_shares);
2975 * This calculates the effective runnable weight for a group entity based on
2976 * the group entity weight calculated above.
2978 * Because of the above approximation (2), our group entity weight is
2979 * an load_avg based ratio (3). This means that it includes blocked load and
2980 * does not represent the runnable weight.
2982 * Approximate the group entity's runnable weight per ratio from the group
2985 * grq->avg.runnable_load_avg
2986 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
2989 * However, analogous to above, since the avg numbers are slow, this leads to
2990 * transients in the from-idle case. Instead we use:
2992 * ge->runnable_weight = ge->load.weight *
2994 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
2995 * ----------------------------------------------------- (8)
2996 * max(grq->avg.load_avg, grq->load.weight)
2998 * Where these max() serve both to use the 'instant' values to fix the slow
2999 * from-idle and avoid the /0 on to-idle, similar to (6).
3001 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
3003 long runnable, load_avg;
3005 load_avg = max(cfs_rq->avg.load_avg,
3006 scale_load_down(cfs_rq->load.weight));
3008 runnable = max(cfs_rq->avg.runnable_load_avg,
3009 scale_load_down(cfs_rq->runnable_weight));
3013 runnable /= load_avg;
3015 return clamp_t(long, runnable, MIN_SHARES, shares);
3017 #endif /* CONFIG_SMP */
3019 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3022 * Recomputes the group entity based on the current state of its group
3025 static void update_cfs_group(struct sched_entity *se)
3027 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3028 long shares, runnable;
3033 if (throttled_hierarchy(gcfs_rq))
3037 runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
3039 if (likely(se->load.weight == shares))
3042 shares = calc_group_shares(gcfs_rq);
3043 runnable = calc_group_runnable(gcfs_rq, shares);
3046 reweight_entity(cfs_rq_of(se), se, shares, runnable);
3049 #else /* CONFIG_FAIR_GROUP_SCHED */
3050 static inline void update_cfs_group(struct sched_entity *se)
3053 #endif /* CONFIG_FAIR_GROUP_SCHED */
3055 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3057 struct rq *rq = rq_of(cfs_rq);
3059 if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) {
3061 * There are a few boundary cases this might miss but it should
3062 * get called often enough that that should (hopefully) not be
3065 * It will not get called when we go idle, because the idle
3066 * thread is a different class (!fair), nor will the utilization
3067 * number include things like RT tasks.
3069 * As is, the util number is not freq-invariant (we'd have to
3070 * implement arch_scale_freq_capacity() for that).
3074 cpufreq_update_util(rq, flags);
3079 #ifdef CONFIG_FAIR_GROUP_SCHED
3081 * update_tg_load_avg - update the tg's load avg
3082 * @cfs_rq: the cfs_rq whose avg changed
3083 * @force: update regardless of how small the difference
3085 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3086 * However, because tg->load_avg is a global value there are performance
3089 * In order to avoid having to look at the other cfs_rq's, we use a
3090 * differential update where we store the last value we propagated. This in
3091 * turn allows skipping updates if the differential is 'small'.
3093 * Updating tg's load_avg is necessary before update_cfs_share().
3095 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3097 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3100 * No need to update load_avg for root_task_group as it is not used.
3102 if (cfs_rq->tg == &root_task_group)
3105 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3106 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3107 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3112 * Called within set_task_rq() right before setting a task's CPU. The
3113 * caller only guarantees p->pi_lock is held; no other assumptions,
3114 * including the state of rq->lock, should be made.
3116 void set_task_rq_fair(struct sched_entity *se,
3117 struct cfs_rq *prev, struct cfs_rq *next)
3119 u64 p_last_update_time;
3120 u64 n_last_update_time;
3122 if (!sched_feat(ATTACH_AGE_LOAD))
3126 * We are supposed to update the task to "current" time, then its up to
3127 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3128 * getting what current time is, so simply throw away the out-of-date
3129 * time. This will result in the wakee task is less decayed, but giving
3130 * the wakee more load sounds not bad.
3132 if (!(se->avg.last_update_time && prev))
3135 #ifndef CONFIG_64BIT
3137 u64 p_last_update_time_copy;
3138 u64 n_last_update_time_copy;
3141 p_last_update_time_copy = prev->load_last_update_time_copy;
3142 n_last_update_time_copy = next->load_last_update_time_copy;
3146 p_last_update_time = prev->avg.last_update_time;
3147 n_last_update_time = next->avg.last_update_time;
3149 } while (p_last_update_time != p_last_update_time_copy ||
3150 n_last_update_time != n_last_update_time_copy);
3153 p_last_update_time = prev->avg.last_update_time;
3154 n_last_update_time = next->avg.last_update_time;
3156 __update_load_avg_blocked_se(p_last_update_time, se);
3157 se->avg.last_update_time = n_last_update_time;
3162 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3163 * propagate its contribution. The key to this propagation is the invariant
3164 * that for each group:
3166 * ge->avg == grq->avg (1)
3168 * _IFF_ we look at the pure running and runnable sums. Because they
3169 * represent the very same entity, just at different points in the hierarchy.
3171 * Per the above update_tg_cfs_util() is trivial and simply copies the running
3172 * sum over (but still wrong, because the group entity and group rq do not have
3173 * their PELT windows aligned).
3175 * However, update_tg_cfs_runnable() is more complex. So we have:
3177 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3179 * And since, like util, the runnable part should be directly transferable,
3180 * the following would _appear_ to be the straight forward approach:
3182 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3184 * And per (1) we have:
3186 * ge->avg.runnable_avg == grq->avg.runnable_avg
3190 * ge->load.weight * grq->avg.load_avg
3191 * ge->avg.load_avg = ----------------------------------- (4)
3194 * Except that is wrong!
3196 * Because while for entities historical weight is not important and we
3197 * really only care about our future and therefore can consider a pure
3198 * runnable sum, runqueues can NOT do this.
3200 * We specifically want runqueues to have a load_avg that includes
3201 * historical weights. Those represent the blocked load, the load we expect
3202 * to (shortly) return to us. This only works by keeping the weights as
3203 * integral part of the sum. We therefore cannot decompose as per (3).
3205 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3206 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3207 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3208 * runnable section of these tasks overlap (or not). If they were to perfectly
3209 * align the rq as a whole would be runnable 2/3 of the time. If however we
3210 * always have at least 1 runnable task, the rq as a whole is always runnable.
3212 * So we'll have to approximate.. :/
3214 * Given the constraint:
3216 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3218 * We can construct a rule that adds runnable to a rq by assuming minimal
3221 * On removal, we'll assume each task is equally runnable; which yields:
3223 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3225 * XXX: only do this for the part of runnable > running ?
3230 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3232 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3234 /* Nothing to update */
3239 * The relation between sum and avg is:
3241 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3243 * however, the PELT windows are not aligned between grq and gse.
3246 /* Set new sched_entity's utilization */
3247 se->avg.util_avg = gcfs_rq->avg.util_avg;
3248 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3250 /* Update parent cfs_rq utilization */
3251 add_positive(&cfs_rq->avg.util_avg, delta);
3252 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3256 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3258 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3259 unsigned long runnable_load_avg, load_avg;
3260 u64 runnable_load_sum, load_sum = 0;
3266 gcfs_rq->prop_runnable_sum = 0;
3268 if (runnable_sum >= 0) {
3270 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3271 * the CPU is saturated running == runnable.
3273 runnable_sum += se->avg.load_sum;
3274 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3277 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3278 * assuming all tasks are equally runnable.
3280 if (scale_load_down(gcfs_rq->load.weight)) {
3281 load_sum = div_s64(gcfs_rq->avg.load_sum,
3282 scale_load_down(gcfs_rq->load.weight));
3285 /* But make sure to not inflate se's runnable */
3286 runnable_sum = min(se->avg.load_sum, load_sum);
3290 * runnable_sum can't be lower than running_sum
3291 * Rescale running sum to be in the same range as runnable sum
3292 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3293 * runnable_sum is in [0 : LOAD_AVG_MAX]
3295 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3296 runnable_sum = max(runnable_sum, running_sum);
3298 load_sum = (s64)se_weight(se) * runnable_sum;
3299 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3301 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3302 delta_avg = load_avg - se->avg.load_avg;
3304 se->avg.load_sum = runnable_sum;
3305 se->avg.load_avg = load_avg;
3306 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3307 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3309 runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3310 runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3311 delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3312 delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3314 se->avg.runnable_load_sum = runnable_sum;
3315 se->avg.runnable_load_avg = runnable_load_avg;
3318 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3319 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3323 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3325 cfs_rq->propagate = 1;
3326 cfs_rq->prop_runnable_sum += runnable_sum;
3329 /* Update task and its cfs_rq load average */
3330 static inline int propagate_entity_load_avg(struct sched_entity *se)
3332 struct cfs_rq *cfs_rq, *gcfs_rq;
3334 if (entity_is_task(se))
3337 gcfs_rq = group_cfs_rq(se);
3338 if (!gcfs_rq->propagate)
3341 gcfs_rq->propagate = 0;
3343 cfs_rq = cfs_rq_of(se);
3345 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3347 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3348 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3350 trace_pelt_cfs_tp(cfs_rq);
3351 trace_pelt_se_tp(se);
3357 * Check if we need to update the load and the utilization of a blocked
3360 static inline bool skip_blocked_update(struct sched_entity *se)
3362 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3365 * If sched_entity still have not zero load or utilization, we have to
3368 if (se->avg.load_avg || se->avg.util_avg)
3372 * If there is a pending propagation, we have to update the load and
3373 * the utilization of the sched_entity:
3375 if (gcfs_rq->propagate)
3379 * Otherwise, the load and the utilization of the sched_entity is
3380 * already zero and there is no pending propagation, so it will be a
3381 * waste of time to try to decay it:
3386 #else /* CONFIG_FAIR_GROUP_SCHED */
3388 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3390 static inline int propagate_entity_load_avg(struct sched_entity *se)
3395 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3397 #endif /* CONFIG_FAIR_GROUP_SCHED */
3400 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3401 * @now: current time, as per cfs_rq_clock_pelt()
3402 * @cfs_rq: cfs_rq to update
3404 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3405 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3406 * post_init_entity_util_avg().
3408 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3410 * Returns true if the load decayed or we removed load.
3412 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3413 * call update_tg_load_avg() when this function returns true.
3416 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3418 unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3419 struct sched_avg *sa = &cfs_rq->avg;
3422 if (cfs_rq->removed.nr) {
3424 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3426 raw_spin_lock(&cfs_rq->removed.lock);
3427 swap(cfs_rq->removed.util_avg, removed_util);
3428 swap(cfs_rq->removed.load_avg, removed_load);
3429 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3430 cfs_rq->removed.nr = 0;
3431 raw_spin_unlock(&cfs_rq->removed.lock);
3434 sub_positive(&sa->load_avg, r);
3435 sub_positive(&sa->load_sum, r * divider);
3438 sub_positive(&sa->util_avg, r);
3439 sub_positive(&sa->util_sum, r * divider);
3441 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3446 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
3448 #ifndef CONFIG_64BIT
3450 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3454 cfs_rq_util_change(cfs_rq, 0);
3460 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3461 * @cfs_rq: cfs_rq to attach to
3462 * @se: sched_entity to attach
3463 * @flags: migration hints
3465 * Must call update_cfs_rq_load_avg() before this, since we rely on
3466 * cfs_rq->avg.last_update_time being current.
3468 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3470 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3473 * When we attach the @se to the @cfs_rq, we must align the decay
3474 * window because without that, really weird and wonderful things can
3479 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3480 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3483 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3484 * period_contrib. This isn't strictly correct, but since we're
3485 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3488 se->avg.util_sum = se->avg.util_avg * divider;
3490 se->avg.load_sum = divider;
3491 if (se_weight(se)) {
3493 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3496 se->avg.runnable_load_sum = se->avg.load_sum;
3498 enqueue_load_avg(cfs_rq, se);
3499 cfs_rq->avg.util_avg += se->avg.util_avg;
3500 cfs_rq->avg.util_sum += se->avg.util_sum;
3502 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3504 cfs_rq_util_change(cfs_rq, flags);
3506 trace_pelt_cfs_tp(cfs_rq);
3510 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3511 * @cfs_rq: cfs_rq to detach from
3512 * @se: sched_entity to detach
3514 * Must call update_cfs_rq_load_avg() before this, since we rely on
3515 * cfs_rq->avg.last_update_time being current.
3517 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3519 dequeue_load_avg(cfs_rq, se);
3520 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3521 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3523 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3525 cfs_rq_util_change(cfs_rq, 0);
3527 trace_pelt_cfs_tp(cfs_rq);
3531 * Optional action to be done while updating the load average
3533 #define UPDATE_TG 0x1
3534 #define SKIP_AGE_LOAD 0x2
3535 #define DO_ATTACH 0x4
3537 /* Update task and its cfs_rq load average */
3538 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3540 u64 now = cfs_rq_clock_pelt(cfs_rq);
3544 * Track task load average for carrying it to new CPU after migrated, and
3545 * track group sched_entity load average for task_h_load calc in migration
3547 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3548 __update_load_avg_se(now, cfs_rq, se);
3550 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3551 decayed |= propagate_entity_load_avg(se);
3553 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3556 * DO_ATTACH means we're here from enqueue_entity().
3557 * !last_update_time means we've passed through
3558 * migrate_task_rq_fair() indicating we migrated.
3560 * IOW we're enqueueing a task on a new CPU.
3562 attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
3563 update_tg_load_avg(cfs_rq, 0);
3565 } else if (decayed && (flags & UPDATE_TG))
3566 update_tg_load_avg(cfs_rq, 0);
3569 #ifndef CONFIG_64BIT
3570 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3572 u64 last_update_time_copy;
3573 u64 last_update_time;
3576 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3578 last_update_time = cfs_rq->avg.last_update_time;
3579 } while (last_update_time != last_update_time_copy);
3581 return last_update_time;
3584 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3586 return cfs_rq->avg.last_update_time;
3591 * Synchronize entity load avg of dequeued entity without locking
3594 static void sync_entity_load_avg(struct sched_entity *se)
3596 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3597 u64 last_update_time;
3599 last_update_time = cfs_rq_last_update_time(cfs_rq);
3600 __update_load_avg_blocked_se(last_update_time, se);
3604 * Task first catches up with cfs_rq, and then subtract
3605 * itself from the cfs_rq (task must be off the queue now).
3607 static void remove_entity_load_avg(struct sched_entity *se)
3609 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3610 unsigned long flags;
3613 * tasks cannot exit without having gone through wake_up_new_task() ->
3614 * post_init_entity_util_avg() which will have added things to the
3615 * cfs_rq, so we can remove unconditionally.
3618 sync_entity_load_avg(se);
3620 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3621 ++cfs_rq->removed.nr;
3622 cfs_rq->removed.util_avg += se->avg.util_avg;
3623 cfs_rq->removed.load_avg += se->avg.load_avg;
3624 cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */
3625 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3628 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3630 return cfs_rq->avg.runnable_load_avg;
3633 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3635 return cfs_rq->avg.load_avg;
3638 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3640 static inline unsigned long task_util(struct task_struct *p)
3642 return READ_ONCE(p->se.avg.util_avg);
3645 static inline unsigned long _task_util_est(struct task_struct *p)
3647 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3649 return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED);
3652 static inline unsigned long task_util_est(struct task_struct *p)
3654 return max(task_util(p), _task_util_est(p));
3657 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3658 struct task_struct *p)
3660 unsigned int enqueued;
3662 if (!sched_feat(UTIL_EST))
3665 /* Update root cfs_rq's estimated utilization */
3666 enqueued = cfs_rq->avg.util_est.enqueued;
3667 enqueued += _task_util_est(p);
3668 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3672 * Check if a (signed) value is within a specified (unsigned) margin,
3673 * based on the observation that:
3675 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3677 * NOTE: this only works when value + maring < INT_MAX.
3679 static inline bool within_margin(int value, int margin)
3681 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3685 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3687 long last_ewma_diff;
3691 if (!sched_feat(UTIL_EST))
3694 /* Update root cfs_rq's estimated utilization */
3695 ue.enqueued = cfs_rq->avg.util_est.enqueued;
3696 ue.enqueued -= min_t(unsigned int, ue.enqueued, _task_util_est(p));
3697 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3700 * Skip update of task's estimated utilization when the task has not
3701 * yet completed an activation, e.g. being migrated.
3707 * If the PELT values haven't changed since enqueue time,
3708 * skip the util_est update.
3710 ue = p->se.avg.util_est;
3711 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3715 * Skip update of task's estimated utilization when its EWMA is
3716 * already ~1% close to its last activation value.
3718 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3719 last_ewma_diff = ue.enqueued - ue.ewma;
3720 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3724 * To avoid overestimation of actual task utilization, skip updates if
3725 * we cannot grant there is idle time in this CPU.
3727 cpu = cpu_of(rq_of(cfs_rq));
3728 if (task_util(p) > capacity_orig_of(cpu))
3732 * Update Task's estimated utilization
3734 * When *p completes an activation we can consolidate another sample
3735 * of the task size. This is done by storing the current PELT value
3736 * as ue.enqueued and by using this value to update the Exponential
3737 * Weighted Moving Average (EWMA):
3739 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
3740 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
3741 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
3742 * = w * ( last_ewma_diff ) + ewma(t-1)
3743 * = w * (last_ewma_diff + ewma(t-1) / w)
3745 * Where 'w' is the weight of new samples, which is configured to be
3746 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3748 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
3749 ue.ewma += last_ewma_diff;
3750 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
3751 WRITE_ONCE(p->se.avg.util_est, ue);
3754 static inline int task_fits_capacity(struct task_struct *p, long capacity)
3756 return capacity * 1024 > task_util_est(p) * capacity_margin;
3759 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
3761 if (!static_branch_unlikely(&sched_asym_cpucapacity))
3765 rq->misfit_task_load = 0;
3769 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
3770 rq->misfit_task_load = 0;
3774 rq->misfit_task_load = task_h_load(p);
3777 #else /* CONFIG_SMP */
3779 #define UPDATE_TG 0x0
3780 #define SKIP_AGE_LOAD 0x0
3781 #define DO_ATTACH 0x0
3783 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3785 cfs_rq_util_change(cfs_rq, 0);
3788 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3791 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
3793 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3795 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3801 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
3804 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
3806 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
3808 #endif /* CONFIG_SMP */
3810 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3812 #ifdef CONFIG_SCHED_DEBUG
3813 s64 d = se->vruntime - cfs_rq->min_vruntime;
3818 if (d > 3*sysctl_sched_latency)
3819 schedstat_inc(cfs_rq->nr_spread_over);
3824 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3826 u64 vruntime = cfs_rq->min_vruntime;
3829 * The 'current' period is already promised to the current tasks,
3830 * however the extra weight of the new task will slow them down a
3831 * little, place the new task so that it fits in the slot that
3832 * stays open at the end.
3834 if (initial && sched_feat(START_DEBIT))
3835 vruntime += sched_vslice(cfs_rq, se);
3837 /* sleeps up to a single latency don't count. */
3839 unsigned long thresh = sysctl_sched_latency;
3842 * Halve their sleep time's effect, to allow
3843 * for a gentler effect of sleepers:
3845 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3851 /* ensure we never gain time by being placed backwards. */
3852 se->vruntime = max_vruntime(se->vruntime, vruntime);
3855 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3857 static inline void check_schedstat_required(void)
3859 #ifdef CONFIG_SCHEDSTATS
3860 if (schedstat_enabled())
3863 /* Force schedstat enabled if a dependent tracepoint is active */
3864 if (trace_sched_stat_wait_enabled() ||
3865 trace_sched_stat_sleep_enabled() ||
3866 trace_sched_stat_iowait_enabled() ||
3867 trace_sched_stat_blocked_enabled() ||
3868 trace_sched_stat_runtime_enabled()) {
3869 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3870 "stat_blocked and stat_runtime require the "
3871 "kernel parameter schedstats=enable or "
3872 "kernel.sched_schedstats=1\n");
3883 * update_min_vruntime()
3884 * vruntime -= min_vruntime
3888 * update_min_vruntime()
3889 * vruntime += min_vruntime
3891 * this way the vruntime transition between RQs is done when both
3892 * min_vruntime are up-to-date.
3896 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3897 * vruntime -= min_vruntime
3901 * update_min_vruntime()
3902 * vruntime += min_vruntime
3904 * this way we don't have the most up-to-date min_vruntime on the originating
3905 * CPU and an up-to-date min_vruntime on the destination CPU.
3909 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3911 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3912 bool curr = cfs_rq->curr == se;
3915 * If we're the current task, we must renormalise before calling
3919 se->vruntime += cfs_rq->min_vruntime;
3921 update_curr(cfs_rq);
3924 * Otherwise, renormalise after, such that we're placed at the current
3925 * moment in time, instead of some random moment in the past. Being
3926 * placed in the past could significantly boost this task to the
3927 * fairness detriment of existing tasks.
3929 if (renorm && !curr)
3930 se->vruntime += cfs_rq->min_vruntime;
3933 * When enqueuing a sched_entity, we must:
3934 * - Update loads to have both entity and cfs_rq synced with now.
