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 bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
280 struct rq *rq = rq_of(cfs_rq);
281 int cpu = cpu_of(rq);
284 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
289 * Ensure we either appear before our parent (if already
290 * enqueued) or force our parent to appear after us when it is
291 * enqueued. The fact that we always enqueue bottom-up
292 * reduces this to two cases and a special case for the root
293 * cfs_rq. Furthermore, it also means that we will always reset
294 * tmp_alone_branch either when the branch is connected
295 * to a tree or when we reach the top of the tree
297 if (cfs_rq->tg->parent &&
298 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
300 * If parent is already on the list, we add the child
301 * just before. Thanks to circular linked property of
302 * the list, this means to put the child at the tail
303 * of the list that starts by parent.
305 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
306 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
308 * The branch is now connected to its tree so we can
309 * reset tmp_alone_branch to the beginning of the
312 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
316 if (!cfs_rq->tg->parent) {
318 * cfs rq without parent should be put
319 * at the tail of the list.
321 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
322 &rq->leaf_cfs_rq_list);
324 * We have reach the top of a tree so we can reset
325 * tmp_alone_branch to the beginning of the list.
327 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
332 * The parent has not already been added so we want to
333 * make sure that it will be put after us.
334 * tmp_alone_branch points to the begin of the branch
335 * where we will add parent.
337 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
339 * update tmp_alone_branch to points to the new begin
342 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
346 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
348 if (cfs_rq->on_list) {
349 struct rq *rq = rq_of(cfs_rq);
352 * With cfs_rq being unthrottled/throttled during an enqueue,
353 * it can happen the tmp_alone_branch points the a leaf that
354 * we finally want to del. In this case, tmp_alone_branch moves
355 * to the prev element but it will point to rq->leaf_cfs_rq_list
356 * at the end of the enqueue.
358 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
359 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
361 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
366 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
368 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
371 /* Iterate thr' all leaf cfs_rq's on a runqueue */
372 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
373 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
376 /* Do the two (enqueued) entities belong to the same group ? */
377 static inline struct cfs_rq *
378 is_same_group(struct sched_entity *se, struct sched_entity *pse)
380 if (se->cfs_rq == pse->cfs_rq)
386 static inline struct sched_entity *parent_entity(struct sched_entity *se)
392 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
394 int se_depth, pse_depth;
397 * preemption test can be made between sibling entities who are in the
398 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
399 * both tasks until we find their ancestors who are siblings of common
403 /* First walk up until both entities are at same depth */
404 se_depth = (*se)->depth;
405 pse_depth = (*pse)->depth;
407 while (se_depth > pse_depth) {
409 *se = parent_entity(*se);
412 while (pse_depth > se_depth) {
414 *pse = parent_entity(*pse);
417 while (!is_same_group(*se, *pse)) {
418 *se = parent_entity(*se);
419 *pse = parent_entity(*pse);
423 #else /* !CONFIG_FAIR_GROUP_SCHED */
425 static inline struct task_struct *task_of(struct sched_entity *se)
427 return container_of(se, struct task_struct, se);
430 #define for_each_sched_entity(se) \
431 for (; se; se = NULL)
433 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
435 return &task_rq(p)->cfs;
438 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
440 struct task_struct *p = task_of(se);
441 struct rq *rq = task_rq(p);
446 /* runqueue "owned" by this group */
447 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
452 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
457 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
461 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
465 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
466 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
468 static inline struct sched_entity *parent_entity(struct sched_entity *se)
474 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
478 #endif /* CONFIG_FAIR_GROUP_SCHED */
480 static __always_inline
481 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
483 /**************************************************************
484 * Scheduling class tree data structure manipulation methods:
487 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
489 s64 delta = (s64)(vruntime - max_vruntime);
491 max_vruntime = vruntime;
496 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
498 s64 delta = (s64)(vruntime - min_vruntime);
500 min_vruntime = vruntime;
505 static inline int entity_before(struct sched_entity *a,
506 struct sched_entity *b)
508 return (s64)(a->vruntime - b->vruntime) < 0;
511 static void update_min_vruntime(struct cfs_rq *cfs_rq)
513 struct sched_entity *curr = cfs_rq->curr;
514 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
516 u64 vruntime = cfs_rq->min_vruntime;
520 vruntime = curr->vruntime;
525 if (leftmost) { /* non-empty tree */
526 struct sched_entity *se;
527 se = rb_entry(leftmost, struct sched_entity, run_node);
530 vruntime = se->vruntime;
532 vruntime = min_vruntime(vruntime, se->vruntime);
535 /* ensure we never gain time by being placed backwards. */
536 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
539 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
544 * Enqueue an entity into the rb-tree:
546 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
548 struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node;
549 struct rb_node *parent = NULL;
550 struct sched_entity *entry;
551 bool leftmost = true;
554 * Find the right place in the rbtree:
558 entry = rb_entry(parent, struct sched_entity, run_node);
560 * We dont care about collisions. Nodes with
561 * the same key stay together.
563 if (entity_before(se, entry)) {
564 link = &parent->rb_left;
566 link = &parent->rb_right;
571 rb_link_node(&se->run_node, parent, link);
572 rb_insert_color_cached(&se->run_node,
573 &cfs_rq->tasks_timeline, leftmost);
576 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
578 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
581 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
583 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
588 return rb_entry(left, struct sched_entity, run_node);
591 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
593 struct rb_node *next = rb_next(&se->run_node);
598 return rb_entry(next, struct sched_entity, run_node);
601 #ifdef CONFIG_SCHED_DEBUG
602 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
604 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
609 return rb_entry(last, struct sched_entity, run_node);
612 /**************************************************************
613 * Scheduling class statistics methods:
616 int sched_proc_update_handler(struct ctl_table *table, int write,
617 void __user *buffer, size_t *lenp,
620 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
621 unsigned int factor = get_update_sysctl_factor();
626 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
627 sysctl_sched_min_granularity);
629 #define WRT_SYSCTL(name) \
630 (normalized_sysctl_##name = sysctl_##name / (factor))
631 WRT_SYSCTL(sched_min_granularity);
632 WRT_SYSCTL(sched_latency);
633 WRT_SYSCTL(sched_wakeup_granularity);
643 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
645 if (unlikely(se->load.weight != NICE_0_LOAD))
646 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
652 * The idea is to set a period in which each task runs once.
654 * When there are too many tasks (sched_nr_latency) we have to stretch
655 * this period because otherwise the slices get too small.
657 * p = (nr <= nl) ? l : l*nr/nl
659 static u64 __sched_period(unsigned long nr_running)
661 if (unlikely(nr_running > sched_nr_latency))
662 return nr_running * sysctl_sched_min_granularity;
664 return sysctl_sched_latency;
668 * We calculate the wall-time slice from the period by taking a part
669 * proportional to the weight.
673 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
675 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
677 for_each_sched_entity(se) {
678 struct load_weight *load;
679 struct load_weight lw;
681 cfs_rq = cfs_rq_of(se);
682 load = &cfs_rq->load;
684 if (unlikely(!se->on_rq)) {
687 update_load_add(&lw, se->load.weight);
690 slice = __calc_delta(slice, se->load.weight, load);
696 * We calculate the vruntime slice of a to-be-inserted task.
700 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
702 return calc_delta_fair(sched_slice(cfs_rq, se), se);
708 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
709 static unsigned long task_h_load(struct task_struct *p);
710 static unsigned long capacity_of(int cpu);
712 /* Give new sched_entity start runnable values to heavy its load in infant time */
713 void init_entity_runnable_average(struct sched_entity *se)
715 struct sched_avg *sa = &se->avg;
717 memset(sa, 0, sizeof(*sa));
720 * Tasks are initialized with full load to be seen as heavy tasks until
721 * they get a chance to stabilize to their real load level.
722 * Group entities are initialized with zero load to reflect the fact that
723 * nothing has been attached to the task group yet.
725 if (entity_is_task(se))
726 sa->runnable_load_avg = sa->load_avg = scale_load_down(se->load.weight);
728 se->runnable_weight = se->load.weight;
730 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
733 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
734 static void attach_entity_cfs_rq(struct sched_entity *se);
737 * With new tasks being created, their initial util_avgs are extrapolated
738 * based on the cfs_rq's current util_avg:
740 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
742 * However, in many cases, the above util_avg does not give a desired
743 * value. Moreover, the sum of the util_avgs may be divergent, such
744 * as when the series is a harmonic series.
746 * To solve this problem, we also cap the util_avg of successive tasks to
747 * only 1/2 of the left utilization budget:
749 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
751 * where n denotes the nth task and cpu_scale the CPU capacity.
753 * For example, for a CPU with 1024 of capacity, a simplest series from
754 * the beginning would be like:
756 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
757 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
759 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
760 * if util_avg > util_avg_cap.
762 void post_init_entity_util_avg(struct task_struct *p)
764 struct sched_entity *se = &p->se;
765 struct cfs_rq *cfs_rq = cfs_rq_of(se);
766 struct sched_avg *sa = &se->avg;
767 long cpu_scale = arch_scale_cpu_capacity(NULL, cpu_of(rq_of(cfs_rq)));
768 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
771 if (cfs_rq->avg.util_avg != 0) {
772 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
773 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
775 if (sa->util_avg > cap)
782 if (p->sched_class != &fair_sched_class) {
784 * For !fair tasks do:
786 update_cfs_rq_load_avg(now, cfs_rq);
787 attach_entity_load_avg(cfs_rq, se, 0);
788 switched_from_fair(rq, p);
790 * such that the next switched_to_fair() has the
793 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
797 attach_entity_cfs_rq(se);
800 #else /* !CONFIG_SMP */
801 void init_entity_runnable_average(struct sched_entity *se)
804 void post_init_entity_util_avg(struct task_struct *p)
807 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
810 #endif /* CONFIG_SMP */
813 * Update the current task's runtime statistics.
815 static void update_curr(struct cfs_rq *cfs_rq)
817 struct sched_entity *curr = cfs_rq->curr;
818 u64 now = rq_clock_task(rq_of(cfs_rq));
824 delta_exec = now - curr->exec_start;
825 if (unlikely((s64)delta_exec <= 0))
828 curr->exec_start = now;
830 schedstat_set(curr->statistics.exec_max,
831 max(delta_exec, curr->statistics.exec_max));
833 curr->sum_exec_runtime += delta_exec;
834 schedstat_add(cfs_rq->exec_clock, delta_exec);
836 curr->vruntime += calc_delta_fair(delta_exec, curr);
837 update_min_vruntime(cfs_rq);
839 if (entity_is_task(curr)) {
840 struct task_struct *curtask = task_of(curr);
842 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
843 cgroup_account_cputime(curtask, delta_exec);
844 account_group_exec_runtime(curtask, delta_exec);
847 account_cfs_rq_runtime(cfs_rq, delta_exec);
850 static void update_curr_fair(struct rq *rq)
852 update_curr(cfs_rq_of(&rq->curr->se));
856 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
858 u64 wait_start, prev_wait_start;
860 if (!schedstat_enabled())
863 wait_start = rq_clock(rq_of(cfs_rq));
864 prev_wait_start = schedstat_val(se->statistics.wait_start);
866 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
867 likely(wait_start > prev_wait_start))
868 wait_start -= prev_wait_start;
870 __schedstat_set(se->statistics.wait_start, wait_start);
874 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
876 struct task_struct *p;
879 if (!schedstat_enabled())
882 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
884 if (entity_is_task(se)) {
886 if (task_on_rq_migrating(p)) {
888 * Preserve migrating task's wait time so wait_start
889 * time stamp can be adjusted to accumulate wait time
890 * prior to migration.
892 __schedstat_set(se->statistics.wait_start, delta);
895 trace_sched_stat_wait(p, delta);
898 __schedstat_set(se->statistics.wait_max,
899 max(schedstat_val(se->statistics.wait_max), delta));
900 __schedstat_inc(se->statistics.wait_count);
901 __schedstat_add(se->statistics.wait_sum, delta);
902 __schedstat_set(se->statistics.wait_start, 0);
906 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
908 struct task_struct *tsk = NULL;
909 u64 sleep_start, block_start;
911 if (!schedstat_enabled())
914 sleep_start = schedstat_val(se->statistics.sleep_start);
915 block_start = schedstat_val(se->statistics.block_start);
917 if (entity_is_task(se))
921 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
926 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
927 __schedstat_set(se->statistics.sleep_max, delta);
929 __schedstat_set(se->statistics.sleep_start, 0);
930 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
933 account_scheduler_latency(tsk, delta >> 10, 1);
934 trace_sched_stat_sleep(tsk, delta);
938 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
943 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
944 __schedstat_set(se->statistics.block_max, delta);
946 __schedstat_set(se->statistics.block_start, 0);
947 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
950 if (tsk->in_iowait) {
951 __schedstat_add(se->statistics.iowait_sum, delta);
952 __schedstat_inc(se->statistics.iowait_count);
953 trace_sched_stat_iowait(tsk, delta);
956 trace_sched_stat_blocked(tsk, delta);
959 * Blocking time is in units of nanosecs, so shift by
960 * 20 to get a milliseconds-range estimation of the
961 * amount of time that the task spent sleeping:
963 if (unlikely(prof_on == SLEEP_PROFILING)) {
964 profile_hits(SLEEP_PROFILING,
965 (void *)get_wchan(tsk),
968 account_scheduler_latency(tsk, delta >> 10, 0);
974 * Task is being enqueued - update stats:
977 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
979 if (!schedstat_enabled())
983 * Are we enqueueing a waiting task? (for current tasks
984 * a dequeue/enqueue event is a NOP)
986 if (se != cfs_rq->curr)
987 update_stats_wait_start(cfs_rq, se);
989 if (flags & ENQUEUE_WAKEUP)
990 update_stats_enqueue_sleeper(cfs_rq, se);
994 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
997 if (!schedstat_enabled())
1001 * Mark the end of the wait period if dequeueing a
1004 if (se != cfs_rq->curr)
1005 update_stats_wait_end(cfs_rq, se);
1007 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1008 struct task_struct *tsk = task_of(se);
1010 if (tsk->state & TASK_INTERRUPTIBLE)
1011 __schedstat_set(se->statistics.sleep_start,
1012 rq_clock(rq_of(cfs_rq)));
1013 if (tsk->state & TASK_UNINTERRUPTIBLE)
1014 __schedstat_set(se->statistics.block_start,
1015 rq_clock(rq_of(cfs_rq)));
1020 * We are picking a new current task - update its stats:
1023 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1026 * We are starting a new run period:
1028 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1031 /**************************************************
1032 * Scheduling class queueing methods:
1035 #ifdef CONFIG_NUMA_BALANCING
1037 * Approximate time to scan a full NUMA task in ms. The task scan period is
1038 * calculated based on the tasks virtual memory size and
1039 * numa_balancing_scan_size.
1041 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1042 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1044 /* Portion of address space to scan in MB */
1045 unsigned int sysctl_numa_balancing_scan_size = 256;
1047 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1048 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1051 refcount_t refcount;
1053 spinlock_t lock; /* nr_tasks, tasks */
1058 struct rcu_head rcu;
1059 unsigned long total_faults;
1060 unsigned long max_faults_cpu;
1062 * Faults_cpu is used to decide whether memory should move
1063 * towards the CPU. As a consequence, these stats are weighted
1064 * more by CPU use than by memory faults.
1066 unsigned long *faults_cpu;
1067 unsigned long faults[0];
1070 static inline unsigned long group_faults_priv(struct numa_group *ng);
1071 static inline unsigned long group_faults_shared(struct numa_group *ng);
1073 static unsigned int task_nr_scan_windows(struct task_struct *p)
1075 unsigned long rss = 0;
1076 unsigned long nr_scan_pages;
1079 * Calculations based on RSS as non-present and empty pages are skipped
1080 * by the PTE scanner and NUMA hinting faults should be trapped based
1083 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1084 rss = get_mm_rss(p->mm);
1086 rss = nr_scan_pages;
1088 rss = round_up(rss, nr_scan_pages);
1089 return rss / nr_scan_pages;
1092 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1093 #define MAX_SCAN_WINDOW 2560
1095 static unsigned int task_scan_min(struct task_struct *p)
1097 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1098 unsigned int scan, floor;
1099 unsigned int windows = 1;
1101 if (scan_size < MAX_SCAN_WINDOW)
1102 windows = MAX_SCAN_WINDOW / scan_size;
1103 floor = 1000 / windows;
1105 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1106 return max_t(unsigned int, floor, scan);
1109 static unsigned int task_scan_start(struct task_struct *p)
1111 unsigned long smin = task_scan_min(p);
1112 unsigned long period = smin;
1114 /* Scale the maximum scan period with the amount of shared memory. */
1115 if (p->numa_group) {
1116 struct numa_group *ng = p->numa_group;
1117 unsigned long shared = group_faults_shared(ng);
1118 unsigned long private = group_faults_priv(ng);
1120 period *= refcount_read(&ng->refcount);
1121 period *= shared + 1;
1122 period /= private + shared + 1;
1125 return max(smin, period);
1128 static unsigned int task_scan_max(struct task_struct *p)
1130 unsigned long smin = task_scan_min(p);
1133 /* Watch for min being lower than max due to floor calculations */
1134 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1136 /* Scale the maximum scan period with the amount of shared memory. */
1137 if (p->numa_group) {
1138 struct numa_group *ng = p->numa_group;
1139 unsigned long shared = group_faults_shared(ng);
1140 unsigned long private = group_faults_priv(ng);
1141 unsigned long period = smax;
1143 period *= refcount_read(&ng->refcount);
1144 period *= shared + 1;
1145 period /= private + shared + 1;
1147 smax = max(smax, period);
1150 return max(smin, smax);
1153 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
1156 struct mm_struct *mm = p->mm;
1159 mm_users = atomic_read(&mm->mm_users);
1160 if (mm_users == 1) {
1161 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1162 mm->numa_scan_seq = 0;
1166 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
1167 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1168 p->numa_work.next = &p->numa_work;
1169 p->numa_faults = NULL;
1170 p->numa_group = NULL;
1171 p->last_task_numa_placement = 0;
1172 p->last_sum_exec_runtime = 0;
1174 /* New address space, reset the preferred nid */
1175 if (!(clone_flags & CLONE_VM)) {
1176 p->numa_preferred_nid = -1;
1181 * New thread, keep existing numa_preferred_nid which should be copied
1182 * already by arch_dup_task_struct but stagger when scans start.
1187 delay = min_t(unsigned int, task_scan_max(current),
1188 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
1189 delay += 2 * TICK_NSEC;
1190 p->node_stamp = delay;
1194 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1196 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1197 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1200 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1202 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1203 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1206 /* Shared or private faults. */
1207 #define NR_NUMA_HINT_FAULT_TYPES 2
1209 /* Memory and CPU locality */
1210 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1212 /* Averaged statistics, and temporary buffers. */
1213 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1215 pid_t task_numa_group_id(struct task_struct *p)
1217 return p->numa_group ? p->numa_group->gid : 0;
1221 * The averaged statistics, shared & private, memory & CPU,
1222 * occupy the first half of the array. The second half of the
1223 * array is for current counters, which are averaged into the
1224 * first set by task_numa_placement.
1226 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1228 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1231 static inline unsigned long task_faults(struct task_struct *p, int nid)
1233 if (!p->numa_faults)
1236 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1237 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1240 static inline unsigned long group_faults(struct task_struct *p, int nid)
1245 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1246 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1249 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1251 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1252 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1255 static inline unsigned long group_faults_priv(struct numa_group *ng)
1257 unsigned long faults = 0;
1260 for_each_online_node(node) {
1261 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1267 static inline unsigned long group_faults_shared(struct numa_group *ng)
1269 unsigned long faults = 0;
1272 for_each_online_node(node) {
1273 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1280 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1281 * considered part of a numa group's pseudo-interleaving set. Migrations
1282 * between these nodes are slowed down, to allow things to settle down.
1284 #define ACTIVE_NODE_FRACTION 3
1286 static bool numa_is_active_node(int nid, struct numa_group *ng)
1288 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1291 /* Handle placement on systems where not all nodes are directly connected. */
1292 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1293 int maxdist, bool task)
1295 unsigned long score = 0;
1299 * All nodes are directly connected, and the same distance
1300 * from each other. No need for fancy placement algorithms.
1302 if (sched_numa_topology_type == NUMA_DIRECT)
1306 * This code is called for each node, introducing N^2 complexity,
1307 * which should be ok given the number of nodes rarely exceeds 8.
1309 for_each_online_node(node) {
1310 unsigned long faults;
1311 int dist = node_distance(nid, node);
1314 * The furthest away nodes in the system are not interesting
1315 * for placement; nid was already counted.
1317 if (dist == sched_max_numa_distance || node == nid)
1321 * On systems with a backplane NUMA topology, compare groups
1322 * of nodes, and move tasks towards the group with the most
1323 * memory accesses. When comparing two nodes at distance
1324 * "hoplimit", only nodes closer by than "hoplimit" are part
1325 * of each group. Skip other nodes.
1327 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1331 /* Add up the faults from nearby nodes. */
1333 faults = task_faults(p, node);
1335 faults = group_faults(p, node);
1338 * On systems with a glueless mesh NUMA topology, there are
1339 * no fixed "groups of nodes". Instead, nodes that are not
1340 * directly connected bounce traffic through intermediate
1341 * nodes; a numa_group can occupy any set of nodes.
1342 * The further away a node is, the less the faults count.
1343 * This seems to result in good task placement.
1345 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1346 faults *= (sched_max_numa_distance - dist);
1347 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1357 * These return the fraction of accesses done by a particular task, or
1358 * task group, on a particular numa node. The group weight is given a
1359 * larger multiplier, in order to group tasks together that are almost
1360 * evenly spread out between numa nodes.
1362 static inline unsigned long task_weight(struct task_struct *p, int nid,
1365 unsigned long faults, total_faults;
1367 if (!p->numa_faults)
1370 total_faults = p->total_numa_faults;
1375 faults = task_faults(p, nid);
1376 faults += score_nearby_nodes(p, nid, dist, true);
1378 return 1000 * faults / total_faults;
1381 static inline unsigned long group_weight(struct task_struct *p, int nid,
1384 unsigned long faults, total_faults;
1389 total_faults = p->numa_group->total_faults;
1394 faults = group_faults(p, nid);
1395 faults += score_nearby_nodes(p, nid, dist, false);
1397 return 1000 * faults / total_faults;
1400 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1401 int src_nid, int dst_cpu)
1403 struct numa_group *ng = p->numa_group;
1404 int dst_nid = cpu_to_node(dst_cpu);
1405 int last_cpupid, this_cpupid;
1407 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1408 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1411 * Allow first faults or private faults to migrate immediately early in
1412 * the lifetime of a task. The magic number 4 is based on waiting for
1413 * two full passes of the "multi-stage node selection" test that is
1416 if ((p->numa_preferred_nid == -1 || p->numa_scan_seq <= 4) &&
1417 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1421 * Multi-stage node selection is used in conjunction with a periodic
1422 * migration fault to build a temporal task<->page relation. By using
1423 * a two-stage filter we remove short/unlikely relations.
1425 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1426 * a task's usage of a particular page (n_p) per total usage of this
1427 * page (n_t) (in a given time-span) to a probability.
1429 * Our periodic faults will sample this probability and getting the
1430 * same result twice in a row, given these samples are fully
1431 * independent, is then given by P(n)^2, provided our sample period
1432 * is sufficiently short compared to the usage pattern.
1434 * This quadric squishes small probabilities, making it less likely we
1435 * act on an unlikely task<->page relation.
1437 if (!cpupid_pid_unset(last_cpupid) &&
1438 cpupid_to_nid(last_cpupid) != dst_nid)
1441 /* Always allow migrate on private faults */
1442 if (cpupid_match_pid(p, last_cpupid))
1445 /* A shared fault, but p->numa_group has not been set up yet. */
1450 * Destination node is much more heavily used than the source
1451 * node? Allow migration.
1453 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1454 ACTIVE_NODE_FRACTION)
1458 * Distribute memory according to CPU & memory use on each node,
1459 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1461 * faults_cpu(dst) 3 faults_cpu(src)
1462 * --------------- * - > ---------------
1463 * faults_mem(dst) 4 faults_mem(src)
1465 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1466 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1469 static unsigned long weighted_cpuload(struct rq *rq);
1470 static unsigned long source_load(int cpu, int type);
1471 static unsigned long target_load(int cpu, int type);
1473 /* Cached statistics for all CPUs within a node */
1477 /* Total compute capacity of CPUs on a node */
1478 unsigned long compute_capacity;
1482 * XXX borrowed from update_sg_lb_stats
1484 static void update_numa_stats(struct numa_stats *ns, int nid)
1488 memset(ns, 0, sizeof(*ns));
1489 for_each_cpu(cpu, cpumask_of_node(nid)) {
1490 struct rq *rq = cpu_rq(cpu);
1492 ns->load += weighted_cpuload(rq);
1493 ns->compute_capacity += capacity_of(cpu);
1498 struct task_numa_env {
1499 struct task_struct *p;
1501 int src_cpu, src_nid;
1502 int dst_cpu, dst_nid;
1504 struct numa_stats src_stats, dst_stats;
1509 struct task_struct *best_task;
1514 static void task_numa_assign(struct task_numa_env *env,
1515 struct task_struct *p, long imp)
1517 struct rq *rq = cpu_rq(env->dst_cpu);
1519 /* Bail out if run-queue part of active NUMA balance. */
1520 if (xchg(&rq->numa_migrate_on, 1))
1524 * Clear previous best_cpu/rq numa-migrate flag, since task now
1525 * found a better CPU to move/swap.
1527 if (env->best_cpu != -1) {
1528 rq = cpu_rq(env->best_cpu);
1529 WRITE_ONCE(rq->numa_migrate_on, 0);
1533 put_task_struct(env->best_task);
1538 env->best_imp = imp;
1539 env->best_cpu = env->dst_cpu;
1542 static bool load_too_imbalanced(long src_load, long dst_load,
1543 struct task_numa_env *env)
1546 long orig_src_load, orig_dst_load;
1547 long src_capacity, dst_capacity;
1550 * The load is corrected for the CPU capacity available on each node.
1553 * ------------ vs ---------
1554 * src_capacity dst_capacity
1556 src_capacity = env->src_stats.compute_capacity;
1557 dst_capacity = env->dst_stats.compute_capacity;
1559 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1561 orig_src_load = env->src_stats.load;
1562 orig_dst_load = env->dst_stats.load;
1564 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1566 /* Would this change make things worse? */
1567 return (imb > old_imb);
1571 * Maximum NUMA importance can be 1998 (2*999);
1572 * SMALLIMP @ 30 would be close to 1998/64.
1573 * Used to deter task migration.
1578 * This checks if the overall compute and NUMA accesses of the system would
1579 * be improved if the source tasks was migrated to the target dst_cpu taking
1580 * into account that it might be best if task running on the dst_cpu should
1581 * be exchanged with the source task
1583 static void task_numa_compare(struct task_numa_env *env,
1584 long taskimp, long groupimp, bool maymove)
1586 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1587 struct task_struct *cur;
1588 long src_load, dst_load;
1590 long imp = env->p->numa_group ? groupimp : taskimp;
1592 int dist = env->dist;
1594 if (READ_ONCE(dst_rq->numa_migrate_on))
1598 cur = task_rcu_dereference(&dst_rq->curr);
1599 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1603 * Because we have preemption enabled we can get migrated around and
1604 * end try selecting ourselves (current == env->p) as a swap candidate.
1610 if (maymove && moveimp >= env->best_imp)
1617 * "imp" is the fault differential for the source task between the
1618 * source and destination node. Calculate the total differential for
1619 * the source task and potential destination task. The more negative
1620 * the value is, the more remote accesses that would be expected to
1621 * be incurred if the tasks were swapped.
1623 /* Skip this swap candidate if cannot move to the source cpu */
1624 if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed))
1628 * If dst and source tasks are in the same NUMA group, or not
1629 * in any group then look only at task weights.
1631 if (cur->numa_group == env->p->numa_group) {
1632 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1633 task_weight(cur, env->dst_nid, dist);
1635 * Add some hysteresis to prevent swapping the
1636 * tasks within a group over tiny differences.
1638 if (cur->numa_group)
1642 * Compare the group weights. If a task is all by itself
1643 * (not part of a group), use the task weight instead.
1645 if (cur->numa_group && env->p->numa_group)
1646 imp += group_weight(cur, env->src_nid, dist) -
1647 group_weight(cur, env->dst_nid, dist);
1649 imp += task_weight(cur, env->src_nid, dist) -
1650 task_weight(cur, env->dst_nid, dist);
1653 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1660 * If the NUMA importance is less than SMALLIMP,
1661 * task migration might only result in ping pong
1662 * of tasks and also hurt performance due to cache
1665 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1669 * In the overloaded case, try and keep the load balanced.
1671 load = task_h_load(env->p) - task_h_load(cur);
1675 dst_load = env->dst_stats.load + load;
1676 src_load = env->src_stats.load - load;
1678 if (load_too_imbalanced(src_load, dst_load, env))
1683 * One idle CPU per node is evaluated for a task numa move.
1684 * Call select_idle_sibling to maybe find a better one.
1688 * select_idle_siblings() uses an per-CPU cpumask that
1689 * can be used from IRQ context.
1691 local_irq_disable();
1692 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1697 task_numa_assign(env, cur, imp);
1702 static void task_numa_find_cpu(struct task_numa_env *env,
1703 long taskimp, long groupimp)
1705 long src_load, dst_load, load;
1706 bool maymove = false;
1709 load = task_h_load(env->p);
1710 dst_load = env->dst_stats.load + load;
1711 src_load = env->src_stats.load - load;
1714 * If the improvement from just moving env->p direction is better
1715 * than swapping tasks around, check if a move is possible.
