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(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 = NUMA_NO_NODE;
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 != NUMA_NO_NODE);
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 != NUMA_NO_NODE);
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 == NUMA_NO_NODE || 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);
1471 /* Cached statistics for all CPUs within a node */
1475 /* Total compute capacity of CPUs on a node */
1476 unsigned long compute_capacity;
1480 * XXX borrowed from update_sg_lb_stats
1482 static void update_numa_stats(struct numa_stats *ns, int nid)
1486 memset(ns, 0, sizeof(*ns));
1487 for_each_cpu(cpu, cpumask_of_node(nid)) {
1488 struct rq *rq = cpu_rq(cpu);
1490 ns->load += weighted_cpuload(rq);
1491 ns->compute_capacity += capacity_of(cpu);
1496 struct task_numa_env {
1497 struct task_struct *p;
1499 int src_cpu, src_nid;
1500 int dst_cpu, dst_nid;
1502 struct numa_stats src_stats, dst_stats;
1507 struct task_struct *best_task;
1512 static void task_numa_assign(struct task_numa_env *env,
1513 struct task_struct *p, long imp)
1515 struct rq *rq = cpu_rq(env->dst_cpu);
1517 /* Bail out if run-queue part of active NUMA balance. */
1518 if (xchg(&rq->numa_migrate_on, 1))
1522 * Clear previous best_cpu/rq numa-migrate flag, since task now
1523 * found a better CPU to move/swap.
1525 if (env->best_cpu != -1) {
1526 rq = cpu_rq(env->best_cpu);
1527 WRITE_ONCE(rq->numa_migrate_on, 0);
1531 put_task_struct(env->best_task);
1536 env->best_imp = imp;
1537 env->best_cpu = env->dst_cpu;
1540 static bool load_too_imbalanced(long src_load, long dst_load,
1541 struct task_numa_env *env)
1544 long orig_src_load, orig_dst_load;
1545 long src_capacity, dst_capacity;
1548 * The load is corrected for the CPU capacity available on each node.
1551 * ------------ vs ---------
1552 * src_capacity dst_capacity
1554 src_capacity = env->src_stats.compute_capacity;
1555 dst_capacity = env->dst_stats.compute_capacity;
1557 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1559 orig_src_load = env->src_stats.load;
1560 orig_dst_load = env->dst_stats.load;
1562 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1564 /* Would this change make things worse? */
1565 return (imb > old_imb);
1569 * Maximum NUMA importance can be 1998 (2*999);
1570 * SMALLIMP @ 30 would be close to 1998/64.
1571 * Used to deter task migration.
1576 * This checks if the overall compute and NUMA accesses of the system would
1577 * be improved if the source tasks was migrated to the target dst_cpu taking
1578 * into account that it might be best if task running on the dst_cpu should
1579 * be exchanged with the source task
1581 static void task_numa_compare(struct task_numa_env *env,
1582 long taskimp, long groupimp, bool maymove)
1584 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1585 struct task_struct *cur;
1586 long src_load, dst_load;
1588 long imp = env->p->numa_group ? groupimp : taskimp;
1590 int dist = env->dist;
1592 if (READ_ONCE(dst_rq->numa_migrate_on))
1596 cur = task_rcu_dereference(&dst_rq->curr);
1597 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1601 * Because we have preemption enabled we can get migrated around and
1602 * end try selecting ourselves (current == env->p) as a swap candidate.
1608 if (maymove && moveimp >= env->best_imp)
1615 * "imp" is the fault differential for the source task between the
1616 * source and destination node. Calculate the total differential for
1617 * the source task and potential destination task. The more negative
1618 * the value is, the more remote accesses that would be expected to
1619 * be incurred if the tasks were swapped.
1621 /* Skip this swap candidate if cannot move to the source cpu */
1622 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
1626 * If dst and source tasks are in the same NUMA group, or not
1627 * in any group then look only at task weights.
1629 if (cur->numa_group == env->p->numa_group) {
1630 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1631 task_weight(cur, env->dst_nid, dist);
1633 * Add some hysteresis to prevent swapping the
1634 * tasks within a group over tiny differences.
1636 if (cur->numa_group)
1640 * Compare the group weights. If a task is all by itself
1641 * (not part of a group), use the task weight instead.