3935 * - Add its load to cfs_rq->runnable_avg
3936 * - For group_entity, update its weight to reflect the new share of
3938 * - Add its new weight to cfs_rq->load.weight
3940 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
3941 update_cfs_group(se);
3942 enqueue_runnable_load_avg(cfs_rq, se);
3943 account_entity_enqueue(cfs_rq, se);
3945 if (flags & ENQUEUE_WAKEUP)
3946 place_entity(cfs_rq, se, 0);
3948 check_schedstat_required();
3949 update_stats_enqueue(cfs_rq, se, flags);
3950 check_spread(cfs_rq, se);
3952 __enqueue_entity(cfs_rq, se);
3955 if (cfs_rq->nr_running == 1) {
3956 list_add_leaf_cfs_rq(cfs_rq);
3957 check_enqueue_throttle(cfs_rq);
3961 static void __clear_buddies_last(struct sched_entity *se)
3963 for_each_sched_entity(se) {
3964 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3965 if (cfs_rq->last != se)
3968 cfs_rq->last = NULL;
3972 static void __clear_buddies_next(struct sched_entity *se)
3974 for_each_sched_entity(se) {
3975 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3976 if (cfs_rq->next != se)
3979 cfs_rq->next = NULL;
3983 static void __clear_buddies_skip(struct sched_entity *se)
3985 for_each_sched_entity(se) {
3986 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3987 if (cfs_rq->skip != se)
3990 cfs_rq->skip = NULL;
3994 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3996 if (cfs_rq->last == se)
3997 __clear_buddies_last(se);
3999 if (cfs_rq->next == se)
4000 __clear_buddies_next(se);
4002 if (cfs_rq->skip == se)
4003 __clear_buddies_skip(se);
4006 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4009 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4012 * Update run-time statistics of the 'current'.
4014 update_curr(cfs_rq);
4017 * When dequeuing a sched_entity, we must:
4018 * - Update loads to have both entity and cfs_rq synced with now.
4019 * - Subtract its load from the cfs_rq->runnable_avg.
4020 * - Subtract its previous weight from cfs_rq->load.weight.
4021 * - For group entity, update its weight to reflect the new share
4022 * of its group cfs_rq.
4024 update_load_avg(cfs_rq, se, UPDATE_TG);
4025 dequeue_runnable_load_avg(cfs_rq, se);
4027 update_stats_dequeue(cfs_rq, se, flags);
4029 clear_buddies(cfs_rq, se);
4031 if (se != cfs_rq->curr)
4032 __dequeue_entity(cfs_rq, se);
4034 account_entity_dequeue(cfs_rq, se);
4037 * Normalize after update_curr(); which will also have moved
4038 * min_vruntime if @se is the one holding it back. But before doing
4039 * update_min_vruntime() again, which will discount @se's position and
4040 * can move min_vruntime forward still more.
4042 if (!(flags & DEQUEUE_SLEEP))
4043 se->vruntime -= cfs_rq->min_vruntime;
4045 /* return excess runtime on last dequeue */
4046 return_cfs_rq_runtime(cfs_rq);
4048 update_cfs_group(se);
4051 * Now advance min_vruntime if @se was the entity holding it back,
4052 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4053 * put back on, and if we advance min_vruntime, we'll be placed back
4054 * further than we started -- ie. we'll be penalized.
4056 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4057 update_min_vruntime(cfs_rq);
4061 * Preempt the current task with a newly woken task if needed:
4064 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4066 unsigned long ideal_runtime, delta_exec;
4067 struct sched_entity *se;
4070 ideal_runtime = sched_slice(cfs_rq, curr);
4071 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4072 if (delta_exec > ideal_runtime) {
4073 resched_curr(rq_of(cfs_rq));
4075 * The current task ran long enough, ensure it doesn't get
4076 * re-elected due to buddy favours.
4078 clear_buddies(cfs_rq, curr);
4083 * Ensure that a task that missed wakeup preemption by a
4084 * narrow margin doesn't have to wait for a full slice.
4085 * This also mitigates buddy induced latencies under load.
4087 if (delta_exec < sysctl_sched_min_granularity)
4090 se = __pick_first_entity(cfs_rq);
4091 delta = curr->vruntime - se->vruntime;
4096 if (delta > ideal_runtime)
4097 resched_curr(rq_of(cfs_rq));
4101 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4103 /* 'current' is not kept within the tree. */
4106 * Any task has to be enqueued before it get to execute on
4107 * a CPU. So account for the time it spent waiting on the
4110 update_stats_wait_end(cfs_rq, se);
4111 __dequeue_entity(cfs_rq, se);
4112 update_load_avg(cfs_rq, se, UPDATE_TG);
4115 update_stats_curr_start(cfs_rq, se);
4119 * Track our maximum slice length, if the CPU's load is at
4120 * least twice that of our own weight (i.e. dont track it
4121 * when there are only lesser-weight tasks around):
4123 if (schedstat_enabled() &&
4124 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
4125 schedstat_set(se->statistics.slice_max,
4126 max((u64)schedstat_val(se->statistics.slice_max),
4127 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4130 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4134 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4137 * Pick the next process, keeping these things in mind, in this order:
4138 * 1) keep things fair between processes/task groups
4139 * 2) pick the "next" process, since someone really wants that to run
4140 * 3) pick the "last" process, for cache locality
4141 * 4) do not run the "skip" process, if something else is available
4143 static struct sched_entity *
4144 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4146 struct sched_entity *left = __pick_first_entity(cfs_rq);
4147 struct sched_entity *se;
4150 * If curr is set we have to see if its left of the leftmost entity
4151 * still in the tree, provided there was anything in the tree at all.
4153 if (!left || (curr && entity_before(curr, left)))
4156 se = left; /* ideally we run the leftmost entity */
4159 * Avoid running the skip buddy, if running something else can
4160 * be done without getting too unfair.
4162 if (cfs_rq->skip == se) {
4163 struct sched_entity *second;
4166 second = __pick_first_entity(cfs_rq);
4168 second = __pick_next_entity(se);
4169 if (!second || (curr && entity_before(curr, second)))
4173 if (second && wakeup_preempt_entity(second, left) < 1)
4178 * Prefer last buddy, try to return the CPU to a preempted task.
4180 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4184 * Someone really wants this to run. If it's not unfair, run it.
4186 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4189 clear_buddies(cfs_rq, se);
4194 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4196 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4199 * If still on the runqueue then deactivate_task()
4200 * was not called and update_curr() has to be done:
4203 update_curr(cfs_rq);
4205 /* throttle cfs_rqs exceeding runtime */
4206 check_cfs_rq_runtime(cfs_rq);
4208 check_spread(cfs_rq, prev);
4211 update_stats_wait_start(cfs_rq, prev);
4212 /* Put 'current' back into the tree. */
4213 __enqueue_entity(cfs_rq, prev);
4214 /* in !on_rq case, update occurred at dequeue */
4215 update_load_avg(cfs_rq, prev, 0);
4217 cfs_rq->curr = NULL;
4221 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4224 * Update run-time statistics of the 'current'.
4226 update_curr(cfs_rq);
4229 * Ensure that runnable average is periodically updated.
4231 update_load_avg(cfs_rq, curr, UPDATE_TG);
4232 update_cfs_group(curr);
4234 #ifdef CONFIG_SCHED_HRTICK
4236 * queued ticks are scheduled to match the slice, so don't bother
4237 * validating it and just reschedule.
4240 resched_curr(rq_of(cfs_rq));
4244 * don't let the period tick interfere with the hrtick preemption
4246 if (!sched_feat(DOUBLE_TICK) &&
4247 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4251 if (cfs_rq->nr_running > 1)
4252 check_preempt_tick(cfs_rq, curr);
4256 /**************************************************
4257 * CFS bandwidth control machinery
4260 #ifdef CONFIG_CFS_BANDWIDTH
4262 #ifdef CONFIG_JUMP_LABEL
4263 static struct static_key __cfs_bandwidth_used;
4265 static inline bool cfs_bandwidth_used(void)
4267 return static_key_false(&__cfs_bandwidth_used);
4270 void cfs_bandwidth_usage_inc(void)
4272 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4275 void cfs_bandwidth_usage_dec(void)
4277 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4279 #else /* CONFIG_JUMP_LABEL */
4280 static bool cfs_bandwidth_used(void)
4285 void cfs_bandwidth_usage_inc(void) {}
4286 void cfs_bandwidth_usage_dec(void) {}
4287 #endif /* CONFIG_JUMP_LABEL */
4290 * default period for cfs group bandwidth.
4291 * default: 0.1s, units: nanoseconds
4293 static inline u64 default_cfs_period(void)
4295 return 100000000ULL;
4298 static inline u64 sched_cfs_bandwidth_slice(void)
4300 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4304 * Replenish runtime according to assigned quota and update expiration time.
4305 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4306 * additional synchronization around rq->lock.
4308 * requires cfs_b->lock
4310 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4314 if (cfs_b->quota == RUNTIME_INF)
4317 now = sched_clock_cpu(smp_processor_id());
4318 cfs_b->runtime = cfs_b->quota;
4319 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4320 cfs_b->expires_seq++;
4323 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4325 return &tg->cfs_bandwidth;
4328 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4329 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4331 if (unlikely(cfs_rq->throttle_count))
4332 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4334 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4337 /* returns 0 on failure to allocate runtime */
4338 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4340 struct task_group *tg = cfs_rq->tg;
4341 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4342 u64 amount = 0, min_amount, expires;
4345 /* note: this is a positive sum as runtime_remaining <= 0 */
4346 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4348 raw_spin_lock(&cfs_b->lock);
4349 if (cfs_b->quota == RUNTIME_INF)
4350 amount = min_amount;
4352 start_cfs_bandwidth(cfs_b);
4354 if (cfs_b->runtime > 0) {
4355 amount = min(cfs_b->runtime, min_amount);
4356 cfs_b->runtime -= amount;
4360 expires_seq = cfs_b->expires_seq;
4361 expires = cfs_b->runtime_expires;
4362 raw_spin_unlock(&cfs_b->lock);
4364 cfs_rq->runtime_remaining += amount;
4366 * we may have advanced our local expiration to account for allowed
4367 * spread between our sched_clock and the one on which runtime was
4370 if (cfs_rq->expires_seq != expires_seq) {
4371 cfs_rq->expires_seq = expires_seq;
4372 cfs_rq->runtime_expires = expires;
4375 return cfs_rq->runtime_remaining > 0;
4379 * Note: This depends on the synchronization provided by sched_clock and the
4380 * fact that rq->clock snapshots this value.
4382 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4384 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4386 /* if the deadline is ahead of our clock, nothing to do */
4387 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4390 if (cfs_rq->runtime_remaining < 0)
4394 * If the local deadline has passed we have to consider the
4395 * possibility that our sched_clock is 'fast' and the global deadline
4396 * has not truly expired.
4398 * Fortunately we can check determine whether this the case by checking
4399 * whether the global deadline(cfs_b->expires_seq) has advanced.
4401 if (cfs_rq->expires_seq == cfs_b->expires_seq) {
4402 /* extend local deadline, drift is bounded above by 2 ticks */
4403 cfs_rq->runtime_expires += TICK_NSEC;
4405 /* global deadline is ahead, expiration has passed */
4406 cfs_rq->runtime_remaining = 0;
4410 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4412 /* dock delta_exec before expiring quota (as it could span periods) */
4413 cfs_rq->runtime_remaining -= delta_exec;
4414 expire_cfs_rq_runtime(cfs_rq);
4416 if (likely(cfs_rq->runtime_remaining > 0))
4420 * if we're unable to extend our runtime we resched so that the active
4421 * hierarchy can be throttled
4423 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4424 resched_curr(rq_of(cfs_rq));
4427 static __always_inline
4428 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4430 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4433 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4436 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4438 return cfs_bandwidth_used() && cfs_rq->throttled;
4441 /* check whether cfs_rq, or any parent, is throttled */
4442 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4444 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4448 * Ensure that neither of the group entities corresponding to src_cpu or
4449 * dest_cpu are members of a throttled hierarchy when performing group
4450 * load-balance operations.
4452 static inline int throttled_lb_pair(struct task_group *tg,
4453 int src_cpu, int dest_cpu)
4455 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4457 src_cfs_rq = tg->cfs_rq[src_cpu];
4458 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4460 return throttled_hierarchy(src_cfs_rq) ||
4461 throttled_hierarchy(dest_cfs_rq);
4464 static int tg_unthrottle_up(struct task_group *tg, void *data)
4466 struct rq *rq = data;
4467 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4469 cfs_rq->throttle_count--;
4470 if (!cfs_rq->throttle_count) {
4471 /* adjust cfs_rq_clock_task() */
4472 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4473 cfs_rq->throttled_clock_task;
4475 /* Add cfs_rq with already running entity in the list */
4476 if (cfs_rq->nr_running >= 1)
4477 list_add_leaf_cfs_rq(cfs_rq);
4483 static int tg_throttle_down(struct task_group *tg, void *data)
4485 struct rq *rq = data;
4486 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4488 /* group is entering throttled state, stop time */
4489 if (!cfs_rq->throttle_count) {
4490 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4491 list_del_leaf_cfs_rq(cfs_rq);
4493 cfs_rq->throttle_count++;
4498 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4500 struct rq *rq = rq_of(cfs_rq);
4501 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4502 struct sched_entity *se;
4503 long task_delta, dequeue = 1;
4506 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4508 /* freeze hierarchy runnable averages while throttled */
4510 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4513 task_delta = cfs_rq->h_nr_running;
4514 for_each_sched_entity(se) {
4515 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4516 /* throttled entity or throttle-on-deactivate */
4521 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4522 qcfs_rq->h_nr_running -= task_delta;
4524 if (qcfs_rq->load.weight)
4529 sub_nr_running(rq, task_delta);
4531 cfs_rq->throttled = 1;
4532 cfs_rq->throttled_clock = rq_clock(rq);
4533 raw_spin_lock(&cfs_b->lock);
4534 empty = list_empty(&cfs_b->throttled_cfs_rq);
4537 * Add to the _head_ of the list, so that an already-started
4538 * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
4539 * not running add to the tail so that later runqueues don't get starved.
4541 if (cfs_b->distribute_running)
4542 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4544 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4547 * If we're the first throttled task, make sure the bandwidth
4551 start_cfs_bandwidth(cfs_b);
4553 raw_spin_unlock(&cfs_b->lock);
4556 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4558 struct rq *rq = rq_of(cfs_rq);
4559 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4560 struct sched_entity *se;
4564 se = cfs_rq->tg->se[cpu_of(rq)];
4566 cfs_rq->throttled = 0;
4568 update_rq_clock(rq);
4570 raw_spin_lock(&cfs_b->lock);
4571 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4572 list_del_rcu(&cfs_rq->throttled_list);
4573 raw_spin_unlock(&cfs_b->lock);
4575 /* update hierarchical throttle state */
4576 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4578 if (!cfs_rq->load.weight)
4581 task_delta = cfs_rq->h_nr_running;
4582 for_each_sched_entity(se) {
4586 cfs_rq = cfs_rq_of(se);
4588 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4589 cfs_rq->h_nr_running += task_delta;
4591 if (cfs_rq_throttled(cfs_rq))
4595 assert_list_leaf_cfs_rq(rq);
4598 add_nr_running(rq, task_delta);
4600 /* Determine whether we need to wake up potentially idle CPU: */
4601 if (rq->curr == rq->idle && rq->cfs.nr_running)
4605 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4606 u64 remaining, u64 expires)
4608 struct cfs_rq *cfs_rq;
4610 u64 starting_runtime = remaining;
4613 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4615 struct rq *rq = rq_of(cfs_rq);
4618 rq_lock_irqsave(rq, &rf);
4619 if (!cfs_rq_throttled(cfs_rq))
4622 runtime = -cfs_rq->runtime_remaining + 1;
4623 if (runtime > remaining)
4624 runtime = remaining;
4625 remaining -= runtime;
4627 cfs_rq->runtime_remaining += runtime;
4628 cfs_rq->runtime_expires = expires;
4630 /* we check whether we're throttled above */
4631 if (cfs_rq->runtime_remaining > 0)
4632 unthrottle_cfs_rq(cfs_rq);
4635 rq_unlock_irqrestore(rq, &rf);
4642 return starting_runtime - remaining;
4646 * Responsible for refilling a task_group's bandwidth and unthrottling its
4647 * cfs_rqs as appropriate. If there has been no activity within the last
4648 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4649 * used to track this state.
4651 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
4653 u64 runtime, runtime_expires;
4656 /* no need to continue the timer with no bandwidth constraint */
4657 if (cfs_b->quota == RUNTIME_INF)
4658 goto out_deactivate;
4660 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4661 cfs_b->nr_periods += overrun;
4664 * idle depends on !throttled (for the case of a large deficit), and if
4665 * we're going inactive then everything else can be deferred
4667 if (cfs_b->idle && !throttled)
4668 goto out_deactivate;
4670 __refill_cfs_bandwidth_runtime(cfs_b);
4673 /* mark as potentially idle for the upcoming period */
4678 /* account preceding periods in which throttling occurred */
4679 cfs_b->nr_throttled += overrun;
4681 runtime_expires = cfs_b->runtime_expires;
4684 * This check is repeated as we are holding onto the new bandwidth while
4685 * we unthrottle. This can potentially race with an unthrottled group
4686 * trying to acquire new bandwidth from the global pool. This can result
4687 * in us over-using our runtime if it is all used during this loop, but
4688 * only by limited amounts in that extreme case.
4690 while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
4691 runtime = cfs_b->runtime;
4692 cfs_b->distribute_running = 1;
4693 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4694 /* we can't nest cfs_b->lock while distributing bandwidth */
4695 runtime = distribute_cfs_runtime(cfs_b, runtime,
4697 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4699 cfs_b->distribute_running = 0;
4700 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4702 lsub_positive(&cfs_b->runtime, runtime);
4706 * While we are ensured activity in the period following an
4707 * unthrottle, this also covers the case in which the new bandwidth is
4708 * insufficient to cover the existing bandwidth deficit. (Forcing the
4709 * timer to remain active while there are any throttled entities.)
4719 /* a cfs_rq won't donate quota below this amount */
4720 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4721 /* minimum remaining period time to redistribute slack quota */
4722 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4723 /* how long we wait to gather additional slack before distributing */
4724 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4727 * Are we near the end of the current quota period?
4729 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4730 * hrtimer base being cleared by hrtimer_start. In the case of
4731 * migrate_hrtimers, base is never cleared, so we are fine.
4733 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4735 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4738 /* if the call-back is running a quota refresh is already occurring */
4739 if (hrtimer_callback_running(refresh_timer))
4742 /* is a quota refresh about to occur? */
4743 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4744 if (remaining < min_expire)
4750 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4752 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4754 /* if there's a quota refresh soon don't bother with slack */
4755 if (runtime_refresh_within(cfs_b, min_left))
4758 /* don't push forwards an existing deferred unthrottle */
4759 if (cfs_b->slack_started)
4761 cfs_b->slack_started = true;
4763 hrtimer_start(&cfs_b->slack_timer,
4764 ns_to_ktime(cfs_bandwidth_slack_period),
4768 /* we know any runtime found here is valid as update_curr() precedes return */
4769 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4771 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4772 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4774 if (slack_runtime <= 0)
4777 raw_spin_lock(&cfs_b->lock);
4778 if (cfs_b->quota != RUNTIME_INF &&
4779 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4780 cfs_b->runtime += slack_runtime;
4782 /* we are under rq->lock, defer unthrottling using a timer */
4783 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4784 !list_empty(&cfs_b->throttled_cfs_rq))
4785 start_cfs_slack_bandwidth(cfs_b);
4787 raw_spin_unlock(&cfs_b->lock);
4789 /* even if it's not valid for return we don't want to try again */
4790 cfs_rq->runtime_remaining -= slack_runtime;
4793 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4795 if (!cfs_bandwidth_used())
4798 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4801 __return_cfs_rq_runtime(cfs_rq);
4805 * This is done with a timer (instead of inline with bandwidth return) since
4806 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4808 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4810 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4811 unsigned long flags;
4814 /* confirm we're still not at a refresh boundary */
4815 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4816 cfs_b->slack_started = false;
4817 if (cfs_b->distribute_running) {
4818 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4822 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4823 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4827 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4828 runtime = cfs_b->runtime;
4830 expires = cfs_b->runtime_expires;
4832 cfs_b->distribute_running = 1;
4834 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4839 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4841 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4842 if (expires == cfs_b->runtime_expires)
4843 lsub_positive(&cfs_b->runtime, runtime);
4844 cfs_b->distribute_running = 0;
4845 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4849 * When a group wakes up we want to make sure that its quota is not already
4850 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4851 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4853 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4855 if (!cfs_bandwidth_used())
4858 /* an active group must be handled by the update_curr()->put() path */
4859 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4862 /* ensure the group is not already throttled */
4863 if (cfs_rq_throttled(cfs_rq))
4866 /* update runtime allocation */
4867 account_cfs_rq_runtime(cfs_rq, 0);
4868 if (cfs_rq->runtime_remaining <= 0)
4869 throttle_cfs_rq(cfs_rq);
4872 static void sync_throttle(struct task_group *tg, int cpu)
4874 struct cfs_rq *pcfs_rq, *cfs_rq;
4876 if (!cfs_bandwidth_used())
4882 cfs_rq = tg->cfs_rq[cpu];
4883 pcfs_rq = tg->parent->cfs_rq[cpu];
4885 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4886 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4889 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4890 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4892 if (!cfs_bandwidth_used())
4895 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4899 * it's possible for a throttled entity to be forced into a running
4900 * state (e.g. set_curr_task), in this case we're finished.