1717 maymove = !load_too_imbalanced(src_load, dst_load, env);
1719 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1720 /* Skip this CPU if the source task cannot migrate */
1721 if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
1725 task_numa_compare(env, taskimp, groupimp, maymove);
1729 static int task_numa_migrate(struct task_struct *p)
1731 struct task_numa_env env = {
1734 .src_cpu = task_cpu(p),
1735 .src_nid = task_node(p),
1737 .imbalance_pct = 112,
1743 struct sched_domain *sd;
1745 unsigned long taskweight, groupweight;
1747 long taskimp, groupimp;
1750 * Pick the lowest SD_NUMA domain, as that would have the smallest
1751 * imbalance and would be the first to start moving tasks about.
1753 * And we want to avoid any moving of tasks about, as that would create
1754 * random movement of tasks -- counter the numa conditions we're trying
1758 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1760 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1764 * Cpusets can break the scheduler domain tree into smaller
1765 * balance domains, some of which do not cross NUMA boundaries.
1766 * Tasks that are "trapped" in such domains cannot be migrated
1767 * elsewhere, so there is no point in (re)trying.
1769 if (unlikely(!sd)) {
1770 sched_setnuma(p, task_node(p));
1774 env.dst_nid = p->numa_preferred_nid;
1775 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1776 taskweight = task_weight(p, env.src_nid, dist);
1777 groupweight = group_weight(p, env.src_nid, dist);
1778 update_numa_stats(&env.src_stats, env.src_nid);
1779 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1780 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1781 update_numa_stats(&env.dst_stats, env.dst_nid);
1783 /* Try to find a spot on the preferred nid. */
1784 task_numa_find_cpu(&env, taskimp, groupimp);
1787 * Look at other nodes in these cases:
1788 * - there is no space available on the preferred_nid
1789 * - the task is part of a numa_group that is interleaved across
1790 * multiple NUMA nodes; in order to better consolidate the group,
1791 * we need to check other locations.
1793 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1794 for_each_online_node(nid) {
1795 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1798 dist = node_distance(env.src_nid, env.dst_nid);
1799 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1801 taskweight = task_weight(p, env.src_nid, dist);
1802 groupweight = group_weight(p, env.src_nid, dist);
1805 /* Only consider nodes where both task and groups benefit */
1806 taskimp = task_weight(p, nid, dist) - taskweight;
1807 groupimp = group_weight(p, nid, dist) - groupweight;
1808 if (taskimp < 0 && groupimp < 0)
1813 update_numa_stats(&env.dst_stats, env.dst_nid);
1814 task_numa_find_cpu(&env, taskimp, groupimp);
1819 * If the task is part of a workload that spans multiple NUMA nodes,
1820 * and is migrating into one of the workload's active nodes, remember
1821 * this node as the task's preferred numa node, so the workload can
1823 * A task that migrated to a second choice node will be better off
1824 * trying for a better one later. Do not set the preferred node here.
1826 if (p->numa_group) {
1827 if (env.best_cpu == -1)
1830 nid = cpu_to_node(env.best_cpu);
1832 if (nid != p->numa_preferred_nid)
1833 sched_setnuma(p, nid);
1836 /* No better CPU than the current one was found. */
1837 if (env.best_cpu == -1)
1840 best_rq = cpu_rq(env.best_cpu);
1841 if (env.best_task == NULL) {
1842 ret = migrate_task_to(p, env.best_cpu);
1843 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1845 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1849 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
1850 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1853 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1854 put_task_struct(env.best_task);
1858 /* Attempt to migrate a task to a CPU on the preferred node. */
1859 static void numa_migrate_preferred(struct task_struct *p)
1861 unsigned long interval = HZ;
1863 /* This task has no NUMA fault statistics yet */
1864 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1867 /* Periodically retry migrating the task to the preferred node */
1868 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1869 p->numa_migrate_retry = jiffies + interval;
1871 /* Success if task is already running on preferred CPU */
1872 if (task_node(p) == p->numa_preferred_nid)
1875 /* Otherwise, try migrate to a CPU on the preferred node */
1876 task_numa_migrate(p);
1880 * Find out how many nodes on the workload is actively running on. Do this by
1881 * tracking the nodes from which NUMA hinting faults are triggered. This can
1882 * be different from the set of nodes where the workload's memory is currently
1885 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1887 unsigned long faults, max_faults = 0;
1888 int nid, active_nodes = 0;
1890 for_each_online_node(nid) {
1891 faults = group_faults_cpu(numa_group, nid);
1892 if (faults > max_faults)
1893 max_faults = faults;
1896 for_each_online_node(nid) {
1897 faults = group_faults_cpu(numa_group, nid);
1898 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1902 numa_group->max_faults_cpu = max_faults;
1903 numa_group->active_nodes = active_nodes;
1907 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1908 * increments. The more local the fault statistics are, the higher the scan
1909 * period will be for the next scan window. If local/(local+remote) ratio is
1910 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1911 * the scan period will decrease. Aim for 70% local accesses.
1913 #define NUMA_PERIOD_SLOTS 10
1914 #define NUMA_PERIOD_THRESHOLD 7
1917 * Increase the scan period (slow down scanning) if the majority of
1918 * our memory is already on our local node, or if the majority of
1919 * the page accesses are shared with other processes.
1920 * Otherwise, decrease the scan period.
1922 static void update_task_scan_period(struct task_struct *p,
1923 unsigned long shared, unsigned long private)
1925 unsigned int period_slot;
1926 int lr_ratio, ps_ratio;
1929 unsigned long remote = p->numa_faults_locality[0];
1930 unsigned long local = p->numa_faults_locality[1];
1933 * If there were no record hinting faults then either the task is
1934 * completely idle or all activity is areas that are not of interest
1935 * to automatic numa balancing. Related to that, if there were failed
1936 * migration then it implies we are migrating too quickly or the local
1937 * node is overloaded. In either case, scan slower
1939 if (local + shared == 0 || p->numa_faults_locality[2]) {
1940 p->numa_scan_period = min(p->numa_scan_period_max,
1941 p->numa_scan_period << 1);
1943 p->mm->numa_next_scan = jiffies +
1944 msecs_to_jiffies(p->numa_scan_period);
1950 * Prepare to scale scan period relative to the current period.
1951 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1952 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1953 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1955 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1956 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1957 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1959 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1961 * Most memory accesses are local. There is no need to
1962 * do fast NUMA scanning, since memory is already local.
1964 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
1967 diff = slot * period_slot;
1968 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
1970 * Most memory accesses are shared with other tasks.
1971 * There is no point in continuing fast NUMA scanning,
1972 * since other tasks may just move the memory elsewhere.
1974 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
1977 diff = slot * period_slot;
1980 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
1981 * yet they are not on the local NUMA node. Speed up
1982 * NUMA scanning to get the memory moved over.
1984 int ratio = max(lr_ratio, ps_ratio);
1985 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1988 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1989 task_scan_min(p), task_scan_max(p));
1990 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1994 * Get the fraction of time the task has been running since the last
1995 * NUMA placement cycle. The scheduler keeps similar statistics, but
1996 * decays those on a 32ms period, which is orders of magnitude off
1997 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1998 * stats only if the task is so new there are no NUMA statistics yet.
2000 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2002 u64 runtime, delta, now;
2003 /* Use the start of this time slice to avoid calculations. */
2004 now = p->se.exec_start;
2005 runtime = p->se.sum_exec_runtime;
2007 if (p->last_task_numa_placement) {
2008 delta = runtime - p->last_sum_exec_runtime;
2009 *period = now - p->last_task_numa_placement;
2011 delta = p->se.avg.load_sum;
2012 *period = LOAD_AVG_MAX;
2015 p->last_sum_exec_runtime = runtime;
2016 p->last_task_numa_placement = now;
2022 * Determine the preferred nid for a task in a numa_group. This needs to
2023 * be done in a way that produces consistent results with group_weight,
2024 * otherwise workloads might not converge.
2026 static int preferred_group_nid(struct task_struct *p, int nid)
2031 /* Direct connections between all NUMA nodes. */
2032 if (sched_numa_topology_type == NUMA_DIRECT)
2036 * On a system with glueless mesh NUMA topology, group_weight
2037 * scores nodes according to the number of NUMA hinting faults on
2038 * both the node itself, and on nearby nodes.
2040 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2041 unsigned long score, max_score = 0;
2042 int node, max_node = nid;
2044 dist = sched_max_numa_distance;
2046 for_each_online_node(node) {
2047 score = group_weight(p, node, dist);
2048 if (score > max_score) {
2057 * Finding the preferred nid in a system with NUMA backplane
2058 * interconnect topology is more involved. The goal is to locate
2059 * tasks from numa_groups near each other in the system, and
2060 * untangle workloads from different sides of the system. This requires
2061 * searching down the hierarchy of node groups, recursively searching
2062 * inside the highest scoring group of nodes. The nodemask tricks
2063 * keep the complexity of the search down.
2065 nodes = node_online_map;
2066 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2067 unsigned long max_faults = 0;
2068 nodemask_t max_group = NODE_MASK_NONE;
2071 /* Are there nodes at this distance from each other? */
2072 if (!find_numa_distance(dist))
2075 for_each_node_mask(a, nodes) {
2076 unsigned long faults = 0;
2077 nodemask_t this_group;
2078 nodes_clear(this_group);
2080 /* Sum group's NUMA faults; includes a==b case. */
2081 for_each_node_mask(b, nodes) {
2082 if (node_distance(a, b) < dist) {
2083 faults += group_faults(p, b);
2084 node_set(b, this_group);
2085 node_clear(b, nodes);
2089 /* Remember the top group. */
2090 if (faults > max_faults) {
2091 max_faults = faults;
2092 max_group = this_group;
2094 * subtle: at the smallest distance there is
2095 * just one node left in each "group", the
2096 * winner is the preferred nid.
2101 /* Next round, evaluate the nodes within max_group. */
2109 static void task_numa_placement(struct task_struct *p)
2111 int seq, nid, max_nid = -1;
2112 unsigned long max_faults = 0;
2113 unsigned long fault_types[2] = { 0, 0 };
2114 unsigned long total_faults;
2115 u64 runtime, period;
2116 spinlock_t *group_lock = NULL;
2119 * The p->mm->numa_scan_seq field gets updated without
2120 * exclusive access. Use READ_ONCE() here to ensure
2121 * that the field is read in a single access:
2123 seq = READ_ONCE(p->mm->numa_scan_seq);
2124 if (p->numa_scan_seq == seq)
2126 p->numa_scan_seq = seq;
2127 p->numa_scan_period_max = task_scan_max(p);
2129 total_faults = p->numa_faults_locality[0] +
2130 p->numa_faults_locality[1];
2131 runtime = numa_get_avg_runtime(p, &period);
2133 /* If the task is part of a group prevent parallel updates to group stats */
2134 if (p->numa_group) {
2135 group_lock = &p->numa_group->lock;
2136 spin_lock_irq(group_lock);
2139 /* Find the node with the highest number of faults */
2140 for_each_online_node(nid) {
2141 /* Keep track of the offsets in numa_faults array */
2142 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2143 unsigned long faults = 0, group_faults = 0;
2146 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2147 long diff, f_diff, f_weight;
2149 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2150 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2151 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2152 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2154 /* Decay existing window, copy faults since last scan */
2155 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2156 fault_types[priv] += p->numa_faults[membuf_idx];
2157 p->numa_faults[membuf_idx] = 0;
2160 * Normalize the faults_from, so all tasks in a group
2161 * count according to CPU use, instead of by the raw
2162 * number of faults. Tasks with little runtime have
2163 * little over-all impact on throughput, and thus their
2164 * faults are less important.
2166 f_weight = div64_u64(runtime << 16, period + 1);
2167 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2169 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2170 p->numa_faults[cpubuf_idx] = 0;
2172 p->numa_faults[mem_idx] += diff;
2173 p->numa_faults[cpu_idx] += f_diff;
2174 faults += p->numa_faults[mem_idx];
2175 p->total_numa_faults += diff;
2176 if (p->numa_group) {
2178 * safe because we can only change our own group
2180 * mem_idx represents the offset for a given
2181 * nid and priv in a specific region because it
2182 * is at the beginning of the numa_faults array.
2184 p->numa_group->faults[mem_idx] += diff;
2185 p->numa_group->faults_cpu[mem_idx] += f_diff;
2186 p->numa_group->total_faults += diff;
2187 group_faults += p->numa_group->faults[mem_idx];
2191 if (!p->numa_group) {
2192 if (faults > max_faults) {
2193 max_faults = faults;
2196 } else if (group_faults > max_faults) {
2197 max_faults = group_faults;
2202 if (p->numa_group) {
2203 numa_group_count_active_nodes(p->numa_group);
2204 spin_unlock_irq(group_lock);
2205 max_nid = preferred_group_nid(p, max_nid);
2209 /* Set the new preferred node */
2210 if (max_nid != p->numa_preferred_nid)
2211 sched_setnuma(p, max_nid);
2214 update_task_scan_period(p, fault_types[0], fault_types[1]);
2217 static inline int get_numa_group(struct numa_group *grp)
2219 return refcount_inc_not_zero(&grp->refcount);
2222 static inline void put_numa_group(struct numa_group *grp)
2224 if (refcount_dec_and_test(&grp->refcount))
2225 kfree_rcu(grp, rcu);
2228 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2231 struct numa_group *grp, *my_grp;
2232 struct task_struct *tsk;
2234 int cpu = cpupid_to_cpu(cpupid);
2237 if (unlikely(!p->numa_group)) {
2238 unsigned int size = sizeof(struct numa_group) +
2239 4*nr_node_ids*sizeof(unsigned long);
2241 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2245 refcount_set(&grp->refcount, 1);
2246 grp->active_nodes = 1;
2247 grp->max_faults_cpu = 0;
2248 spin_lock_init(&grp->lock);
2250 /* Second half of the array tracks nids where faults happen */
2251 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2254 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2255 grp->faults[i] = p->numa_faults[i];
2257 grp->total_faults = p->total_numa_faults;
2260 rcu_assign_pointer(p->numa_group, grp);
2264 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2266 if (!cpupid_match_pid(tsk, cpupid))
2269 grp = rcu_dereference(tsk->numa_group);
2273 my_grp = p->numa_group;
2278 * Only join the other group if its bigger; if we're the bigger group,
2279 * the other task will join us.
2281 if (my_grp->nr_tasks > grp->nr_tasks)
2285 * Tie-break on the grp address.
2287 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2290 /* Always join threads in the same process. */
2291 if (tsk->mm == current->mm)
2294 /* Simple filter to avoid false positives due to PID collisions */
2295 if (flags & TNF_SHARED)
2298 /* Update priv based on whether false sharing was detected */
2301 if (join && !get_numa_group(grp))
2309 BUG_ON(irqs_disabled());
2310 double_lock_irq(&my_grp->lock, &grp->lock);
2312 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2313 my_grp->faults[i] -= p->numa_faults[i];
2314 grp->faults[i] += p->numa_faults[i];
2316 my_grp->total_faults -= p->total_numa_faults;
2317 grp->total_faults += p->total_numa_faults;
2322 spin_unlock(&my_grp->lock);
2323 spin_unlock_irq(&grp->lock);
2325 rcu_assign_pointer(p->numa_group, grp);
2327 put_numa_group(my_grp);
2335 void task_numa_free(struct task_struct *p)
2337 struct numa_group *grp = p->numa_group;
2338 void *numa_faults = p->numa_faults;
2339 unsigned long flags;
2343 spin_lock_irqsave(&grp->lock, flags);
2344 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2345 grp->faults[i] -= p->numa_faults[i];
2346 grp->total_faults -= p->total_numa_faults;
2349 spin_unlock_irqrestore(&grp->lock, flags);
2350 RCU_INIT_POINTER(p->numa_group, NULL);
2351 put_numa_group(grp);
2354 p->numa_faults = NULL;
2359 * Got a PROT_NONE fault for a page on @node.
2361 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2363 struct task_struct *p = current;
2364 bool migrated = flags & TNF_MIGRATED;
2365 int cpu_node = task_node(current);
2366 int local = !!(flags & TNF_FAULT_LOCAL);
2367 struct numa_group *ng;
2370 if (!static_branch_likely(&sched_numa_balancing))
2373 /* for example, ksmd faulting in a user's mm */
2377 /* Allocate buffer to track faults on a per-node basis */
2378 if (unlikely(!p->numa_faults)) {
2379 int size = sizeof(*p->numa_faults) *
2380 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2382 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2383 if (!p->numa_faults)
2386 p->total_numa_faults = 0;
2387 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2391 * First accesses are treated as private, otherwise consider accesses
2392 * to be private if the accessing pid has not changed
2394 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2397 priv = cpupid_match_pid(p, last_cpupid);
2398 if (!priv && !(flags & TNF_NO_GROUP))
2399 task_numa_group(p, last_cpupid, flags, &priv);
2403 * If a workload spans multiple NUMA nodes, a shared fault that
2404 * occurs wholly within the set of nodes that the workload is
2405 * actively using should be counted as local. This allows the
2406 * scan rate to slow down when a workload has settled down.
2409 if (!priv && !local && ng && ng->active_nodes > 1 &&
2410 numa_is_active_node(cpu_node, ng) &&
2411 numa_is_active_node(mem_node, ng))
2415 * Retry to migrate task to preferred node periodically, in case it
2416 * previously failed, or the scheduler moved us.
2418 if (time_after(jiffies, p->numa_migrate_retry)) {
2419 task_numa_placement(p);
2420 numa_migrate_preferred(p);
2424 p->numa_pages_migrated += pages;
2425 if (flags & TNF_MIGRATE_FAIL)
2426 p->numa_faults_locality[2] += pages;
2428 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2429 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2430 p->numa_faults_locality[local] += pages;
2433 static void reset_ptenuma_scan(struct task_struct *p)
2436 * We only did a read acquisition of the mmap sem, so
2437 * p->mm->numa_scan_seq is written to without exclusive access
2438 * and the update is not guaranteed to be atomic. That's not
2439 * much of an issue though, since this is just used for
2440 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2441 * expensive, to avoid any form of compiler optimizations:
2443 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2444 p->mm->numa_scan_offset = 0;
2448 * The expensive part of numa migration is done from task_work context.
2449 * Triggered from task_tick_numa().
2451 void task_numa_work(struct callback_head *work)
2453 unsigned long migrate, next_scan, now = jiffies;
2454 struct task_struct *p = current;
2455 struct mm_struct *mm = p->mm;
2456 u64 runtime = p->se.sum_exec_runtime;
2457 struct vm_area_struct *vma;
2458 unsigned long start, end;
2459 unsigned long nr_pte_updates = 0;
2460 long pages, virtpages;
2462 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2464 work->next = work; /* protect against double add */
2466 * Who cares about NUMA placement when they're dying.
2468 * NOTE: make sure not to dereference p->mm before this check,
2469 * exit_task_work() happens _after_ exit_mm() so we could be called
2470 * without p->mm even though we still had it when we enqueued this
2473 if (p->flags & PF_EXITING)
2476 if (!mm->numa_next_scan) {
2477 mm->numa_next_scan = now +
2478 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2482 * Enforce maximal scan/migration frequency..
2484 migrate = mm->numa_next_scan;
2485 if (time_before(now, migrate))
2488 if (p->numa_scan_period == 0) {
2489 p->numa_scan_period_max = task_scan_max(p);
2490 p->numa_scan_period = task_scan_start(p);
2493 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2494 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2498 * Delay this task enough that another task of this mm will likely win
2499 * the next time around.
2501 p->node_stamp += 2 * TICK_NSEC;
2503 start = mm->numa_scan_offset;
2504 pages = sysctl_numa_balancing_scan_size;
2505 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2506 virtpages = pages * 8; /* Scan up to this much virtual space */
2511 if (!down_read_trylock(&mm->mmap_sem))
2513 vma = find_vma(mm, start);
2515 reset_ptenuma_scan(p);
2519 for (; vma; vma = vma->vm_next) {
2520 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2521 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2526 * Shared library pages mapped by multiple processes are not
2527 * migrated as it is expected they are cache replicated. Avoid
2528 * hinting faults in read-only file-backed mappings or the vdso
2529 * as migrating the pages will be of marginal benefit.
2532 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2536 * Skip inaccessible VMAs to avoid any confusion between
2537 * PROT_NONE and NUMA hinting ptes
2539 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2543 start = max(start, vma->vm_start);
2544 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2545 end = min(end, vma->vm_end);
2546 nr_pte_updates = change_prot_numa(vma, start, end);
2549 * Try to scan sysctl_numa_balancing_size worth of
2550 * hpages that have at least one present PTE that
2551 * is not already pte-numa. If the VMA contains
2552 * areas that are unused or already full of prot_numa
2553 * PTEs, scan up to virtpages, to skip through those
2557 pages -= (end - start) >> PAGE_SHIFT;
2558 virtpages -= (end - start) >> PAGE_SHIFT;
2561 if (pages <= 0 || virtpages <= 0)
2565 } while (end != vma->vm_end);
2570 * It is possible to reach the end of the VMA list but the last few
2571 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2572 * would find the !migratable VMA on the next scan but not reset the
2573 * scanner to the start so check it now.
2576 mm->numa_scan_offset = start;
2578 reset_ptenuma_scan(p);
2579 up_read(&mm->mmap_sem);
2582 * Make sure tasks use at least 32x as much time to run other code
2583 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2584 * Usually update_task_scan_period slows down scanning enough; on an
2585 * overloaded system we need to limit overhead on a per task basis.
2587 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2588 u64 diff = p->se.sum_exec_runtime - runtime;
2589 p->node_stamp += 32 * diff;
2594 * Drive the periodic memory faults..
2596 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2598 struct callback_head *work = &curr->numa_work;
2602 * We don't care about NUMA placement if we don't have memory.
2604 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2608 * Using runtime rather than walltime has the dual advantage that
2609 * we (mostly) drive the selection from busy threads and that the
2610 * task needs to have done some actual work before we bother with
2613 now = curr->se.sum_exec_runtime;
2614 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2616 if (now > curr->node_stamp + period) {
2617 if (!curr->node_stamp)
2618 curr->numa_scan_period = task_scan_start(curr);
2619 curr->node_stamp += period;
2621 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2622 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2623 task_work_add(curr, work, true);
2628 static void update_scan_period(struct task_struct *p, int new_cpu)
2630 int src_nid = cpu_to_node(task_cpu(p));
2631 int dst_nid = cpu_to_node(new_cpu);
2633 if (!static_branch_likely(&sched_numa_balancing))
2636 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2639 if (src_nid == dst_nid)
2643 * Allow resets if faults have been trapped before one scan
2644 * has completed. This is most likely due to a new task that
2645 * is pulled cross-node due to wakeups or load balancing.
2647 if (p->numa_scan_seq) {
2649 * Avoid scan adjustments if moving to the preferred
2650 * node or if the task was not previously running on
2651 * the preferred node.
2653 if (dst_nid == p->numa_preferred_nid ||
2654 (p->numa_preferred_nid != -1 && src_nid != p->numa_preferred_nid))
2658 p->numa_scan_period = task_scan_start(p);
2662 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2666 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2670 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2674 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2678 #endif /* CONFIG_NUMA_BALANCING */
2681 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2683 update_load_add(&cfs_rq->load, se->load.weight);
2684 if (!parent_entity(se))
2685 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2687 if (entity_is_task(se)) {
2688 struct rq *rq = rq_of(cfs_rq);
2690 account_numa_enqueue(rq, task_of(se));
2691 list_add(&se->group_node, &rq->cfs_tasks);
2694 cfs_rq->nr_running++;
2698 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2700 update_load_sub(&cfs_rq->load, se->load.weight);
2701 if (!parent_entity(se))
2702 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2704 if (entity_is_task(se)) {
2705 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2706 list_del_init(&se->group_node);
2709 cfs_rq->nr_running--;
2713 * Signed add and clamp on underflow.
2715 * Explicitly do a load-store to ensure the intermediate value never hits
2716 * memory. This allows lockless observations without ever seeing the negative
2719 #define add_positive(_ptr, _val) do { \
2720 typeof(_ptr) ptr = (_ptr); \
2721 typeof(_val) val = (_val); \
2722 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2726 if (val < 0 && res > var) \
2729 WRITE_ONCE(*ptr, res); \
2733 * Unsigned subtract and clamp on underflow.
2735 * Explicitly do a load-store to ensure the intermediate value never hits
2736 * memory. This allows lockless observations without ever seeing the negative
2739 #define sub_positive(_ptr, _val) do { \
2740 typeof(_ptr) ptr = (_ptr); \
2741 typeof(*ptr) val = (_val); \
2742 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2746 WRITE_ONCE(*ptr, res); \
2750 * Remove and clamp on negative, from a local variable.
2752 * A variant of sub_positive(), which does not use explicit load-store
2753 * and is thus optimized for local variable updates.
2755 #define lsub_positive(_ptr, _val) do { \
2756 typeof(_ptr) ptr = (_ptr); \
2757 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
2762 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2764 cfs_rq->runnable_weight += se->runnable_weight;
2766 cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2767 cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2771 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2773 cfs_rq->runnable_weight -= se->runnable_weight;
2775 sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2776 sub_positive(&cfs_rq->avg.runnable_load_sum,
2777 se_runnable(se) * se->avg.runnable_load_sum);
2781 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2783 cfs_rq->avg.load_avg += se->avg.load_avg;
2784 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2788 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2790 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2791 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2795 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2797 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2799 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2801 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2804 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2805 unsigned long weight, unsigned long runnable)
2808 /* commit outstanding execution time */
2809 if (cfs_rq->curr == se)
2810 update_curr(cfs_rq);
2811 account_entity_dequeue(cfs_rq, se);
2812 dequeue_runnable_load_avg(cfs_rq, se);
2814 dequeue_load_avg(cfs_rq, se);
2816 se->runnable_weight = runnable;
2817 update_load_set(&se->load, weight);
2821 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2823 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2824 se->avg.runnable_load_avg =
2825 div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2829 enqueue_load_avg(cfs_rq, se);
2831 account_entity_enqueue(cfs_rq, se);
2832 enqueue_runnable_load_avg(cfs_rq, se);
2836 void reweight_task(struct task_struct *p, int prio)
2838 struct sched_entity *se = &p->se;
2839 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2840 struct load_weight *load = &se->load;
2841 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2843 reweight_entity(cfs_rq, se, weight, weight);
2844 load->inv_weight = sched_prio_to_wmult[prio];
2847 #ifdef CONFIG_FAIR_GROUP_SCHED
2850 * All this does is approximate the hierarchical proportion which includes that
2851 * global sum we all love to hate.
2853 * That is, the weight of a group entity, is the proportional share of the
2854 * group weight based on the group runqueue weights. That is:
2856 * tg->weight * grq->load.weight
2857 * ge->load.weight = ----------------------------- (1)
2858 * \Sum grq->load.weight
2860 * Now, because computing that sum is prohibitively expensive to compute (been
2861 * there, done that) we approximate it with this average stuff. The average
2862 * moves slower and therefore the approximation is cheaper and more stable.
2864 * So instead of the above, we substitute:
2866 * grq->load.weight -> grq->avg.load_avg (2)
2868 * which yields the following:
2870 * tg->weight * grq->avg.load_avg
2871 * ge->load.weight = ------------------------------ (3)
2874 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2876 * That is shares_avg, and it is right (given the approximation (2)).
2878 * The problem with it is that because the average is slow -- it was designed
2879 * to be exactly that of course -- this leads to transients in boundary
2880 * conditions. In specific, the case where the group was idle and we start the
2881 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2882 * yielding bad latency etc..
2884 * Now, in that special case (1) reduces to:
2886 * tg->weight * grq->load.weight
2887 * ge->load.weight = ----------------------------- = tg->weight (4)
2890 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2892 * So what we do is modify our approximation (3) to approach (4) in the (near)
2897 * tg->weight * grq->load.weight
2898 * --------------------------------------------------- (5)
2899 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2901 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2902 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2905 * tg->weight * grq->load.weight
2906 * ge->load.weight = ----------------------------- (6)
2911 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2912 * max(grq->load.weight, grq->avg.load_avg)
2914 * And that is shares_weight and is icky. In the (near) UP case it approaches
2915 * (4) while in the normal case it approaches (3). It consistently
2916 * overestimates the ge->load.weight and therefore:
2918 * \Sum ge->load.weight >= tg->weight
2922 static long calc_group_shares(struct cfs_rq *cfs_rq)
2924 long tg_weight, tg_shares, load, shares;
2925 struct task_group *tg = cfs_rq->tg;
2927 tg_shares = READ_ONCE(tg->shares);
2929 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
2931 tg_weight = atomic_long_read(&tg->load_avg);
2933 /* Ensure tg_weight >= load */
2934 tg_weight -= cfs_rq->tg_load_avg_contrib;
2937 shares = (tg_shares * load);
2939 shares /= tg_weight;
2942 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2943 * of a group with small tg->shares value. It is a floor value which is
2944 * assigned as a minimum load.weight to the sched_entity representing
2945 * the group on a CPU.
2947 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2948 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2949 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2950 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2953 return clamp_t(long, shares, MIN_SHARES, tg_shares);
2957 * This calculates the effective runnable weight for a group entity based on
2958 * the group entity weight calculated above.
2960 * Because of the above approximation (2), our group entity weight is
2961 * an load_avg based ratio (3). This means that it includes blocked load and
2962 * does not represent the runnable weight.
2964 * Approximate the group entity's runnable weight per ratio from the group
2967 * grq->avg.runnable_load_avg
2968 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
2971 * However, analogous to above, since the avg numbers are slow, this leads to
2972 * transients in the from-idle case. Instead we use:
2974 * ge->runnable_weight = ge->load.weight *
2976 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
2977 * ----------------------------------------------------- (8)
2978 * max(grq->avg.load_avg, grq->load.weight)
2980 * Where these max() serve both to use the 'instant' values to fix the slow
2981 * from-idle and avoid the /0 on to-idle, similar to (6).