1643 if (cur->numa_group && env->p->numa_group)
1644 imp += group_weight(cur, env->src_nid, dist) -
1645 group_weight(cur, env->dst_nid, dist);
1647 imp += task_weight(cur, env->src_nid, dist) -
1648 task_weight(cur, env->dst_nid, dist);
1651 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1658 * If the NUMA importance is less than SMALLIMP,
1659 * task migration might only result in ping pong
1660 * of tasks and also hurt performance due to cache
1663 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1667 * In the overloaded case, try and keep the load balanced.
1669 load = task_h_load(env->p) - task_h_load(cur);
1673 dst_load = env->dst_stats.load + load;
1674 src_load = env->src_stats.load - load;
1676 if (load_too_imbalanced(src_load, dst_load, env))
1681 * One idle CPU per node is evaluated for a task numa move.
1682 * Call select_idle_sibling to maybe find a better one.
1686 * select_idle_siblings() uses an per-CPU cpumask that
1687 * can be used from IRQ context.
1689 local_irq_disable();
1690 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1695 task_numa_assign(env, cur, imp);
1700 static void task_numa_find_cpu(struct task_numa_env *env,
1701 long taskimp, long groupimp)
1703 long src_load, dst_load, load;
1704 bool maymove = false;
1707 load = task_h_load(env->p);
1708 dst_load = env->dst_stats.load + load;
1709 src_load = env->src_stats.load - load;
1712 * If the improvement from just moving env->p direction is better
1713 * than swapping tasks around, check if a move is possible.
1715 maymove = !load_too_imbalanced(src_load, dst_load, env);
1717 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1718 /* Skip this CPU if the source task cannot migrate */
1719 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
1723 task_numa_compare(env, taskimp, groupimp, maymove);
1727 static int task_numa_migrate(struct task_struct *p)
1729 struct task_numa_env env = {
1732 .src_cpu = task_cpu(p),
1733 .src_nid = task_node(p),
1735 .imbalance_pct = 112,
1741 struct sched_domain *sd;
1743 unsigned long taskweight, groupweight;
1745 long taskimp, groupimp;
1748 * Pick the lowest SD_NUMA domain, as that would have the smallest
1749 * imbalance and would be the first to start moving tasks about.
1751 * And we want to avoid any moving of tasks about, as that would create
1752 * random movement of tasks -- counter the numa conditions we're trying
1756 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1758 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1762 * Cpusets can break the scheduler domain tree into smaller
1763 * balance domains, some of which do not cross NUMA boundaries.
1764 * Tasks that are "trapped" in such domains cannot be migrated
1765 * elsewhere, so there is no point in (re)trying.
1767 if (unlikely(!sd)) {
1768 sched_setnuma(p, task_node(p));
1772 env.dst_nid = p->numa_preferred_nid;
1773 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1774 taskweight = task_weight(p, env.src_nid, dist);
1775 groupweight = group_weight(p, env.src_nid, dist);
1776 update_numa_stats(&env.src_stats, env.src_nid);
1777 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1778 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1779 update_numa_stats(&env.dst_stats, env.dst_nid);
1781 /* Try to find a spot on the preferred nid. */
1782 task_numa_find_cpu(&env, taskimp, groupimp);
1785 * Look at other nodes in these cases:
1786 * - there is no space available on the preferred_nid
1787 * - the task is part of a numa_group that is interleaved across
1788 * multiple NUMA nodes; in order to better consolidate the group,
1789 * we need to check other locations.
1791 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1792 for_each_online_node(nid) {
1793 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1796 dist = node_distance(env.src_nid, env.dst_nid);
1797 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1799 taskweight = task_weight(p, env.src_nid, dist);
1800 groupweight = group_weight(p, env.src_nid, dist);
1803 /* Only consider nodes where both task and groups benefit */
1804 taskimp = task_weight(p, nid, dist) - taskweight;
1805 groupimp = group_weight(p, nid, dist) - groupweight;
1806 if (taskimp < 0 && groupimp < 0)
1811 update_numa_stats(&env.dst_stats, env.dst_nid);
1812 task_numa_find_cpu(&env, taskimp, groupimp);
1817 * If the task is part of a workload that spans multiple NUMA nodes,
1818 * and is migrating into one of the workload's active nodes, remember
1819 * this node as the task's preferred numa node, so the workload can
1821 * A task that migrated to a second choice node will be better off
1822 * trying for a better one later. Do not set the preferred node here.