4902 if (cfs_rq_throttled(cfs_rq))
4905 throttle_cfs_rq(cfs_rq);
4909 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4911 struct cfs_bandwidth *cfs_b =
4912 container_of(timer, struct cfs_bandwidth, slack_timer);
4914 do_sched_cfs_slack_timer(cfs_b);
4916 return HRTIMER_NORESTART;
4919 extern const u64 max_cfs_quota_period;
4921 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4923 struct cfs_bandwidth *cfs_b =
4924 container_of(timer, struct cfs_bandwidth, period_timer);
4925 unsigned long flags;
4930 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4932 overrun = hrtimer_forward_now(timer, cfs_b->period);
4937 u64 new, old = ktime_to_ns(cfs_b->period);
4939 new = (old * 147) / 128; /* ~115% */
4940 new = min(new, max_cfs_quota_period);
4942 cfs_b->period = ns_to_ktime(new);
4944 /* since max is 1s, this is limited to 1e9^2, which fits in u64 */
4945 cfs_b->quota *= new;
4946 cfs_b->quota = div64_u64(cfs_b->quota, old);
4948 pr_warn_ratelimited(
4949 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us %lld, cfs_quota_us = %lld)\n",
4951 div_u64(new, NSEC_PER_USEC),
4952 div_u64(cfs_b->quota, NSEC_PER_USEC));
4954 /* reset count so we don't come right back in here */
4958 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
4961 cfs_b->period_active = 0;
4962 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4964 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4967 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4969 raw_spin_lock_init(&cfs_b->lock);
4971 cfs_b->quota = RUNTIME_INF;
4972 cfs_b->period = ns_to_ktime(default_cfs_period());
4974 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4975 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4976 cfs_b->period_timer.function = sched_cfs_period_timer;
4977 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4978 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4979 cfs_b->distribute_running = 0;
4980 cfs_b->slack_started = false;
4983 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4985 cfs_rq->runtime_enabled = 0;
4986 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4989 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4993 lockdep_assert_held(&cfs_b->lock);
4995 if (cfs_b->period_active)
4998 cfs_b->period_active = 1;
4999 overrun = hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5000 cfs_b->runtime_expires += (overrun + 1) * ktime_to_ns(cfs_b->period);
5001 cfs_b->expires_seq++;
5002 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5005 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5007 /* init_cfs_bandwidth() was not called */
5008 if (!cfs_b->throttled_cfs_rq.next)
5011 hrtimer_cancel(&cfs_b->period_timer);
5012 hrtimer_cancel(&cfs_b->slack_timer);
5016 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5018 * The race is harmless, since modifying bandwidth settings of unhooked group
5019 * bits doesn't do much.
5022 /* cpu online calback */
5023 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5025 struct task_group *tg;
5027 lockdep_assert_held(&rq->lock);
5030 list_for_each_entry_rcu(tg, &task_groups, list) {
5031 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5032 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5034 raw_spin_lock(&cfs_b->lock);
5035 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5036 raw_spin_unlock(&cfs_b->lock);
5041 /* cpu offline callback */
5042 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5044 struct task_group *tg;
5046 lockdep_assert_held(&rq->lock);
5049 list_for_each_entry_rcu(tg, &task_groups, list) {
5050 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5052 if (!cfs_rq->runtime_enabled)
5056 * clock_task is not advancing so we just need to make sure
5057 * there's some valid quota amount
5059 cfs_rq->runtime_remaining = 1;
5061 * Offline rq is schedulable till CPU is completely disabled
5062 * in take_cpu_down(), so we prevent new cfs throttling here.
5064 cfs_rq->runtime_enabled = 0;
5066 if (cfs_rq_throttled(cfs_rq))
5067 unthrottle_cfs_rq(cfs_rq);
5072 #else /* CONFIG_CFS_BANDWIDTH */
5074 static inline bool cfs_bandwidth_used(void)
5079 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
5081 return rq_clock_task(rq_of(cfs_rq));
5084 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5085 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5086 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5087 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5088 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5090 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5095 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5100 static inline int throttled_lb_pair(struct task_group *tg,
5101 int src_cpu, int dest_cpu)
5106 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5108 #ifdef CONFIG_FAIR_GROUP_SCHED
5109 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5112 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5116 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5117 static inline void update_runtime_enabled(struct rq *rq) {}
5118 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5120 #endif /* CONFIG_CFS_BANDWIDTH */
5122 /**************************************************
5123 * CFS operations on tasks:
5126 #ifdef CONFIG_SCHED_HRTICK
5127 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5129 struct sched_entity *se = &p->se;
5130 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5132 SCHED_WARN_ON(task_rq(p) != rq);
5134 if (rq->cfs.h_nr_running > 1) {
5135 u64 slice = sched_slice(cfs_rq, se);
5136 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5137 s64 delta = slice - ran;
5144 hrtick_start(rq, delta);
5149 * called from enqueue/dequeue and updates the hrtick when the
5150 * current task is from our class and nr_running is low enough
5153 static void hrtick_update(struct rq *rq)
5155 struct task_struct *curr = rq->curr;
5157 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5160 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5161 hrtick_start_fair(rq, curr);
5163 #else /* !CONFIG_SCHED_HRTICK */
5165 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5169 static inline void hrtick_update(struct rq *rq)
5175 static inline unsigned long cpu_util(int cpu);
5177 static inline bool cpu_overutilized(int cpu)
5179 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5182 static inline void update_overutilized_status(struct rq *rq)
5184 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
5185 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5186 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
5190 static inline void update_overutilized_status(struct rq *rq) { }
5194 * The enqueue_task method is called before nr_running is
5195 * increased. Here we update the fair scheduling stats and
5196 * then put the task into the rbtree:
5199 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5201 struct cfs_rq *cfs_rq;
5202 struct sched_entity *se = &p->se;
5205 * The code below (indirectly) updates schedutil which looks at
5206 * the cfs_rq utilization to select a frequency.
5207 * Let's add the task's estimated utilization to the cfs_rq's
5208 * estimated utilization, before we update schedutil.
5210 util_est_enqueue(&rq->cfs, p);
5213 * If in_iowait is set, the code below may not trigger any cpufreq
5214 * utilization updates, so do it here explicitly with the IOWAIT flag
5218 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5220 for_each_sched_entity(se) {
5223 cfs_rq = cfs_rq_of(se);
5224 enqueue_entity(cfs_rq, se, flags);
5227 * end evaluation on encountering a throttled cfs_rq
5229 * note: in the case of encountering a throttled cfs_rq we will
5230 * post the final h_nr_running increment below.
5232 if (cfs_rq_throttled(cfs_rq))
5234 cfs_rq->h_nr_running++;
5236 flags = ENQUEUE_WAKEUP;
5239 for_each_sched_entity(se) {
5240 cfs_rq = cfs_rq_of(se);
5241 cfs_rq->h_nr_running++;
5243 if (cfs_rq_throttled(cfs_rq))
5246 update_load_avg(cfs_rq, se, UPDATE_TG);
5247 update_cfs_group(se);
5251 add_nr_running(rq, 1);
5253 * Since new tasks are assigned an initial util_avg equal to
5254 * half of the spare capacity of their CPU, tiny tasks have the
5255 * ability to cross the overutilized threshold, which will
5256 * result in the load balancer ruining all the task placement
5257 * done by EAS. As a way to mitigate that effect, do not account
5258 * for the first enqueue operation of new tasks during the
5259 * overutilized flag detection.
5261 * A better way of solving this problem would be to wait for
5262 * the PELT signals of tasks to converge before taking them
5263 * into account, but that is not straightforward to implement,
5264 * and the following generally works well enough in practice.
5266 if (flags & ENQUEUE_WAKEUP)
5267 update_overutilized_status(rq);
5271 if (cfs_bandwidth_used()) {
5273 * When bandwidth control is enabled; the cfs_rq_throttled()
5274 * breaks in the above iteration can result in incomplete
5275 * leaf list maintenance, resulting in triggering the assertion
5278 for_each_sched_entity(se) {
5279 cfs_rq = cfs_rq_of(se);
5281 if (list_add_leaf_cfs_rq(cfs_rq))
5286 assert_list_leaf_cfs_rq(rq);
5291 static void set_next_buddy(struct sched_entity *se);
5294 * The dequeue_task method is called before nr_running is
5295 * decreased. We remove the task from the rbtree and
5296 * update the fair scheduling stats:
5298 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5300 struct cfs_rq *cfs_rq;
5301 struct sched_entity *se = &p->se;
5302 int task_sleep = flags & DEQUEUE_SLEEP;
5304 for_each_sched_entity(se) {
5305 cfs_rq = cfs_rq_of(se);
5306 dequeue_entity(cfs_rq, se, flags);
5309 * end evaluation on encountering a throttled cfs_rq
5311 * note: in the case of encountering a throttled cfs_rq we will
5312 * post the final h_nr_running decrement below.
5314 if (cfs_rq_throttled(cfs_rq))
5316 cfs_rq->h_nr_running--;
5318 /* Don't dequeue parent if it has other entities besides us */
5319 if (cfs_rq->load.weight) {
5320 /* Avoid re-evaluating load for this entity: */
5321 se = parent_entity(se);
5323 * Bias pick_next to pick a task from this cfs_rq, as
5324 * p is sleeping when it is within its sched_slice.
5326 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5330 flags |= DEQUEUE_SLEEP;
5333 for_each_sched_entity(se) {
5334 cfs_rq = cfs_rq_of(se);
5335 cfs_rq->h_nr_running--;
5337 if (cfs_rq_throttled(cfs_rq))
5340 update_load_avg(cfs_rq, se, UPDATE_TG);
5341 update_cfs_group(se);
5345 sub_nr_running(rq, 1);
5347 util_est_dequeue(&rq->cfs, p, task_sleep);
5353 /* Working cpumask for: load_balance, load_balance_newidle. */
5354 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5355 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5357 #ifdef CONFIG_NO_HZ_COMMON
5360 cpumask_var_t idle_cpus_mask;
5362 int has_blocked; /* Idle CPUS has blocked load */
5363 unsigned long next_balance; /* in jiffy units */
5364 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5365 } nohz ____cacheline_aligned;
5367 #endif /* CONFIG_NO_HZ_COMMON */
5369 static unsigned long weighted_cpuload(struct rq *rq)
5371 return cfs_rq_runnable_load_avg(&rq->cfs);
5374 static unsigned long capacity_of(int cpu)
5376 return cpu_rq(cpu)->cpu_capacity;
5379 static unsigned long cpu_avg_load_per_task(int cpu)
5381 struct rq *rq = cpu_rq(cpu);
5382 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5383 unsigned long load_avg = weighted_cpuload(rq);
5386 return load_avg / nr_running;
5391 static void record_wakee(struct task_struct *p)
5394 * Only decay a single time; tasks that have less then 1 wakeup per
5395 * jiffy will not have built up many flips.
5397 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5398 current->wakee_flips >>= 1;
5399 current->wakee_flip_decay_ts = jiffies;
5402 if (current->last_wakee != p) {
5403 current->last_wakee = p;
5404 current->wakee_flips++;
5409 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5411 * A waker of many should wake a different task than the one last awakened
5412 * at a frequency roughly N times higher than one of its wakees.
5414 * In order to determine whether we should let the load spread vs consolidating
5415 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5416 * partner, and a factor of lls_size higher frequency in the other.
5418 * With both conditions met, we can be relatively sure that the relationship is
5419 * non-monogamous, with partner count exceeding socket size.
5421 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5422 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5425 static int wake_wide(struct task_struct *p)
5427 unsigned int master = current->wakee_flips;
5428 unsigned int slave = p->wakee_flips;
5429 int factor = this_cpu_read(sd_llc_size);
5432 swap(master, slave);
5433 if (slave < factor || master < slave * factor)
5439 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5440 * soonest. For the purpose of speed we only consider the waking and previous
5443 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5444 * cache-affine and is (or will be) idle.
5446 * wake_affine_weight() - considers the weight to reflect the average
5447 * scheduling latency of the CPUs. This seems to work
5448 * for the overloaded case.
5451 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5454 * If this_cpu is idle, it implies the wakeup is from interrupt
5455 * context. Only allow the move if cache is shared. Otherwise an
5456 * interrupt intensive workload could force all tasks onto one
5457 * node depending on the IO topology or IRQ affinity settings.
5459 * If the prev_cpu is idle and cache affine then avoid a migration.
5460 * There is no guarantee that the cache hot data from an interrupt
5461 * is more important than cache hot data on the prev_cpu and from
5462 * a cpufreq perspective, it's better to have higher utilisation
5465 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5466 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5468 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5471 return nr_cpumask_bits;
5475 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5476 int this_cpu, int prev_cpu, int sync)
5478 s64 this_eff_load, prev_eff_load;
5479 unsigned long task_load;
5481 this_eff_load = weighted_cpuload(cpu_rq(this_cpu));
5484 unsigned long current_load = task_h_load(current);
5486 if (current_load > this_eff_load)
5489 this_eff_load -= current_load;
5492 task_load = task_h_load(p);
5494 this_eff_load += task_load;
5495 if (sched_feat(WA_BIAS))
5496 this_eff_load *= 100;
5497 this_eff_load *= capacity_of(prev_cpu);
5499 prev_eff_load = weighted_cpuload(cpu_rq(prev_cpu));
5500 prev_eff_load -= task_load;
5501 if (sched_feat(WA_BIAS))
5502 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5503 prev_eff_load *= capacity_of(this_cpu);
5506 * If sync, adjust the weight of prev_eff_load such that if
5507 * prev_eff == this_eff that select_idle_sibling() will consider
5508 * stacking the wakee on top of the waker if no other CPU is
5514 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5517 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5518 int this_cpu, int prev_cpu, int sync)
5520 int target = nr_cpumask_bits;
5522 if (sched_feat(WA_IDLE))
5523 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5525 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5526 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5528 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5529 if (target == nr_cpumask_bits)
5532 schedstat_inc(sd->ttwu_move_affine);
5533 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5537 static unsigned long cpu_util_without(int cpu, struct task_struct *p);
5539 static unsigned long capacity_spare_without(int cpu, struct task_struct *p)
5541 return max_t(long, capacity_of(cpu) - cpu_util_without(cpu, p), 0);
5545 * find_idlest_group finds and returns the least busy CPU group within the
5548 * Assumes p is allowed on at least one CPU in sd.
5550 static struct sched_group *
5551 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5552 int this_cpu, int sd_flag)
5554 struct sched_group *idlest = NULL, *group = sd->groups;
5555 struct sched_group *most_spare_sg = NULL;
5556 unsigned long min_runnable_load = ULONG_MAX;
5557 unsigned long this_runnable_load = ULONG_MAX;
5558 unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5559 unsigned long most_spare = 0, this_spare = 0;
5560 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5561 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5562 (sd->imbalance_pct-100) / 100;
5565 unsigned long load, avg_load, runnable_load;
5566 unsigned long spare_cap, max_spare_cap;
5570 /* Skip over this group if it has no CPUs allowed */
5571 if (!cpumask_intersects(sched_group_span(group),
5575 local_group = cpumask_test_cpu(this_cpu,
5576 sched_group_span(group));
5579 * Tally up the load of all CPUs in the group and find
5580 * the group containing the CPU with most spare capacity.
5586 for_each_cpu(i, sched_group_span(group)) {
5587 load = weighted_cpuload(cpu_rq(i));
5588 runnable_load += load;
5590 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5592 spare_cap = capacity_spare_without(i, p);
5594 if (spare_cap > max_spare_cap)
5595 max_spare_cap = spare_cap;
5598 /* Adjust by relative CPU capacity of the group */
5599 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5600 group->sgc->capacity;
5601 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5602 group->sgc->capacity;
5605 this_runnable_load = runnable_load;
5606 this_avg_load = avg_load;
5607 this_spare = max_spare_cap;
5609 if (min_runnable_load > (runnable_load + imbalance)) {
5611 * The runnable load is significantly smaller
5612 * so we can pick this new CPU:
5614 min_runnable_load = runnable_load;
5615 min_avg_load = avg_load;
5617 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5618 (100*min_avg_load > imbalance_scale*avg_load)) {
5620 * The runnable loads are close so take the
5621 * blocked load into account through avg_load:
5623 min_avg_load = avg_load;
5627 if (most_spare < max_spare_cap) {
5628 most_spare = max_spare_cap;
5629 most_spare_sg = group;
5632 } while (group = group->next, group != sd->groups);
5635 * The cross-over point between using spare capacity or least load
5636 * is too conservative for high utilization tasks on partially
5637 * utilized systems if we require spare_capacity > task_util(p),
5638 * so we allow for some task stuffing by using
5639 * spare_capacity > task_util(p)/2.
5641 * Spare capacity can't be used for fork because the utilization has
5642 * not been set yet, we must first select a rq to compute the initial
5645 if (sd_flag & SD_BALANCE_FORK)
5648 if (this_spare > task_util(p) / 2 &&
5649 imbalance_scale*this_spare > 100*most_spare)
5652 if (most_spare > task_util(p) / 2)
5653 return most_spare_sg;
5660 * When comparing groups across NUMA domains, it's possible for the
5661 * local domain to be very lightly loaded relative to the remote
5662 * domains but "imbalance" skews the comparison making remote CPUs
5663 * look much more favourable. When considering cross-domain, add
5664 * imbalance to the runnable load on the remote node and consider
5667 if ((sd->flags & SD_NUMA) &&
5668 min_runnable_load + imbalance >= this_runnable_load)
5671 if (min_runnable_load > (this_runnable_load + imbalance))
5674 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5675 (100*this_avg_load < imbalance_scale*min_avg_load))
5682 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5685 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5687 unsigned long load, min_load = ULONG_MAX;
5688 unsigned int min_exit_latency = UINT_MAX;
5689 u64 latest_idle_timestamp = 0;
5690 int least_loaded_cpu = this_cpu;
5691 int shallowest_idle_cpu = -1;
5694 /* Check if we have any choice: */
5695 if (group->group_weight == 1)
5696 return cpumask_first(sched_group_span(group));
5698 /* Traverse only the allowed CPUs */
5699 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
5700 if (available_idle_cpu(i)) {
5701 struct rq *rq = cpu_rq(i);
5702 struct cpuidle_state *idle = idle_get_state(rq);
5703 if (idle && idle->exit_latency < min_exit_latency) {
5705 * We give priority to a CPU whose idle state
5706 * has the smallest exit latency irrespective
5707 * of any idle timestamp.
5709 min_exit_latency = idle->exit_latency;
5710 latest_idle_timestamp = rq->idle_stamp;
5711 shallowest_idle_cpu = i;
5712 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5713 rq->idle_stamp > latest_idle_timestamp) {
5715 * If equal or no active idle state, then
5716 * the most recently idled CPU might have
5719 latest_idle_timestamp = rq->idle_stamp;
5720 shallowest_idle_cpu = i;
5722 } else if (shallowest_idle_cpu == -1) {
5723 load = weighted_cpuload(cpu_rq(i));
5724 if (load < min_load) {
5726 least_loaded_cpu = i;
5731 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5734 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5735 int cpu, int prev_cpu, int sd_flag)
5739 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
5743 * We need task's util for capacity_spare_without, sync it up to
5744 * prev_cpu's last_update_time.
5746 if (!(sd_flag & SD_BALANCE_FORK))
5747 sync_entity_load_avg(&p->se);
5750 struct sched_group *group;
5751 struct sched_domain *tmp;
5754 if (!(sd->flags & sd_flag)) {
5759 group = find_idlest_group(sd, p, cpu, sd_flag);
5765 new_cpu = find_idlest_group_cpu(group, p, cpu);
5766 if (new_cpu == cpu) {
5767 /* Now try balancing at a lower domain level of 'cpu': */
5772 /* Now try balancing at a lower domain level of 'new_cpu': */
5774 weight = sd->span_weight;
5776 for_each_domain(cpu, tmp) {
5777 if (weight <= tmp->span_weight)
5779 if (tmp->flags & sd_flag)
5787 #ifdef CONFIG_SCHED_SMT
5788 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5789 EXPORT_SYMBOL_GPL(sched_smt_present);
5791 static inline void set_idle_cores(int cpu, int val)
5793 struct sched_domain_shared *sds;
5795 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5797 WRITE_ONCE(sds->has_idle_cores, val);
5800 static inline bool test_idle_cores(int cpu, bool def)
5802 struct sched_domain_shared *sds;
5804 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5806 return READ_ONCE(sds->has_idle_cores);
5812 * Scans the local SMT mask to see if the entire core is idle, and records this
5813 * information in sd_llc_shared->has_idle_cores.
5815 * Since SMT siblings share all cache levels, inspecting this limited remote
5816 * state should be fairly cheap.
5818 void __update_idle_core(struct rq *rq)
5820 int core = cpu_of(rq);
5824 if (test_idle_cores(core, true))
5827 for_each_cpu(cpu, cpu_smt_mask(core)) {
5831 if (!available_idle_cpu(cpu))
5835 set_idle_cores(core, 1);
5841 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5842 * there are no idle cores left in the system; tracked through
5843 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5845 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5847 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5850 if (!static_branch_likely(&sched_smt_present))
5853 if (!test_idle_cores(target, false))
5856 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
5858 for_each_cpu_wrap(core, cpus, target) {
5861 for_each_cpu(cpu, cpu_smt_mask(core)) {
5862 __cpumask_clear_cpu(cpu, cpus);
5863 if (!available_idle_cpu(cpu))
5872 * Failed to find an idle core; stop looking for one.
5874 set_idle_cores(target, 0);
5880 * Scan the local SMT mask for idle CPUs.
5882 static int select_idle_smt(struct task_struct *p, int target)
5886 if (!static_branch_likely(&sched_smt_present))
5889 for_each_cpu(cpu, cpu_smt_mask(target)) {
5890 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
5892 if (available_idle_cpu(cpu))
5899 #else /* CONFIG_SCHED_SMT */
5901 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5906 static inline int select_idle_smt(struct task_struct *p, int target)
5911 #endif /* CONFIG_SCHED_SMT */
5914 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5915 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5916 * average idle time for this rq (as found in rq->avg_idle).