2983 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
2985 long runnable, load_avg;
2987 load_avg = max(cfs_rq->avg.load_avg,
2988 scale_load_down(cfs_rq->load.weight));
2990 runnable = max(cfs_rq->avg.runnable_load_avg,
2991 scale_load_down(cfs_rq->runnable_weight));
2995 runnable /= load_avg;
2997 return clamp_t(long, runnable, MIN_SHARES, shares);
2999 #endif /* CONFIG_SMP */
3001 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3004 * Recomputes the group entity based on the current state of its group
3007 static void update_cfs_group(struct sched_entity *se)
3009 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3010 long shares, runnable;
3015 if (throttled_hierarchy(gcfs_rq))
3019 runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
3021 if (likely(se->load.weight == shares))
3024 shares = calc_group_shares(gcfs_rq);
3025 runnable = calc_group_runnable(gcfs_rq, shares);
3028 reweight_entity(cfs_rq_of(se), se, shares, runnable);
3031 #else /* CONFIG_FAIR_GROUP_SCHED */
3032 static inline void update_cfs_group(struct sched_entity *se)
3035 #endif /* CONFIG_FAIR_GROUP_SCHED */
3037 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3039 struct rq *rq = rq_of(cfs_rq);
3041 if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) {
3043 * There are a few boundary cases this might miss but it should
3044 * get called often enough that that should (hopefully) not be
3047 * It will not get called when we go idle, because the idle
3048 * thread is a different class (!fair), nor will the utilization
3049 * number include things like RT tasks.
3051 * As is, the util number is not freq-invariant (we'd have to
3052 * implement arch_scale_freq_capacity() for that).
3056 cpufreq_update_util(rq, flags);
3061 #ifdef CONFIG_FAIR_GROUP_SCHED
3063 * update_tg_load_avg - update the tg's load avg
3064 * @cfs_rq: the cfs_rq whose avg changed
3065 * @force: update regardless of how small the difference
3067 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3068 * However, because tg->load_avg is a global value there are performance
3071 * In order to avoid having to look at the other cfs_rq's, we use a
3072 * differential update where we store the last value we propagated. This in
3073 * turn allows skipping updates if the differential is 'small'.
3075 * Updating tg's load_avg is necessary before update_cfs_share().
3077 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3079 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3082 * No need to update load_avg for root_task_group as it is not used.
3084 if (cfs_rq->tg == &root_task_group)
3087 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3088 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3089 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3094 * Called within set_task_rq() right before setting a task's CPU. The
3095 * caller only guarantees p->pi_lock is held; no other assumptions,
3096 * including the state of rq->lock, should be made.
3098 void set_task_rq_fair(struct sched_entity *se,
3099 struct cfs_rq *prev, struct cfs_rq *next)
3101 u64 p_last_update_time;
3102 u64 n_last_update_time;
3104 if (!sched_feat(ATTACH_AGE_LOAD))
3108 * We are supposed to update the task to "current" time, then its up to
3109 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3110 * getting what current time is, so simply throw away the out-of-date
3111 * time. This will result in the wakee task is less decayed, but giving
3112 * the wakee more load sounds not bad.
3114 if (!(se->avg.last_update_time && prev))
3117 #ifndef CONFIG_64BIT
3119 u64 p_last_update_time_copy;
3120 u64 n_last_update_time_copy;
3123 p_last_update_time_copy = prev->load_last_update_time_copy;
3124 n_last_update_time_copy = next->load_last_update_time_copy;
3128 p_last_update_time = prev->avg.last_update_time;
3129 n_last_update_time = next->avg.last_update_time;
3131 } while (p_last_update_time != p_last_update_time_copy ||
3132 n_last_update_time != n_last_update_time_copy);
3135 p_last_update_time = prev->avg.last_update_time;
3136 n_last_update_time = next->avg.last_update_time;
3138 __update_load_avg_blocked_se(p_last_update_time, se);
3139 se->avg.last_update_time = n_last_update_time;
3144 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3145 * propagate its contribution. The key to this propagation is the invariant
3146 * that for each group:
3148 * ge->avg == grq->avg (1)
3150 * _IFF_ we look at the pure running and runnable sums. Because they
3151 * represent the very same entity, just at different points in the hierarchy.
3153 * Per the above update_tg_cfs_util() is trivial and simply copies the running
3154 * sum over (but still wrong, because the group entity and group rq do not have
3155 * their PELT windows aligned).
3157 * However, update_tg_cfs_runnable() is more complex. So we have:
3159 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3161 * And since, like util, the runnable part should be directly transferable,
3162 * the following would _appear_ to be the straight forward approach:
3164 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3166 * And per (1) we have:
3168 * ge->avg.runnable_avg == grq->avg.runnable_avg
3172 * ge->load.weight * grq->avg.load_avg
3173 * ge->avg.load_avg = ----------------------------------- (4)
3176 * Except that is wrong!
3178 * Because while for entities historical weight is not important and we
3179 * really only care about our future and therefore can consider a pure
3180 * runnable sum, runqueues can NOT do this.
3182 * We specifically want runqueues to have a load_avg that includes
3183 * historical weights. Those represent the blocked load, the load we expect
3184 * to (shortly) return to us. This only works by keeping the weights as
3185 * integral part of the sum. We therefore cannot decompose as per (3).
3187 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3188 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3189 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3190 * runnable section of these tasks overlap (or not). If they were to perfectly
3191 * align the rq as a whole would be runnable 2/3 of the time. If however we
3192 * always have at least 1 runnable task, the rq as a whole is always runnable.
3194 * So we'll have to approximate.. :/
3196 * Given the constraint:
3198 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3200 * We can construct a rule that adds runnable to a rq by assuming minimal
3203 * On removal, we'll assume each task is equally runnable; which yields:
3205 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3207 * XXX: only do this for the part of runnable > running ?
3212 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3214 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3216 /* Nothing to update */
3221 * The relation between sum and avg is:
3223 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3225 * however, the PELT windows are not aligned between grq and gse.
3228 /* Set new sched_entity's utilization */
3229 se->avg.util_avg = gcfs_rq->avg.util_avg;
3230 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3232 /* Update parent cfs_rq utilization */
3233 add_positive(&cfs_rq->avg.util_avg, delta);
3234 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3238 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3240 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3241 unsigned long runnable_load_avg, load_avg;
3242 u64 runnable_load_sum, load_sum = 0;
3248 gcfs_rq->prop_runnable_sum = 0;
3250 if (runnable_sum >= 0) {
3252 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3253 * the CPU is saturated running == runnable.
3255 runnable_sum += se->avg.load_sum;
3256 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3259 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3260 * assuming all tasks are equally runnable.
3262 if (scale_load_down(gcfs_rq->load.weight)) {
3263 load_sum = div_s64(gcfs_rq->avg.load_sum,
3264 scale_load_down(gcfs_rq->load.weight));
3267 /* But make sure to not inflate se's runnable */
3268 runnable_sum = min(se->avg.load_sum, load_sum);
3272 * runnable_sum can't be lower than running_sum
3273 * Rescale running sum to be in the same range as runnable sum
3274 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3275 * runnable_sum is in [0 : LOAD_AVG_MAX]
3277 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3278 runnable_sum = max(runnable_sum, running_sum);
3280 load_sum = (s64)se_weight(se) * runnable_sum;
3281 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3283 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3284 delta_avg = load_avg - se->avg.load_avg;
3286 se->avg.load_sum = runnable_sum;
3287 se->avg.load_avg = load_avg;
3288 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3289 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3291 runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3292 runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3293 delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3294 delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3296 se->avg.runnable_load_sum = runnable_sum;
3297 se->avg.runnable_load_avg = runnable_load_avg;
3300 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3301 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3305 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3307 cfs_rq->propagate = 1;
3308 cfs_rq->prop_runnable_sum += runnable_sum;
3311 /* Update task and its cfs_rq load average */
3312 static inline int propagate_entity_load_avg(struct sched_entity *se)
3314 struct cfs_rq *cfs_rq, *gcfs_rq;
3316 if (entity_is_task(se))
3319 gcfs_rq = group_cfs_rq(se);
3320 if (!gcfs_rq->propagate)
3323 gcfs_rq->propagate = 0;
3325 cfs_rq = cfs_rq_of(se);
3327 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3329 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3330 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3336 * Check if we need to update the load and the utilization of a blocked
3339 static inline bool skip_blocked_update(struct sched_entity *se)
3341 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3344 * If sched_entity still have not zero load or utilization, we have to
3347 if (se->avg.load_avg || se->avg.util_avg)
3351 * If there is a pending propagation, we have to update the load and
3352 * the utilization of the sched_entity:
3354 if (gcfs_rq->propagate)
3358 * Otherwise, the load and the utilization of the sched_entity is
3359 * already zero and there is no pending propagation, so it will be a
3360 * waste of time to try to decay it:
3365 #else /* CONFIG_FAIR_GROUP_SCHED */
3367 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3369 static inline int propagate_entity_load_avg(struct sched_entity *se)
3374 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3376 #endif /* CONFIG_FAIR_GROUP_SCHED */
3379 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3380 * @now: current time, as per cfs_rq_clock_pelt()
3381 * @cfs_rq: cfs_rq to update
3383 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3384 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3385 * post_init_entity_util_avg().
3387 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3389 * Returns true if the load decayed or we removed load.
3391 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3392 * call update_tg_load_avg() when this function returns true.
3395 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3397 unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3398 struct sched_avg *sa = &cfs_rq->avg;
3401 if (cfs_rq->removed.nr) {
3403 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3405 raw_spin_lock(&cfs_rq->removed.lock);
3406 swap(cfs_rq->removed.util_avg, removed_util);
3407 swap(cfs_rq->removed.load_avg, removed_load);
3408 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3409 cfs_rq->removed.nr = 0;
3410 raw_spin_unlock(&cfs_rq->removed.lock);
3413 sub_positive(&sa->load_avg, r);
3414 sub_positive(&sa->load_sum, r * divider);
3417 sub_positive(&sa->util_avg, r);
3418 sub_positive(&sa->util_sum, r * divider);
3420 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3425 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
3427 #ifndef CONFIG_64BIT
3429 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3433 cfs_rq_util_change(cfs_rq, 0);
3439 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3440 * @cfs_rq: cfs_rq to attach to
3441 * @se: sched_entity to attach
3442 * @flags: migration hints
3444 * Must call update_cfs_rq_load_avg() before this, since we rely on
3445 * cfs_rq->avg.last_update_time being current.
3447 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3449 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3452 * When we attach the @se to the @cfs_rq, we must align the decay
3453 * window because without that, really weird and wonderful things can
3458 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3459 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3462 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3463 * period_contrib. This isn't strictly correct, but since we're
3464 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3467 se->avg.util_sum = se->avg.util_avg * divider;
3469 se->avg.load_sum = divider;
3470 if (se_weight(se)) {
3472 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3475 se->avg.runnable_load_sum = se->avg.load_sum;
3477 enqueue_load_avg(cfs_rq, se);
3478 cfs_rq->avg.util_avg += se->avg.util_avg;
3479 cfs_rq->avg.util_sum += se->avg.util_sum;
3481 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3483 cfs_rq_util_change(cfs_rq, flags);
3487 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3488 * @cfs_rq: cfs_rq to detach from
3489 * @se: sched_entity to detach
3491 * Must call update_cfs_rq_load_avg() before this, since we rely on
3492 * cfs_rq->avg.last_update_time being current.
3494 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3496 dequeue_load_avg(cfs_rq, se);
3497 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3498 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3500 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3502 cfs_rq_util_change(cfs_rq, 0);
3506 * Optional action to be done while updating the load average
3508 #define UPDATE_TG 0x1
3509 #define SKIP_AGE_LOAD 0x2
3510 #define DO_ATTACH 0x4
3512 /* Update task and its cfs_rq load average */
3513 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3515 u64 now = cfs_rq_clock_pelt(cfs_rq);
3519 * Track task load average for carrying it to new CPU after migrated, and
3520 * track group sched_entity load average for task_h_load calc in migration
3522 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3523 __update_load_avg_se(now, cfs_rq, se);
3525 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3526 decayed |= propagate_entity_load_avg(se);
3528 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3531 * DO_ATTACH means we're here from enqueue_entity().
3532 * !last_update_time means we've passed through
3533 * migrate_task_rq_fair() indicating we migrated.
3535 * IOW we're enqueueing a task on a new CPU.
3537 attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
3538 update_tg_load_avg(cfs_rq, 0);
3540 } else if (decayed && (flags & UPDATE_TG))
3541 update_tg_load_avg(cfs_rq, 0);
3544 #ifndef CONFIG_64BIT
3545 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3547 u64 last_update_time_copy;
3548 u64 last_update_time;
3551 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3553 last_update_time = cfs_rq->avg.last_update_time;
3554 } while (last_update_time != last_update_time_copy);
3556 return last_update_time;
3559 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3561 return cfs_rq->avg.last_update_time;
3566 * Synchronize entity load avg of dequeued entity without locking
3569 void sync_entity_load_avg(struct sched_entity *se)
3571 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3572 u64 last_update_time;
3574 last_update_time = cfs_rq_last_update_time(cfs_rq);
3575 __update_load_avg_blocked_se(last_update_time, se);
3579 * Task first catches up with cfs_rq, and then subtract
3580 * itself from the cfs_rq (task must be off the queue now).
3582 void remove_entity_load_avg(struct sched_entity *se)
3584 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3585 unsigned long flags;
3588 * tasks cannot exit without having gone through wake_up_new_task() ->
3589 * post_init_entity_util_avg() which will have added things to the
3590 * cfs_rq, so we can remove unconditionally.
3593 sync_entity_load_avg(se);
3595 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3596 ++cfs_rq->removed.nr;
3597 cfs_rq->removed.util_avg += se->avg.util_avg;
3598 cfs_rq->removed.load_avg += se->avg.load_avg;
3599 cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */
3600 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3603 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3605 return cfs_rq->avg.runnable_load_avg;
3608 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3610 return cfs_rq->avg.load_avg;
3613 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3615 static inline unsigned long task_util(struct task_struct *p)
3617 return READ_ONCE(p->se.avg.util_avg);
3620 static inline unsigned long _task_util_est(struct task_struct *p)
3622 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3624 return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED);
3627 static inline unsigned long task_util_est(struct task_struct *p)
3629 return max(task_util(p), _task_util_est(p));
3632 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3633 struct task_struct *p)
3635 unsigned int enqueued;
3637 if (!sched_feat(UTIL_EST))
3640 /* Update root cfs_rq's estimated utilization */
3641 enqueued = cfs_rq->avg.util_est.enqueued;
3642 enqueued += _task_util_est(p);
3643 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3647 * Check if a (signed) value is within a specified (unsigned) margin,
3648 * based on the observation that:
3650 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3652 * NOTE: this only works when value + maring < INT_MAX.
3654 static inline bool within_margin(int value, int margin)
3656 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3660 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3662 long last_ewma_diff;
3666 if (!sched_feat(UTIL_EST))
3669 /* Update root cfs_rq's estimated utilization */
3670 ue.enqueued = cfs_rq->avg.util_est.enqueued;
3671 ue.enqueued -= min_t(unsigned int, ue.enqueued, _task_util_est(p));
3672 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3675 * Skip update of task's estimated utilization when the task has not
3676 * yet completed an activation, e.g. being migrated.
3682 * If the PELT values haven't changed since enqueue time,
3683 * skip the util_est update.
3685 ue = p->se.avg.util_est;
3686 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3690 * Skip update of task's estimated utilization when its EWMA is
3691 * already ~1% close to its last activation value.
3693 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3694 last_ewma_diff = ue.enqueued - ue.ewma;
3695 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3699 * To avoid overestimation of actual task utilization, skip updates if
3700 * we cannot grant there is idle time in this CPU.
3702 cpu = cpu_of(rq_of(cfs_rq));
3703 if (task_util(p) > capacity_orig_of(cpu))
3707 * Update Task's estimated utilization
3709 * When *p completes an activation we can consolidate another sample
3710 * of the task size. This is done by storing the current PELT value
3711 * as ue.enqueued and by using this value to update the Exponential
3712 * Weighted Moving Average (EWMA):
3714 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
3715 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
3716 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
3717 * = w * ( last_ewma_diff ) + ewma(t-1)
3718 * = w * (last_ewma_diff + ewma(t-1) / w)
3720 * Where 'w' is the weight of new samples, which is configured to be
3721 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3723 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
3724 ue.ewma += last_ewma_diff;
3725 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
3726 WRITE_ONCE(p->se.avg.util_est, ue);
3729 static inline int task_fits_capacity(struct task_struct *p, long capacity)
3731 return capacity * 1024 > task_util_est(p) * capacity_margin;
3734 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
3736 if (!static_branch_unlikely(&sched_asym_cpucapacity))
3740 rq->misfit_task_load = 0;
3744 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
3745 rq->misfit_task_load = 0;
3749 rq->misfit_task_load = task_h_load(p);
3752 #else /* CONFIG_SMP */
3754 #define UPDATE_TG 0x0
3755 #define SKIP_AGE_LOAD 0x0
3756 #define DO_ATTACH 0x0
3758 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3760 cfs_rq_util_change(cfs_rq, 0);
3763 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3766 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
3768 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3770 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3776 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
3779 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
3781 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
3783 #endif /* CONFIG_SMP */
3785 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3787 #ifdef CONFIG_SCHED_DEBUG
3788 s64 d = se->vruntime - cfs_rq->min_vruntime;
3793 if (d > 3*sysctl_sched_latency)
3794 schedstat_inc(cfs_rq->nr_spread_over);
3799 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3801 u64 vruntime = cfs_rq->min_vruntime;
3804 * The 'current' period is already promised to the current tasks,
3805 * however the extra weight of the new task will slow them down a
3806 * little, place the new task so that it fits in the slot that
3807 * stays open at the end.
3809 if (initial && sched_feat(START_DEBIT))
3810 vruntime += sched_vslice(cfs_rq, se);
3812 /* sleeps up to a single latency don't count. */
3814 unsigned long thresh = sysctl_sched_latency;
3817 * Halve their sleep time's effect, to allow
3818 * for a gentler effect of sleepers:
3820 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3826 /* ensure we never gain time by being placed backwards. */
3827 se->vruntime = max_vruntime(se->vruntime, vruntime);
3830 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3832 static inline void check_schedstat_required(void)
3834 #ifdef CONFIG_SCHEDSTATS
3835 if (schedstat_enabled())
3838 /* Force schedstat enabled if a dependent tracepoint is active */
3839 if (trace_sched_stat_wait_enabled() ||
3840 trace_sched_stat_sleep_enabled() ||
3841 trace_sched_stat_iowait_enabled() ||
3842 trace_sched_stat_blocked_enabled() ||
3843 trace_sched_stat_runtime_enabled()) {
3844 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3845 "stat_blocked and stat_runtime require the "
3846 "kernel parameter schedstats=enable or "
3847 "kernel.sched_schedstats=1\n");
3858 * update_min_vruntime()
3859 * vruntime -= min_vruntime
3863 * update_min_vruntime()
3864 * vruntime += min_vruntime
3866 * this way the vruntime transition between RQs is done when both
3867 * min_vruntime are up-to-date.
3871 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3872 * vruntime -= min_vruntime
3876 * update_min_vruntime()
3877 * vruntime += min_vruntime
3879 * this way we don't have the most up-to-date min_vruntime on the originating
3880 * CPU and an up-to-date min_vruntime on the destination CPU.
3884 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3886 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3887 bool curr = cfs_rq->curr == se;
3890 * If we're the current task, we must renormalise before calling
3894 se->vruntime += cfs_rq->min_vruntime;
3896 update_curr(cfs_rq);
3899 * Otherwise, renormalise after, such that we're placed at the current
3900 * moment in time, instead of some random moment in the past. Being
3901 * placed in the past could significantly boost this task to the
3902 * fairness detriment of existing tasks.
3904 if (renorm && !curr)
3905 se->vruntime += cfs_rq->min_vruntime;
3908 * When enqueuing a sched_entity, we must:
3909 * - Update loads to have both entity and cfs_rq synced with now.
3910 * - Add its load to cfs_rq->runnable_avg
3911 * - For group_entity, update its weight to reflect the new share of
3913 * - Add its new weight to cfs_rq->load.weight
3915 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
3916 update_cfs_group(se);
3917 enqueue_runnable_load_avg(cfs_rq, se);
3918 account_entity_enqueue(cfs_rq, se);
3920 if (flags & ENQUEUE_WAKEUP)
3921 place_entity(cfs_rq, se, 0);
3923 check_schedstat_required();
3924 update_stats_enqueue(cfs_rq, se, flags);
3925 check_spread(cfs_rq, se);
3927 __enqueue_entity(cfs_rq, se);
3930 if (cfs_rq->nr_running == 1) {
3931 list_add_leaf_cfs_rq(cfs_rq);
3932 check_enqueue_throttle(cfs_rq);
3936 static void __clear_buddies_last(struct sched_entity *se)
3938 for_each_sched_entity(se) {
3939 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3940 if (cfs_rq->last != se)
3943 cfs_rq->last = NULL;
3947 static void __clear_buddies_next(struct sched_entity *se)
3949 for_each_sched_entity(se) {
3950 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3951 if (cfs_rq->next != se)
3954 cfs_rq->next = NULL;
3958 static void __clear_buddies_skip(struct sched_entity *se)
3960 for_each_sched_entity(se) {
3961 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3962 if (cfs_rq->skip != se)
3965 cfs_rq->skip = NULL;
3969 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3971 if (cfs_rq->last == se)
3972 __clear_buddies_last(se);
3974 if (cfs_rq->next == se)
3975 __clear_buddies_next(se);
3977 if (cfs_rq->skip == se)
3978 __clear_buddies_skip(se);
3981 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3984 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3987 * Update run-time statistics of the 'current'.
3989 update_curr(cfs_rq);
3992 * When dequeuing a sched_entity, we must:
3993 * - Update loads to have both entity and cfs_rq synced with now.
3994 * - Subtract its load from the cfs_rq->runnable_avg.
3995 * - Subtract its previous weight from cfs_rq->load.weight.
3996 * - For group entity, update its weight to reflect the new share
3997 * of its group cfs_rq.
3999 update_load_avg(cfs_rq, se, UPDATE_TG);
4000 dequeue_runnable_load_avg(cfs_rq, se);
4002 update_stats_dequeue(cfs_rq, se, flags);
4004 clear_buddies(cfs_rq, se);
4006 if (se != cfs_rq->curr)
4007 __dequeue_entity(cfs_rq, se);
4009 account_entity_dequeue(cfs_rq, se);
4012 * Normalize after update_curr(); which will also have moved
4013 * min_vruntime if @se is the one holding it back. But before doing
4014 * update_min_vruntime() again, which will discount @se's position and
4015 * can move min_vruntime forward still more.
4017 if (!(flags & DEQUEUE_SLEEP))
4018 se->vruntime -= cfs_rq->min_vruntime;
4020 /* return excess runtime on last dequeue */
4021 return_cfs_rq_runtime(cfs_rq);
4023 update_cfs_group(se);
4026 * Now advance min_vruntime if @se was the entity holding it back,
4027 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4028 * put back on, and if we advance min_vruntime, we'll be placed back
4029 * further than we started -- ie. we'll be penalized.
4031 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4032 update_min_vruntime(cfs_rq);
4036 * Preempt the current task with a newly woken task if needed:
4039 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4041 unsigned long ideal_runtime, delta_exec;
4042 struct sched_entity *se;
4045 ideal_runtime = sched_slice(cfs_rq, curr);
4046 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4047 if (delta_exec > ideal_runtime) {
4048 resched_curr(rq_of(cfs_rq));
4050 * The current task ran long enough, ensure it doesn't get
4051 * re-elected due to buddy favours.
4053 clear_buddies(cfs_rq, curr);
4058 * Ensure that a task that missed wakeup preemption by a
4059 * narrow margin doesn't have to wait for a full slice.
4060 * This also mitigates buddy induced latencies under load.
4062 if (delta_exec < sysctl_sched_min_granularity)
4065 se = __pick_first_entity(cfs_rq);
4066 delta = curr->vruntime - se->vruntime;
4071 if (delta > ideal_runtime)
4072 resched_curr(rq_of(cfs_rq));
4076 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4078 /* 'current' is not kept within the tree. */
4081 * Any task has to be enqueued before it get to execute on
4082 * a CPU. So account for the time it spent waiting on the
4085 update_stats_wait_end(cfs_rq, se);
4086 __dequeue_entity(cfs_rq, se);
4087 update_load_avg(cfs_rq, se, UPDATE_TG);
4090 update_stats_curr_start(cfs_rq, se);
4094 * Track our maximum slice length, if the CPU's load is at
4095 * least twice that of our own weight (i.e. dont track it
4096 * when there are only lesser-weight tasks around):
4098 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
4099 schedstat_set(se->statistics.slice_max,
4100 max((u64)schedstat_val(se->statistics.slice_max),
4101 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4104 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4108 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4111 * Pick the next process, keeping these things in mind, in this order:
4112 * 1) keep things fair between processes/task groups
4113 * 2) pick the "next" process, since someone really wants that to run
4114 * 3) pick the "last" process, for cache locality
4115 * 4) do not run the "skip" process, if something else is available
4117 static struct sched_entity *
4118 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4120 struct sched_entity *left = __pick_first_entity(cfs_rq);
4121 struct sched_entity *se;
4124 * If curr is set we have to see if its left of the leftmost entity
4125 * still in the tree, provided there was anything in the tree at all.
4127 if (!left || (curr && entity_before(curr, left)))
4130 se = left; /* ideally we run the leftmost entity */
4133 * Avoid running the skip buddy, if running something else can
4134 * be done without getting too unfair.
4136 if (cfs_rq->skip == se) {
4137 struct sched_entity *second;
4140 second = __pick_first_entity(cfs_rq);
4142 second = __pick_next_entity(se);
4143 if (!second || (curr && entity_before(curr, second)))
4147 if (second && wakeup_preempt_entity(second, left) < 1)
4152 * Prefer last buddy, try to return the CPU to a preempted task.
4154 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4158 * Someone really wants this to run. If it's not unfair, run it.
4160 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4163 clear_buddies(cfs_rq, se);
4168 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4170 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4173 * If still on the runqueue then deactivate_task()
4174 * was not called and update_curr() has to be done:
4177 update_curr(cfs_rq);
4179 /* throttle cfs_rqs exceeding runtime */
4180 check_cfs_rq_runtime(cfs_rq);
4182 check_spread(cfs_rq, prev);
4185 update_stats_wait_start(cfs_rq, prev);
4186 /* Put 'current' back into the tree. */
4187 __enqueue_entity(cfs_rq, prev);
4188 /* in !on_rq case, update occurred at dequeue */
4189 update_load_avg(cfs_rq, prev, 0);
4191 cfs_rq->curr = NULL;
4195 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4198 * Update run-time statistics of the 'current'.
4200 update_curr(cfs_rq);
4203 * Ensure that runnable average is periodically updated.
4205 update_load_avg(cfs_rq, curr, UPDATE_TG);
4206 update_cfs_group(curr);
4208 #ifdef CONFIG_SCHED_HRTICK
4210 * queued ticks are scheduled to match the slice, so don't bother
4211 * validating it and just reschedule.
4214 resched_curr(rq_of(cfs_rq));
4218 * don't let the period tick interfere with the hrtick preemption
4220 if (!sched_feat(DOUBLE_TICK) &&
4221 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4225 if (cfs_rq->nr_running > 1)
4226 check_preempt_tick(cfs_rq, curr);
4230 /**************************************************
4231 * CFS bandwidth control machinery
4234 #ifdef CONFIG_CFS_BANDWIDTH
4236 #ifdef CONFIG_JUMP_LABEL
4237 static struct static_key __cfs_bandwidth_used;
4239 static inline bool cfs_bandwidth_used(void)
4241 return static_key_false(&__cfs_bandwidth_used);
4244 void cfs_bandwidth_usage_inc(void)
4246 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4249 void cfs_bandwidth_usage_dec(void)
4251 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4253 #else /* CONFIG_JUMP_LABEL */
4254 static bool cfs_bandwidth_used(void)
4259 void cfs_bandwidth_usage_inc(void) {}
4260 void cfs_bandwidth_usage_dec(void) {}
4261 #endif /* CONFIG_JUMP_LABEL */
4264 * default period for cfs group bandwidth.
4265 * default: 0.1s, units: nanoseconds
4267 static inline u64 default_cfs_period(void)
4269 return 100000000ULL;
4272 static inline u64 sched_cfs_bandwidth_slice(void)
4274 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4278 * Replenish runtime according to assigned quota and update expiration time.
4279 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4280 * additional synchronization around rq->lock.
4282 * requires cfs_b->lock
4284 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4288 if (cfs_b->quota == RUNTIME_INF)
4291 now = sched_clock_cpu(smp_processor_id());
4292 cfs_b->runtime = cfs_b->quota;
4293 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4294 cfs_b->expires_seq++;
4297 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4299 return &tg->cfs_bandwidth;
4302 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4303 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4305 if (unlikely(cfs_rq->throttle_count))
4306 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4308 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4311 /* returns 0 on failure to allocate runtime */
4312 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4314 struct task_group *tg = cfs_rq->tg;
4315 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4316 u64 amount = 0, min_amount, expires;
4319 /* note: this is a positive sum as runtime_remaining <= 0 */
4320 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4322 raw_spin_lock(&cfs_b->lock);
4323 if (cfs_b->quota == RUNTIME_INF)
4324 amount = min_amount;
4326 start_cfs_bandwidth(cfs_b);
4328 if (cfs_b->runtime > 0) {
4329 amount = min(cfs_b->runtime, min_amount);
4330 cfs_b->runtime -= amount;
4334 expires_seq = cfs_b->expires_seq;
4335 expires = cfs_b->runtime_expires;
4336 raw_spin_unlock(&cfs_b->lock);
4338 cfs_rq->runtime_remaining += amount;
4340 * we may have advanced our local expiration to account for allowed
4341 * spread between our sched_clock and the one on which runtime was
4344 if (cfs_rq->expires_seq != expires_seq) {
4345 cfs_rq->expires_seq = expires_seq;
4346 cfs_rq->runtime_expires = expires;
4349 return cfs_rq->runtime_remaining > 0;
4353 * Note: This depends on the synchronization provided by sched_clock and the
4354 * fact that rq->clock snapshots this value.