1824 if (p->numa_group) {
1825 if (env.best_cpu == -1)
1828 nid = cpu_to_node(env.best_cpu);
1830 if (nid != p->numa_preferred_nid)
1831 sched_setnuma(p, nid);
1834 /* No better CPU than the current one was found. */
1835 if (env.best_cpu == -1)
1838 best_rq = cpu_rq(env.best_cpu);
1839 if (env.best_task == NULL) {
1840 ret = migrate_task_to(p, env.best_cpu);
1841 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1843 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1847 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
1848 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1851 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1852 put_task_struct(env.best_task);
1856 /* Attempt to migrate a task to a CPU on the preferred node. */
1857 static void numa_migrate_preferred(struct task_struct *p)
1859 unsigned long interval = HZ;
1861 /* This task has no NUMA fault statistics yet */
1862 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
1865 /* Periodically retry migrating the task to the preferred node */
1866 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1867 p->numa_migrate_retry = jiffies + interval;
1869 /* Success if task is already running on preferred CPU */
1870 if (task_node(p) == p->numa_preferred_nid)
1873 /* Otherwise, try migrate to a CPU on the preferred node */
1874 task_numa_migrate(p);
1878 * Find out how many nodes on the workload is actively running on. Do this by
1879 * tracking the nodes from which NUMA hinting faults are triggered. This can
1880 * be different from the set of nodes where the workload's memory is currently
1883 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1885 unsigned long faults, max_faults = 0;
1886 int nid, active_nodes = 0;
1888 for_each_online_node(nid) {
1889 faults = group_faults_cpu(numa_group, nid);
1890 if (faults > max_faults)
1891 max_faults = faults;
1894 for_each_online_node(nid) {
1895 faults = group_faults_cpu(numa_group, nid);
1896 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1900 numa_group->max_faults_cpu = max_faults;
1901 numa_group->active_nodes = active_nodes;
1905 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1906 * increments. The more local the fault statistics are, the higher the scan
1907 * period will be for the next scan window. If local/(local+remote) ratio is
1908 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1909 * the scan period will decrease. Aim for 70% local accesses.
1911 #define NUMA_PERIOD_SLOTS 10
1912 #define NUMA_PERIOD_THRESHOLD 7
1915 * Increase the scan period (slow down scanning) if the majority of
1916 * our memory is already on our local node, or if the majority of
1917 * the page accesses are shared with other processes.
1918 * Otherwise, decrease the scan period.
1920 static void update_task_scan_period(struct task_struct *p,
1921 unsigned long shared, unsigned long private)
1923 unsigned int period_slot;
1924 int lr_ratio, ps_ratio;
1927 unsigned long remote = p->numa_faults_locality[0];
1928 unsigned long local = p->numa_faults_locality[1];
1931 * If there were no record hinting faults then either the task is
1932 * completely idle or all activity is areas that are not of interest
1933 * to automatic numa balancing. Related to that, if there were failed
1934 * migration then it implies we are migrating too quickly or the local
1935 * node is overloaded. In either case, scan slower
1937 if (local + shared == 0 || p->numa_faults_locality[2]) {
1938 p->numa_scan_period = min(p->numa_scan_period_max,
1939 p->numa_scan_period << 1);
1941 p->mm->numa_next_scan = jiffies +
1942 msecs_to_jiffies(p->numa_scan_period);
1948 * Prepare to scale scan period relative to the current period.
1949 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1950 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1951 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1953 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1954 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1955 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1957 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1959 * Most memory accesses are local. There is no need to
1960 * do fast NUMA scanning, since memory is already local.
1962 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
1965 diff = slot * period_slot;
1966 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
1968 * Most memory accesses are shared with other tasks.
1969 * There is no point in continuing fast NUMA scanning,
1970 * since other tasks may just move the memory elsewhere.
1972 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
1975 diff = slot * period_slot;
1978 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
1979 * yet they are not on the local NUMA node. Speed up
1980 * NUMA scanning to get the memory moved over.