5918 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
5920 struct sched_domain *this_sd;
5921 u64 avg_cost, avg_idle;
5924 int cpu, nr = INT_MAX;
5926 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
5931 * Due to large variance we need a large fuzz factor; hackbench in
5932 * particularly is sensitive here.
5934 avg_idle = this_rq()->avg_idle / 512;
5935 avg_cost = this_sd->avg_scan_cost + 1;
5937 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
5940 if (sched_feat(SIS_PROP)) {
5941 u64 span_avg = sd->span_weight * avg_idle;
5942 if (span_avg > 4*avg_cost)
5943 nr = div_u64(span_avg, avg_cost);
5948 time = local_clock();
5950 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
5953 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
5955 if (available_idle_cpu(cpu))
5959 time = local_clock() - time;
5960 cost = this_sd->avg_scan_cost;
5961 delta = (s64)(time - cost) / 8;
5962 this_sd->avg_scan_cost += delta;
5968 * Try and locate an idle core/thread in the LLC cache domain.
5970 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5972 struct sched_domain *sd;
5973 int i, recent_used_cpu;
5975 if (available_idle_cpu(target))
5979 * If the previous CPU is cache affine and idle, don't be stupid:
5981 if (prev != target && cpus_share_cache(prev, target) && available_idle_cpu(prev))
5984 /* Check a recently used CPU as a potential idle candidate: */
5985 recent_used_cpu = p->recent_used_cpu;
5986 if (recent_used_cpu != prev &&
5987 recent_used_cpu != target &&
5988 cpus_share_cache(recent_used_cpu, target) &&
5989 available_idle_cpu(recent_used_cpu) &&
5990 cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr)) {
5992 * Replace recent_used_cpu with prev as it is a potential
5993 * candidate for the next wake:
5995 p->recent_used_cpu = prev;
5996 return recent_used_cpu;
5999 sd = rcu_dereference(per_cpu(sd_llc, target));
6003 i = select_idle_core(p, sd, target);
6004 if ((unsigned)i < nr_cpumask_bits)
6007 i = select_idle_cpu(p, sd, target);
6008 if ((unsigned)i < nr_cpumask_bits)
6011 i = select_idle_smt(p, target);
6012 if ((unsigned)i < nr_cpumask_bits)
6019 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6020 * @cpu: the CPU to get the utilization of
6022 * The unit of the return value must be the one of capacity so we can compare
6023 * the utilization with the capacity of the CPU that is available for CFS task
6024 * (ie cpu_capacity).
6026 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6027 * recent utilization of currently non-runnable tasks on a CPU. It represents
6028 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6029 * capacity_orig is the cpu_capacity available at the highest frequency
6030 * (arch_scale_freq_capacity()).
6031 * The utilization of a CPU converges towards a sum equal to or less than the
6032 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6033 * the running time on this CPU scaled by capacity_curr.
6035 * The estimated utilization of a CPU is defined to be the maximum between its
6036 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6037 * currently RUNNABLE on that CPU.
6038 * This allows to properly represent the expected utilization of a CPU which
6039 * has just got a big task running since a long sleep period. At the same time
6040 * however it preserves the benefits of the "blocked utilization" in
6041 * describing the potential for other tasks waking up on the same CPU.
6043 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6044 * higher than capacity_orig because of unfortunate rounding in
6045 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6046 * the average stabilizes with the new running time. We need to check that the
6047 * utilization stays within the range of [0..capacity_orig] and cap it if
6048 * necessary. Without utilization capping, a group could be seen as overloaded
6049 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6050 * available capacity. We allow utilization to overshoot capacity_curr (but not
6051 * capacity_orig) as it useful for predicting the capacity required after task
6052 * migrations (scheduler-driven DVFS).
6054 * Return: the (estimated) utilization for the specified CPU
6056 static inline unsigned long cpu_util(int cpu)
6058 struct cfs_rq *cfs_rq;
6061 cfs_rq = &cpu_rq(cpu)->cfs;
6062 util = READ_ONCE(cfs_rq->avg.util_avg);
6064 if (sched_feat(UTIL_EST))
6065 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6067 return min_t(unsigned long, util, capacity_orig_of(cpu));
6071 * cpu_util_without: compute cpu utilization without any contributions from *p
6072 * @cpu: the CPU which utilization is requested
6073 * @p: the task which utilization should be discounted
6075 * The utilization of a CPU is defined by the utilization of tasks currently
6076 * enqueued on that CPU as well as tasks which are currently sleeping after an
6077 * execution on that CPU.
6079 * This method returns the utilization of the specified CPU by discounting the
6080 * utilization of the specified task, whenever the task is currently
6081 * contributing to the CPU utilization.
6083 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6085 struct cfs_rq *cfs_rq;
6088 /* Task has no contribution or is new */
6089 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6090 return cpu_util(cpu);
6092 cfs_rq = &cpu_rq(cpu)->cfs;
6093 util = READ_ONCE(cfs_rq->avg.util_avg);
6095 /* Discount task's util from CPU's util */
6096 lsub_positive(&util, task_util(p));
6101 * a) if *p is the only task sleeping on this CPU, then:
6102 * cpu_util (== task_util) > util_est (== 0)
6103 * and thus we return:
6104 * cpu_util_without = (cpu_util - task_util) = 0
6106 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6108 * cpu_util >= task_util
6109 * cpu_util > util_est (== 0)
6110 * and thus we discount *p's blocked utilization to return:
6111 * cpu_util_without = (cpu_util - task_util) >= 0
6113 * c) if other tasks are RUNNABLE on that CPU and
6114 * util_est > cpu_util
6115 * then we use util_est since it returns a more restrictive
6116 * estimation of the spare capacity on that CPU, by just
6117 * considering the expected utilization of tasks already
6118 * runnable on that CPU.
6120 * Cases a) and b) are covered by the above code, while case c) is
6121 * covered by the following code when estimated utilization is
6124 if (sched_feat(UTIL_EST)) {
6125 unsigned int estimated =
6126 READ_ONCE(cfs_rq->avg.util_est.enqueued);
6129 * Despite the following checks we still have a small window
6130 * for a possible race, when an execl's select_task_rq_fair()
6131 * races with LB's detach_task():
6134 * p->on_rq = TASK_ON_RQ_MIGRATING;
6135 * ---------------------------------- A
6136 * deactivate_task() \
6137 * dequeue_task() + RaceTime
6138 * util_est_dequeue() /
6139 * ---------------------------------- B
6141 * The additional check on "current == p" it's required to
6142 * properly fix the execl regression and it helps in further
6143 * reducing the chances for the above race.
6145 if (unlikely(task_on_rq_queued(p) || current == p))
6146 lsub_positive(&estimated, _task_util_est(p));
6148 util = max(util, estimated);
6152 * Utilization (estimated) can exceed the CPU capacity, thus let's
6153 * clamp to the maximum CPU capacity to ensure consistency with
6154 * the cpu_util call.
6156 return min_t(unsigned long, util, capacity_orig_of(cpu));
6160 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6161 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6163 * In that case WAKE_AFFINE doesn't make sense and we'll let
6164 * BALANCE_WAKE sort things out.
6166 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6168 long min_cap, max_cap;
6170 if (!static_branch_unlikely(&sched_asym_cpucapacity))
6173 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6174 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6176 /* Minimum capacity is close to max, no need to abort wake_affine */
6177 if (max_cap - min_cap < max_cap >> 3)
6180 /* Bring task utilization in sync with prev_cpu */
6181 sync_entity_load_avg(&p->se);
6183 return !task_fits_capacity(p, min_cap);
6187 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6190 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6192 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6193 unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
6196 * If @p migrates from @cpu to another, remove its contribution. Or,
6197 * if @p migrates from another CPU to @cpu, add its contribution. In
6198 * the other cases, @cpu is not impacted by the migration, so the
6199 * util_avg should already be correct.
6201 if (task_cpu(p) == cpu && dst_cpu != cpu)
6202 sub_positive(&util, task_util(p));
6203 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6204 util += task_util(p);
6206 if (sched_feat(UTIL_EST)) {
6207 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6210 * During wake-up, the task isn't enqueued yet and doesn't
6211 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6212 * so just add it (if needed) to "simulate" what will be
6213 * cpu_util() after the task has been enqueued.
6216 util_est += _task_util_est(p);
6218 util = max(util, util_est);
6221 return min(util, capacity_orig_of(cpu));
6225 * compute_energy(): Estimates the energy that would be consumed if @p was
6226 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6227 * landscape of the * CPUs after the task migration, and uses the Energy Model
6228 * to compute what would be the energy if we decided to actually migrate that
6232 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
6234 long util, max_util, sum_util, energy = 0;
6237 for (; pd; pd = pd->next) {
6238 max_util = sum_util = 0;
6240 * The capacity state of CPUs of the current rd can be driven by
6241 * CPUs of another rd if they belong to the same performance
6242 * domain. So, account for the utilization of these CPUs too
6243 * by masking pd with cpu_online_mask instead of the rd span.
6245 * If an entire performance domain is outside of the current rd,
6246 * it will not appear in its pd list and will not be accounted
6247 * by compute_energy().
6249 for_each_cpu_and(cpu, perf_domain_span(pd), cpu_online_mask) {
6250 util = cpu_util_next(cpu, p, dst_cpu);
6251 util = schedutil_energy_util(cpu, util);
6252 max_util = max(util, max_util);
6256 energy += em_pd_energy(pd->em_pd, max_util, sum_util);
6263 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6264 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6265 * spare capacity in each performance domain and uses it as a potential
6266 * candidate to execute the task. Then, it uses the Energy Model to figure
6267 * out which of the CPU candidates is the most energy-efficient.
6269 * The rationale for this heuristic is as follows. In a performance domain,
6270 * all the most energy efficient CPU candidates (according to the Energy
6271 * Model) are those for which we'll request a low frequency. When there are
6272 * several CPUs for which the frequency request will be the same, we don't
6273 * have enough data to break the tie between them, because the Energy Model
6274 * only includes active power costs. With this model, if we assume that
6275 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6276 * the maximum spare capacity in a performance domain is guaranteed to be among
6277 * the best candidates of the performance domain.
6279 * In practice, it could be preferable from an energy standpoint to pack
6280 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6281 * but that could also hurt our chances to go cluster idle, and we have no
6282 * ways to tell with the current Energy Model if this is actually a good
6283 * idea or not. So, find_energy_efficient_cpu() basically favors
6284 * cluster-packing, and spreading inside a cluster. That should at least be
6285 * a good thing for latency, and this is consistent with the idea that most
6286 * of the energy savings of EAS come from the asymmetry of the system, and
6287 * not so much from breaking the tie between identical CPUs. That's also the
6288 * reason why EAS is enabled in the topology code only for systems where
6289 * SD_ASYM_CPUCAPACITY is set.
6291 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6292 * they don't have any useful utilization data yet and it's not possible to
6293 * forecast their impact on energy consumption. Consequently, they will be
6294 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6295 * to be energy-inefficient in some use-cases. The alternative would be to
6296 * bias new tasks towards specific types of CPUs first, or to try to infer
6297 * their util_avg from the parent task, but those heuristics could hurt
6298 * other use-cases too. So, until someone finds a better way to solve this,
6299 * let's keep things simple by re-using the existing slow path.
6302 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6304 unsigned long prev_energy = ULONG_MAX, best_energy = ULONG_MAX;
6305 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6306 int cpu, best_energy_cpu = prev_cpu;
6307 struct perf_domain *head, *pd;
6308 unsigned long cpu_cap, util;
6309 struct sched_domain *sd;
6312 pd = rcu_dereference(rd->pd);
6313 if (!pd || READ_ONCE(rd->overutilized))
6318 * Energy-aware wake-up happens on the lowest sched_domain starting
6319 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6321 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6322 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6327 sync_entity_load_avg(&p->se);
6328 if (!task_util_est(p))
6331 for (; pd; pd = pd->next) {
6332 unsigned long cur_energy, spare_cap, max_spare_cap = 0;
6333 int max_spare_cap_cpu = -1;
6335 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
6336 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6339 /* Skip CPUs that will be overutilized. */
6340 util = cpu_util_next(cpu, p, cpu);
6341 cpu_cap = capacity_of(cpu);
6342 if (cpu_cap * 1024 < util * capacity_margin)
6345 /* Always use prev_cpu as a candidate. */
6346 if (cpu == prev_cpu) {
6347 prev_energy = compute_energy(p, prev_cpu, head);
6348 best_energy = min(best_energy, prev_energy);
6353 * Find the CPU with the maximum spare capacity in
6354 * the performance domain
6356 spare_cap = cpu_cap - util;
6357 if (spare_cap > max_spare_cap) {
6358 max_spare_cap = spare_cap;
6359 max_spare_cap_cpu = cpu;
6363 /* Evaluate the energy impact of using this CPU. */
6364 if (max_spare_cap_cpu >= 0) {
6365 cur_energy = compute_energy(p, max_spare_cap_cpu, head);
6366 if (cur_energy < best_energy) {
6367 best_energy = cur_energy;
6368 best_energy_cpu = max_spare_cap_cpu;
6376 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6377 * least 6% of the energy used by prev_cpu.
6379 if (prev_energy == ULONG_MAX)
6380 return best_energy_cpu;
6382 if ((prev_energy - best_energy) > (prev_energy >> 4))
6383 return best_energy_cpu;
6394 * select_task_rq_fair: Select target runqueue for the waking task in domains
6395 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6396 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6398 * Balances load by selecting the idlest CPU in the idlest group, or under
6399 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6401 * Returns the target CPU number.
6403 * preempt must be disabled.
6406 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6408 struct sched_domain *tmp, *sd = NULL;
6409 int cpu = smp_processor_id();
6410 int new_cpu = prev_cpu;
6411 int want_affine = 0;
6412 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6414 if (sd_flag & SD_BALANCE_WAKE) {
6417 if (sched_energy_enabled()) {
6418 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
6424 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu) &&
6425 cpumask_test_cpu(cpu, p->cpus_ptr);
6429 for_each_domain(cpu, tmp) {
6430 if (!(tmp->flags & SD_LOAD_BALANCE))
6434 * If both 'cpu' and 'prev_cpu' are part of this domain,
6435 * cpu is a valid SD_WAKE_AFFINE target.
6437 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6438 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6439 if (cpu != prev_cpu)
6440 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6442 sd = NULL; /* Prefer wake_affine over balance flags */
6446 if (tmp->flags & sd_flag)
6448 else if (!want_affine)
6454 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6455 } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6458 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6461 current->recent_used_cpu = cpu;
6468 static void detach_entity_cfs_rq(struct sched_entity *se);
6471 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6472 * cfs_rq_of(p) references at time of call are still valid and identify the
6473 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6475 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6478 * As blocked tasks retain absolute vruntime the migration needs to
6479 * deal with this by subtracting the old and adding the new
6480 * min_vruntime -- the latter is done by enqueue_entity() when placing
6481 * the task on the new runqueue.
6483 if (p->state == TASK_WAKING) {
6484 struct sched_entity *se = &p->se;
6485 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6488 #ifndef CONFIG_64BIT
6489 u64 min_vruntime_copy;
6492 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6494 min_vruntime = cfs_rq->min_vruntime;
6495 } while (min_vruntime != min_vruntime_copy);
6497 min_vruntime = cfs_rq->min_vruntime;
6500 se->vruntime -= min_vruntime;
6503 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6505 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6506 * rq->lock and can modify state directly.
6508 lockdep_assert_held(&task_rq(p)->lock);
6509 detach_entity_cfs_rq(&p->se);
6513 * We are supposed to update the task to "current" time, then
6514 * its up to date and ready to go to new CPU/cfs_rq. But we
6515 * have difficulty in getting what current time is, so simply
6516 * throw away the out-of-date time. This will result in the
6517 * wakee task is less decayed, but giving the wakee more load
6520 remove_entity_load_avg(&p->se);
6523 /* Tell new CPU we are migrated */
6524 p->se.avg.last_update_time = 0;
6526 /* We have migrated, no longer consider this task hot */
6527 p->se.exec_start = 0;
6529 update_scan_period(p, new_cpu);
6532 static void task_dead_fair(struct task_struct *p)
6534 remove_entity_load_avg(&p->se);
6536 #endif /* CONFIG_SMP */
6538 static unsigned long wakeup_gran(struct sched_entity *se)
6540 unsigned long gran = sysctl_sched_wakeup_granularity;
6543 * Since its curr running now, convert the gran from real-time
6544 * to virtual-time in his units.
6546 * By using 'se' instead of 'curr' we penalize light tasks, so
6547 * they get preempted easier. That is, if 'se' < 'curr' then
6548 * the resulting gran will be larger, therefore penalizing the
6549 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6550 * be smaller, again penalizing the lighter task.
6552 * This is especially important for buddies when the leftmost
6553 * task is higher priority than the buddy.
6555 return calc_delta_fair(gran, se);
6559 * Should 'se' preempt 'curr'.
6573 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6575 s64 gran, vdiff = curr->vruntime - se->vruntime;
6580 gran = wakeup_gran(se);
6587 static void set_last_buddy(struct sched_entity *se)
6589 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6592 for_each_sched_entity(se) {
6593 if (SCHED_WARN_ON(!se->on_rq))
6595 cfs_rq_of(se)->last = se;
6599 static void set_next_buddy(struct sched_entity *se)
6601 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6604 for_each_sched_entity(se) {
6605 if (SCHED_WARN_ON(!se->on_rq))
6607 cfs_rq_of(se)->next = se;
6611 static void set_skip_buddy(struct sched_entity *se)
6613 for_each_sched_entity(se)
6614 cfs_rq_of(se)->skip = se;
6618 * Preempt the current task with a newly woken task if needed:
6620 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6622 struct task_struct *curr = rq->curr;
6623 struct sched_entity *se = &curr->se, *pse = &p->se;
6624 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6625 int scale = cfs_rq->nr_running >= sched_nr_latency;
6626 int next_buddy_marked = 0;
6628 if (unlikely(se == pse))
6632 * This is possible from callers such as attach_tasks(), in which we
6633 * unconditionally check_prempt_curr() after an enqueue (which may have
6634 * lead to a throttle). This both saves work and prevents false
6635 * next-buddy nomination below.
6637 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6640 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6641 set_next_buddy(pse);
6642 next_buddy_marked = 1;
6646 * We can come here with TIF_NEED_RESCHED already set from new task
6649 * Note: this also catches the edge-case of curr being in a throttled
6650 * group (e.g. via set_curr_task), since update_curr() (in the
6651 * enqueue of curr) will have resulted in resched being set. This
6652 * prevents us from potentially nominating it as a false LAST_BUDDY
6655 if (test_tsk_need_resched(curr))
6658 /* Idle tasks are by definition preempted by non-idle tasks. */
6659 if (unlikely(task_has_idle_policy(curr)) &&
6660 likely(!task_has_idle_policy(p)))
6664 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6665 * is driven by the tick):
6667 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6670 find_matching_se(&se, &pse);
6671 update_curr(cfs_rq_of(se));
6673 if (wakeup_preempt_entity(se, pse) == 1) {
6675 * Bias pick_next to pick the sched entity that is
6676 * triggering this preemption.
6678 if (!next_buddy_marked)
6679 set_next_buddy(pse);
6688 * Only set the backward buddy when the current task is still
6689 * on the rq. This can happen when a wakeup gets interleaved
6690 * with schedule on the ->pre_schedule() or idle_balance()
6691 * point, either of which can * drop the rq lock.
6693 * Also, during early boot the idle thread is in the fair class,
6694 * for obvious reasons its a bad idea to schedule back to it.
6696 if (unlikely(!se->on_rq || curr == rq->idle))
6699 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6703 static struct task_struct *
6704 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6706 struct cfs_rq *cfs_rq = &rq->cfs;
6707 struct sched_entity *se;
6708 struct task_struct *p;
6712 if (!cfs_rq->nr_running)
6715 #ifdef CONFIG_FAIR_GROUP_SCHED
6716 if (prev->sched_class != &fair_sched_class)
6720 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6721 * likely that a next task is from the same cgroup as the current.
6723 * Therefore attempt to avoid putting and setting the entire cgroup
6724 * hierarchy, only change the part that actually changes.
6728 struct sched_entity *curr = cfs_rq->curr;
6731 * Since we got here without doing put_prev_entity() we also
6732 * have to consider cfs_rq->curr. If it is still a runnable
6733 * entity, update_curr() will update its vruntime, otherwise
6734 * forget we've ever seen it.
6738 update_curr(cfs_rq);
6743 * This call to check_cfs_rq_runtime() will do the
6744 * throttle and dequeue its entity in the parent(s).
6745 * Therefore the nr_running test will indeed
6748 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6751 if (!cfs_rq->nr_running)
6758 se = pick_next_entity(cfs_rq, curr);
6759 cfs_rq = group_cfs_rq(se);
6765 * Since we haven't yet done put_prev_entity and if the selected task
6766 * is a different task than we started out with, try and touch the
6767 * least amount of cfs_rqs.
6770 struct sched_entity *pse = &prev->se;
6772 while (!(cfs_rq = is_same_group(se, pse))) {
6773 int se_depth = se->depth;
6774 int pse_depth = pse->depth;
6776 if (se_depth <= pse_depth) {
6777 put_prev_entity(cfs_rq_of(pse), pse);
6778 pse = parent_entity(pse);
6780 if (se_depth >= pse_depth) {
6781 set_next_entity(cfs_rq_of(se), se);
6782 se = parent_entity(se);
6786 put_prev_entity(cfs_rq, pse);
6787 set_next_entity(cfs_rq, se);
6794 put_prev_task(rq, prev);
6797 se = pick_next_entity(cfs_rq, NULL);
6798 set_next_entity(cfs_rq, se);
6799 cfs_rq = group_cfs_rq(se);
6804 done: __maybe_unused;
6807 * Move the next running task to the front of
6808 * the list, so our cfs_tasks list becomes MRU
6811 list_move(&p->se.group_node, &rq->cfs_tasks);
6814 if (hrtick_enabled(rq))
6815 hrtick_start_fair(rq, p);
6817 update_misfit_status(p, rq);
6822 update_misfit_status(NULL, rq);
6823 new_tasks = idle_balance(rq, rf);
6826 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6827 * possible for any higher priority task to appear. In that case we
6828 * must re-start the pick_next_entity() loop.