4356 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4358 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4360 /* if the deadline is ahead of our clock, nothing to do */
4361 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4364 if (cfs_rq->runtime_remaining < 0)
4368 * If the local deadline has passed we have to consider the
4369 * possibility that our sched_clock is 'fast' and the global deadline
4370 * has not truly expired.
4372 * Fortunately we can check determine whether this the case by checking
4373 * whether the global deadline(cfs_b->expires_seq) has advanced.
4375 if (cfs_rq->expires_seq == cfs_b->expires_seq) {
4376 /* extend local deadline, drift is bounded above by 2 ticks */
4377 cfs_rq->runtime_expires += TICK_NSEC;
4379 /* global deadline is ahead, expiration has passed */
4380 cfs_rq->runtime_remaining = 0;
4384 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4386 /* dock delta_exec before expiring quota (as it could span periods) */
4387 cfs_rq->runtime_remaining -= delta_exec;
4388 expire_cfs_rq_runtime(cfs_rq);
4390 if (likely(cfs_rq->runtime_remaining > 0))
4394 * if we're unable to extend our runtime we resched so that the active
4395 * hierarchy can be throttled
4397 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4398 resched_curr(rq_of(cfs_rq));
4401 static __always_inline
4402 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4404 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4407 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4410 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4412 return cfs_bandwidth_used() && cfs_rq->throttled;
4415 /* check whether cfs_rq, or any parent, is throttled */
4416 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4418 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4422 * Ensure that neither of the group entities corresponding to src_cpu or
4423 * dest_cpu are members of a throttled hierarchy when performing group
4424 * load-balance operations.
4426 static inline int throttled_lb_pair(struct task_group *tg,
4427 int src_cpu, int dest_cpu)
4429 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4431 src_cfs_rq = tg->cfs_rq[src_cpu];
4432 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4434 return throttled_hierarchy(src_cfs_rq) ||
4435 throttled_hierarchy(dest_cfs_rq);
4438 static int tg_unthrottle_up(struct task_group *tg, void *data)
4440 struct rq *rq = data;
4441 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4443 cfs_rq->throttle_count--;
4444 if (!cfs_rq->throttle_count) {
4445 /* adjust cfs_rq_clock_task() */
4446 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4447 cfs_rq->throttled_clock_task;
4449 /* Add cfs_rq with already running entity in the list */
4450 if (cfs_rq->nr_running >= 1)
4451 list_add_leaf_cfs_rq(cfs_rq);
4457 static int tg_throttle_down(struct task_group *tg, void *data)
4459 struct rq *rq = data;
4460 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4462 /* group is entering throttled state, stop time */
4463 if (!cfs_rq->throttle_count) {
4464 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4465 list_del_leaf_cfs_rq(cfs_rq);
4467 cfs_rq->throttle_count++;
4472 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4474 struct rq *rq = rq_of(cfs_rq);
4475 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4476 struct sched_entity *se;
4477 long task_delta, dequeue = 1;
4480 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4482 /* freeze hierarchy runnable averages while throttled */
4484 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4487 task_delta = cfs_rq->h_nr_running;
4488 for_each_sched_entity(se) {
4489 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4490 /* throttled entity or throttle-on-deactivate */
4495 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4496 qcfs_rq->h_nr_running -= task_delta;
4498 if (qcfs_rq->load.weight)
4503 sub_nr_running(rq, task_delta);
4505 cfs_rq->throttled = 1;
4506 cfs_rq->throttled_clock = rq_clock(rq);
4507 raw_spin_lock(&cfs_b->lock);
4508 empty = list_empty(&cfs_b->throttled_cfs_rq);
4511 * Add to the _head_ of the list, so that an already-started
4512 * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
4513 * not running add to the tail so that later runqueues don't get starved.
4515 if (cfs_b->distribute_running)
4516 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4518 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4521 * If we're the first throttled task, make sure the bandwidth
4525 start_cfs_bandwidth(cfs_b);
4527 raw_spin_unlock(&cfs_b->lock);
4530 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4532 struct rq *rq = rq_of(cfs_rq);
4533 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4534 struct sched_entity *se;
4538 se = cfs_rq->tg->se[cpu_of(rq)];
4540 cfs_rq->throttled = 0;
4542 update_rq_clock(rq);
4544 raw_spin_lock(&cfs_b->lock);
4545 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4546 list_del_rcu(&cfs_rq->throttled_list);
4547 raw_spin_unlock(&cfs_b->lock);
4549 /* update hierarchical throttle state */
4550 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4552 if (!cfs_rq->load.weight)
4555 task_delta = cfs_rq->h_nr_running;
4556 for_each_sched_entity(se) {
4560 cfs_rq = cfs_rq_of(se);
4562 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4563 cfs_rq->h_nr_running += task_delta;
4565 if (cfs_rq_throttled(cfs_rq))
4569 assert_list_leaf_cfs_rq(rq);
4572 add_nr_running(rq, task_delta);
4574 /* Determine whether we need to wake up potentially idle CPU: */
4575 if (rq->curr == rq->idle && rq->cfs.nr_running)
4579 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4580 u64 remaining, u64 expires)
4582 struct cfs_rq *cfs_rq;
4584 u64 starting_runtime = remaining;
4587 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4589 struct rq *rq = rq_of(cfs_rq);
4592 rq_lock_irqsave(rq, &rf);
4593 if (!cfs_rq_throttled(cfs_rq))
4596 runtime = -cfs_rq->runtime_remaining + 1;
4597 if (runtime > remaining)
4598 runtime = remaining;
4599 remaining -= runtime;
4601 cfs_rq->runtime_remaining += runtime;
4602 cfs_rq->runtime_expires = expires;
4604 /* we check whether we're throttled above */
4605 if (cfs_rq->runtime_remaining > 0)
4606 unthrottle_cfs_rq(cfs_rq);
4609 rq_unlock_irqrestore(rq, &rf);
4616 return starting_runtime - remaining;
4620 * Responsible for refilling a task_group's bandwidth and unthrottling its
4621 * cfs_rqs as appropriate. If there has been no activity within the last
4622 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4623 * used to track this state.
4625 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
4627 u64 runtime, runtime_expires;
4630 /* no need to continue the timer with no bandwidth constraint */
4631 if (cfs_b->quota == RUNTIME_INF)
4632 goto out_deactivate;
4634 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4635 cfs_b->nr_periods += overrun;
4638 * idle depends on !throttled (for the case of a large deficit), and if
4639 * we're going inactive then everything else can be deferred
4641 if (cfs_b->idle && !throttled)
4642 goto out_deactivate;
4644 __refill_cfs_bandwidth_runtime(cfs_b);
4647 /* mark as potentially idle for the upcoming period */
4652 /* account preceding periods in which throttling occurred */
4653 cfs_b->nr_throttled += overrun;
4655 runtime_expires = cfs_b->runtime_expires;
4658 * This check is repeated as we are holding onto the new bandwidth while
4659 * we unthrottle. This can potentially race with an unthrottled group
4660 * trying to acquire new bandwidth from the global pool. This can result
4661 * in us over-using our runtime if it is all used during this loop, but
4662 * only by limited amounts in that extreme case.
4664 while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
4665 runtime = cfs_b->runtime;
4666 cfs_b->distribute_running = 1;
4667 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4668 /* we can't nest cfs_b->lock while distributing bandwidth */
4669 runtime = distribute_cfs_runtime(cfs_b, runtime,
4671 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4673 cfs_b->distribute_running = 0;
4674 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4676 lsub_positive(&cfs_b->runtime, runtime);
4680 * While we are ensured activity in the period following an
4681 * unthrottle, this also covers the case in which the new bandwidth is
4682 * insufficient to cover the existing bandwidth deficit. (Forcing the
4683 * timer to remain active while there are any throttled entities.)
4693 /* a cfs_rq won't donate quota below this amount */
4694 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4695 /* minimum remaining period time to redistribute slack quota */
4696 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4697 /* how long we wait to gather additional slack before distributing */
4698 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4701 * Are we near the end of the current quota period?
4703 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4704 * hrtimer base being cleared by hrtimer_start. In the case of
4705 * migrate_hrtimers, base is never cleared, so we are fine.
4707 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4709 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4712 /* if the call-back is running a quota refresh is already occurring */
4713 if (hrtimer_callback_running(refresh_timer))
4716 /* is a quota refresh about to occur? */
4717 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4718 if (remaining < min_expire)
4724 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4726 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4728 /* if there's a quota refresh soon don't bother with slack */
4729 if (runtime_refresh_within(cfs_b, min_left))
4732 hrtimer_start(&cfs_b->slack_timer,
4733 ns_to_ktime(cfs_bandwidth_slack_period),
4737 /* we know any runtime found here is valid as update_curr() precedes return */
4738 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4740 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4741 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4743 if (slack_runtime <= 0)
4746 raw_spin_lock(&cfs_b->lock);
4747 if (cfs_b->quota != RUNTIME_INF &&
4748 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4749 cfs_b->runtime += slack_runtime;
4751 /* we are under rq->lock, defer unthrottling using a timer */
4752 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4753 !list_empty(&cfs_b->throttled_cfs_rq))
4754 start_cfs_slack_bandwidth(cfs_b);
4756 raw_spin_unlock(&cfs_b->lock);
4758 /* even if it's not valid for return we don't want to try again */
4759 cfs_rq->runtime_remaining -= slack_runtime;
4762 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4764 if (!cfs_bandwidth_used())
4767 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4770 __return_cfs_rq_runtime(cfs_rq);
4774 * This is done with a timer (instead of inline with bandwidth return) since
4775 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4777 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4779 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4780 unsigned long flags;
4783 /* confirm we're still not at a refresh boundary */
4784 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4785 if (cfs_b->distribute_running) {
4786 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4790 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4791 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4795 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4796 runtime = cfs_b->runtime;
4798 expires = cfs_b->runtime_expires;
4800 cfs_b->distribute_running = 1;
4802 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4807 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4809 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4810 if (expires == cfs_b->runtime_expires)
4811 lsub_positive(&cfs_b->runtime, runtime);
4812 cfs_b->distribute_running = 0;
4813 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4817 * When a group wakes up we want to make sure that its quota is not already
4818 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4819 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4821 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4823 if (!cfs_bandwidth_used())
4826 /* an active group must be handled by the update_curr()->put() path */
4827 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4830 /* ensure the group is not already throttled */
4831 if (cfs_rq_throttled(cfs_rq))
4834 /* update runtime allocation */
4835 account_cfs_rq_runtime(cfs_rq, 0);
4836 if (cfs_rq->runtime_remaining <= 0)
4837 throttle_cfs_rq(cfs_rq);
4840 static void sync_throttle(struct task_group *tg, int cpu)
4842 struct cfs_rq *pcfs_rq, *cfs_rq;
4844 if (!cfs_bandwidth_used())
4850 cfs_rq = tg->cfs_rq[cpu];
4851 pcfs_rq = tg->parent->cfs_rq[cpu];
4853 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4854 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4857 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4858 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4860 if (!cfs_bandwidth_used())
4863 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4867 * it's possible for a throttled entity to be forced into a running
4868 * state (e.g. set_curr_task), in this case we're finished.
4870 if (cfs_rq_throttled(cfs_rq))
4873 throttle_cfs_rq(cfs_rq);
4877 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4879 struct cfs_bandwidth *cfs_b =
4880 container_of(timer, struct cfs_bandwidth, slack_timer);
4882 do_sched_cfs_slack_timer(cfs_b);
4884 return HRTIMER_NORESTART;
4887 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4889 struct cfs_bandwidth *cfs_b =
4890 container_of(timer, struct cfs_bandwidth, period_timer);
4891 unsigned long flags;
4895 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4897 overrun = hrtimer_forward_now(timer, cfs_b->period);
4901 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
4904 cfs_b->period_active = 0;
4905 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4907 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4910 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4912 raw_spin_lock_init(&cfs_b->lock);
4914 cfs_b->quota = RUNTIME_INF;
4915 cfs_b->period = ns_to_ktime(default_cfs_period());
4917 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4918 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4919 cfs_b->period_timer.function = sched_cfs_period_timer;
4920 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4921 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4922 cfs_b->distribute_running = 0;
4925 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4927 cfs_rq->runtime_enabled = 0;
4928 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4931 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4935 lockdep_assert_held(&cfs_b->lock);
4937 if (cfs_b->period_active)
4940 cfs_b->period_active = 1;
4941 overrun = hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4942 cfs_b->runtime_expires += (overrun + 1) * ktime_to_ns(cfs_b->period);
4943 cfs_b->expires_seq++;
4944 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4947 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4949 /* init_cfs_bandwidth() was not called */
4950 if (!cfs_b->throttled_cfs_rq.next)
4953 hrtimer_cancel(&cfs_b->period_timer);
4954 hrtimer_cancel(&cfs_b->slack_timer);
4958 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
4960 * The race is harmless, since modifying bandwidth settings of unhooked group
4961 * bits doesn't do much.
4964 /* cpu online calback */
4965 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4967 struct task_group *tg;
4969 lockdep_assert_held(&rq->lock);
4972 list_for_each_entry_rcu(tg, &task_groups, list) {
4973 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
4974 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4976 raw_spin_lock(&cfs_b->lock);
4977 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4978 raw_spin_unlock(&cfs_b->lock);
4983 /* cpu offline callback */
4984 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4986 struct task_group *tg;
4988 lockdep_assert_held(&rq->lock);
4991 list_for_each_entry_rcu(tg, &task_groups, list) {
4992 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4994 if (!cfs_rq->runtime_enabled)
4998 * clock_task is not advancing so we just need to make sure
4999 * there's some valid quota amount
5001 cfs_rq->runtime_remaining = 1;
5003 * Offline rq is schedulable till CPU is completely disabled
5004 * in take_cpu_down(), so we prevent new cfs throttling here.
5006 cfs_rq->runtime_enabled = 0;
5008 if (cfs_rq_throttled(cfs_rq))
5009 unthrottle_cfs_rq(cfs_rq);
5014 #else /* CONFIG_CFS_BANDWIDTH */
5016 static inline bool cfs_bandwidth_used(void)
5021 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
5023 return rq_clock_task(rq_of(cfs_rq));
5026 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5027 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5028 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5029 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5030 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5032 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5037 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5042 static inline int throttled_lb_pair(struct task_group *tg,
5043 int src_cpu, int dest_cpu)
5048 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5050 #ifdef CONFIG_FAIR_GROUP_SCHED
5051 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5054 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5058 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5059 static inline void update_runtime_enabled(struct rq *rq) {}
5060 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5062 #endif /* CONFIG_CFS_BANDWIDTH */
5064 /**************************************************
5065 * CFS operations on tasks:
5068 #ifdef CONFIG_SCHED_HRTICK
5069 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5071 struct sched_entity *se = &p->se;
5072 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5074 SCHED_WARN_ON(task_rq(p) != rq);
5076 if (rq->cfs.h_nr_running > 1) {
5077 u64 slice = sched_slice(cfs_rq, se);
5078 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5079 s64 delta = slice - ran;
5086 hrtick_start(rq, delta);
5091 * called from enqueue/dequeue and updates the hrtick when the
5092 * current task is from our class and nr_running is low enough
5095 static void hrtick_update(struct rq *rq)
5097 struct task_struct *curr = rq->curr;
5099 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5102 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5103 hrtick_start_fair(rq, curr);
5105 #else /* !CONFIG_SCHED_HRTICK */
5107 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5111 static inline void hrtick_update(struct rq *rq)
5117 static inline unsigned long cpu_util(int cpu);
5118 static unsigned long capacity_of(int cpu);
5120 static inline bool cpu_overutilized(int cpu)
5122 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5125 static inline void update_overutilized_status(struct rq *rq)
5127 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu))
5128 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5131 static inline void update_overutilized_status(struct rq *rq) { }
5135 * The enqueue_task method is called before nr_running is
5136 * increased. Here we update the fair scheduling stats and
5137 * then put the task into the rbtree:
5140 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5142 struct cfs_rq *cfs_rq;
5143 struct sched_entity *se = &p->se;
5146 * The code below (indirectly) updates schedutil which looks at
5147 * the cfs_rq utilization to select a frequency.
5148 * Let's add the task's estimated utilization to the cfs_rq's
5149 * estimated utilization, before we update schedutil.
5151 util_est_enqueue(&rq->cfs, p);
5154 * If in_iowait is set, the code below may not trigger any cpufreq
5155 * utilization updates, so do it here explicitly with the IOWAIT flag
5159 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5161 for_each_sched_entity(se) {
5164 cfs_rq = cfs_rq_of(se);
5165 enqueue_entity(cfs_rq, se, flags);
5168 * end evaluation on encountering a throttled cfs_rq
5170 * note: in the case of encountering a throttled cfs_rq we will
5171 * post the final h_nr_running increment below.
5173 if (cfs_rq_throttled(cfs_rq))
5175 cfs_rq->h_nr_running++;
5177 flags = ENQUEUE_WAKEUP;
5180 for_each_sched_entity(se) {
5181 cfs_rq = cfs_rq_of(se);
5182 cfs_rq->h_nr_running++;
5184 if (cfs_rq_throttled(cfs_rq))
5187 update_load_avg(cfs_rq, se, UPDATE_TG);
5188 update_cfs_group(se);
5192 add_nr_running(rq, 1);
5194 * Since new tasks are assigned an initial util_avg equal to
5195 * half of the spare capacity of their CPU, tiny tasks have the
5196 * ability to cross the overutilized threshold, which will
5197 * result in the load balancer ruining all the task placement
5198 * done by EAS. As a way to mitigate that effect, do not account
5199 * for the first enqueue operation of new tasks during the
5200 * overutilized flag detection.
5202 * A better way of solving this problem would be to wait for
5203 * the PELT signals of tasks to converge before taking them
5204 * into account, but that is not straightforward to implement,
5205 * and the following generally works well enough in practice.
5207 if (flags & ENQUEUE_WAKEUP)
5208 update_overutilized_status(rq);
5212 if (cfs_bandwidth_used()) {
5214 * When bandwidth control is enabled; the cfs_rq_throttled()
5215 * breaks in the above iteration can result in incomplete
5216 * leaf list maintenance, resulting in triggering the assertion
5219 for_each_sched_entity(se) {
5220 cfs_rq = cfs_rq_of(se);
5222 if (list_add_leaf_cfs_rq(cfs_rq))
5227 assert_list_leaf_cfs_rq(rq);
5232 static void set_next_buddy(struct sched_entity *se);
5235 * The dequeue_task method is called before nr_running is
5236 * decreased. We remove the task from the rbtree and
5237 * update the fair scheduling stats:
5239 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5241 struct cfs_rq *cfs_rq;
5242 struct sched_entity *se = &p->se;
5243 int task_sleep = flags & DEQUEUE_SLEEP;
5245 for_each_sched_entity(se) {
5246 cfs_rq = cfs_rq_of(se);
5247 dequeue_entity(cfs_rq, se, flags);
5250 * end evaluation on encountering a throttled cfs_rq
5252 * note: in the case of encountering a throttled cfs_rq we will
5253 * post the final h_nr_running decrement below.
5255 if (cfs_rq_throttled(cfs_rq))
5257 cfs_rq->h_nr_running--;
5259 /* Don't dequeue parent if it has other entities besides us */
5260 if (cfs_rq->load.weight) {
5261 /* Avoid re-evaluating load for this entity: */
5262 se = parent_entity(se);
5264 * Bias pick_next to pick a task from this cfs_rq, as
5265 * p is sleeping when it is within its sched_slice.
5267 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5271 flags |= DEQUEUE_SLEEP;
5274 for_each_sched_entity(se) {
5275 cfs_rq = cfs_rq_of(se);
5276 cfs_rq->h_nr_running--;
5278 if (cfs_rq_throttled(cfs_rq))
5281 update_load_avg(cfs_rq, se, UPDATE_TG);
5282 update_cfs_group(se);
5286 sub_nr_running(rq, 1);
5288 util_est_dequeue(&rq->cfs, p, task_sleep);
5294 /* Working cpumask for: load_balance, load_balance_newidle. */
5295 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5296 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5298 #ifdef CONFIG_NO_HZ_COMMON
5300 * per rq 'load' arrray crap; XXX kill this.
5304 * The exact cpuload calculated at every tick would be:
5306 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5308 * If a CPU misses updates for n ticks (as it was idle) and update gets
5309 * called on the n+1-th tick when CPU may be busy, then we have:
5311 * load_n = (1 - 1/2^i)^n * load_0
5312 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
5314 * decay_load_missed() below does efficient calculation of
5316 * load' = (1 - 1/2^i)^n * load
5318 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5319 * This allows us to precompute the above in said factors, thereby allowing the
5320 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5321 * fixed_power_int())
5323 * The calculation is approximated on a 128 point scale.
5325 #define DEGRADE_SHIFT 7
5327 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5328 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5329 { 0, 0, 0, 0, 0, 0, 0, 0 },
5330 { 64, 32, 8, 0, 0, 0, 0, 0 },
5331 { 96, 72, 40, 12, 1, 0, 0, 0 },
5332 { 112, 98, 75, 43, 15, 1, 0, 0 },
5333 { 120, 112, 98, 76, 45, 16, 2, 0 }
5337 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5338 * would be when CPU is idle and so we just decay the old load without
5339 * adding any new load.
5341 static unsigned long
5342 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5346 if (!missed_updates)
5349 if (missed_updates >= degrade_zero_ticks[idx])
5353 return load >> missed_updates;
5355 while (missed_updates) {
5356 if (missed_updates % 2)
5357 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5359 missed_updates >>= 1;
5366 cpumask_var_t idle_cpus_mask;
5368 int has_blocked; /* Idle CPUS has blocked load */
5369 unsigned long next_balance; /* in jiffy units */
5370 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5371 } nohz ____cacheline_aligned;
5373 #endif /* CONFIG_NO_HZ_COMMON */
5376 * __cpu_load_update - update the rq->cpu_load[] statistics
5377 * @this_rq: The rq to update statistics for
5378 * @this_load: The current load
5379 * @pending_updates: The number of missed updates
5381 * Update rq->cpu_load[] statistics. This function is usually called every
5382 * scheduler tick (TICK_NSEC).
5384 * This function computes a decaying average:
5386 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5388 * Because of NOHZ it might not get called on every tick which gives need for
5389 * the @pending_updates argument.
5391 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5392 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5393 * = A * (A * load[i]_n-2 + B) + B
5394 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5395 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5396 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5397 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5398 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5400 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5401 * any change in load would have resulted in the tick being turned back on.
5403 * For regular NOHZ, this reduces to:
5405 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5407 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5410 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5411 unsigned long pending_updates)
5413 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5416 this_rq->nr_load_updates++;
5418 /* Update our load: */
5419 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5420 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5421 unsigned long old_load, new_load;
5423 /* scale is effectively 1 << i now, and >> i divides by scale */
5425 old_load = this_rq->cpu_load[i];
5426 #ifdef CONFIG_NO_HZ_COMMON
5427 old_load = decay_load_missed(old_load, pending_updates - 1, i);
5428 if (tickless_load) {
5429 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5431 * old_load can never be a negative value because a
5432 * decayed tickless_load cannot be greater than the
5433 * original tickless_load.
5435 old_load += tickless_load;
5438 new_load = this_load;
5440 * Round up the averaging division if load is increasing. This
5441 * prevents us from getting stuck on 9 if the load is 10, for
5444 if (new_load > old_load)
5445 new_load += scale - 1;
5447 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5451 /* Used instead of source_load when we know the type == 0 */
5452 static unsigned long weighted_cpuload(struct rq *rq)
5454 return cfs_rq_runnable_load_avg(&rq->cfs);
5457 #ifdef CONFIG_NO_HZ_COMMON
5459 * There is no sane way to deal with nohz on smp when using jiffies because the
5460 * CPU doing the jiffies update might drift wrt the CPU doing the jiffy reading
5461 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5463 * Therefore we need to avoid the delta approach from the regular tick when
5464 * possible since that would seriously skew the load calculation. This is why we
5465 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5466 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5467 * loop exit, nohz_idle_balance, nohz full exit...)
5469 * This means we might still be one tick off for nohz periods.
5472 static void cpu_load_update_nohz(struct rq *this_rq,
5473 unsigned long curr_jiffies,
5476 unsigned long pending_updates;
5478 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5479 if (pending_updates) {
5480 this_rq->last_load_update_tick = curr_jiffies;
5482 * In the regular NOHZ case, we were idle, this means load 0.
5483 * In the NOHZ_FULL case, we were non-idle, we should consider
5484 * its weighted load.
5486 cpu_load_update(this_rq, load, pending_updates);
5491 * Called from nohz_idle_balance() to update the load ratings before doing the
5494 static void cpu_load_update_idle(struct rq *this_rq)
5497 * bail if there's load or we're actually up-to-date.
5499 if (weighted_cpuload(this_rq))
5502 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5506 * Record CPU load on nohz entry so we know the tickless load to account
5507 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5508 * than other cpu_load[idx] but it should be fine as cpu_load readers
5509 * shouldn't rely into synchronized cpu_load[*] updates.
5511 void cpu_load_update_nohz_start(void)
5513 struct rq *this_rq = this_rq();
5516 * This is all lockless but should be fine. If weighted_cpuload changes
5517 * concurrently we'll exit nohz. And cpu_load write can race with
5518 * cpu_load_update_idle() but both updater would be writing the same.
5520 this_rq->cpu_load[0] = weighted_cpuload(this_rq);
5524 * Account the tickless load in the end of a nohz frame.
5526 void cpu_load_update_nohz_stop(void)
5528 unsigned long curr_jiffies = READ_ONCE(jiffies);
5529 struct rq *this_rq = this_rq();
5533 if (curr_jiffies == this_rq->last_load_update_tick)
5536 load = weighted_cpuload(this_rq);
5537 rq_lock(this_rq, &rf);
5538 update_rq_clock(this_rq);
5539 cpu_load_update_nohz(this_rq, curr_jiffies, load);
5540 rq_unlock(this_rq, &rf);
5542 #else /* !CONFIG_NO_HZ_COMMON */
5543 static inline void cpu_load_update_nohz(struct rq *this_rq,
5544 unsigned long curr_jiffies,
5545 unsigned long load) { }
5546 #endif /* CONFIG_NO_HZ_COMMON */
5548 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5550 #ifdef CONFIG_NO_HZ_COMMON
5551 /* See the mess around cpu_load_update_nohz(). */
5552 this_rq->last_load_update_tick = READ_ONCE(jiffies);
5554 cpu_load_update(this_rq, load, 1);
5558 * Called from scheduler_tick()
5560 void cpu_load_update_active(struct rq *this_rq)
5562 unsigned long load = weighted_cpuload(this_rq);
5564 if (tick_nohz_tick_stopped())
5565 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5567 cpu_load_update_periodic(this_rq, load);
5571 * Return a low guess at the load of a migration-source CPU weighted
5572 * according to the scheduling class and "nice" value.
5574 * We want to under-estimate the load of migration sources, to
5575 * balance conservatively.
5577 static unsigned long source_load(int cpu, int type)
5579 struct rq *rq = cpu_rq(cpu);
5580 unsigned long total = weighted_cpuload(rq);
5582 if (type == 0 || !sched_feat(LB_BIAS))
5585 return min(rq->cpu_load[type-1], total);
5589 * Return a high guess at the load of a migration-target CPU weighted
5590 * according to the scheduling class and "nice" value.
5592 static unsigned long target_load(int cpu, int type)
5594 struct rq *rq = cpu_rq(cpu);
5595 unsigned long total = weighted_cpuload(rq);
5597 if (type == 0 || !sched_feat(LB_BIAS))
5600 return max(rq->cpu_load[type-1], total);
5603 static unsigned long capacity_of(int cpu)
5605 return cpu_rq(cpu)->cpu_capacity;
5608 static unsigned long cpu_avg_load_per_task(int cpu)
5610 struct rq *rq = cpu_rq(cpu);
5611 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5612 unsigned long load_avg = weighted_cpuload(rq);
5615 return load_avg / nr_running;
5620 static void record_wakee(struct task_struct *p)
5623 * Only decay a single time; tasks that have less then 1 wakeup per
5624 * jiffy will not have built up many flips.
5626 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5627 current->wakee_flips >>= 1;
5628 current->wakee_flip_decay_ts = jiffies;
5631 if (current->last_wakee != p) {
5632 current->last_wakee = p;
5633 current->wakee_flips++;
5638 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5640 * A waker of many should wake a different task than the one last awakened
5641 * at a frequency roughly N times higher than one of its wakees.
5643 * In order to determine whether we should let the load spread vs consolidating
5644 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5645 * partner, and a factor of lls_size higher frequency in the other.
5647 * With both conditions met, we can be relatively sure that the relationship is
5648 * non-monogamous, with partner count exceeding socket size.
5650 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5651 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5654 static int wake_wide(struct task_struct *p)
5656 unsigned int master = current->wakee_flips;
5657 unsigned int slave = p->wakee_flips;
5658 int factor = this_cpu_read(sd_llc_size);
5661 swap(master, slave);
5662 if (slave < factor || master < slave * factor)
5668 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5669 * soonest. For the purpose of speed we only consider the waking and previous
5672 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5673 * cache-affine and is (or will be) idle.