1982 int ratio = max(lr_ratio, ps_ratio);
1983 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1986 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1987 task_scan_min(p), task_scan_max(p));
1988 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1992 * Get the fraction of time the task has been running since the last
1993 * NUMA placement cycle. The scheduler keeps similar statistics, but
1994 * decays those on a 32ms period, which is orders of magnitude off
1995 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1996 * stats only if the task is so new there are no NUMA statistics yet.
1998 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2000 u64 runtime, delta, now;
2001 /* Use the start of this time slice to avoid calculations. */
2002 now = p->se.exec_start;
2003 runtime = p->se.sum_exec_runtime;
2005 if (p->last_task_numa_placement) {
2006 delta = runtime - p->last_sum_exec_runtime;
2007 *period = now - p->last_task_numa_placement;
2009 /* Avoid time going backwards, prevent potential divide error: */
2010 if (unlikely((s64)*period < 0))
2013 delta = p->se.avg.load_sum;
2014 *period = LOAD_AVG_MAX;
2017 p->last_sum_exec_runtime = runtime;
2018 p->last_task_numa_placement = now;
2024 * Determine the preferred nid for a task in a numa_group. This needs to
2025 * be done in a way that produces consistent results with group_weight,
2026 * otherwise workloads might not converge.
2028 static int preferred_group_nid(struct task_struct *p, int nid)
2033 /* Direct connections between all NUMA nodes. */
2034 if (sched_numa_topology_type == NUMA_DIRECT)
2038 * On a system with glueless mesh NUMA topology, group_weight
2039 * scores nodes according to the number of NUMA hinting faults on
2040 * both the node itself, and on nearby nodes.
2042 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2043 unsigned long score, max_score = 0;
2044 int node, max_node = nid;
2046 dist = sched_max_numa_distance;
2048 for_each_online_node(node) {
2049 score = group_weight(p, node, dist);
2050 if (score > max_score) {
2059 * Finding the preferred nid in a system with NUMA backplane
2060 * interconnect topology is more involved. The goal is to locate
2061 * tasks from numa_groups near each other in the system, and
2062 * untangle workloads from different sides of the system. This requires
2063 * searching down the hierarchy of node groups, recursively searching
2064 * inside the highest scoring group of nodes. The nodemask tricks
2065 * keep the complexity of the search down.
2067 nodes = node_online_map;
2068 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2069 unsigned long max_faults = 0;
2070 nodemask_t max_group = NODE_MASK_NONE;
2073 /* Are there nodes at this distance from each other? */
2074 if (!find_numa_distance(dist))
2077 for_each_node_mask(a, nodes) {
2078 unsigned long faults = 0;
2079 nodemask_t this_group;
2080 nodes_clear(this_group);
2082 /* Sum group's NUMA faults; includes a==b case. */
2083 for_each_node_mask(b, nodes) {
2084 if (node_distance(a, b) < dist) {
2085 faults += group_faults(p, b);
2086 node_set(b, this_group);
2087 node_clear(b, nodes);
2091 /* Remember the top group. */
2092 if (faults > max_faults) {
2093 max_faults = faults;
2094 max_group = this_group;
2096 * subtle: at the smallest distance there is
2097 * just one node left in each "group", the
2098 * winner is the preferred nid.