6837 * rq is about to be idle, check if we need to update the
6838 * lost_idle_time of clock_pelt
6840 update_idle_rq_clock_pelt(rq);
6846 * Account for a descheduled task:
6848 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6850 struct sched_entity *se = &prev->se;
6851 struct cfs_rq *cfs_rq;
6853 for_each_sched_entity(se) {
6854 cfs_rq = cfs_rq_of(se);
6855 put_prev_entity(cfs_rq, se);
6860 * sched_yield() is very simple
6862 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6864 static void yield_task_fair(struct rq *rq)
6866 struct task_struct *curr = rq->curr;
6867 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6868 struct sched_entity *se = &curr->se;
6871 * Are we the only task in the tree?
6873 if (unlikely(rq->nr_running == 1))
6876 clear_buddies(cfs_rq, se);
6878 if (curr->policy != SCHED_BATCH) {
6879 update_rq_clock(rq);
6881 * Update run-time statistics of the 'current'.
6883 update_curr(cfs_rq);
6885 * Tell update_rq_clock() that we've just updated,
6886 * so we don't do microscopic update in schedule()
6887 * and double the fastpath cost.
6889 rq_clock_skip_update(rq);
6895 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6897 struct sched_entity *se = &p->se;
6899 /* throttled hierarchies are not runnable */
6900 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6903 /* Tell the scheduler that we'd really like pse to run next. */
6906 yield_task_fair(rq);
6912 /**************************************************
6913 * Fair scheduling class load-balancing methods.
6917 * The purpose of load-balancing is to achieve the same basic fairness the
6918 * per-CPU scheduler provides, namely provide a proportional amount of compute
6919 * time to each task. This is expressed in the following equation:
6921 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6923 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
6924 * W_i,0 is defined as:
6926 * W_i,0 = \Sum_j w_i,j (2)
6928 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
6929 * is derived from the nice value as per sched_prio_to_weight[].
6931 * The weight average is an exponential decay average of the instantaneous
6934 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6936 * C_i is the compute capacity of CPU i, typically it is the
6937 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6938 * can also include other factors [XXX].
6940 * To achieve this balance we define a measure of imbalance which follows
6941 * directly from (1):
6943 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6945 * We them move tasks around to minimize the imbalance. In the continuous
6946 * function space it is obvious this converges, in the discrete case we get
6947 * a few fun cases generally called infeasible weight scenarios.
6950 * - infeasible weights;
6951 * - local vs global optima in the discrete case. ]
6956 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6957 * for all i,j solution, we create a tree of CPUs that follows the hardware
6958 * topology where each level pairs two lower groups (or better). This results
6959 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
6960 * tree to only the first of the previous level and we decrease the frequency
6961 * of load-balance at each level inv. proportional to the number of CPUs in
6967 * \Sum { --- * --- * 2^i } = O(n) (5)
6969 * `- size of each group
6970 * | | `- number of CPUs doing load-balance
6972 * `- sum over all levels
6974 * Coupled with a limit on how many tasks we can migrate every balance pass,
6975 * this makes (5) the runtime complexity of the balancer.
6977 * An important property here is that each CPU is still (indirectly) connected
6978 * to every other CPU in at most O(log n) steps:
6980 * The adjacency matrix of the resulting graph is given by:
6983 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6986 * And you'll find that:
6988 * A^(log_2 n)_i,j != 0 for all i,j (7)
6990 * Showing there's indeed a path between every CPU in at most O(log n) steps.
6991 * The task movement gives a factor of O(m), giving a convergence complexity
6994 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6999 * In order to avoid CPUs going idle while there's still work to do, new idle
7000 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7001 * tree itself instead of relying on other CPUs to bring it work.
7003 * This adds some complexity to both (5) and (8) but it reduces the total idle
7011 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7014 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7019 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7021 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7023 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7026 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7027 * rewrite all of this once again.]
7030 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7032 enum fbq_type { regular, remote, all };
7041 #define LBF_ALL_PINNED 0x01
7042 #define LBF_NEED_BREAK 0x02
7043 #define LBF_DST_PINNED 0x04
7044 #define LBF_SOME_PINNED 0x08
7045 #define LBF_NOHZ_STATS 0x10
7046 #define LBF_NOHZ_AGAIN 0x20
7049 struct sched_domain *sd;
7057 struct cpumask *dst_grpmask;
7059 enum cpu_idle_type idle;
7061 /* The set of CPUs under consideration for load-balancing */
7062 struct cpumask *cpus;
7067 unsigned int loop_break;
7068 unsigned int loop_max;
7070 enum fbq_type fbq_type;
7071 enum group_type src_grp_type;
7072 struct list_head tasks;
7076 * Is this task likely cache-hot:
7078 static int task_hot(struct task_struct *p, struct lb_env *env)
7082 lockdep_assert_held(&env->src_rq->lock);
7084 if (p->sched_class != &fair_sched_class)
7087 if (unlikely(task_has_idle_policy(p)))
7091 * Buddy candidates are cache hot:
7093 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7094 (&p->se == cfs_rq_of(&p->se)->next ||
7095 &p->se == cfs_rq_of(&p->se)->last))
7098 if (sysctl_sched_migration_cost == -1)
7100 if (sysctl_sched_migration_cost == 0)
7103 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7105 return delta < (s64)sysctl_sched_migration_cost;
7108 #ifdef CONFIG_NUMA_BALANCING
7110 * Returns 1, if task migration degrades locality
7111 * Returns 0, if task migration improves locality i.e migration preferred.
7112 * Returns -1, if task migration is not affected by locality.
7114 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7116 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7117 unsigned long src_weight, dst_weight;
7118 int src_nid, dst_nid, dist;
7120 if (!static_branch_likely(&sched_numa_balancing))
7123 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7126 src_nid = cpu_to_node(env->src_cpu);
7127 dst_nid = cpu_to_node(env->dst_cpu);
7129 if (src_nid == dst_nid)
7132 /* Migrating away from the preferred node is always bad. */
7133 if (src_nid == p->numa_preferred_nid) {
7134 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7140 /* Encourage migration to the preferred node. */
7141 if (dst_nid == p->numa_preferred_nid)
7144 /* Leaving a core idle is often worse than degrading locality. */
7145 if (env->idle == CPU_IDLE)
7148 dist = node_distance(src_nid, dst_nid);
7150 src_weight = group_weight(p, src_nid, dist);
7151 dst_weight = group_weight(p, dst_nid, dist);
7153 src_weight = task_weight(p, src_nid, dist);
7154 dst_weight = task_weight(p, dst_nid, dist);
7157 return dst_weight < src_weight;
7161 static inline int migrate_degrades_locality(struct task_struct *p,
7169 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7172 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7176 lockdep_assert_held(&env->src_rq->lock);
7179 * We do not migrate tasks that are:
7180 * 1) throttled_lb_pair, or
7181 * 2) cannot be migrated to this CPU due to cpus_ptr, or
7182 * 3) running (obviously), or
7183 * 4) are cache-hot on their current CPU.
7185 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7188 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
7191 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7193 env->flags |= LBF_SOME_PINNED;
7196 * Remember if this task can be migrated to any other CPU in
7197 * our sched_group. We may want to revisit it if we couldn't
7198 * meet load balance goals by pulling other tasks on src_cpu.
7200 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7201 * already computed one in current iteration.
7203 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7206 /* Prevent to re-select dst_cpu via env's CPUs: */
7207 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7208 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
7209 env->flags |= LBF_DST_PINNED;
7210 env->new_dst_cpu = cpu;
7218 /* Record that we found atleast one task that could run on dst_cpu */
7219 env->flags &= ~LBF_ALL_PINNED;
7221 if (task_running(env->src_rq, p)) {
7222 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7227 * Aggressive migration if:
7228 * 1) destination numa is preferred
7229 * 2) task is cache cold, or
7230 * 3) too many balance attempts have failed.
7232 tsk_cache_hot = migrate_degrades_locality(p, env);
7233 if (tsk_cache_hot == -1)
7234 tsk_cache_hot = task_hot(p, env);
7236 if (tsk_cache_hot <= 0 ||
7237 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7238 if (tsk_cache_hot == 1) {
7239 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7240 schedstat_inc(p->se.statistics.nr_forced_migrations);
7245 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7250 * detach_task() -- detach the task for the migration specified in env
7252 static void detach_task(struct task_struct *p, struct lb_env *env)
7254 lockdep_assert_held(&env->src_rq->lock);
7256 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7257 set_task_cpu(p, env->dst_cpu);
7261 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7262 * part of active balancing operations within "domain".
7264 * Returns a task if successful and NULL otherwise.
7266 static struct task_struct *detach_one_task(struct lb_env *env)
7268 struct task_struct *p;
7270 lockdep_assert_held(&env->src_rq->lock);
7272 list_for_each_entry_reverse(p,
7273 &env->src_rq->cfs_tasks, se.group_node) {
7274 if (!can_migrate_task(p, env))
7277 detach_task(p, env);
7280 * Right now, this is only the second place where
7281 * lb_gained[env->idle] is updated (other is detach_tasks)
7282 * so we can safely collect stats here rather than
7283 * inside detach_tasks().
7285 schedstat_inc(env->sd->lb_gained[env->idle]);
7291 static const unsigned int sched_nr_migrate_break = 32;
7294 * detach_tasks() -- tries to detach up to imbalance weighted load from
7295 * busiest_rq, as part of a balancing operation within domain "sd".
7297 * Returns number of detached tasks if successful and 0 otherwise.
7299 static int detach_tasks(struct lb_env *env)
7301 struct list_head *tasks = &env->src_rq->cfs_tasks;
7302 struct task_struct *p;
7306 lockdep_assert_held(&env->src_rq->lock);
7308 if (env->imbalance <= 0)
7311 while (!list_empty(tasks)) {
7313 * We don't want to steal all, otherwise we may be treated likewise,
7314 * which could at worst lead to a livelock crash.
7316 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7319 p = list_last_entry(tasks, struct task_struct, se.group_node);
7322 /* We've more or less seen every task there is, call it quits */
7323 if (env->loop > env->loop_max)
7326 /* take a breather every nr_migrate tasks */
7327 if (env->loop > env->loop_break) {
7328 env->loop_break += sched_nr_migrate_break;
7329 env->flags |= LBF_NEED_BREAK;
7333 if (!can_migrate_task(p, env))
7336 load = task_h_load(p);
7338 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7341 if ((load / 2) > env->imbalance)
7344 detach_task(p, env);
7345 list_add(&p->se.group_node, &env->tasks);
7348 env->imbalance -= load;
7350 #ifdef CONFIG_PREEMPT
7352 * NEWIDLE balancing is a source of latency, so preemptible
7353 * kernels will stop after the first task is detached to minimize
7354 * the critical section.
7356 if (env->idle == CPU_NEWLY_IDLE)
7361 * We only want to steal up to the prescribed amount of
7364 if (env->imbalance <= 0)
7369 list_move(&p->se.group_node, tasks);
7373 * Right now, this is one of only two places we collect this stat
7374 * so we can safely collect detach_one_task() stats here rather
7375 * than inside detach_one_task().
7377 schedstat_add(env->sd->lb_gained[env->idle], detached);
7383 * attach_task() -- attach the task detached by detach_task() to its new rq.
7385 static void attach_task(struct rq *rq, struct task_struct *p)
7387 lockdep_assert_held(&rq->lock);
7389 BUG_ON(task_rq(p) != rq);
7390 activate_task(rq, p, ENQUEUE_NOCLOCK);
7391 check_preempt_curr(rq, p, 0);
7395 * attach_one_task() -- attaches the task returned from detach_one_task() to
7398 static void attach_one_task(struct rq *rq, struct task_struct *p)
7403 update_rq_clock(rq);
7409 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7412 static void attach_tasks(struct lb_env *env)
7414 struct list_head *tasks = &env->tasks;
7415 struct task_struct *p;
7418 rq_lock(env->dst_rq, &rf);
7419 update_rq_clock(env->dst_rq);
7421 while (!list_empty(tasks)) {
7422 p = list_first_entry(tasks, struct task_struct, se.group_node);
7423 list_del_init(&p->se.group_node);
7425 attach_task(env->dst_rq, p);
7428 rq_unlock(env->dst_rq, &rf);
7431 #ifdef CONFIG_NO_HZ_COMMON
7432 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7434 if (cfs_rq->avg.load_avg)
7437 if (cfs_rq->avg.util_avg)
7443 static inline bool others_have_blocked(struct rq *rq)
7445 if (READ_ONCE(rq->avg_rt.util_avg))
7448 if (READ_ONCE(rq->avg_dl.util_avg))
7451 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7452 if (READ_ONCE(rq->avg_irq.util_avg))
7459 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
7461 rq->last_blocked_load_update_tick = jiffies;
7464 rq->has_blocked_load = 0;
7467 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
7468 static inline bool others_have_blocked(struct rq *rq) { return false; }
7469 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
7472 #ifdef CONFIG_FAIR_GROUP_SCHED
7474 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7476 if (cfs_rq->load.weight)
7479 if (cfs_rq->avg.load_sum)
7482 if (cfs_rq->avg.util_sum)
7485 if (cfs_rq->avg.runnable_load_sum)
7491 static void update_blocked_averages(int cpu)
7493 struct rq *rq = cpu_rq(cpu);
7494 struct cfs_rq *cfs_rq, *pos;
7495 const struct sched_class *curr_class;
7499 rq_lock_irqsave(rq, &rf);
7500 update_rq_clock(rq);
7503 * Iterates the task_group tree in a bottom up fashion, see
7504 * list_add_leaf_cfs_rq() for details.
7506 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7507 struct sched_entity *se;
7509 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq))
7510 update_tg_load_avg(cfs_rq, 0);
7512 /* Propagate pending load changes to the parent, if any: */
7513 se = cfs_rq->tg->se[cpu];
7514 if (se && !skip_blocked_update(se))
7515 update_load_avg(cfs_rq_of(se), se, 0);
7518 * There can be a lot of idle CPU cgroups. Don't let fully
7519 * decayed cfs_rqs linger on the list.
7521 if (cfs_rq_is_decayed(cfs_rq))
7522 list_del_leaf_cfs_rq(cfs_rq);
7524 /* Don't need periodic decay once load/util_avg are null */
7525 if (cfs_rq_has_blocked(cfs_rq))
7529 curr_class = rq->curr->sched_class;
7530 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class);
7531 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class);
7532 update_irq_load_avg(rq, 0);
7533 /* Don't need periodic decay once load/util_avg are null */
7534 if (others_have_blocked(rq))
7537 update_blocked_load_status(rq, !done);
7538 rq_unlock_irqrestore(rq, &rf);
7542 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7543 * This needs to be done in a top-down fashion because the load of a child
7544 * group is a fraction of its parents load.
7546 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7548 struct rq *rq = rq_of(cfs_rq);
7549 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7550 unsigned long now = jiffies;
7553 if (cfs_rq->last_h_load_update == now)
7556 WRITE_ONCE(cfs_rq->h_load_next, NULL);
7557 for_each_sched_entity(se) {
7558 cfs_rq = cfs_rq_of(se);
7559 WRITE_ONCE(cfs_rq->h_load_next, se);
7560 if (cfs_rq->last_h_load_update == now)
7565 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7566 cfs_rq->last_h_load_update = now;
7569 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
7570 load = cfs_rq->h_load;
7571 load = div64_ul(load * se->avg.load_avg,
7572 cfs_rq_load_avg(cfs_rq) + 1);
7573 cfs_rq = group_cfs_rq(se);
7574 cfs_rq->h_load = load;
7575 cfs_rq->last_h_load_update = now;
7579 static unsigned long task_h_load(struct task_struct *p)
7581 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7583 update_cfs_rq_h_load(cfs_rq);
7584 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7585 cfs_rq_load_avg(cfs_rq) + 1);
7588 static inline void update_blocked_averages(int cpu)
7590 struct rq *rq = cpu_rq(cpu);
7591 struct cfs_rq *cfs_rq = &rq->cfs;
7592 const struct sched_class *curr_class;
7595 rq_lock_irqsave(rq, &rf);
7596 update_rq_clock(rq);
7597 update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
7599 curr_class = rq->curr->sched_class;
7600 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class);
7601 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class);
7602 update_irq_load_avg(rq, 0);
7603 update_blocked_load_status(rq, cfs_rq_has_blocked(cfs_rq) || others_have_blocked(rq));
7604 rq_unlock_irqrestore(rq, &rf);
7607 static unsigned long task_h_load(struct task_struct *p)
7609 return p->se.avg.load_avg;
7613 /********** Helpers for find_busiest_group ************************/
7616 * sg_lb_stats - stats of a sched_group required for load_balancing
7618 struct sg_lb_stats {
7619 unsigned long avg_load; /*Avg load across the CPUs of the group */
7620 unsigned long group_load; /* Total load over the CPUs of the group */
7621 unsigned long load_per_task;
7622 unsigned long group_capacity;
7623 unsigned long group_util; /* Total utilization of the group */
7624 unsigned int sum_nr_running; /* Nr tasks running in the group */
7625 unsigned int idle_cpus;
7626 unsigned int group_weight;
7627 enum group_type group_type;
7628 int group_no_capacity;
7629 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
7630 #ifdef CONFIG_NUMA_BALANCING
7631 unsigned int nr_numa_running;
7632 unsigned int nr_preferred_running;
7637 * sd_lb_stats - Structure to store the statistics of a sched_domain
7638 * during load balancing.
7640 struct sd_lb_stats {
7641 struct sched_group *busiest; /* Busiest group in this sd */
7642 struct sched_group *local; /* Local group in this sd */
7643 unsigned long total_running;
7644 unsigned long total_load; /* Total load of all groups in sd */
7645 unsigned long total_capacity; /* Total capacity of all groups in sd */
7646 unsigned long avg_load; /* Average load across all groups in sd */
7648 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7649 struct sg_lb_stats local_stat; /* Statistics of the local group */
7652 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7655 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7656 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7657 * We must however clear busiest_stat::avg_load because
7658 * update_sd_pick_busiest() reads this before assignment.
7660 *sds = (struct sd_lb_stats){
7663 .total_running = 0UL,
7665 .total_capacity = 0UL,
7668 .sum_nr_running = 0,
7669 .group_type = group_other,
7674 static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu)
7676 struct rq *rq = cpu_rq(cpu);
7677 unsigned long max = arch_scale_cpu_capacity(cpu);
7678 unsigned long used, free;
7681 irq = cpu_util_irq(rq);
7683 if (unlikely(irq >= max))
7686 used = READ_ONCE(rq->avg_rt.util_avg);
7687 used += READ_ONCE(rq->avg_dl.util_avg);
7689 if (unlikely(used >= max))
7694 return scale_irq_capacity(free, irq, max);
7697 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7699 unsigned long capacity = scale_rt_capacity(sd, cpu);
7700 struct sched_group *sdg = sd->groups;
7702 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
7707 cpu_rq(cpu)->cpu_capacity = capacity;
7708 sdg->sgc->capacity = capacity;
7709 sdg->sgc->min_capacity = capacity;
7710 sdg->sgc->max_capacity = capacity;
7713 void update_group_capacity(struct sched_domain *sd, int cpu)
7715 struct sched_domain *child = sd->child;
7716 struct sched_group *group, *sdg = sd->groups;
7717 unsigned long capacity, min_capacity, max_capacity;
7718 unsigned long interval;
7720 interval = msecs_to_jiffies(sd->balance_interval);
7721 interval = clamp(interval, 1UL, max_load_balance_interval);
7722 sdg->sgc->next_update = jiffies + interval;
7725 update_cpu_capacity(sd, cpu);
7730 min_capacity = ULONG_MAX;
7733 if (child->flags & SD_OVERLAP) {
7735 * SD_OVERLAP domains cannot assume that child groups
7736 * span the current group.
7739 for_each_cpu(cpu, sched_group_span(sdg)) {
7740 struct sched_group_capacity *sgc;
7741 struct rq *rq = cpu_rq(cpu);
7744 * build_sched_domains() -> init_sched_groups_capacity()
7745 * gets here before we've attached the domains to the
7748 * Use capacity_of(), which is set irrespective of domains
7749 * in update_cpu_capacity().
7751 * This avoids capacity from being 0 and
7752 * causing divide-by-zero issues on boot.
7754 if (unlikely(!rq->sd)) {
7755 capacity += capacity_of(cpu);
7757 sgc = rq->sd->groups->sgc;
7758 capacity += sgc->capacity;
7761 min_capacity = min(capacity, min_capacity);
7762 max_capacity = max(capacity, max_capacity);
7766 * !SD_OVERLAP domains can assume that child groups
7767 * span the current group.
7770 group = child->groups;
7772 struct sched_group_capacity *sgc = group->sgc;
7774 capacity += sgc->capacity;
7775 min_capacity = min(sgc->min_capacity, min_capacity);
7776 max_capacity = max(sgc->max_capacity, max_capacity);
7777 group = group->next;
7778 } while (group != child->groups);
7781 sdg->sgc->capacity = capacity;
7782 sdg->sgc->min_capacity = min_capacity;
7783 sdg->sgc->max_capacity = max_capacity;
7787 * Check whether the capacity of the rq has been noticeably reduced by side
7788 * activity. The imbalance_pct is used for the threshold.