5675 * wake_affine_weight() - considers the weight to reflect the average
5676 * scheduling latency of the CPUs. This seems to work
5677 * for the overloaded case.
5680 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5683 * If this_cpu is idle, it implies the wakeup is from interrupt
5684 * context. Only allow the move if cache is shared. Otherwise an
5685 * interrupt intensive workload could force all tasks onto one
5686 * node depending on the IO topology or IRQ affinity settings.
5688 * If the prev_cpu is idle and cache affine then avoid a migration.
5689 * There is no guarantee that the cache hot data from an interrupt
5690 * is more important than cache hot data on the prev_cpu and from
5691 * a cpufreq perspective, it's better to have higher utilisation
5694 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5695 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5697 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5700 return nr_cpumask_bits;
5704 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5705 int this_cpu, int prev_cpu, int sync)
5707 s64 this_eff_load, prev_eff_load;
5708 unsigned long task_load;
5710 this_eff_load = target_load(this_cpu, sd->wake_idx);
5713 unsigned long current_load = task_h_load(current);
5715 if (current_load > this_eff_load)
5718 this_eff_load -= current_load;
5721 task_load = task_h_load(p);
5723 this_eff_load += task_load;
5724 if (sched_feat(WA_BIAS))
5725 this_eff_load *= 100;
5726 this_eff_load *= capacity_of(prev_cpu);
5728 prev_eff_load = source_load(prev_cpu, sd->wake_idx);
5729 prev_eff_load -= task_load;
5730 if (sched_feat(WA_BIAS))
5731 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5732 prev_eff_load *= capacity_of(this_cpu);
5735 * If sync, adjust the weight of prev_eff_load such that if
5736 * prev_eff == this_eff that select_idle_sibling() will consider
5737 * stacking the wakee on top of the waker if no other CPU is
5743 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5746 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5747 int this_cpu, int prev_cpu, int sync)
5749 int target = nr_cpumask_bits;
5751 if (sched_feat(WA_IDLE))
5752 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5754 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5755 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5757 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5758 if (target == nr_cpumask_bits)
5761 schedstat_inc(sd->ttwu_move_affine);
5762 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5766 static unsigned long cpu_util_without(int cpu, struct task_struct *p);
5768 static unsigned long capacity_spare_without(int cpu, struct task_struct *p)
5770 return max_t(long, capacity_of(cpu) - cpu_util_without(cpu, p), 0);
5774 * find_idlest_group finds and returns the least busy CPU group within the
5777 * Assumes p is allowed on at least one CPU in sd.
5779 static struct sched_group *
5780 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5781 int this_cpu, int sd_flag)
5783 struct sched_group *idlest = NULL, *group = sd->groups;
5784 struct sched_group *most_spare_sg = NULL;
5785 unsigned long min_runnable_load = ULONG_MAX;
5786 unsigned long this_runnable_load = ULONG_MAX;
5787 unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5788 unsigned long most_spare = 0, this_spare = 0;
5789 int load_idx = sd->forkexec_idx;
5790 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5791 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5792 (sd->imbalance_pct-100) / 100;
5794 if (sd_flag & SD_BALANCE_WAKE)
5795 load_idx = sd->wake_idx;
5798 unsigned long load, avg_load, runnable_load;
5799 unsigned long spare_cap, max_spare_cap;
5803 /* Skip over this group if it has no CPUs allowed */
5804 if (!cpumask_intersects(sched_group_span(group),
5808 local_group = cpumask_test_cpu(this_cpu,
5809 sched_group_span(group));
5812 * Tally up the load of all CPUs in the group and find
5813 * the group containing the CPU with most spare capacity.
5819 for_each_cpu(i, sched_group_span(group)) {
5820 /* Bias balancing toward CPUs of our domain */
5822 load = source_load(i, load_idx);
5824 load = target_load(i, load_idx);
5826 runnable_load += load;
5828 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5830 spare_cap = capacity_spare_without(i, p);
5832 if (spare_cap > max_spare_cap)
5833 max_spare_cap = spare_cap;
5836 /* Adjust by relative CPU capacity of the group */
5837 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5838 group->sgc->capacity;
5839 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5840 group->sgc->capacity;
5843 this_runnable_load = runnable_load;
5844 this_avg_load = avg_load;
5845 this_spare = max_spare_cap;
5847 if (min_runnable_load > (runnable_load + imbalance)) {
5849 * The runnable load is significantly smaller
5850 * so we can pick this new CPU:
5852 min_runnable_load = runnable_load;
5853 min_avg_load = avg_load;
5855 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5856 (100*min_avg_load > imbalance_scale*avg_load)) {
5858 * The runnable loads are close so take the
5859 * blocked load into account through avg_load:
5861 min_avg_load = avg_load;
5865 if (most_spare < max_spare_cap) {
5866 most_spare = max_spare_cap;
5867 most_spare_sg = group;
5870 } while (group = group->next, group != sd->groups);
5873 * The cross-over point between using spare capacity or least load
5874 * is too conservative for high utilization tasks on partially
5875 * utilized systems if we require spare_capacity > task_util(p),
5876 * so we allow for some task stuffing by using
5877 * spare_capacity > task_util(p)/2.
5879 * Spare capacity can't be used for fork because the utilization has
5880 * not been set yet, we must first select a rq to compute the initial
5883 if (sd_flag & SD_BALANCE_FORK)
5886 if (this_spare > task_util(p) / 2 &&
5887 imbalance_scale*this_spare > 100*most_spare)
5890 if (most_spare > task_util(p) / 2)
5891 return most_spare_sg;
5898 * When comparing groups across NUMA domains, it's possible for the
5899 * local domain to be very lightly loaded relative to the remote
5900 * domains but "imbalance" skews the comparison making remote CPUs
5901 * look much more favourable. When considering cross-domain, add
5902 * imbalance to the runnable load on the remote node and consider
5905 if ((sd->flags & SD_NUMA) &&
5906 min_runnable_load + imbalance >= this_runnable_load)
5909 if (min_runnable_load > (this_runnable_load + imbalance))
5912 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5913 (100*this_avg_load < imbalance_scale*min_avg_load))
5920 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5923 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5925 unsigned long load, min_load = ULONG_MAX;
5926 unsigned int min_exit_latency = UINT_MAX;
5927 u64 latest_idle_timestamp = 0;
5928 int least_loaded_cpu = this_cpu;
5929 int shallowest_idle_cpu = -1;
5932 /* Check if we have any choice: */
5933 if (group->group_weight == 1)
5934 return cpumask_first(sched_group_span(group));
5936 /* Traverse only the allowed CPUs */
5937 for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
5938 if (available_idle_cpu(i)) {
5939 struct rq *rq = cpu_rq(i);
5940 struct cpuidle_state *idle = idle_get_state(rq);
5941 if (idle && idle->exit_latency < min_exit_latency) {
5943 * We give priority to a CPU whose idle state
5944 * has the smallest exit latency irrespective
5945 * of any idle timestamp.
5947 min_exit_latency = idle->exit_latency;
5948 latest_idle_timestamp = rq->idle_stamp;
5949 shallowest_idle_cpu = i;
5950 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5951 rq->idle_stamp > latest_idle_timestamp) {
5953 * If equal or no active idle state, then
5954 * the most recently idled CPU might have
5957 latest_idle_timestamp = rq->idle_stamp;
5958 shallowest_idle_cpu = i;
5960 } else if (shallowest_idle_cpu == -1) {
5961 load = weighted_cpuload(cpu_rq(i));
5962 if (load < min_load) {
5964 least_loaded_cpu = i;
5969 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5972 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5973 int cpu, int prev_cpu, int sd_flag)
5977 if (!cpumask_intersects(sched_domain_span(sd), &p->cpus_allowed))
5981 * We need task's util for capacity_spare_without, sync it up to
5982 * prev_cpu's last_update_time.
5984 if (!(sd_flag & SD_BALANCE_FORK))
5985 sync_entity_load_avg(&p->se);
5988 struct sched_group *group;
5989 struct sched_domain *tmp;
5992 if (!(sd->flags & sd_flag)) {
5997 group = find_idlest_group(sd, p, cpu, sd_flag);
6003 new_cpu = find_idlest_group_cpu(group, p, cpu);
6004 if (new_cpu == cpu) {
6005 /* Now try balancing at a lower domain level of 'cpu': */
6010 /* Now try balancing at a lower domain level of 'new_cpu': */
6012 weight = sd->span_weight;
6014 for_each_domain(cpu, tmp) {
6015 if (weight <= tmp->span_weight)
6017 if (tmp->flags & sd_flag)
6025 #ifdef CONFIG_SCHED_SMT
6026 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
6027 EXPORT_SYMBOL_GPL(sched_smt_present);
6029 static inline void set_idle_cores(int cpu, int val)
6031 struct sched_domain_shared *sds;
6033 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6035 WRITE_ONCE(sds->has_idle_cores, val);
6038 static inline bool test_idle_cores(int cpu, bool def)
6040 struct sched_domain_shared *sds;
6042 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6044 return READ_ONCE(sds->has_idle_cores);
6050 * Scans the local SMT mask to see if the entire core is idle, and records this
6051 * information in sd_llc_shared->has_idle_cores.
6053 * Since SMT siblings share all cache levels, inspecting this limited remote
6054 * state should be fairly cheap.
6056 void __update_idle_core(struct rq *rq)
6058 int core = cpu_of(rq);
6062 if (test_idle_cores(core, true))
6065 for_each_cpu(cpu, cpu_smt_mask(core)) {
6069 if (!available_idle_cpu(cpu))
6073 set_idle_cores(core, 1);
6079 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6080 * there are no idle cores left in the system; tracked through
6081 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6083 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6085 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6088 if (!static_branch_likely(&sched_smt_present))
6091 if (!test_idle_cores(target, false))
6094 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
6096 for_each_cpu_wrap(core, cpus, target) {
6099 for_each_cpu(cpu, cpu_smt_mask(core)) {
6100 __cpumask_clear_cpu(cpu, cpus);
6101 if (!available_idle_cpu(cpu))
6110 * Failed to find an idle core; stop looking for one.
6112 set_idle_cores(target, 0);
6118 * Scan the local SMT mask for idle CPUs.
6120 static int select_idle_smt(struct task_struct *p, int target)
6124 if (!static_branch_likely(&sched_smt_present))
6127 for_each_cpu(cpu, cpu_smt_mask(target)) {
6128 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6130 if (available_idle_cpu(cpu))
6137 #else /* CONFIG_SCHED_SMT */
6139 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6144 static inline int select_idle_smt(struct task_struct *p, int target)
6149 #endif /* CONFIG_SCHED_SMT */
6152 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6153 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6154 * average idle time for this rq (as found in rq->avg_idle).
6156 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
6158 struct sched_domain *this_sd;
6159 u64 avg_cost, avg_idle;
6162 int cpu, nr = INT_MAX;
6164 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6169 * Due to large variance we need a large fuzz factor; hackbench in
6170 * particularly is sensitive here.
6172 avg_idle = this_rq()->avg_idle / 512;
6173 avg_cost = this_sd->avg_scan_cost + 1;
6175 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
6178 if (sched_feat(SIS_PROP)) {
6179 u64 span_avg = sd->span_weight * avg_idle;
6180 if (span_avg > 4*avg_cost)
6181 nr = div_u64(span_avg, avg_cost);
6186 time = local_clock();
6188 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
6191 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6193 if (available_idle_cpu(cpu))
6197 time = local_clock() - time;
6198 cost = this_sd->avg_scan_cost;
6199 delta = (s64)(time - cost) / 8;
6200 this_sd->avg_scan_cost += delta;
6206 * Try and locate an idle core/thread in the LLC cache domain.
6208 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6210 struct sched_domain *sd;
6211 int i, recent_used_cpu;
6213 if (available_idle_cpu(target))
6217 * If the previous CPU is cache affine and idle, don't be stupid:
6219 if (prev != target && cpus_share_cache(prev, target) && available_idle_cpu(prev))
6222 /* Check a recently used CPU as a potential idle candidate: */
6223 recent_used_cpu = p->recent_used_cpu;
6224 if (recent_used_cpu != prev &&
6225 recent_used_cpu != target &&
6226 cpus_share_cache(recent_used_cpu, target) &&
6227 available_idle_cpu(recent_used_cpu) &&
6228 cpumask_test_cpu(p->recent_used_cpu, &p->cpus_allowed)) {
6230 * Replace recent_used_cpu with prev as it is a potential
6231 * candidate for the next wake:
6233 p->recent_used_cpu = prev;
6234 return recent_used_cpu;
6237 sd = rcu_dereference(per_cpu(sd_llc, target));
6241 i = select_idle_core(p, sd, target);
6242 if ((unsigned)i < nr_cpumask_bits)
6245 i = select_idle_cpu(p, sd, target);
6246 if ((unsigned)i < nr_cpumask_bits)
6249 i = select_idle_smt(p, target);
6250 if ((unsigned)i < nr_cpumask_bits)
6257 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6258 * @cpu: the CPU to get the utilization of
6260 * The unit of the return value must be the one of capacity so we can compare
6261 * the utilization with the capacity of the CPU that is available for CFS task
6262 * (ie cpu_capacity).
6264 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6265 * recent utilization of currently non-runnable tasks on a CPU. It represents
6266 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6267 * capacity_orig is the cpu_capacity available at the highest frequency
6268 * (arch_scale_freq_capacity()).
6269 * The utilization of a CPU converges towards a sum equal to or less than the
6270 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6271 * the running time on this CPU scaled by capacity_curr.
6273 * The estimated utilization of a CPU is defined to be the maximum between its
6274 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6275 * currently RUNNABLE on that CPU.
6276 * This allows to properly represent the expected utilization of a CPU which
6277 * has just got a big task running since a long sleep period. At the same time
6278 * however it preserves the benefits of the "blocked utilization" in
6279 * describing the potential for other tasks waking up on the same CPU.
6281 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6282 * higher than capacity_orig because of unfortunate rounding in
6283 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6284 * the average stabilizes with the new running time. We need to check that the
6285 * utilization stays within the range of [0..capacity_orig] and cap it if
6286 * necessary. Without utilization capping, a group could be seen as overloaded
6287 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6288 * available capacity. We allow utilization to overshoot capacity_curr (but not
6289 * capacity_orig) as it useful for predicting the capacity required after task
6290 * migrations (scheduler-driven DVFS).
6292 * Return: the (estimated) utilization for the specified CPU
6294 static inline unsigned long cpu_util(int cpu)
6296 struct cfs_rq *cfs_rq;
6299 cfs_rq = &cpu_rq(cpu)->cfs;
6300 util = READ_ONCE(cfs_rq->avg.util_avg);
6302 if (sched_feat(UTIL_EST))
6303 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6305 return min_t(unsigned long, util, capacity_orig_of(cpu));
6309 * cpu_util_without: compute cpu utilization without any contributions from *p
6310 * @cpu: the CPU which utilization is requested
6311 * @p: the task which utilization should be discounted
6313 * The utilization of a CPU is defined by the utilization of tasks currently
6314 * enqueued on that CPU as well as tasks which are currently sleeping after an
6315 * execution on that CPU.
6317 * This method returns the utilization of the specified CPU by discounting the
6318 * utilization of the specified task, whenever the task is currently
6319 * contributing to the CPU utilization.
6321 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6323 struct cfs_rq *cfs_rq;
6326 /* Task has no contribution or is new */
6327 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6328 return cpu_util(cpu);
6330 cfs_rq = &cpu_rq(cpu)->cfs;
6331 util = READ_ONCE(cfs_rq->avg.util_avg);
6333 /* Discount task's util from CPU's util */
6334 lsub_positive(&util, task_util(p));
6339 * a) if *p is the only task sleeping on this CPU, then:
6340 * cpu_util (== task_util) > util_est (== 0)
6341 * and thus we return:
6342 * cpu_util_without = (cpu_util - task_util) = 0
6344 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6346 * cpu_util >= task_util
6347 * cpu_util > util_est (== 0)
6348 * and thus we discount *p's blocked utilization to return:
6349 * cpu_util_without = (cpu_util - task_util) >= 0
6351 * c) if other tasks are RUNNABLE on that CPU and
6352 * util_est > cpu_util
6353 * then we use util_est since it returns a more restrictive
6354 * estimation of the spare capacity on that CPU, by just
6355 * considering the expected utilization of tasks already
6356 * runnable on that CPU.
6358 * Cases a) and b) are covered by the above code, while case c) is
6359 * covered by the following code when estimated utilization is
6362 if (sched_feat(UTIL_EST)) {
6363 unsigned int estimated =
6364 READ_ONCE(cfs_rq->avg.util_est.enqueued);
6367 * Despite the following checks we still have a small window
6368 * for a possible race, when an execl's select_task_rq_fair()
6369 * races with LB's detach_task():
6372 * p->on_rq = TASK_ON_RQ_MIGRATING;
6373 * ---------------------------------- A
6374 * deactivate_task() \
6375 * dequeue_task() + RaceTime
6376 * util_est_dequeue() /
6377 * ---------------------------------- B
6379 * The additional check on "current == p" it's required to
6380 * properly fix the execl regression and it helps in further
6381 * reducing the chances for the above race.
6383 if (unlikely(task_on_rq_queued(p) || current == p))
6384 lsub_positive(&estimated, _task_util_est(p));
6386 util = max(util, estimated);
6390 * Utilization (estimated) can exceed the CPU capacity, thus let's
6391 * clamp to the maximum CPU capacity to ensure consistency with
6392 * the cpu_util call.
6394 return min_t(unsigned long, util, capacity_orig_of(cpu));
6398 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6399 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6401 * In that case WAKE_AFFINE doesn't make sense and we'll let
6402 * BALANCE_WAKE sort things out.
6404 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6406 long min_cap, max_cap;
6408 if (!static_branch_unlikely(&sched_asym_cpucapacity))
6411 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6412 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6414 /* Minimum capacity is close to max, no need to abort wake_affine */
6415 if (max_cap - min_cap < max_cap >> 3)
6418 /* Bring task utilization in sync with prev_cpu */
6419 sync_entity_load_avg(&p->se);
6421 return !task_fits_capacity(p, min_cap);
6425 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6428 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6430 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6431 unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
6434 * If @p migrates from @cpu to another, remove its contribution. Or,
6435 * if @p migrates from another CPU to @cpu, add its contribution. In
6436 * the other cases, @cpu is not impacted by the migration, so the
6437 * util_avg should already be correct.
6439 if (task_cpu(p) == cpu && dst_cpu != cpu)
6440 sub_positive(&util, task_util(p));
6441 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6442 util += task_util(p);
6444 if (sched_feat(UTIL_EST)) {
6445 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6448 * During wake-up, the task isn't enqueued yet and doesn't
6449 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6450 * so just add it (if needed) to "simulate" what will be
6451 * cpu_util() after the task has been enqueued.
6454 util_est += _task_util_est(p);
6456 util = max(util, util_est);
6459 return min(util, capacity_orig_of(cpu));
6463 * compute_energy(): Estimates the energy that would be consumed if @p was
6464 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6465 * landscape of the * CPUs after the task migration, and uses the Energy Model
6466 * to compute what would be the energy if we decided to actually migrate that
6470 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
6472 long util, max_util, sum_util, energy = 0;
6475 for (; pd; pd = pd->next) {
6476 max_util = sum_util = 0;
6478 * The capacity state of CPUs of the current rd can be driven by
6479 * CPUs of another rd if they belong to the same performance
6480 * domain. So, account for the utilization of these CPUs too
6481 * by masking pd with cpu_online_mask instead of the rd span.
6483 * If an entire performance domain is outside of the current rd,
6484 * it will not appear in its pd list and will not be accounted
6485 * by compute_energy().
6487 for_each_cpu_and(cpu, perf_domain_span(pd), cpu_online_mask) {
6488 util = cpu_util_next(cpu, p, dst_cpu);
6489 util = schedutil_energy_util(cpu, util);
6490 max_util = max(util, max_util);
6494 energy += em_pd_energy(pd->em_pd, max_util, sum_util);
6501 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6502 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6503 * spare capacity in each performance domain and uses it as a potential
6504 * candidate to execute the task. Then, it uses the Energy Model to figure
6505 * out which of the CPU candidates is the most energy-efficient.
6507 * The rationale for this heuristic is as follows. In a performance domain,
6508 * all the most energy efficient CPU candidates (according to the Energy
6509 * Model) are those for which we'll request a low frequency. When there are
6510 * several CPUs for which the frequency request will be the same, we don't
6511 * have enough data to break the tie between them, because the Energy Model
6512 * only includes active power costs. With this model, if we assume that
6513 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6514 * the maximum spare capacity in a performance domain is guaranteed to be among
6515 * the best candidates of the performance domain.
6517 * In practice, it could be preferable from an energy standpoint to pack
6518 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6519 * but that could also hurt our chances to go cluster idle, and we have no
6520 * ways to tell with the current Energy Model if this is actually a good
6521 * idea or not. So, find_energy_efficient_cpu() basically favors
6522 * cluster-packing, and spreading inside a cluster. That should at least be
6523 * a good thing for latency, and this is consistent with the idea that most
6524 * of the energy savings of EAS come from the asymmetry of the system, and
6525 * not so much from breaking the tie between identical CPUs. That's also the
6526 * reason why EAS is enabled in the topology code only for systems where
6527 * SD_ASYM_CPUCAPACITY is set.
6529 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6530 * they don't have any useful utilization data yet and it's not possible to
6531 * forecast their impact on energy consumption. Consequently, they will be
6532 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6533 * to be energy-inefficient in some use-cases. The alternative would be to
6534 * bias new tasks towards specific types of CPUs first, or to try to infer
6535 * their util_avg from the parent task, but those heuristics could hurt
6536 * other use-cases too. So, until someone finds a better way to solve this,
6537 * let's keep things simple by re-using the existing slow path.
6540 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6542 unsigned long prev_energy = ULONG_MAX, best_energy = ULONG_MAX;
6543 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6544 int cpu, best_energy_cpu = prev_cpu;
6545 struct perf_domain *head, *pd;
6546 unsigned long cpu_cap, util;
6547 struct sched_domain *sd;
6550 pd = rcu_dereference(rd->pd);
6551 if (!pd || READ_ONCE(rd->overutilized))
6556 * Energy-aware wake-up happens on the lowest sched_domain starting
6557 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6559 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6560 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6565 sync_entity_load_avg(&p->se);
6566 if (!task_util_est(p))
6569 for (; pd; pd = pd->next) {
6570 unsigned long cur_energy, spare_cap, max_spare_cap = 0;
6571 int max_spare_cap_cpu = -1;
6573 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
6574 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6577 /* Skip CPUs that will be overutilized. */
6578 util = cpu_util_next(cpu, p, cpu);
6579 cpu_cap = capacity_of(cpu);
6580 if (cpu_cap * 1024 < util * capacity_margin)
6583 /* Always use prev_cpu as a candidate. */
6584 if (cpu == prev_cpu) {
6585 prev_energy = compute_energy(p, prev_cpu, head);
6586 best_energy = min(best_energy, prev_energy);
6591 * Find the CPU with the maximum spare capacity in
6592 * the performance domain
6594 spare_cap = cpu_cap - util;
6595 if (spare_cap > max_spare_cap) {
6596 max_spare_cap = spare_cap;
6597 max_spare_cap_cpu = cpu;
6601 /* Evaluate the energy impact of using this CPU. */
6602 if (max_spare_cap_cpu >= 0) {
6603 cur_energy = compute_energy(p, max_spare_cap_cpu, head);
6604 if (cur_energy < best_energy) {
6605 best_energy = cur_energy;
6606 best_energy_cpu = max_spare_cap_cpu;
6614 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6615 * least 6% of the energy used by prev_cpu.
6617 if (prev_energy == ULONG_MAX)
6618 return best_energy_cpu;
6620 if ((prev_energy - best_energy) > (prev_energy >> 4))
6621 return best_energy_cpu;
6632 * select_task_rq_fair: Select target runqueue for the waking task in domains
6633 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6634 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6636 * Balances load by selecting the idlest CPU in the idlest group, or under
6637 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6639 * Returns the target CPU number.
6641 * preempt must be disabled.
6644 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6646 struct sched_domain *tmp, *sd = NULL;
6647 int cpu = smp_processor_id();
6648 int new_cpu = prev_cpu;
6649 int want_affine = 0;
6650 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6652 if (sd_flag & SD_BALANCE_WAKE) {
6655 if (sched_energy_enabled()) {
6656 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
6662 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu) &&
6663 cpumask_test_cpu(cpu, &p->cpus_allowed);
6667 for_each_domain(cpu, tmp) {
6668 if (!(tmp->flags & SD_LOAD_BALANCE))
6672 * If both 'cpu' and 'prev_cpu' are part of this domain,
6673 * cpu is a valid SD_WAKE_AFFINE target.
6675 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6676 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6677 if (cpu != prev_cpu)
6678 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6680 sd = NULL; /* Prefer wake_affine over balance flags */
6684 if (tmp->flags & sd_flag)
6686 else if (!want_affine)
6692 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6693 } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6696 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6699 current->recent_used_cpu = cpu;
6706 static void detach_entity_cfs_rq(struct sched_entity *se);
6709 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6710 * cfs_rq_of(p) references at time of call are still valid and identify the
6711 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6713 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6716 * As blocked tasks retain absolute vruntime the migration needs to
6717 * deal with this by subtracting the old and adding the new
6718 * min_vruntime -- the latter is done by enqueue_entity() when placing
6719 * the task on the new runqueue.
6721 if (p->state == TASK_WAKING) {
6722 struct sched_entity *se = &p->se;
6723 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6726 #ifndef CONFIG_64BIT
6727 u64 min_vruntime_copy;
6730 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6732 min_vruntime = cfs_rq->min_vruntime;
6733 } while (min_vruntime != min_vruntime_copy);
6735 min_vruntime = cfs_rq->min_vruntime;
6738 se->vruntime -= min_vruntime;
6741 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6743 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6744 * rq->lock and can modify state directly.
6746 lockdep_assert_held(&task_rq(p)->lock);
6747 detach_entity_cfs_rq(&p->se);
6751 * We are supposed to update the task to "current" time, then
6752 * its up to date and ready to go to new CPU/cfs_rq. But we
6753 * have difficulty in getting what current time is, so simply
6754 * throw away the out-of-date time. This will result in the
6755 * wakee task is less decayed, but giving the wakee more load
6758 remove_entity_load_avg(&p->se);
6761 /* Tell new CPU we are migrated */
6762 p->se.avg.last_update_time = 0;
6764 /* We have migrated, no longer consider this task hot */
6765 p->se.exec_start = 0;
6767 update_scan_period(p, new_cpu);
6770 static void task_dead_fair(struct task_struct *p)
6772 remove_entity_load_avg(&p->se);
6774 #endif /* CONFIG_SMP */
6776 static unsigned long wakeup_gran(struct sched_entity *se)
6778 unsigned long gran = sysctl_sched_wakeup_granularity;
6781 * Since its curr running now, convert the gran from real-time
6782 * to virtual-time in his units.
6784 * By using 'se' instead of 'curr' we penalize light tasks, so
6785 * they get preempted easier. That is, if 'se' < 'curr' then
6786 * the resulting gran will be larger, therefore penalizing the
6787 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6788 * be smaller, again penalizing the lighter task.
6790 * This is especially important for buddies when the leftmost
6791 * task is higher priority than the buddy.
6793 return calc_delta_fair(gran, se);
6797 * Should 'se' preempt 'curr'.
6811 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6813 s64 gran, vdiff = curr->vruntime - se->vruntime;
6818 gran = wakeup_gran(se);
6825 static void set_last_buddy(struct sched_entity *se)
6827 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6830 for_each_sched_entity(se) {
6831 if (SCHED_WARN_ON(!se->on_rq))
6833 cfs_rq_of(se)->last = se;
6837 static void set_next_buddy(struct sched_entity *se)
6839 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6842 for_each_sched_entity(se) {
6843 if (SCHED_WARN_ON(!se->on_rq))
6845 cfs_rq_of(se)->next = se;
6849 static void set_skip_buddy(struct sched_entity *se)
6851 for_each_sched_entity(se)
6852 cfs_rq_of(se)->skip = se;
6856 * Preempt the current task with a newly woken task if needed:
6858 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6860 struct task_struct *curr = rq->curr;
6861 struct sched_entity *se = &curr->se, *pse = &p->se;
6862 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6863 int scale = cfs_rq->nr_running >= sched_nr_latency;
6864 int next_buddy_marked = 0;
6866 if (unlikely(se == pse))
6870 * This is possible from callers such as attach_tasks(), in which we
6871 * unconditionally check_prempt_curr() after an enqueue (which may have
6872 * lead to a throttle). This both saves work and prevents false
6873 * next-buddy nomination below.
6875 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6878 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6879 set_next_buddy(pse);
6880 next_buddy_marked = 1;
6884 * We can come here with TIF_NEED_RESCHED already set from new task
6887 * Note: this also catches the edge-case of curr being in a throttled
6888 * group (e.g. via set_curr_task), since update_curr() (in the
6889 * enqueue of curr) will have resulted in resched being set. This
6890 * prevents us from potentially nominating it as a false LAST_BUDDY
6893 if (test_tsk_need_resched(curr))
6896 /* Idle tasks are by definition preempted by non-idle tasks. */
6897 if (unlikely(task_has_idle_policy(curr)) &&
6898 likely(!task_has_idle_policy(p)))
6902 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6903 * is driven by the tick):
6905 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6908 find_matching_se(&se, &pse);
6909 update_curr(cfs_rq_of(se));
6911 if (wakeup_preempt_entity(se, pse) == 1) {
6913 * Bias pick_next to pick the sched entity that is
6914 * triggering this preemption.