2103 /* Next round, evaluate the nodes within max_group. */
2111 static void task_numa_placement(struct task_struct *p)
2113 int seq, nid, max_nid = NUMA_NO_NODE;
2114 unsigned long max_faults = 0;
2115 unsigned long fault_types[2] = { 0, 0 };
2116 unsigned long total_faults;
2117 u64 runtime, period;
2118 spinlock_t *group_lock = NULL;
2121 * The p->mm->numa_scan_seq field gets updated without
2122 * exclusive access. Use READ_ONCE() here to ensure
2123 * that the field is read in a single access:
2125 seq = READ_ONCE(p->mm->numa_scan_seq);
2126 if (p->numa_scan_seq == seq)
2128 p->numa_scan_seq = seq;
2129 p->numa_scan_period_max = task_scan_max(p);
2131 total_faults = p->numa_faults_locality[0] +
2132 p->numa_faults_locality[1];
2133 runtime = numa_get_avg_runtime(p, &period);
2135 /* If the task is part of a group prevent parallel updates to group stats */
2136 if (p->numa_group) {
2137 group_lock = &p->numa_group->lock;
2138 spin_lock_irq(group_lock);
2141 /* Find the node with the highest number of faults */
2142 for_each_online_node(nid) {
2143 /* Keep track of the offsets in numa_faults array */
2144 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2145 unsigned long faults = 0, group_faults = 0;
2148 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2149 long diff, f_diff, f_weight;
2151 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2152 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2153 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2154 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2156 /* Decay existing window, copy faults since last scan */
2157 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2158 fault_types[priv] += p->numa_faults[membuf_idx];
2159 p->numa_faults[membuf_idx] = 0;
2162 * Normalize the faults_from, so all tasks in a group
2163 * count according to CPU use, instead of by the raw
2164 * number of faults. Tasks with little runtime have
2165 * little over-all impact on throughput, and thus their
2166 * faults are less important.
2168 f_weight = div64_u64(runtime << 16, period + 1);
2169 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2171 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2172 p->numa_faults[cpubuf_idx] = 0;
2174 p->numa_faults[mem_idx] += diff;
2175 p->numa_faults[cpu_idx] += f_diff;
2176 faults += p->numa_faults[mem_idx];
2177 p->total_numa_faults += diff;
2178 if (p->numa_group) {
2180 * safe because we can only change our own group
2182 * mem_idx represents the offset for a given
2183 * nid and priv in a specific region because it
2184 * is at the beginning of the numa_faults array.
2186 p->numa_group->faults[mem_idx] += diff;
2187 p->numa_group->faults_cpu[mem_idx] += f_diff;
2188 p->numa_group->total_faults += diff;
2189 group_faults += p->numa_group->faults[mem_idx];
2193 if (!p->numa_group) {
2194 if (faults > max_faults) {
2195 max_faults = faults;
2198 } else if (group_faults > max_faults) {
2199 max_faults = group_faults;
2204 if (p->numa_group) {
2205 numa_group_count_active_nodes(p->numa_group);
2206 spin_unlock_irq(group_lock);
2207 max_nid = preferred_group_nid(p, max_nid);
2211 /* Set the new preferred node */
2212 if (max_nid != p->numa_preferred_nid)
2213 sched_setnuma(p, max_nid);
2216 update_task_scan_period(p, fault_types[0], fault_types[1]);
2219 static inline int get_numa_group(struct numa_group *grp)
2221 return refcount_inc_not_zero(&grp->refcount);
2224 static inline void put_numa_group(struct numa_group *grp)
2226 if (refcount_dec_and_test(&grp->refcount))
2227 kfree_rcu(grp, rcu);
2230 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2233 struct numa_group *grp, *my_grp;
2234 struct task_struct *tsk;
2236 int cpu = cpupid_to_cpu(cpupid);
2239 if (unlikely(!p->numa_group)) {
2240 unsigned int size = sizeof(struct numa_group) +
2241 4*nr_node_ids*sizeof(unsigned long);
2243 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2247 refcount_set(&grp->refcount, 1);
2248 grp->active_nodes = 1;
2249 grp->max_faults_cpu = 0;
2250 spin_lock_init(&grp->lock);
2252 /* Second half of the array tracks nids where faults happen */
2253 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2256 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2257 grp->faults[i] = p->numa_faults[i];
2259 grp->total_faults = p->total_numa_faults;
2262 rcu_assign_pointer(p->numa_group, grp);
2266 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2268 if (!cpupid_match_pid(tsk, cpupid))
2271 grp = rcu_dereference(tsk->numa_group);
2275 my_grp = p->numa_group;
2280 * Only join the other group if its bigger; if we're the bigger group,
2281 * the other task will join us.
2283 if (my_grp->nr_tasks > grp->nr_tasks)
2287 * Tie-break on the grp address.