7789 * Return true is the capacity is reduced
7792 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7794 return ((rq->cpu_capacity * sd->imbalance_pct) <
7795 (rq->cpu_capacity_orig * 100));
7799 * Check whether a rq has a misfit task and if it looks like we can actually
7800 * help that task: we can migrate the task to a CPU of higher capacity, or
7801 * the task's current CPU is heavily pressured.
7803 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
7805 return rq->misfit_task_load &&
7806 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
7807 check_cpu_capacity(rq, sd));
7811 * Group imbalance indicates (and tries to solve) the problem where balancing
7812 * groups is inadequate due to ->cpus_ptr constraints.
7814 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
7815 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
7818 * { 0 1 2 3 } { 4 5 6 7 }
7821 * If we were to balance group-wise we'd place two tasks in the first group and
7822 * two tasks in the second group. Clearly this is undesired as it will overload
7823 * cpu 3 and leave one of the CPUs in the second group unused.
7825 * The current solution to this issue is detecting the skew in the first group
7826 * by noticing the lower domain failed to reach balance and had difficulty
7827 * moving tasks due to affinity constraints.
7829 * When this is so detected; this group becomes a candidate for busiest; see
7830 * update_sd_pick_busiest(). And calculate_imbalance() and
7831 * find_busiest_group() avoid some of the usual balance conditions to allow it
7832 * to create an effective group imbalance.
7834 * This is a somewhat tricky proposition since the next run might not find the
7835 * group imbalance and decide the groups need to be balanced again. A most
7836 * subtle and fragile situation.
7839 static inline int sg_imbalanced(struct sched_group *group)
7841 return group->sgc->imbalance;
7845 * group_has_capacity returns true if the group has spare capacity that could
7846 * be used by some tasks.
7847 * We consider that a group has spare capacity if the * number of task is
7848 * smaller than the number of CPUs or if the utilization is lower than the
7849 * available capacity for CFS tasks.
7850 * For the latter, we use a threshold to stabilize the state, to take into
7851 * account the variance of the tasks' load and to return true if the available
7852 * capacity in meaningful for the load balancer.
7853 * As an example, an available capacity of 1% can appear but it doesn't make
7854 * any benefit for the load balance.
7857 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7859 if (sgs->sum_nr_running < sgs->group_weight)
7862 if ((sgs->group_capacity * 100) >
7863 (sgs->group_util * env->sd->imbalance_pct))
7870 * group_is_overloaded returns true if the group has more tasks than it can
7872 * group_is_overloaded is not equals to !group_has_capacity because a group
7873 * with the exact right number of tasks, has no more spare capacity but is not
7874 * overloaded so both group_has_capacity and group_is_overloaded return
7878 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7880 if (sgs->sum_nr_running <= sgs->group_weight)
7883 if ((sgs->group_capacity * 100) <
7884 (sgs->group_util * env->sd->imbalance_pct))
7891 * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller
7892 * per-CPU capacity than sched_group ref.
7895 group_smaller_min_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7897 return sg->sgc->min_capacity * capacity_margin <
7898 ref->sgc->min_capacity * 1024;
7902 * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller
7903 * per-CPU capacity_orig than sched_group ref.
7906 group_smaller_max_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7908 return sg->sgc->max_capacity * capacity_margin <
7909 ref->sgc->max_capacity * 1024;
7913 group_type group_classify(struct sched_group *group,
7914 struct sg_lb_stats *sgs)
7916 if (sgs->group_no_capacity)
7917 return group_overloaded;
7919 if (sg_imbalanced(group))
7920 return group_imbalanced;
7922 if (sgs->group_misfit_task_load)
7923 return group_misfit_task;
7928 static bool update_nohz_stats(struct rq *rq, bool force)
7930 #ifdef CONFIG_NO_HZ_COMMON
7931 unsigned int cpu = rq->cpu;
7933 if (!rq->has_blocked_load)
7936 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
7939 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
7942 update_blocked_averages(cpu);
7944 return rq->has_blocked_load;
7951 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7952 * @env: The load balancing environment.
7953 * @group: sched_group whose statistics are to be updated.
7954 * @sgs: variable to hold the statistics for this group.
7955 * @sg_status: Holds flag indicating the status of the sched_group
7957 static inline void update_sg_lb_stats(struct lb_env *env,
7958 struct sched_group *group,
7959 struct sg_lb_stats *sgs,
7964 memset(sgs, 0, sizeof(*sgs));
7966 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
7967 struct rq *rq = cpu_rq(i);
7969 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
7970 env->flags |= LBF_NOHZ_AGAIN;
7972 sgs->group_load += weighted_cpuload(rq);
7973 sgs->group_util += cpu_util(i);
7974 sgs->sum_nr_running += rq->cfs.h_nr_running;
7976 nr_running = rq->nr_running;
7978 *sg_status |= SG_OVERLOAD;
7980 if (cpu_overutilized(i))
7981 *sg_status |= SG_OVERUTILIZED;
7983 #ifdef CONFIG_NUMA_BALANCING
7984 sgs->nr_numa_running += rq->nr_numa_running;
7985 sgs->nr_preferred_running += rq->nr_preferred_running;
7988 * No need to call idle_cpu() if nr_running is not 0
7990 if (!nr_running && idle_cpu(i))
7993 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
7994 sgs->group_misfit_task_load < rq->misfit_task_load) {
7995 sgs->group_misfit_task_load = rq->misfit_task_load;
7996 *sg_status |= SG_OVERLOAD;
8000 /* Adjust by relative CPU capacity of the group */
8001 sgs->group_capacity = group->sgc->capacity;
8002 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
8004 if (sgs->sum_nr_running)
8005 sgs->load_per_task = sgs->group_load / sgs->sum_nr_running;
8007 sgs->group_weight = group->group_weight;
8009 sgs->group_no_capacity = group_is_overloaded(env, sgs);
8010 sgs->group_type = group_classify(group, sgs);
8014 * update_sd_pick_busiest - return 1 on busiest group
8015 * @env: The load balancing environment.
8016 * @sds: sched_domain statistics
8017 * @sg: sched_group candidate to be checked for being the busiest
8018 * @sgs: sched_group statistics
8020 * Determine if @sg is a busier group than the previously selected
8023 * Return: %true if @sg is a busier group than the previously selected
8024 * busiest group. %false otherwise.
8026 static bool update_sd_pick_busiest(struct lb_env *env,
8027 struct sd_lb_stats *sds,
8028 struct sched_group *sg,
8029 struct sg_lb_stats *sgs)
8031 struct sg_lb_stats *busiest = &sds->busiest_stat;
8034 * Don't try to pull misfit tasks we can't help.
8035 * We can use max_capacity here as reduction in capacity on some
8036 * CPUs in the group should either be possible to resolve
8037 * internally or be covered by avg_load imbalance (eventually).
8039 if (sgs->group_type == group_misfit_task &&
8040 (!group_smaller_max_cpu_capacity(sg, sds->local) ||
8041 !group_has_capacity(env, &sds->local_stat)))
8044 if (sgs->group_type > busiest->group_type)
8047 if (sgs->group_type < busiest->group_type)
8050 if (sgs->avg_load <= busiest->avg_load)
8053 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
8057 * Candidate sg has no more than one task per CPU and
8058 * has higher per-CPU capacity. Migrating tasks to less
8059 * capable CPUs may harm throughput. Maximize throughput,
8060 * power/energy consequences are not considered.
8062 if (sgs->sum_nr_running <= sgs->group_weight &&
8063 group_smaller_min_cpu_capacity(sds->local, sg))
8067 * If we have more than one misfit sg go with the biggest misfit.
8069 if (sgs->group_type == group_misfit_task &&
8070 sgs->group_misfit_task_load < busiest->group_misfit_task_load)
8074 /* This is the busiest node in its class. */
8075 if (!(env->sd->flags & SD_ASYM_PACKING))
8078 /* No ASYM_PACKING if target CPU is already busy */
8079 if (env->idle == CPU_NOT_IDLE)
8082 * ASYM_PACKING needs to move all the work to the highest
8083 * prority CPUs in the group, therefore mark all groups
8084 * of lower priority than ourself as busy.
8086 if (sgs->sum_nr_running &&
8087 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
8091 /* Prefer to move from lowest priority CPU's work */
8092 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
8093 sg->asym_prefer_cpu))
8100 #ifdef CONFIG_NUMA_BALANCING
8101 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8103 if (sgs->sum_nr_running > sgs->nr_numa_running)
8105 if (sgs->sum_nr_running > sgs->nr_preferred_running)
8110 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8112 if (rq->nr_running > rq->nr_numa_running)
8114 if (rq->nr_running > rq->nr_preferred_running)
8119 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8124 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8128 #endif /* CONFIG_NUMA_BALANCING */
8131 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8132 * @env: The load balancing environment.
8133 * @sds: variable to hold the statistics for this sched_domain.
8135 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8137 struct sched_domain *child = env->sd->child;
8138 struct sched_group *sg = env->sd->groups;
8139 struct sg_lb_stats *local = &sds->local_stat;
8140 struct sg_lb_stats tmp_sgs;
8141 bool prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
8144 #ifdef CONFIG_NO_HZ_COMMON
8145 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
8146 env->flags |= LBF_NOHZ_STATS;
8150 struct sg_lb_stats *sgs = &tmp_sgs;
8153 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
8158 if (env->idle != CPU_NEWLY_IDLE ||
8159 time_after_eq(jiffies, sg->sgc->next_update))
8160 update_group_capacity(env->sd, env->dst_cpu);
8163 update_sg_lb_stats(env, sg, sgs, &sg_status);
8169 * In case the child domain prefers tasks go to siblings
8170 * first, lower the sg capacity so that we'll try
8171 * and move all the excess tasks away. We lower the capacity
8172 * of a group only if the local group has the capacity to fit
8173 * these excess tasks. The extra check prevents the case where
8174 * you always pull from the heaviest group when it is already
8175 * under-utilized (possible with a large weight task outweighs
8176 * the tasks on the system).
8178 if (prefer_sibling && sds->local &&
8179 group_has_capacity(env, local) &&
8180 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
8181 sgs->group_no_capacity = 1;
8182 sgs->group_type = group_classify(sg, sgs);
8185 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8187 sds->busiest_stat = *sgs;
8191 /* Now, start updating sd_lb_stats */
8192 sds->total_running += sgs->sum_nr_running;
8193 sds->total_load += sgs->group_load;
8194 sds->total_capacity += sgs->group_capacity;
8197 } while (sg != env->sd->groups);
8199 #ifdef CONFIG_NO_HZ_COMMON
8200 if ((env->flags & LBF_NOHZ_AGAIN) &&
8201 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8203 WRITE_ONCE(nohz.next_blocked,
8204 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8208 if (env->sd->flags & SD_NUMA)
8209 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8211 if (!env->sd->parent) {
8212 struct root_domain *rd = env->dst_rq->rd;
8214 /* update overload indicator if we are at root domain */
8215 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
8217 /* Update over-utilization (tipping point, U >= 0) indicator */
8218 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
8219 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
8220 } else if (sg_status & SG_OVERUTILIZED) {
8221 struct root_domain *rd = env->dst_rq->rd;
8223 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
8224 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
8229 * check_asym_packing - Check to see if the group is packed into the
8232 * This is primarily intended to used at the sibling level. Some
8233 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8234 * case of POWER7, it can move to lower SMT modes only when higher
8235 * threads are idle. When in lower SMT modes, the threads will
8236 * perform better since they share less core resources. Hence when we
8237 * have idle threads, we want them to be the higher ones.
8239 * This packing function is run on idle threads. It checks to see if
8240 * the busiest CPU in this domain (core in the P7 case) has a higher
8241 * CPU number than the packing function is being run on. Here we are
8242 * assuming lower CPU number will be equivalent to lower a SMT thread
8245 * Return: 1 when packing is required and a task should be moved to
8246 * this CPU. The amount of the imbalance is returned in env->imbalance.
8248 * @env: The load balancing environment.
8249 * @sds: Statistics of the sched_domain which is to be packed
8251 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8255 if (!(env->sd->flags & SD_ASYM_PACKING))
8258 if (env->idle == CPU_NOT_IDLE)
8264 busiest_cpu = sds->busiest->asym_prefer_cpu;
8265 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
8268 env->imbalance = sds->busiest_stat.group_load;
8274 * fix_small_imbalance - Calculate the minor imbalance that exists
8275 * amongst the groups of a sched_domain, during
8277 * @env: The load balancing environment.
8278 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8281 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8283 unsigned long tmp, capa_now = 0, capa_move = 0;
8284 unsigned int imbn = 2;
8285 unsigned long scaled_busy_load_per_task;
8286 struct sg_lb_stats *local, *busiest;
8288 local = &sds->local_stat;
8289 busiest = &sds->busiest_stat;
8291 if (!local->sum_nr_running)
8292 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8293 else if (busiest->load_per_task > local->load_per_task)
8296 scaled_busy_load_per_task =
8297 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8298 busiest->group_capacity;
8300 if (busiest->avg_load + scaled_busy_load_per_task >=
8301 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8302 env->imbalance = busiest->load_per_task;
8307 * OK, we don't have enough imbalance to justify moving tasks,
8308 * however we may be able to increase total CPU capacity used by
8312 capa_now += busiest->group_capacity *
8313 min(busiest->load_per_task, busiest->avg_load);
8314 capa_now += local->group_capacity *
8315 min(local->load_per_task, local->avg_load);
8316 capa_now /= SCHED_CAPACITY_SCALE;
8318 /* Amount of load we'd subtract */
8319 if (busiest->avg_load > scaled_busy_load_per_task) {
8320 capa_move += busiest->group_capacity *
8321 min(busiest->load_per_task,
8322 busiest->avg_load - scaled_busy_load_per_task);
8325 /* Amount of load we'd add */
8326 if (busiest->avg_load * busiest->group_capacity <
8327 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8328 tmp = (busiest->avg_load * busiest->group_capacity) /
8329 local->group_capacity;
8331 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8332 local->group_capacity;
8334 capa_move += local->group_capacity *
8335 min(local->load_per_task, local->avg_load + tmp);
8336 capa_move /= SCHED_CAPACITY_SCALE;
8338 /* Move if we gain throughput */
8339 if (capa_move > capa_now)
8340 env->imbalance = busiest->load_per_task;
8344 * calculate_imbalance - Calculate the amount of imbalance present within the
8345 * groups of a given sched_domain during load balance.
8346 * @env: load balance environment
8347 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8349 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8351 unsigned long max_pull, load_above_capacity = ~0UL;
8352 struct sg_lb_stats *local, *busiest;
8354 local = &sds->local_stat;
8355 busiest = &sds->busiest_stat;
8357 if (busiest->group_type == group_imbalanced) {
8359 * In the group_imb case we cannot rely on group-wide averages
8360 * to ensure CPU-load equilibrium, look at wider averages. XXX
8362 busiest->load_per_task =
8363 min(busiest->load_per_task, sds->avg_load);
8367 * Avg load of busiest sg can be less and avg load of local sg can
8368 * be greater than avg load across all sgs of sd because avg load
8369 * factors in sg capacity and sgs with smaller group_type are
8370 * skipped when updating the busiest sg:
8372 if (busiest->group_type != group_misfit_task &&
8373 (busiest->avg_load <= sds->avg_load ||
8374 local->avg_load >= sds->avg_load)) {
8376 return fix_small_imbalance(env, sds);
8380 * If there aren't any idle CPUs, avoid creating some.
8382 if (busiest->group_type == group_overloaded &&
8383 local->group_type == group_overloaded) {
8384 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
8385 if (load_above_capacity > busiest->group_capacity) {
8386 load_above_capacity -= busiest->group_capacity;
8387 load_above_capacity *= scale_load_down(NICE_0_LOAD);
8388 load_above_capacity /= busiest->group_capacity;
8390 load_above_capacity = ~0UL;
8394 * We're trying to get all the CPUs to the average_load, so we don't
8395 * want to push ourselves above the average load, nor do we wish to
8396 * reduce the max loaded CPU below the average load. At the same time,
8397 * we also don't want to reduce the group load below the group
8398 * capacity. Thus we look for the minimum possible imbalance.
8400 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8402 /* How much load to actually move to equalise the imbalance */
8403 env->imbalance = min(
8404 max_pull * busiest->group_capacity,
8405 (sds->avg_load - local->avg_load) * local->group_capacity
8406 ) / SCHED_CAPACITY_SCALE;
8408 /* Boost imbalance to allow misfit task to be balanced. */
8409 if (busiest->group_type == group_misfit_task) {
8410 env->imbalance = max_t(long, env->imbalance,
8411 busiest->group_misfit_task_load);
8415 * if *imbalance is less than the average load per runnable task
8416 * there is no guarantee that any tasks will be moved so we'll have
8417 * a think about bumping its value to force at least one task to be
8420 if (env->imbalance < busiest->load_per_task)
8421 return fix_small_imbalance(env, sds);
8424 /******* find_busiest_group() helpers end here *********************/
8427 * find_busiest_group - Returns the busiest group within the sched_domain
8428 * if there is an imbalance.
8430 * Also calculates the amount of weighted load which should be moved
8431 * to restore balance.
8433 * @env: The load balancing environment.
8435 * Return: - The busiest group if imbalance exists.
8437 static struct sched_group *find_busiest_group(struct lb_env *env)
8439 struct sg_lb_stats *local, *busiest;
8440 struct sd_lb_stats sds;
8442 init_sd_lb_stats(&sds);
8445 * Compute the various statistics relavent for load balancing at
8448 update_sd_lb_stats(env, &sds);
8450 if (sched_energy_enabled()) {
8451 struct root_domain *rd = env->dst_rq->rd;
8453 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
8457 local = &sds.local_stat;
8458 busiest = &sds.busiest_stat;
8460 /* ASYM feature bypasses nice load balance check */
8461 if (check_asym_packing(env, &sds))
8464 /* There is no busy sibling group to pull tasks from */
8465 if (!sds.busiest || busiest->sum_nr_running == 0)
8468 /* XXX broken for overlapping NUMA groups */
8469 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8470 / sds.total_capacity;
8473 * If the busiest group is imbalanced the below checks don't
8474 * work because they assume all things are equal, which typically
8475 * isn't true due to cpus_ptr constraints and the like.
8477 if (busiest->group_type == group_imbalanced)
8481 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8482 * capacities from resulting in underutilization due to avg_load.
8484 if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8485 busiest->group_no_capacity)
8488 /* Misfit tasks should be dealt with regardless of the avg load */
8489 if (busiest->group_type == group_misfit_task)
8493 * If the local group is busier than the selected busiest group
8494 * don't try and pull any tasks.
8496 if (local->avg_load >= busiest->avg_load)
8500 * Don't pull any tasks if this group is already above the domain
8503 if (local->avg_load >= sds.avg_load)
8506 if (env->idle == CPU_IDLE) {
8508 * This CPU is idle. If the busiest group is not overloaded
8509 * and there is no imbalance between this and busiest group
8510 * wrt idle CPUs, it is balanced. The imbalance becomes
8511 * significant if the diff is greater than 1 otherwise we
8512 * might end up to just move the imbalance on another group
8514 if ((busiest->group_type != group_overloaded) &&
8515 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8519 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8520 * imbalance_pct to be conservative.
8522 if (100 * busiest->avg_load <=
8523 env->sd->imbalance_pct * local->avg_load)
8528 /* Looks like there is an imbalance. Compute it */
8529 env->src_grp_type = busiest->group_type;
8530 calculate_imbalance(env, &sds);
8531 return env->imbalance ? sds.busiest : NULL;
8539 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
8541 static struct rq *find_busiest_queue(struct lb_env *env,
8542 struct sched_group *group)
8544 struct rq *busiest = NULL, *rq;
8545 unsigned long busiest_load = 0, busiest_capacity = 1;
8548 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8549 unsigned long capacity, wl;
8553 rt = fbq_classify_rq(rq);
8556 * We classify groups/runqueues into three groups:
8557 * - regular: there are !numa tasks
8558 * - remote: there are numa tasks that run on the 'wrong' node
8559 * - all: there is no distinction
8561 * In order to avoid migrating ideally placed numa tasks,
8562 * ignore those when there's better options.
8564 * If we ignore the actual busiest queue to migrate another
8565 * task, the next balance pass can still reduce the busiest
8566 * queue by moving tasks around inside the node.
8568 * If we cannot move enough load due to this classification
8569 * the next pass will adjust the group classification and
8570 * allow migration of more tasks.
8572 * Both cases only affect the total convergence complexity.
8574 if (rt > env->fbq_type)
8578 * For ASYM_CPUCAPACITY domains with misfit tasks we simply
8579 * seek the "biggest" misfit task.
8581 if (env->src_grp_type == group_misfit_task) {
8582 if (rq->misfit_task_load > busiest_load) {
8583 busiest_load = rq->misfit_task_load;
8590 capacity = capacity_of(i);
8593 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
8594 * eventually lead to active_balancing high->low capacity.
8595 * Higher per-CPU capacity is considered better than balancing
8598 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8599 capacity_of(env->dst_cpu) < capacity &&
8600 rq->nr_running == 1)
8603 wl = weighted_cpuload(rq);
8606 * When comparing with imbalance, use weighted_cpuload()
8607 * which is not scaled with the CPU capacity.
8610 if (rq->nr_running == 1 && wl > env->imbalance &&
8611 !check_cpu_capacity(rq, env->sd))
8615 * For the load comparisons with the other CPU's, consider
8616 * the weighted_cpuload() scaled with the CPU capacity, so
8617 * that the load can be moved away from the CPU that is
8618 * potentially running at a lower capacity.