6916 if (!next_buddy_marked)
6917 set_next_buddy(pse);
6926 * Only set the backward buddy when the current task is still
6927 * on the rq. This can happen when a wakeup gets interleaved
6928 * with schedule on the ->pre_schedule() or idle_balance()
6929 * point, either of which can * drop the rq lock.
6931 * Also, during early boot the idle thread is in the fair class,
6932 * for obvious reasons its a bad idea to schedule back to it.
6934 if (unlikely(!se->on_rq || curr == rq->idle))
6937 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6941 static struct task_struct *
6942 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6944 struct cfs_rq *cfs_rq = &rq->cfs;
6945 struct sched_entity *se;
6946 struct task_struct *p;
6950 if (!cfs_rq->nr_running)
6953 #ifdef CONFIG_FAIR_GROUP_SCHED
6954 if (prev->sched_class != &fair_sched_class)
6958 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6959 * likely that a next task is from the same cgroup as the current.
6961 * Therefore attempt to avoid putting and setting the entire cgroup
6962 * hierarchy, only change the part that actually changes.
6966 struct sched_entity *curr = cfs_rq->curr;
6969 * Since we got here without doing put_prev_entity() we also
6970 * have to consider cfs_rq->curr. If it is still a runnable
6971 * entity, update_curr() will update its vruntime, otherwise
6972 * forget we've ever seen it.
6976 update_curr(cfs_rq);
6981 * This call to check_cfs_rq_runtime() will do the
6982 * throttle and dequeue its entity in the parent(s).
6983 * Therefore the nr_running test will indeed
6986 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6989 if (!cfs_rq->nr_running)
6996 se = pick_next_entity(cfs_rq, curr);
6997 cfs_rq = group_cfs_rq(se);
7003 * Since we haven't yet done put_prev_entity and if the selected task
7004 * is a different task than we started out with, try and touch the
7005 * least amount of cfs_rqs.
7008 struct sched_entity *pse = &prev->se;
7010 while (!(cfs_rq = is_same_group(se, pse))) {
7011 int se_depth = se->depth;
7012 int pse_depth = pse->depth;
7014 if (se_depth <= pse_depth) {
7015 put_prev_entity(cfs_rq_of(pse), pse);
7016 pse = parent_entity(pse);
7018 if (se_depth >= pse_depth) {
7019 set_next_entity(cfs_rq_of(se), se);
7020 se = parent_entity(se);
7024 put_prev_entity(cfs_rq, pse);
7025 set_next_entity(cfs_rq, se);
7032 put_prev_task(rq, prev);
7035 se = pick_next_entity(cfs_rq, NULL);
7036 set_next_entity(cfs_rq, se);
7037 cfs_rq = group_cfs_rq(se);
7042 done: __maybe_unused;
7045 * Move the next running task to the front of
7046 * the list, so our cfs_tasks list becomes MRU
7049 list_move(&p->se.group_node, &rq->cfs_tasks);
7052 if (hrtick_enabled(rq))
7053 hrtick_start_fair(rq, p);
7055 update_misfit_status(p, rq);
7060 update_misfit_status(NULL, rq);
7061 new_tasks = idle_balance(rq, rf);
7064 * Because idle_balance() releases (and re-acquires) rq->lock, it is
7065 * possible for any higher priority task to appear. In that case we
7066 * must re-start the pick_next_entity() loop.
7075 * rq is about to be idle, check if we need to update the
7076 * lost_idle_time of clock_pelt
7078 update_idle_rq_clock_pelt(rq);
7084 * Account for a descheduled task:
7086 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
7088 struct sched_entity *se = &prev->se;
7089 struct cfs_rq *cfs_rq;
7091 for_each_sched_entity(se) {
7092 cfs_rq = cfs_rq_of(se);
7093 put_prev_entity(cfs_rq, se);
7098 * sched_yield() is very simple
7100 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7102 static void yield_task_fair(struct rq *rq)
7104 struct task_struct *curr = rq->curr;
7105 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7106 struct sched_entity *se = &curr->se;
7109 * Are we the only task in the tree?
7111 if (unlikely(rq->nr_running == 1))
7114 clear_buddies(cfs_rq, se);
7116 if (curr->policy != SCHED_BATCH) {
7117 update_rq_clock(rq);
7119 * Update run-time statistics of the 'current'.
7121 update_curr(cfs_rq);
7123 * Tell update_rq_clock() that we've just updated,
7124 * so we don't do microscopic update in schedule()
7125 * and double the fastpath cost.
7127 rq_clock_skip_update(rq);
7133 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
7135 struct sched_entity *se = &p->se;
7137 /* throttled hierarchies are not runnable */
7138 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
7141 /* Tell the scheduler that we'd really like pse to run next. */
7144 yield_task_fair(rq);
7150 /**************************************************
7151 * Fair scheduling class load-balancing methods.
7155 * The purpose of load-balancing is to achieve the same basic fairness the
7156 * per-CPU scheduler provides, namely provide a proportional amount of compute
7157 * time to each task. This is expressed in the following equation:
7159 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7161 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7162 * W_i,0 is defined as:
7164 * W_i,0 = \Sum_j w_i,j (2)
7166 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7167 * is derived from the nice value as per sched_prio_to_weight[].
7169 * The weight average is an exponential decay average of the instantaneous
7172 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7174 * C_i is the compute capacity of CPU i, typically it is the
7175 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7176 * can also include other factors [XXX].
7178 * To achieve this balance we define a measure of imbalance which follows
7179 * directly from (1):
7181 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7183 * We them move tasks around to minimize the imbalance. In the continuous
7184 * function space it is obvious this converges, in the discrete case we get
7185 * a few fun cases generally called infeasible weight scenarios.
7188 * - infeasible weights;
7189 * - local vs global optima in the discrete case. ]
7194 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7195 * for all i,j solution, we create a tree of CPUs that follows the hardware
7196 * topology where each level pairs two lower groups (or better). This results
7197 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7198 * tree to only the first of the previous level and we decrease the frequency
7199 * of load-balance at each level inv. proportional to the number of CPUs in
7205 * \Sum { --- * --- * 2^i } = O(n) (5)
7207 * `- size of each group
7208 * | | `- number of CPUs doing load-balance
7210 * `- sum over all levels
7212 * Coupled with a limit on how many tasks we can migrate every balance pass,
7213 * this makes (5) the runtime complexity of the balancer.
7215 * An important property here is that each CPU is still (indirectly) connected
7216 * to every other CPU in at most O(log n) steps:
7218 * The adjacency matrix of the resulting graph is given by:
7221 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7224 * And you'll find that:
7226 * A^(log_2 n)_i,j != 0 for all i,j (7)
7228 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7229 * The task movement gives a factor of O(m), giving a convergence complexity
7232 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7237 * In order to avoid CPUs going idle while there's still work to do, new idle
7238 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7239 * tree itself instead of relying on other CPUs to bring it work.
7241 * This adds some complexity to both (5) and (8) but it reduces the total idle
7249 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7252 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7257 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7259 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7261 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7264 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7265 * rewrite all of this once again.]
7268 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7270 enum fbq_type { regular, remote, all };
7279 #define LBF_ALL_PINNED 0x01
7280 #define LBF_NEED_BREAK 0x02
7281 #define LBF_DST_PINNED 0x04
7282 #define LBF_SOME_PINNED 0x08
7283 #define LBF_NOHZ_STATS 0x10
7284 #define LBF_NOHZ_AGAIN 0x20
7287 struct sched_domain *sd;
7295 struct cpumask *dst_grpmask;
7297 enum cpu_idle_type idle;
7299 /* The set of CPUs under consideration for load-balancing */
7300 struct cpumask *cpus;
7305 unsigned int loop_break;
7306 unsigned int loop_max;
7308 enum fbq_type fbq_type;
7309 enum group_type src_grp_type;
7310 struct list_head tasks;
7314 * Is this task likely cache-hot:
7316 static int task_hot(struct task_struct *p, struct lb_env *env)
7320 lockdep_assert_held(&env->src_rq->lock);
7322 if (p->sched_class != &fair_sched_class)
7325 if (unlikely(task_has_idle_policy(p)))
7329 * Buddy candidates are cache hot:
7331 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7332 (&p->se == cfs_rq_of(&p->se)->next ||
7333 &p->se == cfs_rq_of(&p->se)->last))
7336 if (sysctl_sched_migration_cost == -1)
7338 if (sysctl_sched_migration_cost == 0)
7341 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7343 return delta < (s64)sysctl_sched_migration_cost;
7346 #ifdef CONFIG_NUMA_BALANCING
7348 * Returns 1, if task migration degrades locality
7349 * Returns 0, if task migration improves locality i.e migration preferred.
7350 * Returns -1, if task migration is not affected by locality.
7352 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7354 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7355 unsigned long src_weight, dst_weight;
7356 int src_nid, dst_nid, dist;
7358 if (!static_branch_likely(&sched_numa_balancing))
7361 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7364 src_nid = cpu_to_node(env->src_cpu);
7365 dst_nid = cpu_to_node(env->dst_cpu);
7367 if (src_nid == dst_nid)
7370 /* Migrating away from the preferred node is always bad. */
7371 if (src_nid == p->numa_preferred_nid) {
7372 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7378 /* Encourage migration to the preferred node. */
7379 if (dst_nid == p->numa_preferred_nid)
7382 /* Leaving a core idle is often worse than degrading locality. */
7383 if (env->idle == CPU_IDLE)
7386 dist = node_distance(src_nid, dst_nid);
7388 src_weight = group_weight(p, src_nid, dist);
7389 dst_weight = group_weight(p, dst_nid, dist);
7391 src_weight = task_weight(p, src_nid, dist);
7392 dst_weight = task_weight(p, dst_nid, dist);
7395 return dst_weight < src_weight;
7399 static inline int migrate_degrades_locality(struct task_struct *p,
7407 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7410 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7414 lockdep_assert_held(&env->src_rq->lock);
7417 * We do not migrate tasks that are:
7418 * 1) throttled_lb_pair, or
7419 * 2) cannot be migrated to this CPU due to cpus_allowed, or
7420 * 3) running (obviously), or
7421 * 4) are cache-hot on their current CPU.
7423 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7426 if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) {
7429 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7431 env->flags |= LBF_SOME_PINNED;
7434 * Remember if this task can be migrated to any other CPU in
7435 * our sched_group. We may want to revisit it if we couldn't
7436 * meet load balance goals by pulling other tasks on src_cpu.
7438 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7439 * already computed one in current iteration.
7441 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7444 /* Prevent to re-select dst_cpu via env's CPUs: */
7445 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7446 if (cpumask_test_cpu(cpu, &p->cpus_allowed)) {
7447 env->flags |= LBF_DST_PINNED;
7448 env->new_dst_cpu = cpu;
7456 /* Record that we found atleast one task that could run on dst_cpu */
7457 env->flags &= ~LBF_ALL_PINNED;
7459 if (task_running(env->src_rq, p)) {
7460 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7465 * Aggressive migration if:
7466 * 1) destination numa is preferred
7467 * 2) task is cache cold, or
7468 * 3) too many balance attempts have failed.
7470 tsk_cache_hot = migrate_degrades_locality(p, env);
7471 if (tsk_cache_hot == -1)
7472 tsk_cache_hot = task_hot(p, env);
7474 if (tsk_cache_hot <= 0 ||
7475 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7476 if (tsk_cache_hot == 1) {
7477 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7478 schedstat_inc(p->se.statistics.nr_forced_migrations);
7483 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7488 * detach_task() -- detach the task for the migration specified in env
7490 static void detach_task(struct task_struct *p, struct lb_env *env)
7492 lockdep_assert_held(&env->src_rq->lock);
7494 p->on_rq = TASK_ON_RQ_MIGRATING;
7495 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7496 set_task_cpu(p, env->dst_cpu);
7500 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7501 * part of active balancing operations within "domain".
7503 * Returns a task if successful and NULL otherwise.
7505 static struct task_struct *detach_one_task(struct lb_env *env)
7507 struct task_struct *p;
7509 lockdep_assert_held(&env->src_rq->lock);
7511 list_for_each_entry_reverse(p,
7512 &env->src_rq->cfs_tasks, se.group_node) {
7513 if (!can_migrate_task(p, env))
7516 detach_task(p, env);
7519 * Right now, this is only the second place where
7520 * lb_gained[env->idle] is updated (other is detach_tasks)
7521 * so we can safely collect stats here rather than
7522 * inside detach_tasks().
7524 schedstat_inc(env->sd->lb_gained[env->idle]);
7530 static const unsigned int sched_nr_migrate_break = 32;
7533 * detach_tasks() -- tries to detach up to imbalance weighted load from
7534 * busiest_rq, as part of a balancing operation within domain "sd".
7536 * Returns number of detached tasks if successful and 0 otherwise.
7538 static int detach_tasks(struct lb_env *env)
7540 struct list_head *tasks = &env->src_rq->cfs_tasks;
7541 struct task_struct *p;
7545 lockdep_assert_held(&env->src_rq->lock);
7547 if (env->imbalance <= 0)
7550 while (!list_empty(tasks)) {
7552 * We don't want to steal all, otherwise we may be treated likewise,
7553 * which could at worst lead to a livelock crash.
7555 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7558 p = list_last_entry(tasks, struct task_struct, se.group_node);
7561 /* We've more or less seen every task there is, call it quits */
7562 if (env->loop > env->loop_max)
7565 /* take a breather every nr_migrate tasks */
7566 if (env->loop > env->loop_break) {
7567 env->loop_break += sched_nr_migrate_break;
7568 env->flags |= LBF_NEED_BREAK;
7572 if (!can_migrate_task(p, env))
7575 load = task_h_load(p);
7577 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7580 if ((load / 2) > env->imbalance)
7583 detach_task(p, env);
7584 list_add(&p->se.group_node, &env->tasks);
7587 env->imbalance -= load;
7589 #ifdef CONFIG_PREEMPT
7591 * NEWIDLE balancing is a source of latency, so preemptible
7592 * kernels will stop after the first task is detached to minimize
7593 * the critical section.
7595 if (env->idle == CPU_NEWLY_IDLE)
7600 * We only want to steal up to the prescribed amount of
7603 if (env->imbalance <= 0)
7608 list_move(&p->se.group_node, tasks);
7612 * Right now, this is one of only two places we collect this stat
7613 * so we can safely collect detach_one_task() stats here rather
7614 * than inside detach_one_task().
7616 schedstat_add(env->sd->lb_gained[env->idle], detached);
7622 * attach_task() -- attach the task detached by detach_task() to its new rq.
7624 static void attach_task(struct rq *rq, struct task_struct *p)
7626 lockdep_assert_held(&rq->lock);
7628 BUG_ON(task_rq(p) != rq);
7629 activate_task(rq, p, ENQUEUE_NOCLOCK);
7630 p->on_rq = TASK_ON_RQ_QUEUED;
7631 check_preempt_curr(rq, p, 0);
7635 * attach_one_task() -- attaches the task returned from detach_one_task() to
7638 static void attach_one_task(struct rq *rq, struct task_struct *p)
7643 update_rq_clock(rq);
7649 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7652 static void attach_tasks(struct lb_env *env)
7654 struct list_head *tasks = &env->tasks;
7655 struct task_struct *p;
7658 rq_lock(env->dst_rq, &rf);
7659 update_rq_clock(env->dst_rq);
7661 while (!list_empty(tasks)) {
7662 p = list_first_entry(tasks, struct task_struct, se.group_node);
7663 list_del_init(&p->se.group_node);
7665 attach_task(env->dst_rq, p);
7668 rq_unlock(env->dst_rq, &rf);
7671 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7673 if (cfs_rq->avg.load_avg)
7676 if (cfs_rq->avg.util_avg)
7682 static inline bool others_have_blocked(struct rq *rq)
7684 if (READ_ONCE(rq->avg_rt.util_avg))
7687 if (READ_ONCE(rq->avg_dl.util_avg))
7690 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7691 if (READ_ONCE(rq->avg_irq.util_avg))
7698 #ifdef CONFIG_FAIR_GROUP_SCHED
7700 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7702 if (cfs_rq->load.weight)
7705 if (cfs_rq->avg.load_sum)
7708 if (cfs_rq->avg.util_sum)
7711 if (cfs_rq->avg.runnable_load_sum)
7717 static void update_blocked_averages(int cpu)
7719 struct rq *rq = cpu_rq(cpu);
7720 struct cfs_rq *cfs_rq, *pos;
7721 const struct sched_class *curr_class;
7725 rq_lock_irqsave(rq, &rf);
7726 update_rq_clock(rq);
7729 * Iterates the task_group tree in a bottom up fashion, see
7730 * list_add_leaf_cfs_rq() for details.
7732 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7733 struct sched_entity *se;
7735 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq))
7736 update_tg_load_avg(cfs_rq, 0);
7738 /* Propagate pending load changes to the parent, if any: */
7739 se = cfs_rq->tg->se[cpu];
7740 if (se && !skip_blocked_update(se))
7741 update_load_avg(cfs_rq_of(se), se, 0);
7744 * There can be a lot of idle CPU cgroups. Don't let fully
7745 * decayed cfs_rqs linger on the list.
7747 if (cfs_rq_is_decayed(cfs_rq))
7748 list_del_leaf_cfs_rq(cfs_rq);
7750 /* Don't need periodic decay once load/util_avg are null */
7751 if (cfs_rq_has_blocked(cfs_rq))
7755 curr_class = rq->curr->sched_class;
7756 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class);
7757 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class);
7758 update_irq_load_avg(rq, 0);
7759 /* Don't need periodic decay once load/util_avg are null */
7760 if (others_have_blocked(rq))
7763 #ifdef CONFIG_NO_HZ_COMMON
7764 rq->last_blocked_load_update_tick = jiffies;
7766 rq->has_blocked_load = 0;
7768 rq_unlock_irqrestore(rq, &rf);
7772 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7773 * This needs to be done in a top-down fashion because the load of a child
7774 * group is a fraction of its parents load.
7776 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7778 struct rq *rq = rq_of(cfs_rq);
7779 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7780 unsigned long now = jiffies;
7783 if (cfs_rq->last_h_load_update == now)
7786 cfs_rq->h_load_next = NULL;
7787 for_each_sched_entity(se) {
7788 cfs_rq = cfs_rq_of(se);
7789 cfs_rq->h_load_next = se;
7790 if (cfs_rq->last_h_load_update == now)
7795 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7796 cfs_rq->last_h_load_update = now;
7799 while ((se = cfs_rq->h_load_next) != NULL) {
7800 load = cfs_rq->h_load;
7801 load = div64_ul(load * se->avg.load_avg,
7802 cfs_rq_load_avg(cfs_rq) + 1);
7803 cfs_rq = group_cfs_rq(se);
7804 cfs_rq->h_load = load;
7805 cfs_rq->last_h_load_update = now;
7809 static unsigned long task_h_load(struct task_struct *p)
7811 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7813 update_cfs_rq_h_load(cfs_rq);
7814 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7815 cfs_rq_load_avg(cfs_rq) + 1);
7818 static inline void update_blocked_averages(int cpu)
7820 struct rq *rq = cpu_rq(cpu);
7821 struct cfs_rq *cfs_rq = &rq->cfs;
7822 const struct sched_class *curr_class;
7825 rq_lock_irqsave(rq, &rf);
7826 update_rq_clock(rq);
7827 update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
7829 curr_class = rq->curr->sched_class;
7830 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class);
7831 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class);
7832 update_irq_load_avg(rq, 0);
7833 #ifdef CONFIG_NO_HZ_COMMON
7834 rq->last_blocked_load_update_tick = jiffies;
7835 if (!cfs_rq_has_blocked(cfs_rq) && !others_have_blocked(rq))
7836 rq->has_blocked_load = 0;
7838 rq_unlock_irqrestore(rq, &rf);
7841 static unsigned long task_h_load(struct task_struct *p)
7843 return p->se.avg.load_avg;
7847 /********** Helpers for find_busiest_group ************************/
7850 * sg_lb_stats - stats of a sched_group required for load_balancing
7852 struct sg_lb_stats {
7853 unsigned long avg_load; /*Avg load across the CPUs of the group */
7854 unsigned long group_load; /* Total load over the CPUs of the group */
7855 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7856 unsigned long load_per_task;
7857 unsigned long group_capacity;
7858 unsigned long group_util; /* Total utilization of the group */
7859 unsigned int sum_nr_running; /* Nr tasks running in the group */
7860 unsigned int idle_cpus;
7861 unsigned int group_weight;
7862 enum group_type group_type;
7863 int group_no_capacity;
7864 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
7865 #ifdef CONFIG_NUMA_BALANCING
7866 unsigned int nr_numa_running;
7867 unsigned int nr_preferred_running;
7872 * sd_lb_stats - Structure to store the statistics of a sched_domain
7873 * during load balancing.
7875 struct sd_lb_stats {
7876 struct sched_group *busiest; /* Busiest group in this sd */
7877 struct sched_group *local; /* Local group in this sd */
7878 unsigned long total_running;
7879 unsigned long total_load; /* Total load of all groups in sd */
7880 unsigned long total_capacity; /* Total capacity of all groups in sd */
7881 unsigned long avg_load; /* Average load across all groups in sd */
7883 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7884 struct sg_lb_stats local_stat; /* Statistics of the local group */
7887 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7890 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7891 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7892 * We must however clear busiest_stat::avg_load because
7893 * update_sd_pick_busiest() reads this before assignment.
7895 *sds = (struct sd_lb_stats){
7898 .total_running = 0UL,
7900 .total_capacity = 0UL,
7903 .sum_nr_running = 0,
7904 .group_type = group_other,
7910 * get_sd_load_idx - Obtain the load index for a given sched domain.
7911 * @sd: The sched_domain whose load_idx is to be obtained.
7912 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7914 * Return: The load index.
7916 static inline int get_sd_load_idx(struct sched_domain *sd,
7917 enum cpu_idle_type idle)
7923 load_idx = sd->busy_idx;
7926 case CPU_NEWLY_IDLE:
7927 load_idx = sd->newidle_idx;
7930 load_idx = sd->idle_idx;
7937 static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu)
7939 struct rq *rq = cpu_rq(cpu);
7940 unsigned long max = arch_scale_cpu_capacity(sd, cpu);
7941 unsigned long used, free;
7944 irq = cpu_util_irq(rq);
7946 if (unlikely(irq >= max))
7949 used = READ_ONCE(rq->avg_rt.util_avg);
7950 used += READ_ONCE(rq->avg_dl.util_avg);
7952 if (unlikely(used >= max))
7957 return scale_irq_capacity(free, irq, max);
7960 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7962 unsigned long capacity = scale_rt_capacity(sd, cpu);
7963 struct sched_group *sdg = sd->groups;
7965 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(sd, cpu);
7970 cpu_rq(cpu)->cpu_capacity = capacity;
7971 sdg->sgc->capacity = capacity;
7972 sdg->sgc->min_capacity = capacity;
7973 sdg->sgc->max_capacity = capacity;
7976 void update_group_capacity(struct sched_domain *sd, int cpu)
7978 struct sched_domain *child = sd->child;
7979 struct sched_group *group, *sdg = sd->groups;
7980 unsigned long capacity, min_capacity, max_capacity;
7981 unsigned long interval;
7983 interval = msecs_to_jiffies(sd->balance_interval);
7984 interval = clamp(interval, 1UL, max_load_balance_interval);
7985 sdg->sgc->next_update = jiffies + interval;
7988 update_cpu_capacity(sd, cpu);
7993 min_capacity = ULONG_MAX;
7996 if (child->flags & SD_OVERLAP) {
7998 * SD_OVERLAP domains cannot assume that child groups
7999 * span the current group.
8002 for_each_cpu(cpu, sched_group_span(sdg)) {
8003 struct sched_group_capacity *sgc;
8004 struct rq *rq = cpu_rq(cpu);
8007 * build_sched_domains() -> init_sched_groups_capacity()
8008 * gets here before we've attached the domains to the
8011 * Use capacity_of(), which is set irrespective of domains
8012 * in update_cpu_capacity().
8014 * This avoids capacity from being 0 and
8015 * causing divide-by-zero issues on boot.
8017 if (unlikely(!rq->sd)) {
8018 capacity += capacity_of(cpu);
8020 sgc = rq->sd->groups->sgc;
8021 capacity += sgc->capacity;
8024 min_capacity = min(capacity, min_capacity);
8025 max_capacity = max(capacity, max_capacity);
8029 * !SD_OVERLAP domains can assume that child groups
8030 * span the current group.
8033 group = child->groups;
8035 struct sched_group_capacity *sgc = group->sgc;
8037 capacity += sgc->capacity;
8038 min_capacity = min(sgc->min_capacity, min_capacity);
8039 max_capacity = max(sgc->max_capacity, max_capacity);
8040 group = group->next;
8041 } while (group != child->groups);
8044 sdg->sgc->capacity = capacity;
8045 sdg->sgc->min_capacity = min_capacity;
8046 sdg->sgc->max_capacity = max_capacity;
8050 * Check whether the capacity of the rq has been noticeably reduced by side
8051 * activity. The imbalance_pct is used for the threshold.
8052 * Return true is the capacity is reduced
8055 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
8057 return ((rq->cpu_capacity * sd->imbalance_pct) <
8058 (rq->cpu_capacity_orig * 100));
8062 * Check whether a rq has a misfit task and if it looks like we can actually
8063 * help that task: we can migrate the task to a CPU of higher capacity, or
8064 * the task's current CPU is heavily pressured.
8066 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
8068 return rq->misfit_task_load &&
8069 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
8070 check_cpu_capacity(rq, sd));
8074 * Group imbalance indicates (and tries to solve) the problem where balancing
8075 * groups is inadequate due to ->cpus_allowed constraints.
8077 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
8078 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
8081 * { 0 1 2 3 } { 4 5 6 7 }
8084 * If we were to balance group-wise we'd place two tasks in the first group and
8085 * two tasks in the second group. Clearly this is undesired as it will overload
8086 * cpu 3 and leave one of the CPUs in the second group unused.
8088 * The current solution to this issue is detecting the skew in the first group
8089 * by noticing the lower domain failed to reach balance and had difficulty
8090 * moving tasks due to affinity constraints.
8092 * When this is so detected; this group becomes a candidate for busiest; see
8093 * update_sd_pick_busiest(). And calculate_imbalance() and
8094 * find_busiest_group() avoid some of the usual balance conditions to allow it
8095 * to create an effective group imbalance.
8097 * This is a somewhat tricky proposition since the next run might not find the
8098 * group imbalance and decide the groups need to be balanced again. A most
8099 * subtle and fragile situation.
8102 static inline int sg_imbalanced(struct sched_group *group)
8104 return group->sgc->imbalance;
8108 * group_has_capacity returns true if the group has spare capacity that could
8109 * be used by some tasks.
8110 * We consider that a group has spare capacity if the * number of task is
8111 * smaller than the number of CPUs or if the utilization is lower than the
8112 * available capacity for CFS tasks.
8113 * For the latter, we use a threshold to stabilize the state, to take into
8114 * account the variance of the tasks' load and to return true if the available
8115 * capacity in meaningful for the load balancer.
8116 * As an example, an available capacity of 1% can appear but it doesn't make
8117 * any benefit for the load balance.
8120 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
8122 if (sgs->sum_nr_running < sgs->group_weight)
8125 if ((sgs->group_capacity * 100) >
8126 (sgs->group_util * env->sd->imbalance_pct))
8133 * group_is_overloaded returns true if the group has more tasks than it can
8135 * group_is_overloaded is not equals to !group_has_capacity because a group
8136 * with the exact right number of tasks, has no more spare capacity but is not
8137 * overloaded so both group_has_capacity and group_is_overloaded return
8141 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
8143 if (sgs->sum_nr_running <= sgs->group_weight)
8146 if ((sgs->group_capacity * 100) <
8147 (sgs->group_util * env->sd->imbalance_pct))
8154 * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller
8155 * per-CPU capacity than sched_group ref.
8158 group_smaller_min_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
8160 return sg->sgc->min_capacity * capacity_margin <
8161 ref->sgc->min_capacity * 1024;
8165 * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller
8166 * per-CPU capacity_orig than sched_group ref.
8169 group_smaller_max_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
8171 return sg->sgc->max_capacity * capacity_margin <
8172 ref->sgc->max_capacity * 1024;
8176 group_type group_classify(struct sched_group *group,
8177 struct sg_lb_stats *sgs)
8179 if (sgs->group_no_capacity)
8180 return group_overloaded;
8182 if (sg_imbalanced(group))
8183 return group_imbalanced;
8185 if (sgs->group_misfit_task_load)
8186 return group_misfit_task;
8191 static bool update_nohz_stats(struct rq *rq, bool force)
8193 #ifdef CONFIG_NO_HZ_COMMON
8194 unsigned int cpu = rq->cpu;
8196 if (!rq->has_blocked_load)
8199 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
8202 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
8205 update_blocked_averages(cpu);
8207 return rq->has_blocked_load;
8214 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8215 * @env: The load balancing environment.
8216 * @group: sched_group whose statistics are to be updated.
8217 * @sgs: variable to hold the statistics for this group.