2289 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2292 /* Always join threads in the same process. */
2293 if (tsk->mm == current->mm)
2296 /* Simple filter to avoid false positives due to PID collisions */
2297 if (flags & TNF_SHARED)
2300 /* Update priv based on whether false sharing was detected */
2303 if (join && !get_numa_group(grp))
2311 BUG_ON(irqs_disabled());
2312 double_lock_irq(&my_grp->lock, &grp->lock);
2314 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2315 my_grp->faults[i] -= p->numa_faults[i];
2316 grp->faults[i] += p->numa_faults[i];
2318 my_grp->total_faults -= p->total_numa_faults;
2319 grp->total_faults += p->total_numa_faults;
2324 spin_unlock(&my_grp->lock);
2325 spin_unlock_irq(&grp->lock);
2327 rcu_assign_pointer(p->numa_group, grp);
2329 put_numa_group(my_grp);
2337 void task_numa_free(struct task_struct *p)
2339 struct numa_group *grp = p->numa_group;
2340 void *numa_faults = p->numa_faults;
2341 unsigned long flags;
2345 spin_lock_irqsave(&grp->lock, flags);
2346 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2347 grp->faults[i] -= p->numa_faults[i];
2348 grp->total_faults -= p->total_numa_faults;
2351 spin_unlock_irqrestore(&grp->lock, flags);
2352 RCU_INIT_POINTER(p->numa_group, NULL);
2353 put_numa_group(grp);
2356 p->numa_faults = NULL;
2361 * Got a PROT_NONE fault for a page on @node.
2363 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2365 struct task_struct *p = current;
2366 bool migrated = flags & TNF_MIGRATED;
2367 int cpu_node = task_node(current);
2368 int local = !!(flags & TNF_FAULT_LOCAL);
2369 struct numa_group *ng;
2372 if (!static_branch_likely(&sched_numa_balancing))
2375 /* for example, ksmd faulting in a user's mm */
2379 /* Allocate buffer to track faults on a per-node basis */
2380 if (unlikely(!p->numa_faults)) {
2381 int size = sizeof(*p->numa_faults) *
2382 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2384 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2385 if (!p->numa_faults)
2388 p->total_numa_faults = 0;
2389 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2393 * First accesses are treated as private, otherwise consider accesses
2394 * to be private if the accessing pid has not changed
2396 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2399 priv = cpupid_match_pid(p, last_cpupid);
2400 if (!priv && !(flags & TNF_NO_GROUP))
2401 task_numa_group(p, last_cpupid, flags, &priv);
2405 * If a workload spans multiple NUMA nodes, a shared fault that
2406 * occurs wholly within the set of nodes that the workload is
2407 * actively using should be counted as local. This allows the
2408 * scan rate to slow down when a workload has settled down.
2411 if (!priv && !local && ng && ng->active_nodes > 1 &&
2412 numa_is_active_node(cpu_node, ng) &&
2413 numa_is_active_node(mem_node, ng))
2417 * Retry to migrate task to preferred node periodically, in case it
2418 * previously failed, or the scheduler moved us.
2420 if (time_after(jiffies, p->numa_migrate_retry)) {
2421 task_numa_placement(p);
2422 numa_migrate_preferred(p);
2426 p->numa_pages_migrated += pages;
2427 if (flags & TNF_MIGRATE_FAIL)
2428 p->numa_faults_locality[2] += pages;
2430 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2431 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2432 p->numa_faults_locality[local] += pages;
2435 static void reset_ptenuma_scan(struct task_struct *p)
2438 * We only did a read acquisition of the mmap sem, so
2439 * p->mm->numa_scan_seq is written to without exclusive access
2440 * and the update is not guaranteed to be atomic. That's not
2441 * much of an issue though, since this is just used for
2442 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2443 * expensive, to avoid any form of compiler optimizations:
2445 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2446 p->mm->numa_scan_offset = 0;
2450 * The expensive part of numa migration is done from task_work context.
2451 * Triggered from task_tick_numa().
2453 void task_numa_work(struct callback_head *work)
2455 unsigned long migrate, next_scan, now = jiffies;
2456 struct task_struct *p = current;
2457 struct mm_struct *mm = p->mm;
2458 u64 runtime = p->se.sum_exec_runtime;
2459 struct vm_area_struct *vma;
2460 unsigned long start, end;
2461 unsigned long nr_pte_updates = 0;
2462 long pages, virtpages;
2464 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2466 work->next = work; /* protect against double add */
2468 * Who cares about NUMA placement when they're dying.