8620 * Thus we're looking for max(wl_i / capacity_i), crosswise
8621 * multiplication to rid ourselves of the division works out
8622 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8623 * our previous maximum.
8625 if (wl * busiest_capacity > busiest_load * capacity) {
8627 busiest_capacity = capacity;
8636 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8637 * so long as it is large enough.
8639 #define MAX_PINNED_INTERVAL 512
8642 asym_active_balance(struct lb_env *env)
8645 * ASYM_PACKING needs to force migrate tasks from busy but
8646 * lower priority CPUs in order to pack all tasks in the
8647 * highest priority CPUs.
8649 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
8650 sched_asym_prefer(env->dst_cpu, env->src_cpu);
8654 voluntary_active_balance(struct lb_env *env)
8656 struct sched_domain *sd = env->sd;
8658 if (asym_active_balance(env))
8662 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8663 * It's worth migrating the task if the src_cpu's capacity is reduced
8664 * because of other sched_class or IRQs if more capacity stays
8665 * available on dst_cpu.
8667 if ((env->idle != CPU_NOT_IDLE) &&
8668 (env->src_rq->cfs.h_nr_running == 1)) {
8669 if ((check_cpu_capacity(env->src_rq, sd)) &&
8670 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8674 if (env->src_grp_type == group_misfit_task)
8680 static int need_active_balance(struct lb_env *env)
8682 struct sched_domain *sd = env->sd;
8684 if (voluntary_active_balance(env))
8687 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8690 static int active_load_balance_cpu_stop(void *data);
8692 static int should_we_balance(struct lb_env *env)
8694 struct sched_group *sg = env->sd->groups;
8695 int cpu, balance_cpu = -1;
8698 * Ensure the balancing environment is consistent; can happen
8699 * when the softirq triggers 'during' hotplug.
8701 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8705 * In the newly idle case, we will allow all the CPUs
8706 * to do the newly idle load balance.
8708 if (env->idle == CPU_NEWLY_IDLE)
8711 /* Try to find first idle CPU */
8712 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8720 if (balance_cpu == -1)
8721 balance_cpu = group_balance_cpu(sg);
8724 * First idle CPU or the first CPU(busiest) in this sched group
8725 * is eligible for doing load balancing at this and above domains.
8727 return balance_cpu == env->dst_cpu;
8731 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8732 * tasks if there is an imbalance.
8734 static int load_balance(int this_cpu, struct rq *this_rq,
8735 struct sched_domain *sd, enum cpu_idle_type idle,
8736 int *continue_balancing)
8738 int ld_moved, cur_ld_moved, active_balance = 0;
8739 struct sched_domain *sd_parent = sd->parent;
8740 struct sched_group *group;
8743 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8745 struct lb_env env = {
8747 .dst_cpu = this_cpu,
8749 .dst_grpmask = sched_group_span(sd->groups),
8751 .loop_break = sched_nr_migrate_break,
8754 .tasks = LIST_HEAD_INIT(env.tasks),
8757 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8759 schedstat_inc(sd->lb_count[idle]);
8762 if (!should_we_balance(&env)) {
8763 *continue_balancing = 0;
8767 group = find_busiest_group(&env);
8769 schedstat_inc(sd->lb_nobusyg[idle]);
8773 busiest = find_busiest_queue(&env, group);
8775 schedstat_inc(sd->lb_nobusyq[idle]);
8779 BUG_ON(busiest == env.dst_rq);
8781 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8783 env.src_cpu = busiest->cpu;
8784 env.src_rq = busiest;
8787 if (busiest->nr_running > 1) {
8789 * Attempt to move tasks. If find_busiest_group has found
8790 * an imbalance but busiest->nr_running <= 1, the group is
8791 * still unbalanced. ld_moved simply stays zero, so it is
8792 * correctly treated as an imbalance.
8794 env.flags |= LBF_ALL_PINNED;
8795 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8798 rq_lock_irqsave(busiest, &rf);
8799 update_rq_clock(busiest);
8802 * cur_ld_moved - load moved in current iteration
8803 * ld_moved - cumulative load moved across iterations
8805 cur_ld_moved = detach_tasks(&env);
8808 * We've detached some tasks from busiest_rq. Every
8809 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8810 * unlock busiest->lock, and we are able to be sure
8811 * that nobody can manipulate the tasks in parallel.
8812 * See task_rq_lock() family for the details.
8815 rq_unlock(busiest, &rf);
8819 ld_moved += cur_ld_moved;
8822 local_irq_restore(rf.flags);
8824 if (env.flags & LBF_NEED_BREAK) {
8825 env.flags &= ~LBF_NEED_BREAK;
8830 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8831 * us and move them to an alternate dst_cpu in our sched_group
8832 * where they can run. The upper limit on how many times we
8833 * iterate on same src_cpu is dependent on number of CPUs in our
8836 * This changes load balance semantics a bit on who can move
8837 * load to a given_cpu. In addition to the given_cpu itself
8838 * (or a ilb_cpu acting on its behalf where given_cpu is
8839 * nohz-idle), we now have balance_cpu in a position to move
8840 * load to given_cpu. In rare situations, this may cause
8841 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8842 * _independently_ and at _same_ time to move some load to
8843 * given_cpu) causing exceess load to be moved to given_cpu.
8844 * This however should not happen so much in practice and
8845 * moreover subsequent load balance cycles should correct the
8846 * excess load moved.
8848 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8850 /* Prevent to re-select dst_cpu via env's CPUs */
8851 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
8853 env.dst_rq = cpu_rq(env.new_dst_cpu);
8854 env.dst_cpu = env.new_dst_cpu;
8855 env.flags &= ~LBF_DST_PINNED;
8857 env.loop_break = sched_nr_migrate_break;
8860 * Go back to "more_balance" rather than "redo" since we
8861 * need to continue with same src_cpu.
8867 * We failed to reach balance because of affinity.
8870 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8872 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8873 *group_imbalance = 1;
8876 /* All tasks on this runqueue were pinned by CPU affinity */
8877 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8878 __cpumask_clear_cpu(cpu_of(busiest), cpus);
8880 * Attempting to continue load balancing at the current
8881 * sched_domain level only makes sense if there are
8882 * active CPUs remaining as possible busiest CPUs to
8883 * pull load from which are not contained within the
8884 * destination group that is receiving any migrated
8887 if (!cpumask_subset(cpus, env.dst_grpmask)) {
8889 env.loop_break = sched_nr_migrate_break;
8892 goto out_all_pinned;
8897 schedstat_inc(sd->lb_failed[idle]);
8899 * Increment the failure counter only on periodic balance.
8900 * We do not want newidle balance, which can be very
8901 * frequent, pollute the failure counter causing
8902 * excessive cache_hot migrations and active balances.
8904 if (idle != CPU_NEWLY_IDLE)
8905 sd->nr_balance_failed++;
8907 if (need_active_balance(&env)) {
8908 unsigned long flags;
8910 raw_spin_lock_irqsave(&busiest->lock, flags);
8913 * Don't kick the active_load_balance_cpu_stop,
8914 * if the curr task on busiest CPU can't be
8915 * moved to this_cpu:
8917 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
8918 raw_spin_unlock_irqrestore(&busiest->lock,
8920 env.flags |= LBF_ALL_PINNED;
8921 goto out_one_pinned;
8925 * ->active_balance synchronizes accesses to
8926 * ->active_balance_work. Once set, it's cleared
8927 * only after active load balance is finished.
8929 if (!busiest->active_balance) {
8930 busiest->active_balance = 1;
8931 busiest->push_cpu = this_cpu;
8934 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8936 if (active_balance) {
8937 stop_one_cpu_nowait(cpu_of(busiest),
8938 active_load_balance_cpu_stop, busiest,
8939 &busiest->active_balance_work);
8942 /* We've kicked active balancing, force task migration. */
8943 sd->nr_balance_failed = sd->cache_nice_tries+1;
8946 sd->nr_balance_failed = 0;
8948 if (likely(!active_balance) || voluntary_active_balance(&env)) {
8949 /* We were unbalanced, so reset the balancing interval */
8950 sd->balance_interval = sd->min_interval;
8953 * If we've begun active balancing, start to back off. This
8954 * case may not be covered by the all_pinned logic if there
8955 * is only 1 task on the busy runqueue (because we don't call
8958 if (sd->balance_interval < sd->max_interval)
8959 sd->balance_interval *= 2;
8966 * We reach balance although we may have faced some affinity
8967 * constraints. Clear the imbalance flag if it was set.
8970 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8972 if (*group_imbalance)
8973 *group_imbalance = 0;
8978 * We reach balance because all tasks are pinned at this level so
8979 * we can't migrate them. Let the imbalance flag set so parent level
8980 * can try to migrate them.
8982 schedstat_inc(sd->lb_balanced[idle]);
8984 sd->nr_balance_failed = 0;
8990 * idle_balance() disregards balance intervals, so we could repeatedly
8991 * reach this code, which would lead to balance_interval skyrocketting
8992 * in a short amount of time. Skip the balance_interval increase logic
8995 if (env.idle == CPU_NEWLY_IDLE)
8998 /* tune up the balancing interval */
8999 if ((env.flags & LBF_ALL_PINNED &&
9000 sd->balance_interval < MAX_PINNED_INTERVAL) ||
9001 sd->balance_interval < sd->max_interval)
9002 sd->balance_interval *= 2;
9007 static inline unsigned long
9008 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
9010 unsigned long interval = sd->balance_interval;
9013 interval *= sd->busy_factor;
9015 /* scale ms to jiffies */
9016 interval = msecs_to_jiffies(interval);
9017 interval = clamp(interval, 1UL, max_load_balance_interval);
9023 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
9025 unsigned long interval, next;
9027 /* used by idle balance, so cpu_busy = 0 */
9028 interval = get_sd_balance_interval(sd, 0);
9029 next = sd->last_balance + interval;
9031 if (time_after(*next_balance, next))
9032 *next_balance = next;
9036 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9037 * running tasks off the busiest CPU onto idle CPUs. It requires at
9038 * least 1 task to be running on each physical CPU where possible, and
9039 * avoids physical / logical imbalances.
9041 static int active_load_balance_cpu_stop(void *data)
9043 struct rq *busiest_rq = data;
9044 int busiest_cpu = cpu_of(busiest_rq);
9045 int target_cpu = busiest_rq->push_cpu;
9046 struct rq *target_rq = cpu_rq(target_cpu);
9047 struct sched_domain *sd;
9048 struct task_struct *p = NULL;
9051 rq_lock_irq(busiest_rq, &rf);
9053 * Between queueing the stop-work and running it is a hole in which
9054 * CPUs can become inactive. We should not move tasks from or to
9057 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
9060 /* Make sure the requested CPU hasn't gone down in the meantime: */
9061 if (unlikely(busiest_cpu != smp_processor_id() ||
9062 !busiest_rq->active_balance))
9065 /* Is there any task to move? */
9066 if (busiest_rq->nr_running <= 1)
9070 * This condition is "impossible", if it occurs
9071 * we need to fix it. Originally reported by
9072 * Bjorn Helgaas on a 128-CPU setup.
9074 BUG_ON(busiest_rq == target_rq);
9076 /* Search for an sd spanning us and the target CPU. */
9078 for_each_domain(target_cpu, sd) {
9079 if ((sd->flags & SD_LOAD_BALANCE) &&
9080 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9085 struct lb_env env = {
9087 .dst_cpu = target_cpu,
9088 .dst_rq = target_rq,
9089 .src_cpu = busiest_rq->cpu,
9090 .src_rq = busiest_rq,
9093 * can_migrate_task() doesn't need to compute new_dst_cpu
9094 * for active balancing. Since we have CPU_IDLE, but no
9095 * @dst_grpmask we need to make that test go away with lying
9098 .flags = LBF_DST_PINNED,
9101 schedstat_inc(sd->alb_count);
9102 update_rq_clock(busiest_rq);
9104 p = detach_one_task(&env);
9106 schedstat_inc(sd->alb_pushed);
9107 /* Active balancing done, reset the failure counter. */
9108 sd->nr_balance_failed = 0;
9110 schedstat_inc(sd->alb_failed);
9115 busiest_rq->active_balance = 0;
9116 rq_unlock(busiest_rq, &rf);
9119 attach_one_task(target_rq, p);
9126 static DEFINE_SPINLOCK(balancing);
9129 * Scale the max load_balance interval with the number of CPUs in the system.
9130 * This trades load-balance latency on larger machines for less cross talk.
9132 void update_max_interval(void)
9134 max_load_balance_interval = HZ*num_online_cpus()/10;
9138 * It checks each scheduling domain to see if it is due to be balanced,
9139 * and initiates a balancing operation if so.
9141 * Balancing parameters are set up in init_sched_domains.
9143 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9145 int continue_balancing = 1;
9147 unsigned long interval;
9148 struct sched_domain *sd;
9149 /* Earliest time when we have to do rebalance again */
9150 unsigned long next_balance = jiffies + 60*HZ;
9151 int update_next_balance = 0;
9152 int need_serialize, need_decay = 0;
9156 for_each_domain(cpu, sd) {
9158 * Decay the newidle max times here because this is a regular
9159 * visit to all the domains. Decay ~1% per second.
9161 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9162 sd->max_newidle_lb_cost =
9163 (sd->max_newidle_lb_cost * 253) / 256;
9164 sd->next_decay_max_lb_cost = jiffies + HZ;
9167 max_cost += sd->max_newidle_lb_cost;
9169 if (!(sd->flags & SD_LOAD_BALANCE))
9173 * Stop the load balance at this level. There is another
9174 * CPU in our sched group which is doing load balancing more
9177 if (!continue_balancing) {
9183 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9185 need_serialize = sd->flags & SD_SERIALIZE;
9186 if (need_serialize) {
9187 if (!spin_trylock(&balancing))
9191 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9192 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9194 * The LBF_DST_PINNED logic could have changed
9195 * env->dst_cpu, so we can't know our idle
9196 * state even if we migrated tasks. Update it.
9198 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9200 sd->last_balance = jiffies;
9201 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9204 spin_unlock(&balancing);
9206 if (time_after(next_balance, sd->last_balance + interval)) {
9207 next_balance = sd->last_balance + interval;
9208 update_next_balance = 1;
9213 * Ensure the rq-wide value also decays but keep it at a
9214 * reasonable floor to avoid funnies with rq->avg_idle.
9216 rq->max_idle_balance_cost =
9217 max((u64)sysctl_sched_migration_cost, max_cost);
9222 * next_balance will be updated only when there is a need.
9223 * When the cpu is attached to null domain for ex, it will not be
9226 if (likely(update_next_balance)) {
9227 rq->next_balance = next_balance;
9229 #ifdef CONFIG_NO_HZ_COMMON
9231 * If this CPU has been elected to perform the nohz idle
9232 * balance. Other idle CPUs have already rebalanced with
9233 * nohz_idle_balance() and nohz.next_balance has been
9234 * updated accordingly. This CPU is now running the idle load
9235 * balance for itself and we need to update the
9236 * nohz.next_balance accordingly.
9238 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9239 nohz.next_balance = rq->next_balance;
9244 static inline int on_null_domain(struct rq *rq)
9246 return unlikely(!rcu_dereference_sched(rq->sd));
9249 #ifdef CONFIG_NO_HZ_COMMON
9251 * idle load balancing details
9252 * - When one of the busy CPUs notice that there may be an idle rebalancing
9253 * needed, they will kick the idle load balancer, which then does idle
9254 * load balancing for all the idle CPUs.
9255 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
9259 static inline int find_new_ilb(void)
9263 for_each_cpu_and(ilb, nohz.idle_cpus_mask,
9264 housekeeping_cpumask(HK_FLAG_MISC)) {
9273 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
9274 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
9276 static void kick_ilb(unsigned int flags)
9280 nohz.next_balance++;
9282 ilb_cpu = find_new_ilb();
9284 if (ilb_cpu >= nr_cpu_ids)
9287 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
9288 if (flags & NOHZ_KICK_MASK)
9292 * Use smp_send_reschedule() instead of resched_cpu().
9293 * This way we generate a sched IPI on the target CPU which
9294 * is idle. And the softirq performing nohz idle load balance
9295 * will be run before returning from the IPI.
9297 smp_send_reschedule(ilb_cpu);
9301 * Current decision point for kicking the idle load balancer in the presence
9302 * of idle CPUs in the system.
9304 static void nohz_balancer_kick(struct rq *rq)
9306 unsigned long now = jiffies;
9307 struct sched_domain_shared *sds;
9308 struct sched_domain *sd;
9309 int nr_busy, i, cpu = rq->cpu;
9310 unsigned int flags = 0;
9312 if (unlikely(rq->idle_balance))
9316 * We may be recently in ticked or tickless idle mode. At the first
9317 * busy tick after returning from idle, we will update the busy stats.
9319 nohz_balance_exit_idle(rq);
9322 * None are in tickless mode and hence no need for NOHZ idle load
9325 if (likely(!atomic_read(&nohz.nr_cpus)))
9328 if (READ_ONCE(nohz.has_blocked) &&
9329 time_after(now, READ_ONCE(nohz.next_blocked)))
9330 flags = NOHZ_STATS_KICK;
9332 if (time_before(now, nohz.next_balance))
9335 if (rq->nr_running >= 2) {
9336 flags = NOHZ_KICK_MASK;
9342 sd = rcu_dereference(rq->sd);
9345 * If there's a CFS task and the current CPU has reduced
9346 * capacity; kick the ILB to see if there's a better CPU to run
9349 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
9350 flags = NOHZ_KICK_MASK;
9355 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
9358 * When ASYM_PACKING; see if there's a more preferred CPU
9359 * currently idle; in which case, kick the ILB to move tasks
9362 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
9363 if (sched_asym_prefer(i, cpu)) {
9364 flags = NOHZ_KICK_MASK;
9370 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
9373 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
9374 * to run the misfit task on.
9376 if (check_misfit_status(rq, sd)) {
9377 flags = NOHZ_KICK_MASK;
9382 * For asymmetric systems, we do not want to nicely balance
9383 * cache use, instead we want to embrace asymmetry and only
9384 * ensure tasks have enough CPU capacity.
9386 * Skip the LLC logic because it's not relevant in that case.
9391 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9394 * If there is an imbalance between LLC domains (IOW we could
9395 * increase the overall cache use), we need some less-loaded LLC
9396 * domain to pull some load. Likewise, we may need to spread
9397 * load within the current LLC domain (e.g. packed SMT cores but
9398 * other CPUs are idle). We can't really know from here how busy
9399 * the others are - so just get a nohz balance going if it looks
9400 * like this LLC domain has tasks we could move.
9402 nr_busy = atomic_read(&sds->nr_busy_cpus);
9404 flags = NOHZ_KICK_MASK;
9415 static void set_cpu_sd_state_busy(int cpu)
9417 struct sched_domain *sd;
9420 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9422 if (!sd || !sd->nohz_idle)
9426 atomic_inc(&sd->shared->nr_busy_cpus);
9431 void nohz_balance_exit_idle(struct rq *rq)
9433 SCHED_WARN_ON(rq != this_rq());
9435 if (likely(!rq->nohz_tick_stopped))
9438 rq->nohz_tick_stopped = 0;
9439 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
9440 atomic_dec(&nohz.nr_cpus);
9442 set_cpu_sd_state_busy(rq->cpu);
9445 static void set_cpu_sd_state_idle(int cpu)
9447 struct sched_domain *sd;
9450 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9452 if (!sd || sd->nohz_idle)
9456 atomic_dec(&sd->shared->nr_busy_cpus);
9462 * This routine will record that the CPU is going idle with tick stopped.
9463 * This info will be used in performing idle load balancing in the future.
9465 void nohz_balance_enter_idle(int cpu)
9467 struct rq *rq = cpu_rq(cpu);
9469 SCHED_WARN_ON(cpu != smp_processor_id());
9471 /* If this CPU is going down, then nothing needs to be done: */
9472 if (!cpu_active(cpu))
9475 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9476 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9480 * Can be set safely without rq->lock held
9481 * If a clear happens, it will have evaluated last additions because
9482 * rq->lock is held during the check and the clear
9484 rq->has_blocked_load = 1;
9487 * The tick is still stopped but load could have been added in the
9488 * meantime. We set the nohz.has_blocked flag to trig a check of the
9489 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
9490 * of nohz.has_blocked can only happen after checking the new load
9492 if (rq->nohz_tick_stopped)
9495 /* If we're a completely isolated CPU, we don't play: */
9496 if (on_null_domain(rq))
9499 rq->nohz_tick_stopped = 1;
9501 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9502 atomic_inc(&nohz.nr_cpus);
9505 * Ensures that if nohz_idle_balance() fails to observe our
9506 * @idle_cpus_mask store, it must observe the @has_blocked
9509 smp_mb__after_atomic();
9511 set_cpu_sd_state_idle(cpu);
9515 * Each time a cpu enter idle, we assume that it has blocked load and
9516 * enable the periodic update of the load of idle cpus
9518 WRITE_ONCE(nohz.has_blocked, 1);
9522 * Internal function that runs load balance for all idle cpus. The load balance
9523 * can be a simple update of blocked load or a complete load balance with
9524 * tasks movement depending of flags.
9525 * The function returns false if the loop has stopped before running
9526 * through all idle CPUs.