8218 * @sg_status: Holds flag indicating the status of the sched_group
8220 static inline void update_sg_lb_stats(struct lb_env *env,
8221 struct sched_group *group,
8222 struct sg_lb_stats *sgs,
8225 int local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(group));
8226 int load_idx = get_sd_load_idx(env->sd, env->idle);
8230 memset(sgs, 0, sizeof(*sgs));
8232 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8233 struct rq *rq = cpu_rq(i);
8235 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
8236 env->flags |= LBF_NOHZ_AGAIN;
8238 /* Bias balancing toward CPUs of our domain: */
8240 load = target_load(i, load_idx);
8242 load = source_load(i, load_idx);
8244 sgs->group_load += load;
8245 sgs->group_util += cpu_util(i);
8246 sgs->sum_nr_running += rq->cfs.h_nr_running;
8248 nr_running = rq->nr_running;
8250 *sg_status |= SG_OVERLOAD;
8252 if (cpu_overutilized(i))
8253 *sg_status |= SG_OVERUTILIZED;
8255 #ifdef CONFIG_NUMA_BALANCING
8256 sgs->nr_numa_running += rq->nr_numa_running;
8257 sgs->nr_preferred_running += rq->nr_preferred_running;
8259 sgs->sum_weighted_load += weighted_cpuload(rq);
8261 * No need to call idle_cpu() if nr_running is not 0
8263 if (!nr_running && idle_cpu(i))
8266 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8267 sgs->group_misfit_task_load < rq->misfit_task_load) {
8268 sgs->group_misfit_task_load = rq->misfit_task_load;
8269 *sg_status |= SG_OVERLOAD;
8273 /* Adjust by relative CPU capacity of the group */
8274 sgs->group_capacity = group->sgc->capacity;
8275 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
8277 if (sgs->sum_nr_running)
8278 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
8280 sgs->group_weight = group->group_weight;
8282 sgs->group_no_capacity = group_is_overloaded(env, sgs);
8283 sgs->group_type = group_classify(group, sgs);
8287 * update_sd_pick_busiest - return 1 on busiest group
8288 * @env: The load balancing environment.
8289 * @sds: sched_domain statistics
8290 * @sg: sched_group candidate to be checked for being the busiest
8291 * @sgs: sched_group statistics
8293 * Determine if @sg is a busier group than the previously selected
8296 * Return: %true if @sg is a busier group than the previously selected
8297 * busiest group. %false otherwise.
8299 static bool update_sd_pick_busiest(struct lb_env *env,
8300 struct sd_lb_stats *sds,
8301 struct sched_group *sg,
8302 struct sg_lb_stats *sgs)
8304 struct sg_lb_stats *busiest = &sds->busiest_stat;
8307 * Don't try to pull misfit tasks we can't help.
8308 * We can use max_capacity here as reduction in capacity on some
8309 * CPUs in the group should either be possible to resolve
8310 * internally or be covered by avg_load imbalance (eventually).
8312 if (sgs->group_type == group_misfit_task &&
8313 (!group_smaller_max_cpu_capacity(sg, sds->local) ||
8314 !group_has_capacity(env, &sds->local_stat)))
8317 if (sgs->group_type > busiest->group_type)
8320 if (sgs->group_type < busiest->group_type)
8323 if (sgs->avg_load <= busiest->avg_load)
8326 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
8330 * Candidate sg has no more than one task per CPU and
8331 * has higher per-CPU capacity. Migrating tasks to less
8332 * capable CPUs may harm throughput. Maximize throughput,
8333 * power/energy consequences are not considered.
8335 if (sgs->sum_nr_running <= sgs->group_weight &&
8336 group_smaller_min_cpu_capacity(sds->local, sg))
8340 * If we have more than one misfit sg go with the biggest misfit.
8342 if (sgs->group_type == group_misfit_task &&
8343 sgs->group_misfit_task_load < busiest->group_misfit_task_load)
8347 /* This is the busiest node in its class. */
8348 if (!(env->sd->flags & SD_ASYM_PACKING))
8351 /* No ASYM_PACKING if target CPU is already busy */
8352 if (env->idle == CPU_NOT_IDLE)
8355 * ASYM_PACKING needs to move all the work to the highest
8356 * prority CPUs in the group, therefore mark all groups
8357 * of lower priority than ourself as busy.
8359 if (sgs->sum_nr_running &&
8360 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
8364 /* Prefer to move from lowest priority CPU's work */
8365 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
8366 sg->asym_prefer_cpu))
8373 #ifdef CONFIG_NUMA_BALANCING
8374 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8376 if (sgs->sum_nr_running > sgs->nr_numa_running)
8378 if (sgs->sum_nr_running > sgs->nr_preferred_running)
8383 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8385 if (rq->nr_running > rq->nr_numa_running)
8387 if (rq->nr_running > rq->nr_preferred_running)
8392 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8397 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8401 #endif /* CONFIG_NUMA_BALANCING */
8404 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8405 * @env: The load balancing environment.
8406 * @sds: variable to hold the statistics for this sched_domain.
8408 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8410 struct sched_domain *child = env->sd->child;
8411 struct sched_group *sg = env->sd->groups;
8412 struct sg_lb_stats *local = &sds->local_stat;
8413 struct sg_lb_stats tmp_sgs;
8414 bool prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
8417 #ifdef CONFIG_NO_HZ_COMMON
8418 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
8419 env->flags |= LBF_NOHZ_STATS;
8423 struct sg_lb_stats *sgs = &tmp_sgs;
8426 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
8431 if (env->idle != CPU_NEWLY_IDLE ||
8432 time_after_eq(jiffies, sg->sgc->next_update))
8433 update_group_capacity(env->sd, env->dst_cpu);
8436 update_sg_lb_stats(env, sg, sgs, &sg_status);
8442 * In case the child domain prefers tasks go to siblings
8443 * first, lower the sg capacity so that we'll try
8444 * and move all the excess tasks away. We lower the capacity
8445 * of a group only if the local group has the capacity to fit
8446 * these excess tasks. The extra check prevents the case where
8447 * you always pull from the heaviest group when it is already
8448 * under-utilized (possible with a large weight task outweighs
8449 * the tasks on the system).
8451 if (prefer_sibling && sds->local &&
8452 group_has_capacity(env, local) &&
8453 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
8454 sgs->group_no_capacity = 1;
8455 sgs->group_type = group_classify(sg, sgs);
8458 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8460 sds->busiest_stat = *sgs;
8464 /* Now, start updating sd_lb_stats */
8465 sds->total_running += sgs->sum_nr_running;
8466 sds->total_load += sgs->group_load;
8467 sds->total_capacity += sgs->group_capacity;
8470 } while (sg != env->sd->groups);
8472 #ifdef CONFIG_NO_HZ_COMMON
8473 if ((env->flags & LBF_NOHZ_AGAIN) &&
8474 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8476 WRITE_ONCE(nohz.next_blocked,
8477 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8481 if (env->sd->flags & SD_NUMA)
8482 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8484 if (!env->sd->parent) {
8485 struct root_domain *rd = env->dst_rq->rd;
8487 /* update overload indicator if we are at root domain */
8488 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
8490 /* Update over-utilization (tipping point, U >= 0) indicator */
8491 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
8492 } else if (sg_status & SG_OVERUTILIZED) {
8493 WRITE_ONCE(env->dst_rq->rd->overutilized, SG_OVERUTILIZED);
8498 * check_asym_packing - Check to see if the group is packed into the
8501 * This is primarily intended to used at the sibling level. Some
8502 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8503 * case of POWER7, it can move to lower SMT modes only when higher
8504 * threads are idle. When in lower SMT modes, the threads will
8505 * perform better since they share less core resources. Hence when we
8506 * have idle threads, we want them to be the higher ones.
8508 * This packing function is run on idle threads. It checks to see if
8509 * the busiest CPU in this domain (core in the P7 case) has a higher
8510 * CPU number than the packing function is being run on. Here we are
8511 * assuming lower CPU number will be equivalent to lower a SMT thread
8514 * Return: 1 when packing is required and a task should be moved to
8515 * this CPU. The amount of the imbalance is returned in env->imbalance.
8517 * @env: The load balancing environment.
8518 * @sds: Statistics of the sched_domain which is to be packed
8520 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8524 if (!(env->sd->flags & SD_ASYM_PACKING))
8527 if (env->idle == CPU_NOT_IDLE)
8533 busiest_cpu = sds->busiest->asym_prefer_cpu;
8534 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
8537 env->imbalance = sds->busiest_stat.group_load;
8543 * fix_small_imbalance - Calculate the minor imbalance that exists
8544 * amongst the groups of a sched_domain, during
8546 * @env: The load balancing environment.
8547 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8550 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8552 unsigned long tmp, capa_now = 0, capa_move = 0;
8553 unsigned int imbn = 2;
8554 unsigned long scaled_busy_load_per_task;
8555 struct sg_lb_stats *local, *busiest;
8557 local = &sds->local_stat;
8558 busiest = &sds->busiest_stat;
8560 if (!local->sum_nr_running)
8561 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8562 else if (busiest->load_per_task > local->load_per_task)
8565 scaled_busy_load_per_task =
8566 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8567 busiest->group_capacity;
8569 if (busiest->avg_load + scaled_busy_load_per_task >=
8570 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8571 env->imbalance = busiest->load_per_task;
8576 * OK, we don't have enough imbalance to justify moving tasks,
8577 * however we may be able to increase total CPU capacity used by
8581 capa_now += busiest->group_capacity *
8582 min(busiest->load_per_task, busiest->avg_load);
8583 capa_now += local->group_capacity *
8584 min(local->load_per_task, local->avg_load);
8585 capa_now /= SCHED_CAPACITY_SCALE;
8587 /* Amount of load we'd subtract */
8588 if (busiest->avg_load > scaled_busy_load_per_task) {
8589 capa_move += busiest->group_capacity *
8590 min(busiest->load_per_task,
8591 busiest->avg_load - scaled_busy_load_per_task);
8594 /* Amount of load we'd add */
8595 if (busiest->avg_load * busiest->group_capacity <
8596 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8597 tmp = (busiest->avg_load * busiest->group_capacity) /
8598 local->group_capacity;
8600 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8601 local->group_capacity;
8603 capa_move += local->group_capacity *
8604 min(local->load_per_task, local->avg_load + tmp);
8605 capa_move /= SCHED_CAPACITY_SCALE;
8607 /* Move if we gain throughput */
8608 if (capa_move > capa_now)
8609 env->imbalance = busiest->load_per_task;
8613 * calculate_imbalance - Calculate the amount of imbalance present within the
8614 * groups of a given sched_domain during load balance.
8615 * @env: load balance environment
8616 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8618 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8620 unsigned long max_pull, load_above_capacity = ~0UL;
8621 struct sg_lb_stats *local, *busiest;
8623 local = &sds->local_stat;
8624 busiest = &sds->busiest_stat;
8626 if (busiest->group_type == group_imbalanced) {
8628 * In the group_imb case we cannot rely on group-wide averages
8629 * to ensure CPU-load equilibrium, look at wider averages. XXX
8631 busiest->load_per_task =
8632 min(busiest->load_per_task, sds->avg_load);
8636 * Avg load of busiest sg can be less and avg load of local sg can
8637 * be greater than avg load across all sgs of sd because avg load
8638 * factors in sg capacity and sgs with smaller group_type are
8639 * skipped when updating the busiest sg:
8641 if (busiest->group_type != group_misfit_task &&
8642 (busiest->avg_load <= sds->avg_load ||
8643 local->avg_load >= sds->avg_load)) {
8645 return fix_small_imbalance(env, sds);
8649 * If there aren't any idle CPUs, avoid creating some.
8651 if (busiest->group_type == group_overloaded &&
8652 local->group_type == group_overloaded) {
8653 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
8654 if (load_above_capacity > busiest->group_capacity) {
8655 load_above_capacity -= busiest->group_capacity;
8656 load_above_capacity *= scale_load_down(NICE_0_LOAD);
8657 load_above_capacity /= busiest->group_capacity;
8659 load_above_capacity = ~0UL;
8663 * We're trying to get all the CPUs to the average_load, so we don't
8664 * want to push ourselves above the average load, nor do we wish to
8665 * reduce the max loaded CPU below the average load. At the same time,
8666 * we also don't want to reduce the group load below the group
8667 * capacity. Thus we look for the minimum possible imbalance.
8669 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8671 /* How much load to actually move to equalise the imbalance */
8672 env->imbalance = min(
8673 max_pull * busiest->group_capacity,
8674 (sds->avg_load - local->avg_load) * local->group_capacity
8675 ) / SCHED_CAPACITY_SCALE;
8677 /* Boost imbalance to allow misfit task to be balanced. */
8678 if (busiest->group_type == group_misfit_task) {
8679 env->imbalance = max_t(long, env->imbalance,
8680 busiest->group_misfit_task_load);
8684 * if *imbalance is less than the average load per runnable task
8685 * there is no guarantee that any tasks will be moved so we'll have
8686 * a think about bumping its value to force at least one task to be
8689 if (env->imbalance < busiest->load_per_task)
8690 return fix_small_imbalance(env, sds);
8693 /******* find_busiest_group() helpers end here *********************/
8696 * find_busiest_group - Returns the busiest group within the sched_domain
8697 * if there is an imbalance.
8699 * Also calculates the amount of weighted load which should be moved
8700 * to restore balance.
8702 * @env: The load balancing environment.
8704 * Return: - The busiest group if imbalance exists.
8706 static struct sched_group *find_busiest_group(struct lb_env *env)
8708 struct sg_lb_stats *local, *busiest;
8709 struct sd_lb_stats sds;
8711 init_sd_lb_stats(&sds);
8714 * Compute the various statistics relavent for load balancing at
8717 update_sd_lb_stats(env, &sds);
8719 if (sched_energy_enabled()) {
8720 struct root_domain *rd = env->dst_rq->rd;
8722 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
8726 local = &sds.local_stat;
8727 busiest = &sds.busiest_stat;
8729 /* ASYM feature bypasses nice load balance check */
8730 if (check_asym_packing(env, &sds))
8733 /* There is no busy sibling group to pull tasks from */
8734 if (!sds.busiest || busiest->sum_nr_running == 0)
8737 /* XXX broken for overlapping NUMA groups */
8738 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8739 / sds.total_capacity;
8742 * If the busiest group is imbalanced the below checks don't
8743 * work because they assume all things are equal, which typically
8744 * isn't true due to cpus_allowed constraints and the like.
8746 if (busiest->group_type == group_imbalanced)
8750 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8751 * capacities from resulting in underutilization due to avg_load.
8753 if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8754 busiest->group_no_capacity)
8757 /* Misfit tasks should be dealt with regardless of the avg load */
8758 if (busiest->group_type == group_misfit_task)
8762 * If the local group is busier than the selected busiest group
8763 * don't try and pull any tasks.
8765 if (local->avg_load >= busiest->avg_load)
8769 * Don't pull any tasks if this group is already above the domain
8772 if (local->avg_load >= sds.avg_load)
8775 if (env->idle == CPU_IDLE) {
8777 * This CPU is idle. If the busiest group is not overloaded
8778 * and there is no imbalance between this and busiest group
8779 * wrt idle CPUs, it is balanced. The imbalance becomes
8780 * significant if the diff is greater than 1 otherwise we
8781 * might end up to just move the imbalance on another group
8783 if ((busiest->group_type != group_overloaded) &&
8784 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8788 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8789 * imbalance_pct to be conservative.
8791 if (100 * busiest->avg_load <=
8792 env->sd->imbalance_pct * local->avg_load)
8797 /* Looks like there is an imbalance. Compute it */
8798 env->src_grp_type = busiest->group_type;
8799 calculate_imbalance(env, &sds);
8800 return env->imbalance ? sds.busiest : NULL;
8808 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
8810 static struct rq *find_busiest_queue(struct lb_env *env,
8811 struct sched_group *group)
8813 struct rq *busiest = NULL, *rq;
8814 unsigned long busiest_load = 0, busiest_capacity = 1;
8817 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8818 unsigned long capacity, wl;
8822 rt = fbq_classify_rq(rq);
8825 * We classify groups/runqueues into three groups:
8826 * - regular: there are !numa tasks
8827 * - remote: there are numa tasks that run on the 'wrong' node
8828 * - all: there is no distinction
8830 * In order to avoid migrating ideally placed numa tasks,
8831 * ignore those when there's better options.
8833 * If we ignore the actual busiest queue to migrate another
8834 * task, the next balance pass can still reduce the busiest
8835 * queue by moving tasks around inside the node.
8837 * If we cannot move enough load due to this classification
8838 * the next pass will adjust the group classification and
8839 * allow migration of more tasks.
8841 * Both cases only affect the total convergence complexity.
8843 if (rt > env->fbq_type)
8847 * For ASYM_CPUCAPACITY domains with misfit tasks we simply
8848 * seek the "biggest" misfit task.
8850 if (env->src_grp_type == group_misfit_task) {
8851 if (rq->misfit_task_load > busiest_load) {
8852 busiest_load = rq->misfit_task_load;
8859 capacity = capacity_of(i);
8862 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
8863 * eventually lead to active_balancing high->low capacity.
8864 * Higher per-CPU capacity is considered better than balancing
8867 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8868 capacity_of(env->dst_cpu) < capacity &&
8869 rq->nr_running == 1)
8872 wl = weighted_cpuload(rq);
8875 * When comparing with imbalance, use weighted_cpuload()
8876 * which is not scaled with the CPU capacity.
8879 if (rq->nr_running == 1 && wl > env->imbalance &&
8880 !check_cpu_capacity(rq, env->sd))
8884 * For the load comparisons with the other CPU's, consider
8885 * the weighted_cpuload() scaled with the CPU capacity, so
8886 * that the load can be moved away from the CPU that is
8887 * potentially running at a lower capacity.
8889 * Thus we're looking for max(wl_i / capacity_i), crosswise
8890 * multiplication to rid ourselves of the division works out
8891 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8892 * our previous maximum.
8894 if (wl * busiest_capacity > busiest_load * capacity) {
8896 busiest_capacity = capacity;
8905 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8906 * so long as it is large enough.
8908 #define MAX_PINNED_INTERVAL 512
8911 asym_active_balance(struct lb_env *env)
8914 * ASYM_PACKING needs to force migrate tasks from busy but
8915 * lower priority CPUs in order to pack all tasks in the
8916 * highest priority CPUs.
8918 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
8919 sched_asym_prefer(env->dst_cpu, env->src_cpu);
8923 voluntary_active_balance(struct lb_env *env)
8925 struct sched_domain *sd = env->sd;
8927 if (asym_active_balance(env))
8931 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8932 * It's worth migrating the task if the src_cpu's capacity is reduced
8933 * because of other sched_class or IRQs if more capacity stays
8934 * available on dst_cpu.
8936 if ((env->idle != CPU_NOT_IDLE) &&
8937 (env->src_rq->cfs.h_nr_running == 1)) {
8938 if ((check_cpu_capacity(env->src_rq, sd)) &&
8939 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8943 if (env->src_grp_type == group_misfit_task)
8949 static int need_active_balance(struct lb_env *env)
8951 struct sched_domain *sd = env->sd;
8953 if (voluntary_active_balance(env))
8956 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8959 static int active_load_balance_cpu_stop(void *data);
8961 static int should_we_balance(struct lb_env *env)
8963 struct sched_group *sg = env->sd->groups;
8964 int cpu, balance_cpu = -1;
8967 * Ensure the balancing environment is consistent; can happen
8968 * when the softirq triggers 'during' hotplug.
8970 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8974 * In the newly idle case, we will allow all the CPUs
8975 * to do the newly idle load balance.
8977 if (env->idle == CPU_NEWLY_IDLE)
8980 /* Try to find first idle CPU */
8981 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8989 if (balance_cpu == -1)
8990 balance_cpu = group_balance_cpu(sg);
8993 * First idle CPU or the first CPU(busiest) in this sched group
8994 * is eligible for doing load balancing at this and above domains.
8996 return balance_cpu == env->dst_cpu;
9000 * Check this_cpu to ensure it is balanced within domain. Attempt to move
9001 * tasks if there is an imbalance.
9003 static int load_balance(int this_cpu, struct rq *this_rq,
9004 struct sched_domain *sd, enum cpu_idle_type idle,
9005 int *continue_balancing)
9007 int ld_moved, cur_ld_moved, active_balance = 0;
9008 struct sched_domain *sd_parent = sd->parent;
9009 struct sched_group *group;
9012 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
9014 struct lb_env env = {
9016 .dst_cpu = this_cpu,
9018 .dst_grpmask = sched_group_span(sd->groups),
9020 .loop_break = sched_nr_migrate_break,
9023 .tasks = LIST_HEAD_INIT(env.tasks),
9026 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
9028 schedstat_inc(sd->lb_count[idle]);
9031 if (!should_we_balance(&env)) {
9032 *continue_balancing = 0;
9036 group = find_busiest_group(&env);
9038 schedstat_inc(sd->lb_nobusyg[idle]);
9042 busiest = find_busiest_queue(&env, group);
9044 schedstat_inc(sd->lb_nobusyq[idle]);
9048 BUG_ON(busiest == env.dst_rq);
9050 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
9052 env.src_cpu = busiest->cpu;
9053 env.src_rq = busiest;
9056 if (busiest->nr_running > 1) {
9058 * Attempt to move tasks. If find_busiest_group has found
9059 * an imbalance but busiest->nr_running <= 1, the group is
9060 * still unbalanced. ld_moved simply stays zero, so it is
9061 * correctly treated as an imbalance.
9063 env.flags |= LBF_ALL_PINNED;
9064 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
9067 rq_lock_irqsave(busiest, &rf);
9068 update_rq_clock(busiest);
9071 * cur_ld_moved - load moved in current iteration
9072 * ld_moved - cumulative load moved across iterations
9074 cur_ld_moved = detach_tasks(&env);
9077 * We've detached some tasks from busiest_rq. Every
9078 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
9079 * unlock busiest->lock, and we are able to be sure
9080 * that nobody can manipulate the tasks in parallel.
9081 * See task_rq_lock() family for the details.
9084 rq_unlock(busiest, &rf);
9088 ld_moved += cur_ld_moved;
9091 local_irq_restore(rf.flags);
9093 if (env.flags & LBF_NEED_BREAK) {
9094 env.flags &= ~LBF_NEED_BREAK;
9099 * Revisit (affine) tasks on src_cpu that couldn't be moved to
9100 * us and move them to an alternate dst_cpu in our sched_group
9101 * where they can run. The upper limit on how many times we
9102 * iterate on same src_cpu is dependent on number of CPUs in our
9105 * This changes load balance semantics a bit on who can move
9106 * load to a given_cpu. In addition to the given_cpu itself
9107 * (or a ilb_cpu acting on its behalf where given_cpu is
9108 * nohz-idle), we now have balance_cpu in a position to move
9109 * load to given_cpu. In rare situations, this may cause
9110 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
9111 * _independently_ and at _same_ time to move some load to
9112 * given_cpu) causing exceess load to be moved to given_cpu.
9113 * This however should not happen so much in practice and
9114 * moreover subsequent load balance cycles should correct the
9115 * excess load moved.
9117 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
9119 /* Prevent to re-select dst_cpu via env's CPUs */
9120 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
9122 env.dst_rq = cpu_rq(env.new_dst_cpu);
9123 env.dst_cpu = env.new_dst_cpu;
9124 env.flags &= ~LBF_DST_PINNED;
9126 env.loop_break = sched_nr_migrate_break;
9129 * Go back to "more_balance" rather than "redo" since we
9130 * need to continue with same src_cpu.
9136 * We failed to reach balance because of affinity.
9139 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9141 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
9142 *group_imbalance = 1;
9145 /* All tasks on this runqueue were pinned by CPU affinity */
9146 if (unlikely(env.flags & LBF_ALL_PINNED)) {
9147 __cpumask_clear_cpu(cpu_of(busiest), cpus);
9149 * Attempting to continue load balancing at the current
9150 * sched_domain level only makes sense if there are
9151 * active CPUs remaining as possible busiest CPUs to
9152 * pull load from which are not contained within the
9153 * destination group that is receiving any migrated
9156 if (!cpumask_subset(cpus, env.dst_grpmask)) {
9158 env.loop_break = sched_nr_migrate_break;
9161 goto out_all_pinned;
9166 schedstat_inc(sd->lb_failed[idle]);
9168 * Increment the failure counter only on periodic balance.
9169 * We do not want newidle balance, which can be very
9170 * frequent, pollute the failure counter causing
9171 * excessive cache_hot migrations and active balances.
9173 if (idle != CPU_NEWLY_IDLE)
9174 sd->nr_balance_failed++;
9176 if (need_active_balance(&env)) {
9177 unsigned long flags;
9179 raw_spin_lock_irqsave(&busiest->lock, flags);
9182 * Don't kick the active_load_balance_cpu_stop,
9183 * if the curr task on busiest CPU can't be
9184 * moved to this_cpu:
9186 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
9187 raw_spin_unlock_irqrestore(&busiest->lock,
9189 env.flags |= LBF_ALL_PINNED;
9190 goto out_one_pinned;
9194 * ->active_balance synchronizes accesses to
9195 * ->active_balance_work. Once set, it's cleared
9196 * only after active load balance is finished.
9198 if (!busiest->active_balance) {
9199 busiest->active_balance = 1;
9200 busiest->push_cpu = this_cpu;
9203 raw_spin_unlock_irqrestore(&busiest->lock, flags);
9205 if (active_balance) {
9206 stop_one_cpu_nowait(cpu_of(busiest),
9207 active_load_balance_cpu_stop, busiest,
9208 &busiest->active_balance_work);
9211 /* We've kicked active balancing, force task migration. */
9212 sd->nr_balance_failed = sd->cache_nice_tries+1;
9215 sd->nr_balance_failed = 0;
9217 if (likely(!active_balance) || voluntary_active_balance(&env)) {
9218 /* We were unbalanced, so reset the balancing interval */
9219 sd->balance_interval = sd->min_interval;
9222 * If we've begun active balancing, start to back off. This
9223 * case may not be covered by the all_pinned logic if there
9224 * is only 1 task on the busy runqueue (because we don't call
9227 if (sd->balance_interval < sd->max_interval)
9228 sd->balance_interval *= 2;
9235 * We reach balance although we may have faced some affinity
9236 * constraints. Clear the imbalance flag if it was set.
9239 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9241 if (*group_imbalance)
9242 *group_imbalance = 0;
9247 * We reach balance because all tasks are pinned at this level so
9248 * we can't migrate them. Let the imbalance flag set so parent level
9249 * can try to migrate them.
9251 schedstat_inc(sd->lb_balanced[idle]);
9253 sd->nr_balance_failed = 0;
9259 * idle_balance() disregards balance intervals, so we could repeatedly
9260 * reach this code, which would lead to balance_interval skyrocketting
9261 * in a short amount of time. Skip the balance_interval increase logic
9264 if (env.idle == CPU_NEWLY_IDLE)
9267 /* tune up the balancing interval */
9268 if ((env.flags & LBF_ALL_PINNED &&
9269 sd->balance_interval < MAX_PINNED_INTERVAL) ||
9270 sd->balance_interval < sd->max_interval)
9271 sd->balance_interval *= 2;
9276 static inline unsigned long
9277 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
9279 unsigned long interval = sd->balance_interval;
9282 interval *= sd->busy_factor;
9284 /* scale ms to jiffies */
9285 interval = msecs_to_jiffies(interval);
9286 interval = clamp(interval, 1UL, max_load_balance_interval);
9292 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
9294 unsigned long interval, next;
9296 /* used by idle balance, so cpu_busy = 0 */
9297 interval = get_sd_balance_interval(sd, 0);
9298 next = sd->last_balance + interval;
9300 if (time_after(*next_balance, next))
9301 *next_balance = next;
9305 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9306 * running tasks off the busiest CPU onto idle CPUs. It requires at
9307 * least 1 task to be running on each physical CPU where possible, and
9308 * avoids physical / logical imbalances.
9310 static int active_load_balance_cpu_stop(void *data)
9312 struct rq *busiest_rq = data;
9313 int busiest_cpu = cpu_of(busiest_rq);
9314 int target_cpu = busiest_rq->push_cpu;
9315 struct rq *target_rq = cpu_rq(target_cpu);
9316 struct sched_domain *sd;
9317 struct task_struct *p = NULL;
9320 rq_lock_irq(busiest_rq, &rf);
9322 * Between queueing the stop-work and running it is a hole in which
9323 * CPUs can become inactive. We should not move tasks from or to
9326 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
9329 /* Make sure the requested CPU hasn't gone down in the meantime: */
9330 if (unlikely(busiest_cpu != smp_processor_id() ||
9331 !busiest_rq->active_balance))
9334 /* Is there any task to move? */
9335 if (busiest_rq->nr_running <= 1)
9339 * This condition is "impossible", if it occurs
9340 * we need to fix it. Originally reported by
9341 * Bjorn Helgaas on a 128-CPU setup.
9343 BUG_ON(busiest_rq == target_rq);
9345 /* Search for an sd spanning us and the target CPU. */
9347 for_each_domain(target_cpu, sd) {
9348 if ((sd->flags & SD_LOAD_BALANCE) &&
9349 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9354 struct lb_env env = {
9356 .dst_cpu = target_cpu,
9357 .dst_rq = target_rq,
9358 .src_cpu = busiest_rq->cpu,
9359 .src_rq = busiest_rq,
9362 * can_migrate_task() doesn't need to compute new_dst_cpu
9363 * for active balancing. Since we have CPU_IDLE, but no
9364 * @dst_grpmask we need to make that test go away with lying
9367 .flags = LBF_DST_PINNED,
9370 schedstat_inc(sd->alb_count);
9371 update_rq_clock(busiest_rq);
9373 p = detach_one_task(&env);
9375 schedstat_inc(sd->alb_pushed);
9376 /* Active balancing done, reset the failure counter. */
9377 sd->nr_balance_failed = 0;
9379 schedstat_inc(sd->alb_failed);
9384 busiest_rq->active_balance = 0;
9385 rq_unlock(busiest_rq, &rf);
9388 attach_one_task(target_rq, p);
9395 static DEFINE_SPINLOCK(balancing);
9398 * Scale the max load_balance interval with the number of CPUs in the system.
9399 * This trades load-balance latency on larger machines for less cross talk.