2470 * NOTE: make sure not to dereference p->mm before this check,
2471 * exit_task_work() happens _after_ exit_mm() so we could be called
2472 * without p->mm even though we still had it when we enqueued this
2475 if (p->flags & PF_EXITING)
2478 if (!mm->numa_next_scan) {
2479 mm->numa_next_scan = now +
2480 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2484 * Enforce maximal scan/migration frequency..
2486 migrate = mm->numa_next_scan;
2487 if (time_before(now, migrate))
2490 if (p->numa_scan_period == 0) {
2491 p->numa_scan_period_max = task_scan_max(p);
2492 p->numa_scan_period = task_scan_start(p);
2495 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2496 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2500 * Delay this task enough that another task of this mm will likely win
2501 * the next time around.
2503 p->node_stamp += 2 * TICK_NSEC;
2505 start = mm->numa_scan_offset;
2506 pages = sysctl_numa_balancing_scan_size;
2507 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2508 virtpages = pages * 8; /* Scan up to this much virtual space */
2513 if (!down_read_trylock(&mm->mmap_sem))
2515 vma = find_vma(mm, start);
2517 reset_ptenuma_scan(p);
2521 for (; vma; vma = vma->vm_next) {
2522 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2523 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2528 * Shared library pages mapped by multiple processes are not
2529 * migrated as it is expected they are cache replicated. Avoid
2530 * hinting faults in read-only file-backed mappings or the vdso
2531 * as migrating the pages will be of marginal benefit.
2534 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2538 * Skip inaccessible VMAs to avoid any confusion between
2539 * PROT_NONE and NUMA hinting ptes
2541 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2545 start = max(start, vma->vm_start);
2546 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2547 end = min(end, vma->vm_end);
2548 nr_pte_updates = change_prot_numa(vma, start, end);
2551 * Try to scan sysctl_numa_balancing_size worth of
2552 * hpages that have at least one present PTE that
2553 * is not already pte-numa. If the VMA contains
2554 * areas that are unused or already full of prot_numa
2555 * PTEs, scan up to virtpages, to skip through those
2559 pages -= (end - start) >> PAGE_SHIFT;
2560 virtpages -= (end - start) >> PAGE_SHIFT;
2563 if (pages <= 0 || virtpages <= 0)
2567 } while (end != vma->vm_end);
2572 * It is possible to reach the end of the VMA list but the last few
2573 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2574 * would find the !migratable VMA on the next scan but not reset the
2575 * scanner to the start so check it now.
2578 mm->numa_scan_offset = start;
2580 reset_ptenuma_scan(p);
2581 up_read(&mm->mmap_sem);
2584 * Make sure tasks use at least 32x as much time to run other code
2585 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2586 * Usually update_task_scan_period slows down scanning enough; on an
2587 * overloaded system we need to limit overhead on a per task basis.
2589 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2590 u64 diff = p->se.sum_exec_runtime - runtime;
2591 p->node_stamp += 32 * diff;
2596 * Drive the periodic memory faults..
2598 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2600 struct callback_head *work = &curr->numa_work;
2604 * We don't care about NUMA placement if we don't have memory.
2606 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2610 * Using runtime rather than walltime has the dual advantage that
2611 * we (mostly) drive the selection from busy threads and that the
2612 * task needs to have done some actual work before we bother with
2615 now = curr->se.sum_exec_runtime;
2616 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2618 if (now > curr->node_stamp + period) {
2619 if (!curr->node_stamp)
2620 curr->numa_scan_period = task_scan_start(curr);
2621 curr->node_stamp += period;
2623 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2624 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2625 task_work_add(curr, work, true);
2630 static void update_scan_period(struct task_struct *p, int new_cpu)
2632 int src_nid = cpu_to_node(task_cpu(p));
2633 int dst_nid = cpu_to_node(new_cpu);
2635 if (!static_branch_likely(&sched_numa_balancing))
2638 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2641 if (src_nid == dst_nid)
2645 * Allow resets if faults have been trapped before one scan
2646 * has completed. This is most likely due to a new task that
2647 * is pulled cross-node due to wakeups or load balancing.
2649 if (p->numa_scan_seq) {
2651 * Avoid scan adjustments if moving to the preferred
2652 * node or if the task was not previously running on
2653 * the preferred node.