9528 static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
9529 enum cpu_idle_type idle)
9531 /* Earliest time when we have to do rebalance again */
9532 unsigned long now = jiffies;
9533 unsigned long next_balance = now + 60*HZ;
9534 bool has_blocked_load = false;
9535 int update_next_balance = 0;
9536 int this_cpu = this_rq->cpu;
9541 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
9544 * We assume there will be no idle load after this update and clear
9545 * the has_blocked flag. If a cpu enters idle in the mean time, it will
9546 * set the has_blocked flag and trig another update of idle load.
9547 * Because a cpu that becomes idle, is added to idle_cpus_mask before
9548 * setting the flag, we are sure to not clear the state and not
9549 * check the load of an idle cpu.
9551 WRITE_ONCE(nohz.has_blocked, 0);
9554 * Ensures that if we miss the CPU, we must see the has_blocked
9555 * store from nohz_balance_enter_idle().
9559 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9560 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9564 * If this CPU gets work to do, stop the load balancing
9565 * work being done for other CPUs. Next load
9566 * balancing owner will pick it up.
9568 if (need_resched()) {
9569 has_blocked_load = true;
9573 rq = cpu_rq(balance_cpu);
9575 has_blocked_load |= update_nohz_stats(rq, true);
9578 * If time for next balance is due,
9581 if (time_after_eq(jiffies, rq->next_balance)) {
9584 rq_lock_irqsave(rq, &rf);
9585 update_rq_clock(rq);
9586 rq_unlock_irqrestore(rq, &rf);
9588 if (flags & NOHZ_BALANCE_KICK)
9589 rebalance_domains(rq, CPU_IDLE);
9592 if (time_after(next_balance, rq->next_balance)) {
9593 next_balance = rq->next_balance;
9594 update_next_balance = 1;
9598 /* Newly idle CPU doesn't need an update */
9599 if (idle != CPU_NEWLY_IDLE) {
9600 update_blocked_averages(this_cpu);
9601 has_blocked_load |= this_rq->has_blocked_load;
9604 if (flags & NOHZ_BALANCE_KICK)
9605 rebalance_domains(this_rq, CPU_IDLE);
9607 WRITE_ONCE(nohz.next_blocked,
9608 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
9610 /* The full idle balance loop has been done */
9614 /* There is still blocked load, enable periodic update */
9615 if (has_blocked_load)
9616 WRITE_ONCE(nohz.has_blocked, 1);
9619 * next_balance will be updated only when there is a need.
9620 * When the CPU is attached to null domain for ex, it will not be
9623 if (likely(update_next_balance))
9624 nohz.next_balance = next_balance;
9630 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9631 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9633 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9635 int this_cpu = this_rq->cpu;
9638 if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
9641 if (idle != CPU_IDLE) {
9642 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9646 /* could be _relaxed() */
9647 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9648 if (!(flags & NOHZ_KICK_MASK))
9651 _nohz_idle_balance(this_rq, flags, idle);
9656 static void nohz_newidle_balance(struct rq *this_rq)
9658 int this_cpu = this_rq->cpu;
9661 * This CPU doesn't want to be disturbed by scheduler
9664 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
9667 /* Will wake up very soon. No time for doing anything else*/
9668 if (this_rq->avg_idle < sysctl_sched_migration_cost)
9671 /* Don't need to update blocked load of idle CPUs*/
9672 if (!READ_ONCE(nohz.has_blocked) ||
9673 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
9676 raw_spin_unlock(&this_rq->lock);
9678 * This CPU is going to be idle and blocked load of idle CPUs
9679 * need to be updated. Run the ilb locally as it is a good
9680 * candidate for ilb instead of waking up another idle CPU.
9681 * Kick an normal ilb if we failed to do the update.
9683 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
9684 kick_ilb(NOHZ_STATS_KICK);
9685 raw_spin_lock(&this_rq->lock);
9688 #else /* !CONFIG_NO_HZ_COMMON */
9689 static inline void nohz_balancer_kick(struct rq *rq) { }
9691 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9696 static inline void nohz_newidle_balance(struct rq *this_rq) { }
9697 #endif /* CONFIG_NO_HZ_COMMON */
9700 * idle_balance is called by schedule() if this_cpu is about to become
9701 * idle. Attempts to pull tasks from other CPUs.
9703 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
9705 unsigned long next_balance = jiffies + HZ;
9706 int this_cpu = this_rq->cpu;
9707 struct sched_domain *sd;
9708 int pulled_task = 0;
9712 * We must set idle_stamp _before_ calling idle_balance(), such that we
9713 * measure the duration of idle_balance() as idle time.
9715 this_rq->idle_stamp = rq_clock(this_rq);
9718 * Do not pull tasks towards !active CPUs...
9720 if (!cpu_active(this_cpu))
9724 * This is OK, because current is on_cpu, which avoids it being picked
9725 * for load-balance and preemption/IRQs are still disabled avoiding
9726 * further scheduler activity on it and we're being very careful to
9727 * re-start the picking loop.
9729 rq_unpin_lock(this_rq, rf);
9731 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
9732 !READ_ONCE(this_rq->rd->overload)) {
9735 sd = rcu_dereference_check_sched_domain(this_rq->sd);
9737 update_next_balance(sd, &next_balance);
9740 nohz_newidle_balance(this_rq);
9745 raw_spin_unlock(&this_rq->lock);
9747 update_blocked_averages(this_cpu);
9749 for_each_domain(this_cpu, sd) {
9750 int continue_balancing = 1;
9751 u64 t0, domain_cost;
9753 if (!(sd->flags & SD_LOAD_BALANCE))
9756 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
9757 update_next_balance(sd, &next_balance);
9761 if (sd->flags & SD_BALANCE_NEWIDLE) {
9762 t0 = sched_clock_cpu(this_cpu);
9764 pulled_task = load_balance(this_cpu, this_rq,
9766 &continue_balancing);
9768 domain_cost = sched_clock_cpu(this_cpu) - t0;
9769 if (domain_cost > sd->max_newidle_lb_cost)
9770 sd->max_newidle_lb_cost = domain_cost;
9772 curr_cost += domain_cost;
9775 update_next_balance(sd, &next_balance);
9778 * Stop searching for tasks to pull if there are
9779 * now runnable tasks on this rq.
9781 if (pulled_task || this_rq->nr_running > 0)
9786 raw_spin_lock(&this_rq->lock);
9788 if (curr_cost > this_rq->max_idle_balance_cost)
9789 this_rq->max_idle_balance_cost = curr_cost;
9793 * While browsing the domains, we released the rq lock, a task could
9794 * have been enqueued in the meantime. Since we're not going idle,
9795 * pretend we pulled a task.
9797 if (this_rq->cfs.h_nr_running && !pulled_task)
9800 /* Move the next balance forward */
9801 if (time_after(this_rq->next_balance, next_balance))
9802 this_rq->next_balance = next_balance;
9804 /* Is there a task of a high priority class? */
9805 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
9809 this_rq->idle_stamp = 0;
9811 rq_repin_lock(this_rq, rf);
9817 * run_rebalance_domains is triggered when needed from the scheduler tick.
9818 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9820 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9822 struct rq *this_rq = this_rq();
9823 enum cpu_idle_type idle = this_rq->idle_balance ?
9824 CPU_IDLE : CPU_NOT_IDLE;
9827 * If this CPU has a pending nohz_balance_kick, then do the
9828 * balancing on behalf of the other idle CPUs whose ticks are
9829 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9830 * give the idle CPUs a chance to load balance. Else we may
9831 * load balance only within the local sched_domain hierarchy
9832 * and abort nohz_idle_balance altogether if we pull some load.
9834 if (nohz_idle_balance(this_rq, idle))
9837 /* normal load balance */
9838 update_blocked_averages(this_rq->cpu);
9839 rebalance_domains(this_rq, idle);
9843 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9845 void trigger_load_balance(struct rq *rq)
9847 /* Don't need to rebalance while attached to NULL domain */
9848 if (unlikely(on_null_domain(rq)))
9851 if (time_after_eq(jiffies, rq->next_balance))
9852 raise_softirq(SCHED_SOFTIRQ);
9854 nohz_balancer_kick(rq);
9857 static void rq_online_fair(struct rq *rq)
9861 update_runtime_enabled(rq);
9864 static void rq_offline_fair(struct rq *rq)
9868 /* Ensure any throttled groups are reachable by pick_next_task */
9869 unthrottle_offline_cfs_rqs(rq);
9872 #endif /* CONFIG_SMP */
9875 * scheduler tick hitting a task of our scheduling class.
9877 * NOTE: This function can be called remotely by the tick offload that
9878 * goes along full dynticks. Therefore no local assumption can be made
9879 * and everything must be accessed through the @rq and @curr passed in
9882 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9884 struct cfs_rq *cfs_rq;
9885 struct sched_entity *se = &curr->se;
9887 for_each_sched_entity(se) {
9888 cfs_rq = cfs_rq_of(se);
9889 entity_tick(cfs_rq, se, queued);
9892 if (static_branch_unlikely(&sched_numa_balancing))
9893 task_tick_numa(rq, curr);
9895 update_misfit_status(curr, rq);
9896 update_overutilized_status(task_rq(curr));
9900 * called on fork with the child task as argument from the parent's context
9901 * - child not yet on the tasklist
9902 * - preemption disabled
9904 static void task_fork_fair(struct task_struct *p)
9906 struct cfs_rq *cfs_rq;
9907 struct sched_entity *se = &p->se, *curr;
9908 struct rq *rq = this_rq();
9912 update_rq_clock(rq);
9914 cfs_rq = task_cfs_rq(current);
9915 curr = cfs_rq->curr;
9917 update_curr(cfs_rq);
9918 se->vruntime = curr->vruntime;
9920 place_entity(cfs_rq, se, 1);
9922 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9924 * Upon rescheduling, sched_class::put_prev_task() will place
9925 * 'current' within the tree based on its new key value.
9927 swap(curr->vruntime, se->vruntime);
9931 se->vruntime -= cfs_rq->min_vruntime;
9936 * Priority of the task has changed. Check to see if we preempt
9940 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9942 if (!task_on_rq_queued(p))
9946 * Reschedule if we are currently running on this runqueue and
9947 * our priority decreased, or if we are not currently running on
9948 * this runqueue and our priority is higher than the current's
9950 if (rq->curr == p) {
9951 if (p->prio > oldprio)
9954 check_preempt_curr(rq, p, 0);
9957 static inline bool vruntime_normalized(struct task_struct *p)
9959 struct sched_entity *se = &p->se;
9962 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9963 * the dequeue_entity(.flags=0) will already have normalized the
9970 * When !on_rq, vruntime of the task has usually NOT been normalized.
9971 * But there are some cases where it has already been normalized:
9973 * - A forked child which is waiting for being woken up by
9974 * wake_up_new_task().
9975 * - A task which has been woken up by try_to_wake_up() and
9976 * waiting for actually being woken up by sched_ttwu_pending().
9978 if (!se->sum_exec_runtime ||
9979 (p->state == TASK_WAKING && p->sched_remote_wakeup))
9985 #ifdef CONFIG_FAIR_GROUP_SCHED
9987 * Propagate the changes of the sched_entity across the tg tree to make it
9988 * visible to the root
9990 static void propagate_entity_cfs_rq(struct sched_entity *se)
9992 struct cfs_rq *cfs_rq;
9994 /* Start to propagate at parent */
9997 for_each_sched_entity(se) {
9998 cfs_rq = cfs_rq_of(se);
10000 if (cfs_rq_throttled(cfs_rq))
10003 update_load_avg(cfs_rq, se, UPDATE_TG);
10007 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
10010 static void detach_entity_cfs_rq(struct sched_entity *se)
10012 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10014 /* Catch up with the cfs_rq and remove our load when we leave */
10015 update_load_avg(cfs_rq, se, 0);
10016 detach_entity_load_avg(cfs_rq, se);
10017 update_tg_load_avg(cfs_rq, false);
10018 propagate_entity_cfs_rq(se);
10021 static void attach_entity_cfs_rq(struct sched_entity *se)
10023 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10025 #ifdef CONFIG_FAIR_GROUP_SCHED
10027 * Since the real-depth could have been changed (only FAIR
10028 * class maintain depth value), reset depth properly.
10030 se->depth = se->parent ? se->parent->depth + 1 : 0;
10033 /* Synchronize entity with its cfs_rq */
10034 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
10035 attach_entity_load_avg(cfs_rq, se, 0);
10036 update_tg_load_avg(cfs_rq, false);
10037 propagate_entity_cfs_rq(se);
10040 static void detach_task_cfs_rq(struct task_struct *p)
10042 struct sched_entity *se = &p->se;
10043 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10045 if (!vruntime_normalized(p)) {
10047 * Fix up our vruntime so that the current sleep doesn't
10048 * cause 'unlimited' sleep bonus.
10050 place_entity(cfs_rq, se, 0);
10051 se->vruntime -= cfs_rq->min_vruntime;
10054 detach_entity_cfs_rq(se);
10057 static void attach_task_cfs_rq(struct task_struct *p)
10059 struct sched_entity *se = &p->se;
10060 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10062 attach_entity_cfs_rq(se);
10064 if (!vruntime_normalized(p))
10065 se->vruntime += cfs_rq->min_vruntime;
10068 static void switched_from_fair(struct rq *rq, struct task_struct *p)
10070 detach_task_cfs_rq(p);
10073 static void switched_to_fair(struct rq *rq, struct task_struct *p)
10075 attach_task_cfs_rq(p);
10077 if (task_on_rq_queued(p)) {
10079 * We were most likely switched from sched_rt, so
10080 * kick off the schedule if running, otherwise just see
10081 * if we can still preempt the current task.
10086 check_preempt_curr(rq, p, 0);
10090 /* Account for a task changing its policy or group.
10092 * This routine is mostly called to set cfs_rq->curr field when a task
10093 * migrates between groups/classes.
10095 static void set_curr_task_fair(struct rq *rq)
10097 struct sched_entity *se = &rq->curr->se;
10099 for_each_sched_entity(se) {
10100 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10102 set_next_entity(cfs_rq, se);
10103 /* ensure bandwidth has been allocated on our new cfs_rq */
10104 account_cfs_rq_runtime(cfs_rq, 0);
10108 void init_cfs_rq(struct cfs_rq *cfs_rq)
10110 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
10111 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
10112 #ifndef CONFIG_64BIT
10113 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
10116 raw_spin_lock_init(&cfs_rq->removed.lock);
10120 #ifdef CONFIG_FAIR_GROUP_SCHED
10121 static void task_set_group_fair(struct task_struct *p)
10123 struct sched_entity *se = &p->se;
10125 set_task_rq(p, task_cpu(p));
10126 se->depth = se->parent ? se->parent->depth + 1 : 0;
10129 static void task_move_group_fair(struct task_struct *p)
10131 detach_task_cfs_rq(p);
10132 set_task_rq(p, task_cpu(p));
10135 /* Tell se's cfs_rq has been changed -- migrated */
10136 p->se.avg.last_update_time = 0;
10138 attach_task_cfs_rq(p);
10141 static void task_change_group_fair(struct task_struct *p, int type)
10144 case TASK_SET_GROUP:
10145 task_set_group_fair(p);
10148 case TASK_MOVE_GROUP:
10149 task_move_group_fair(p);
10154 void free_fair_sched_group(struct task_group *tg)
10158 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
10160 for_each_possible_cpu(i) {
10162 kfree(tg->cfs_rq[i]);
10171 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10173 struct sched_entity *se;
10174 struct cfs_rq *cfs_rq;
10177 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
10180 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
10184 tg->shares = NICE_0_LOAD;
10186 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
10188 for_each_possible_cpu(i) {
10189 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
10190 GFP_KERNEL, cpu_to_node(i));
10194 se = kzalloc_node(sizeof(struct sched_entity),
10195 GFP_KERNEL, cpu_to_node(i));
10199 init_cfs_rq(cfs_rq);
10200 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
10201 init_entity_runnable_average(se);
10212 void online_fair_sched_group(struct task_group *tg)
10214 struct sched_entity *se;
10218 for_each_possible_cpu(i) {
10222 raw_spin_lock_irq(&rq->lock);
10223 update_rq_clock(rq);
10224 attach_entity_cfs_rq(se);
10225 sync_throttle(tg, i);
10226 raw_spin_unlock_irq(&rq->lock);
10230 void unregister_fair_sched_group(struct task_group *tg)
10232 unsigned long flags;
10236 for_each_possible_cpu(cpu) {
10238 remove_entity_load_avg(tg->se[cpu]);
10241 * Only empty task groups can be destroyed; so we can speculatively
10242 * check on_list without danger of it being re-added.
10244 if (!tg->cfs_rq[cpu]->on_list)
10249 raw_spin_lock_irqsave(&rq->lock, flags);
10250 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
10251 raw_spin_unlock_irqrestore(&rq->lock, flags);
10255 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
10256 struct sched_entity *se, int cpu,
10257 struct sched_entity *parent)
10259 struct rq *rq = cpu_rq(cpu);
10263 init_cfs_rq_runtime(cfs_rq);
10265 tg->cfs_rq[cpu] = cfs_rq;
10268 /* se could be NULL for root_task_group */
10273 se->cfs_rq = &rq->cfs;
10276 se->cfs_rq = parent->my_q;
10277 se->depth = parent->depth + 1;
10281 /* guarantee group entities always have weight */
10282 update_load_set(&se->load, NICE_0_LOAD);
10283 se->parent = parent;
10286 static DEFINE_MUTEX(shares_mutex);
10288 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10293 * We can't change the weight of the root cgroup.
10298 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10300 mutex_lock(&shares_mutex);
10301 if (tg->shares == shares)
10304 tg->shares = shares;
10305 for_each_possible_cpu(i) {
10306 struct rq *rq = cpu_rq(i);
10307 struct sched_entity *se = tg->se[i];
10308 struct rq_flags rf;
10310 /* Propagate contribution to hierarchy */
10311 rq_lock_irqsave(rq, &rf);
10312 update_rq_clock(rq);
10313 for_each_sched_entity(se) {
10314 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
10315 update_cfs_group(se);
10317 rq_unlock_irqrestore(rq, &rf);
10321 mutex_unlock(&shares_mutex);
10324 #else /* CONFIG_FAIR_GROUP_SCHED */
10326 void free_fair_sched_group(struct task_group *tg) { }
10328 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10333 void online_fair_sched_group(struct task_group *tg) { }
10335 void unregister_fair_sched_group(struct task_group *tg) { }
10337 #endif /* CONFIG_FAIR_GROUP_SCHED */
10340 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10342 struct sched_entity *se = &task->se;
10343 unsigned int rr_interval = 0;
10346 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10349 if (rq->cfs.load.weight)
10350 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10352 return rr_interval;
10356 * All the scheduling class methods:
10358 const struct sched_class fair_sched_class = {
10359 .next = &idle_sched_class,
10360 .enqueue_task = enqueue_task_fair,
10361 .dequeue_task = dequeue_task_fair,
10362 .yield_task = yield_task_fair,
10363 .yield_to_task = yield_to_task_fair,
10365 .check_preempt_curr = check_preempt_wakeup,
10367 .pick_next_task = pick_next_task_fair,
10368 .put_prev_task = put_prev_task_fair,
10371 .select_task_rq = select_task_rq_fair,
10372 .migrate_task_rq = migrate_task_rq_fair,
10374 .rq_online = rq_online_fair,
10375 .rq_offline = rq_offline_fair,
10377 .task_dead = task_dead_fair,
10378 .set_cpus_allowed = set_cpus_allowed_common,
10381 .set_curr_task = set_curr_task_fair,
10382 .task_tick = task_tick_fair,
10383 .task_fork = task_fork_fair,
10385 .prio_changed = prio_changed_fair,
10386 .switched_from = switched_from_fair,
10387 .switched_to = switched_to_fair,
10389 .get_rr_interval = get_rr_interval_fair,
10391 .update_curr = update_curr_fair,
10393 #ifdef CONFIG_FAIR_GROUP_SCHED
10394 .task_change_group = task_change_group_fair,
10398 #ifdef CONFIG_SCHED_DEBUG
10399 void print_cfs_stats(struct seq_file *m, int cpu)
10401 struct cfs_rq *cfs_rq, *pos;
10404 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
10405 print_cfs_rq(m, cpu, cfs_rq);
10409 #ifdef CONFIG_NUMA_BALANCING
10410 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10413 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10415 for_each_online_node(node) {
10416 if (p->numa_faults) {
10417 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10418 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10420 if (p->numa_group) {
10421 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
10422 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
10424 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10427 #endif /* CONFIG_NUMA_BALANCING */
10428 #endif /* CONFIG_SCHED_DEBUG */
10430 __init void init_sched_fair_class(void)
10433 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10435 #ifdef CONFIG_NO_HZ_COMMON
10436 nohz.next_balance = jiffies;
10437 nohz.next_blocked = jiffies;
10438 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
10445 * Helper functions to facilitate extracting info from tracepoints.
10448 const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq)
10451 return cfs_rq ? &cfs_rq->avg : NULL;
10456 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg);
10458 char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len)
10462 strlcpy(str, "(null)", len);
10467 cfs_rq_tg_path(cfs_rq, str, len);
10470 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path);
10472 int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq)
10474 return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1;
10476 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu);
10478 const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq)
10481 return rq ? &rq->avg_rt : NULL;
10486 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt);
10488 const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq)
10491 return rq ? &rq->avg_dl : NULL;
10496 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl);
10498 const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq)
10500 #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ)
10501 return rq ? &rq->avg_irq : NULL;
10506 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq);
10508 int sched_trace_rq_cpu(struct rq *rq)
10510 return rq ? cpu_of(rq) : -1;
10512 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu);
10514 const struct cpumask *sched_trace_rd_span(struct root_domain *rd)
10517 return rd ? rd->span : NULL;
10522 EXPORT_SYMBOL_GPL(sched_trace_rd_span);