9401 void update_max_interval(void)
9403 max_load_balance_interval = HZ*num_online_cpus()/10;
9407 * It checks each scheduling domain to see if it is due to be balanced,
9408 * and initiates a balancing operation if so.
9410 * Balancing parameters are set up in init_sched_domains.
9412 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9414 int continue_balancing = 1;
9416 unsigned long interval;
9417 struct sched_domain *sd;
9418 /* Earliest time when we have to do rebalance again */
9419 unsigned long next_balance = jiffies + 60*HZ;
9420 int update_next_balance = 0;
9421 int need_serialize, need_decay = 0;
9425 for_each_domain(cpu, sd) {
9427 * Decay the newidle max times here because this is a regular
9428 * visit to all the domains. Decay ~1% per second.
9430 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9431 sd->max_newidle_lb_cost =
9432 (sd->max_newidle_lb_cost * 253) / 256;
9433 sd->next_decay_max_lb_cost = jiffies + HZ;
9436 max_cost += sd->max_newidle_lb_cost;
9438 if (!(sd->flags & SD_LOAD_BALANCE))
9442 * Stop the load balance at this level. There is another
9443 * CPU in our sched group which is doing load balancing more
9446 if (!continue_balancing) {
9452 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9454 need_serialize = sd->flags & SD_SERIALIZE;
9455 if (need_serialize) {
9456 if (!spin_trylock(&balancing))
9460 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9461 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9463 * The LBF_DST_PINNED logic could have changed
9464 * env->dst_cpu, so we can't know our idle
9465 * state even if we migrated tasks. Update it.
9467 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9469 sd->last_balance = jiffies;
9470 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9473 spin_unlock(&balancing);
9475 if (time_after(next_balance, sd->last_balance + interval)) {
9476 next_balance = sd->last_balance + interval;
9477 update_next_balance = 1;
9482 * Ensure the rq-wide value also decays but keep it at a
9483 * reasonable floor to avoid funnies with rq->avg_idle.
9485 rq->max_idle_balance_cost =
9486 max((u64)sysctl_sched_migration_cost, max_cost);
9491 * next_balance will be updated only when there is a need.
9492 * When the cpu is attached to null domain for ex, it will not be
9495 if (likely(update_next_balance)) {
9496 rq->next_balance = next_balance;
9498 #ifdef CONFIG_NO_HZ_COMMON
9500 * If this CPU has been elected to perform the nohz idle
9501 * balance. Other idle CPUs have already rebalanced with
9502 * nohz_idle_balance() and nohz.next_balance has been
9503 * updated accordingly. This CPU is now running the idle load
9504 * balance for itself and we need to update the
9505 * nohz.next_balance accordingly.
9507 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9508 nohz.next_balance = rq->next_balance;
9513 static inline int on_null_domain(struct rq *rq)
9515 return unlikely(!rcu_dereference_sched(rq->sd));
9518 #ifdef CONFIG_NO_HZ_COMMON
9520 * idle load balancing details
9521 * - When one of the busy CPUs notice that there may be an idle rebalancing
9522 * needed, they will kick the idle load balancer, which then does idle
9523 * load balancing for all the idle CPUs.
9526 static inline int find_new_ilb(void)
9528 int ilb = cpumask_first(nohz.idle_cpus_mask);
9530 if (ilb < nr_cpu_ids && idle_cpu(ilb))
9537 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
9538 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
9539 * CPU (if there is one).
9541 static void kick_ilb(unsigned int flags)
9545 nohz.next_balance++;
9547 ilb_cpu = find_new_ilb();
9549 if (ilb_cpu >= nr_cpu_ids)
9552 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
9553 if (flags & NOHZ_KICK_MASK)
9557 * Use smp_send_reschedule() instead of resched_cpu().
9558 * This way we generate a sched IPI on the target CPU which
9559 * is idle. And the softirq performing nohz idle load balance
9560 * will be run before returning from the IPI.
9562 smp_send_reschedule(ilb_cpu);
9566 * Current decision point for kicking the idle load balancer in the presence
9567 * of idle CPUs in the system.
9569 static void nohz_balancer_kick(struct rq *rq)
9571 unsigned long now = jiffies;
9572 struct sched_domain_shared *sds;
9573 struct sched_domain *sd;
9574 int nr_busy, i, cpu = rq->cpu;
9575 unsigned int flags = 0;
9577 if (unlikely(rq->idle_balance))
9581 * We may be recently in ticked or tickless idle mode. At the first
9582 * busy tick after returning from idle, we will update the busy stats.
9584 nohz_balance_exit_idle(rq);
9587 * None are in tickless mode and hence no need for NOHZ idle load
9590 if (likely(!atomic_read(&nohz.nr_cpus)))
9593 if (READ_ONCE(nohz.has_blocked) &&
9594 time_after(now, READ_ONCE(nohz.next_blocked)))
9595 flags = NOHZ_STATS_KICK;
9597 if (time_before(now, nohz.next_balance))
9600 if (rq->nr_running >= 2) {
9601 flags = NOHZ_KICK_MASK;
9606 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9609 * If there is an imbalance between LLC domains (IOW we could
9610 * increase the overall cache use), we need some less-loaded LLC
9611 * domain to pull some load. Likewise, we may need to spread
9612 * load within the current LLC domain (e.g. packed SMT cores but
9613 * other CPUs are idle). We can't really know from here how busy
9614 * the others are - so just get a nohz balance going if it looks
9615 * like this LLC domain has tasks we could move.
9617 nr_busy = atomic_read(&sds->nr_busy_cpus);
9619 flags = NOHZ_KICK_MASK;
9625 sd = rcu_dereference(rq->sd);
9628 * If there's a CFS task and the current CPU has reduced
9629 * capacity; kick the ILB to see if there's a better CPU to run
9632 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
9633 flags = NOHZ_KICK_MASK;
9638 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
9641 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
9642 * to run the misfit task on.
9644 if (check_misfit_status(rq, sd)) {
9645 flags = NOHZ_KICK_MASK;
9650 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
9653 * When ASYM_PACKING; see if there's a more preferred CPU
9654 * currently idle; in which case, kick the ILB to move tasks
9657 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
9658 if (sched_asym_prefer(i, cpu)) {
9659 flags = NOHZ_KICK_MASK;
9671 static void set_cpu_sd_state_busy(int cpu)
9673 struct sched_domain *sd;
9676 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9678 if (!sd || !sd->nohz_idle)
9682 atomic_inc(&sd->shared->nr_busy_cpus);
9687 void nohz_balance_exit_idle(struct rq *rq)
9689 SCHED_WARN_ON(rq != this_rq());
9691 if (likely(!rq->nohz_tick_stopped))
9694 rq->nohz_tick_stopped = 0;
9695 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
9696 atomic_dec(&nohz.nr_cpus);
9698 set_cpu_sd_state_busy(rq->cpu);
9701 static void set_cpu_sd_state_idle(int cpu)
9703 struct sched_domain *sd;
9706 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9708 if (!sd || sd->nohz_idle)
9712 atomic_dec(&sd->shared->nr_busy_cpus);
9718 * This routine will record that the CPU is going idle with tick stopped.
9719 * This info will be used in performing idle load balancing in the future.
9721 void nohz_balance_enter_idle(int cpu)
9723 struct rq *rq = cpu_rq(cpu);
9725 SCHED_WARN_ON(cpu != smp_processor_id());
9727 /* If this CPU is going down, then nothing needs to be done: */
9728 if (!cpu_active(cpu))
9731 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9732 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9736 * Can be set safely without rq->lock held
9737 * If a clear happens, it will have evaluated last additions because
9738 * rq->lock is held during the check and the clear
9740 rq->has_blocked_load = 1;
9743 * The tick is still stopped but load could have been added in the
9744 * meantime. We set the nohz.has_blocked flag to trig a check of the
9745 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
9746 * of nohz.has_blocked can only happen after checking the new load
9748 if (rq->nohz_tick_stopped)
9751 /* If we're a completely isolated CPU, we don't play: */
9752 if (on_null_domain(rq))
9755 rq->nohz_tick_stopped = 1;
9757 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9758 atomic_inc(&nohz.nr_cpus);
9761 * Ensures that if nohz_idle_balance() fails to observe our
9762 * @idle_cpus_mask store, it must observe the @has_blocked
9765 smp_mb__after_atomic();
9767 set_cpu_sd_state_idle(cpu);
9771 * Each time a cpu enter idle, we assume that it has blocked load and
9772 * enable the periodic update of the load of idle cpus
9774 WRITE_ONCE(nohz.has_blocked, 1);
9778 * Internal function that runs load balance for all idle cpus. The load balance
9779 * can be a simple update of blocked load or a complete load balance with
9780 * tasks movement depending of flags.
9781 * The function returns false if the loop has stopped before running
9782 * through all idle CPUs.
9784 static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
9785 enum cpu_idle_type idle)
9787 /* Earliest time when we have to do rebalance again */
9788 unsigned long now = jiffies;
9789 unsigned long next_balance = now + 60*HZ;
9790 bool has_blocked_load = false;
9791 int update_next_balance = 0;
9792 int this_cpu = this_rq->cpu;
9797 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
9800 * We assume there will be no idle load after this update and clear
9801 * the has_blocked flag. If a cpu enters idle in the mean time, it will
9802 * set the has_blocked flag and trig another update of idle load.
9803 * Because a cpu that becomes idle, is added to idle_cpus_mask before
9804 * setting the flag, we are sure to not clear the state and not
9805 * check the load of an idle cpu.
9807 WRITE_ONCE(nohz.has_blocked, 0);
9810 * Ensures that if we miss the CPU, we must see the has_blocked
9811 * store from nohz_balance_enter_idle().
9815 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9816 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9820 * If this CPU gets work to do, stop the load balancing
9821 * work being done for other CPUs. Next load
9822 * balancing owner will pick it up.
9824 if (need_resched()) {
9825 has_blocked_load = true;
9829 rq = cpu_rq(balance_cpu);
9831 has_blocked_load |= update_nohz_stats(rq, true);
9834 * If time for next balance is due,
9837 if (time_after_eq(jiffies, rq->next_balance)) {
9840 rq_lock_irqsave(rq, &rf);
9841 update_rq_clock(rq);
9842 cpu_load_update_idle(rq);
9843 rq_unlock_irqrestore(rq, &rf);
9845 if (flags & NOHZ_BALANCE_KICK)
9846 rebalance_domains(rq, CPU_IDLE);
9849 if (time_after(next_balance, rq->next_balance)) {
9850 next_balance = rq->next_balance;
9851 update_next_balance = 1;
9855 /* Newly idle CPU doesn't need an update */
9856 if (idle != CPU_NEWLY_IDLE) {
9857 update_blocked_averages(this_cpu);
9858 has_blocked_load |= this_rq->has_blocked_load;
9861 if (flags & NOHZ_BALANCE_KICK)
9862 rebalance_domains(this_rq, CPU_IDLE);
9864 WRITE_ONCE(nohz.next_blocked,
9865 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
9867 /* The full idle balance loop has been done */
9871 /* There is still blocked load, enable periodic update */
9872 if (has_blocked_load)
9873 WRITE_ONCE(nohz.has_blocked, 1);
9876 * next_balance will be updated only when there is a need.
9877 * When the CPU is attached to null domain for ex, it will not be
9880 if (likely(update_next_balance))
9881 nohz.next_balance = next_balance;
9887 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9888 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9890 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9892 int this_cpu = this_rq->cpu;
9895 if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
9898 if (idle != CPU_IDLE) {
9899 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9903 /* could be _relaxed() */
9904 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9905 if (!(flags & NOHZ_KICK_MASK))
9908 _nohz_idle_balance(this_rq, flags, idle);
9913 static void nohz_newidle_balance(struct rq *this_rq)
9915 int this_cpu = this_rq->cpu;
9918 * This CPU doesn't want to be disturbed by scheduler
9921 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
9924 /* Will wake up very soon. No time for doing anything else*/
9925 if (this_rq->avg_idle < sysctl_sched_migration_cost)
9928 /* Don't need to update blocked load of idle CPUs*/
9929 if (!READ_ONCE(nohz.has_blocked) ||
9930 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
9933 raw_spin_unlock(&this_rq->lock);
9935 * This CPU is going to be idle and blocked load of idle CPUs
9936 * need to be updated. Run the ilb locally as it is a good
9937 * candidate for ilb instead of waking up another idle CPU.
9938 * Kick an normal ilb if we failed to do the update.
9940 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
9941 kick_ilb(NOHZ_STATS_KICK);
9942 raw_spin_lock(&this_rq->lock);
9945 #else /* !CONFIG_NO_HZ_COMMON */
9946 static inline void nohz_balancer_kick(struct rq *rq) { }
9948 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9953 static inline void nohz_newidle_balance(struct rq *this_rq) { }
9954 #endif /* CONFIG_NO_HZ_COMMON */
9957 * idle_balance is called by schedule() if this_cpu is about to become
9958 * idle. Attempts to pull tasks from other CPUs.
9960 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
9962 unsigned long next_balance = jiffies + HZ;
9963 int this_cpu = this_rq->cpu;
9964 struct sched_domain *sd;
9965 int pulled_task = 0;
9969 * We must set idle_stamp _before_ calling idle_balance(), such that we
9970 * measure the duration of idle_balance() as idle time.
9972 this_rq->idle_stamp = rq_clock(this_rq);
9975 * Do not pull tasks towards !active CPUs...
9977 if (!cpu_active(this_cpu))
9981 * This is OK, because current is on_cpu, which avoids it being picked
9982 * for load-balance and preemption/IRQs are still disabled avoiding
9983 * further scheduler activity on it and we're being very careful to
9984 * re-start the picking loop.
9986 rq_unpin_lock(this_rq, rf);
9988 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
9989 !READ_ONCE(this_rq->rd->overload)) {
9992 sd = rcu_dereference_check_sched_domain(this_rq->sd);
9994 update_next_balance(sd, &next_balance);
9997 nohz_newidle_balance(this_rq);
10002 raw_spin_unlock(&this_rq->lock);
10004 update_blocked_averages(this_cpu);
10006 for_each_domain(this_cpu, sd) {
10007 int continue_balancing = 1;
10008 u64 t0, domain_cost;
10010 if (!(sd->flags & SD_LOAD_BALANCE))
10013 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
10014 update_next_balance(sd, &next_balance);
10018 if (sd->flags & SD_BALANCE_NEWIDLE) {
10019 t0 = sched_clock_cpu(this_cpu);
10021 pulled_task = load_balance(this_cpu, this_rq,
10022 sd, CPU_NEWLY_IDLE,
10023 &continue_balancing);
10025 domain_cost = sched_clock_cpu(this_cpu) - t0;
10026 if (domain_cost > sd->max_newidle_lb_cost)
10027 sd->max_newidle_lb_cost = domain_cost;
10029 curr_cost += domain_cost;
10032 update_next_balance(sd, &next_balance);
10035 * Stop searching for tasks to pull if there are
10036 * now runnable tasks on this rq.
10038 if (pulled_task || this_rq->nr_running > 0)
10043 raw_spin_lock(&this_rq->lock);
10045 if (curr_cost > this_rq->max_idle_balance_cost)
10046 this_rq->max_idle_balance_cost = curr_cost;
10050 * While browsing the domains, we released the rq lock, a task could
10051 * have been enqueued in the meantime. Since we're not going idle,
10052 * pretend we pulled a task.
10054 if (this_rq->cfs.h_nr_running && !pulled_task)
10057 /* Move the next balance forward */
10058 if (time_after(this_rq->next_balance, next_balance))
10059 this_rq->next_balance = next_balance;
10061 /* Is there a task of a high priority class? */
10062 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
10066 this_rq->idle_stamp = 0;
10068 rq_repin_lock(this_rq, rf);
10070 return pulled_task;
10074 * run_rebalance_domains is triggered when needed from the scheduler tick.
10075 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
10077 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
10079 struct rq *this_rq = this_rq();
10080 enum cpu_idle_type idle = this_rq->idle_balance ?
10081 CPU_IDLE : CPU_NOT_IDLE;
10084 * If this CPU has a pending nohz_balance_kick, then do the
10085 * balancing on behalf of the other idle CPUs whose ticks are
10086 * stopped. Do nohz_idle_balance *before* rebalance_domains to
10087 * give the idle CPUs a chance to load balance. Else we may
10088 * load balance only within the local sched_domain hierarchy
10089 * and abort nohz_idle_balance altogether if we pull some load.
10091 if (nohz_idle_balance(this_rq, idle))
10094 /* normal load balance */
10095 update_blocked_averages(this_rq->cpu);
10096 rebalance_domains(this_rq, idle);
10100 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
10102 void trigger_load_balance(struct rq *rq)
10104 /* Don't need to rebalance while attached to NULL domain */
10105 if (unlikely(on_null_domain(rq)))
10108 if (time_after_eq(jiffies, rq->next_balance))
10109 raise_softirq(SCHED_SOFTIRQ);
10111 nohz_balancer_kick(rq);
10114 static void rq_online_fair(struct rq *rq)
10118 update_runtime_enabled(rq);
10121 static void rq_offline_fair(struct rq *rq)
10125 /* Ensure any throttled groups are reachable by pick_next_task */
10126 unthrottle_offline_cfs_rqs(rq);
10129 #endif /* CONFIG_SMP */
10132 * scheduler tick hitting a task of our scheduling class.
10134 * NOTE: This function can be called remotely by the tick offload that
10135 * goes along full dynticks. Therefore no local assumption can be made
10136 * and everything must be accessed through the @rq and @curr passed in
10139 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
10141 struct cfs_rq *cfs_rq;
10142 struct sched_entity *se = &curr->se;
10144 for_each_sched_entity(se) {
10145 cfs_rq = cfs_rq_of(se);
10146 entity_tick(cfs_rq, se, queued);
10149 if (static_branch_unlikely(&sched_numa_balancing))
10150 task_tick_numa(rq, curr);
10152 update_misfit_status(curr, rq);
10153 update_overutilized_status(task_rq(curr));
10157 * called on fork with the child task as argument from the parent's context
10158 * - child not yet on the tasklist
10159 * - preemption disabled
10161 static void task_fork_fair(struct task_struct *p)
10163 struct cfs_rq *cfs_rq;
10164 struct sched_entity *se = &p->se, *curr;
10165 struct rq *rq = this_rq();
10166 struct rq_flags rf;
10169 update_rq_clock(rq);
10171 cfs_rq = task_cfs_rq(current);
10172 curr = cfs_rq->curr;
10174 update_curr(cfs_rq);
10175 se->vruntime = curr->vruntime;
10177 place_entity(cfs_rq, se, 1);
10179 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
10181 * Upon rescheduling, sched_class::put_prev_task() will place
10182 * 'current' within the tree based on its new key value.
10184 swap(curr->vruntime, se->vruntime);
10188 se->vruntime -= cfs_rq->min_vruntime;
10189 rq_unlock(rq, &rf);
10193 * Priority of the task has changed. Check to see if we preempt
10194 * the current task.
10197 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
10199 if (!task_on_rq_queued(p))
10203 * Reschedule if we are currently running on this runqueue and
10204 * our priority decreased, or if we are not currently running on
10205 * this runqueue and our priority is higher than the current's
10207 if (rq->curr == p) {
10208 if (p->prio > oldprio)
10211 check_preempt_curr(rq, p, 0);
10214 static inline bool vruntime_normalized(struct task_struct *p)
10216 struct sched_entity *se = &p->se;
10219 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
10220 * the dequeue_entity(.flags=0) will already have normalized the
10227 * When !on_rq, vruntime of the task has usually NOT been normalized.
10228 * But there are some cases where it has already been normalized:
10230 * - A forked child which is waiting for being woken up by
10231 * wake_up_new_task().
10232 * - A task which has been woken up by try_to_wake_up() and
10233 * waiting for actually being woken up by sched_ttwu_pending().
10235 if (!se->sum_exec_runtime ||
10236 (p->state == TASK_WAKING && p->sched_remote_wakeup))
10242 #ifdef CONFIG_FAIR_GROUP_SCHED
10244 * Propagate the changes of the sched_entity across the tg tree to make it
10245 * visible to the root
10247 static void propagate_entity_cfs_rq(struct sched_entity *se)
10249 struct cfs_rq *cfs_rq;
10251 /* Start to propagate at parent */
10254 for_each_sched_entity(se) {
10255 cfs_rq = cfs_rq_of(se);
10257 if (cfs_rq_throttled(cfs_rq))
10260 update_load_avg(cfs_rq, se, UPDATE_TG);
10264 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
10267 static void detach_entity_cfs_rq(struct sched_entity *se)
10269 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10271 /* Catch up with the cfs_rq and remove our load when we leave */
10272 update_load_avg(cfs_rq, se, 0);
10273 detach_entity_load_avg(cfs_rq, se);
10274 update_tg_load_avg(cfs_rq, false);
10275 propagate_entity_cfs_rq(se);
10278 static void attach_entity_cfs_rq(struct sched_entity *se)
10280 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10282 #ifdef CONFIG_FAIR_GROUP_SCHED
10284 * Since the real-depth could have been changed (only FAIR
10285 * class maintain depth value), reset depth properly.
10287 se->depth = se->parent ? se->parent->depth + 1 : 0;
10290 /* Synchronize entity with its cfs_rq */
10291 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
10292 attach_entity_load_avg(cfs_rq, se, 0);
10293 update_tg_load_avg(cfs_rq, false);
10294 propagate_entity_cfs_rq(se);
10297 static void detach_task_cfs_rq(struct task_struct *p)
10299 struct sched_entity *se = &p->se;
10300 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10302 if (!vruntime_normalized(p)) {
10304 * Fix up our vruntime so that the current sleep doesn't
10305 * cause 'unlimited' sleep bonus.
10307 place_entity(cfs_rq, se, 0);
10308 se->vruntime -= cfs_rq->min_vruntime;
10311 detach_entity_cfs_rq(se);
10314 static void attach_task_cfs_rq(struct task_struct *p)
10316 struct sched_entity *se = &p->se;
10317 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10319 attach_entity_cfs_rq(se);
10321 if (!vruntime_normalized(p))
10322 se->vruntime += cfs_rq->min_vruntime;
10325 static void switched_from_fair(struct rq *rq, struct task_struct *p)
10327 detach_task_cfs_rq(p);
10330 static void switched_to_fair(struct rq *rq, struct task_struct *p)
10332 attach_task_cfs_rq(p);
10334 if (task_on_rq_queued(p)) {
10336 * We were most likely switched from sched_rt, so
10337 * kick off the schedule if running, otherwise just see
10338 * if we can still preempt the current task.
10343 check_preempt_curr(rq, p, 0);
10347 /* Account for a task changing its policy or group.
10349 * This routine is mostly called to set cfs_rq->curr field when a task
10350 * migrates between groups/classes.
10352 static void set_curr_task_fair(struct rq *rq)
10354 struct sched_entity *se = &rq->curr->se;
10356 for_each_sched_entity(se) {
10357 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10359 set_next_entity(cfs_rq, se);
10360 /* ensure bandwidth has been allocated on our new cfs_rq */
10361 account_cfs_rq_runtime(cfs_rq, 0);
10365 void init_cfs_rq(struct cfs_rq *cfs_rq)
10367 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
10368 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
10369 #ifndef CONFIG_64BIT
10370 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
10373 raw_spin_lock_init(&cfs_rq->removed.lock);
10377 #ifdef CONFIG_FAIR_GROUP_SCHED
10378 static void task_set_group_fair(struct task_struct *p)
10380 struct sched_entity *se = &p->se;
10382 set_task_rq(p, task_cpu(p));
10383 se->depth = se->parent ? se->parent->depth + 1 : 0;
10386 static void task_move_group_fair(struct task_struct *p)
10388 detach_task_cfs_rq(p);
10389 set_task_rq(p, task_cpu(p));
10392 /* Tell se's cfs_rq has been changed -- migrated */
10393 p->se.avg.last_update_time = 0;
10395 attach_task_cfs_rq(p);
10398 static void task_change_group_fair(struct task_struct *p, int type)
10401 case TASK_SET_GROUP:
10402 task_set_group_fair(p);
10405 case TASK_MOVE_GROUP:
10406 task_move_group_fair(p);
10411 void free_fair_sched_group(struct task_group *tg)
10415 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
10417 for_each_possible_cpu(i) {
10419 kfree(tg->cfs_rq[i]);
10428 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10430 struct sched_entity *se;
10431 struct cfs_rq *cfs_rq;
10434 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
10437 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
10441 tg->shares = NICE_0_LOAD;
10443 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
10445 for_each_possible_cpu(i) {
10446 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
10447 GFP_KERNEL, cpu_to_node(i));
10451 se = kzalloc_node(sizeof(struct sched_entity),
10452 GFP_KERNEL, cpu_to_node(i));
10456 init_cfs_rq(cfs_rq);
10457 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
10458 init_entity_runnable_average(se);
10469 void online_fair_sched_group(struct task_group *tg)
10471 struct sched_entity *se;
10475 for_each_possible_cpu(i) {
10479 raw_spin_lock_irq(&rq->lock);
10480 update_rq_clock(rq);
10481 attach_entity_cfs_rq(se);
10482 sync_throttle(tg, i);
10483 raw_spin_unlock_irq(&rq->lock);
10487 void unregister_fair_sched_group(struct task_group *tg)
10489 unsigned long flags;
10493 for_each_possible_cpu(cpu) {
10495 remove_entity_load_avg(tg->se[cpu]);
10498 * Only empty task groups can be destroyed; so we can speculatively
10499 * check on_list without danger of it being re-added.
10501 if (!tg->cfs_rq[cpu]->on_list)
10506 raw_spin_lock_irqsave(&rq->lock, flags);
10507 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
10508 raw_spin_unlock_irqrestore(&rq->lock, flags);
10512 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
10513 struct sched_entity *se, int cpu,
10514 struct sched_entity *parent)
10516 struct rq *rq = cpu_rq(cpu);
10520 init_cfs_rq_runtime(cfs_rq);
10522 tg->cfs_rq[cpu] = cfs_rq;
10525 /* se could be NULL for root_task_group */
10530 se->cfs_rq = &rq->cfs;
10533 se->cfs_rq = parent->my_q;
10534 se->depth = parent->depth + 1;
10538 /* guarantee group entities always have weight */
10539 update_load_set(&se->load, NICE_0_LOAD);
10540 se->parent = parent;
10543 static DEFINE_MUTEX(shares_mutex);
10545 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10550 * We can't change the weight of the root cgroup.
10555 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10557 mutex_lock(&shares_mutex);
10558 if (tg->shares == shares)
10561 tg->shares = shares;
10562 for_each_possible_cpu(i) {
10563 struct rq *rq = cpu_rq(i);
10564 struct sched_entity *se = tg->se[i];
10565 struct rq_flags rf;
10567 /* Propagate contribution to hierarchy */
10568 rq_lock_irqsave(rq, &rf);
10569 update_rq_clock(rq);
10570 for_each_sched_entity(se) {
10571 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
10572 update_cfs_group(se);
10574 rq_unlock_irqrestore(rq, &rf);
10578 mutex_unlock(&shares_mutex);
10581 #else /* CONFIG_FAIR_GROUP_SCHED */
10583 void free_fair_sched_group(struct task_group *tg) { }
10585 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10590 void online_fair_sched_group(struct task_group *tg) { }
10592 void unregister_fair_sched_group(struct task_group *tg) { }
10594 #endif /* CONFIG_FAIR_GROUP_SCHED */
10597 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10599 struct sched_entity *se = &task->se;
10600 unsigned int rr_interval = 0;
10603 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10606 if (rq->cfs.load.weight)
10607 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10609 return rr_interval;
10613 * All the scheduling class methods:
10615 const struct sched_class fair_sched_class = {
10616 .next = &idle_sched_class,
10617 .enqueue_task = enqueue_task_fair,
10618 .dequeue_task = dequeue_task_fair,
10619 .yield_task = yield_task_fair,
10620 .yield_to_task = yield_to_task_fair,
10622 .check_preempt_curr = check_preempt_wakeup,
10624 .pick_next_task = pick_next_task_fair,
10625 .put_prev_task = put_prev_task_fair,
10628 .select_task_rq = select_task_rq_fair,
10629 .migrate_task_rq = migrate_task_rq_fair,
10631 .rq_online = rq_online_fair,
10632 .rq_offline = rq_offline_fair,
10634 .task_dead = task_dead_fair,
10635 .set_cpus_allowed = set_cpus_allowed_common,
10638 .set_curr_task = set_curr_task_fair,
10639 .task_tick = task_tick_fair,
10640 .task_fork = task_fork_fair,
10642 .prio_changed = prio_changed_fair,
10643 .switched_from = switched_from_fair,
10644 .switched_to = switched_to_fair,
10646 .get_rr_interval = get_rr_interval_fair,
10648 .update_curr = update_curr_fair,
10650 #ifdef CONFIG_FAIR_GROUP_SCHED
10651 .task_change_group = task_change_group_fair,
10655 #ifdef CONFIG_SCHED_DEBUG
10656 void print_cfs_stats(struct seq_file *m, int cpu)
10658 struct cfs_rq *cfs_rq, *pos;
10661 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
10662 print_cfs_rq(m, cpu, cfs_rq);
10666 #ifdef CONFIG_NUMA_BALANCING
10667 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10670 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10672 for_each_online_node(node) {
10673 if (p->numa_faults) {
10674 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10675 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10677 if (p->numa_group) {
10678 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
10679 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
10681 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10684 #endif /* CONFIG_NUMA_BALANCING */
10685 #endif /* CONFIG_SCHED_DEBUG */
10687 __init void init_sched_fair_class(void)
10690 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10692 #ifdef CONFIG_NO_HZ_COMMON
10693 nohz.next_balance = jiffies;
10694 nohz.next_blocked = jiffies;
10695 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);