2655 if (dst_nid == p->numa_preferred_nid ||
2656 (p->numa_preferred_nid != NUMA_NO_NODE &&
2657 src_nid != p->numa_preferred_nid))
2661 p->numa_scan_period = task_scan_start(p);
2665 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2669 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2673 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2677 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2681 #endif /* CONFIG_NUMA_BALANCING */
2684 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2686 update_load_add(&cfs_rq->load, se->load.weight);
2688 if (entity_is_task(se)) {
2689 struct rq *rq = rq_of(cfs_rq);
2691 account_numa_enqueue(rq, task_of(se));
2692 list_add(&se->group_node, &rq->cfs_tasks);
2695 cfs_rq->nr_running++;
2699 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2701 update_load_sub(&cfs_rq->load, se->load.weight);
2703 if (entity_is_task(se)) {
2704 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2705 list_del_init(&se->group_node);
2708 cfs_rq->nr_running--;
2712 * Signed add and clamp on underflow.
2714 * Explicitly do a load-store to ensure the intermediate value never hits
2715 * memory. This allows lockless observations without ever seeing the negative
2718 #define add_positive(_ptr, _val) do { \
2719 typeof(_ptr) ptr = (_ptr); \
2720 typeof(_val) val = (_val); \
2721 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2725 if (val < 0 && res > var) \
2728 WRITE_ONCE(*ptr, res); \
2732 * Unsigned subtract and clamp on underflow.
2734 * Explicitly do a load-store to ensure the intermediate value never hits
2735 * memory. This allows lockless observations without ever seeing the negative
2738 #define sub_positive(_ptr, _val) do { \
2739 typeof(_ptr) ptr = (_ptr); \
2740 typeof(*ptr) val = (_val); \
2741 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2745 WRITE_ONCE(*ptr, res); \
2749 * Remove and clamp on negative, from a local variable.
2751 * A variant of sub_positive(), which does not use explicit load-store
2752 * and is thus optimized for local variable updates.
2754 #define lsub_positive(_ptr, _val) do { \
2755 typeof(_ptr) ptr = (_ptr); \
2756 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
2761 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2763 cfs_rq->runnable_weight += se->runnable_weight;
2765 cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2766 cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2770 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2772 cfs_rq->runnable_weight -= se->runnable_weight;
2774 sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2775 sub_positive(&cfs_rq->avg.runnable_load_sum,
2776 se_runnable(se) * se->avg.runnable_load_sum);
2780 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2782 cfs_rq->avg.load_avg += se->avg.load_avg;
2783 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2787 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2789 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2790 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2794 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2796 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2798 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2800 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2803 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2804 unsigned long weight, unsigned long runnable)
2807 /* commit outstanding execution time */
2808 if (cfs_rq->curr == se)
2809 update_curr(cfs_rq);
2810 account_entity_dequeue(cfs_rq, se);
2811 dequeue_runnable_load_avg(cfs_rq, se);
2813 dequeue_load_avg(cfs_rq, se);
2815 se->runnable_weight = runnable;
2816 update_load_set(&se->load, weight);
2820 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2822 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2823 se->avg.runnable_load_avg =
2824 div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2828 enqueue_load_avg(cfs_rq, se);
2830 account_entity_enqueue(cfs_rq, se);
2831 enqueue_runnable_load_avg(cfs_rq, se);
2835 void reweight_task(struct task_struct *p, int prio)
2837 struct sched_entity *se = &p->se;
2838 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2839 struct load_weight *load = &se->load;
2840 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2842 reweight_entity(cfs_rq, se, weight, weight);
2843 load->inv_weight = sched_prio_to_wmult[prio];
2846 #ifdef CONFIG_FAIR_GROUP_SCHED
2849 * All this does is approximate the hierarchical proportion which includes that
2850 * global sum we all love to hate.
2852 * That is, the weight of a group entity, is the proportional share of the
2853 * group weight based on the group runqueue weights. That is:
2855 * tg->weight * grq->load.weight
2856 * ge->load.weight = ----------------------------- (1)
2857 * \Sum grq->load.weight
2859 * Now, because computing that sum is prohibitively expensive to compute (been
2860 * there, done that) we approximate it with this average stuff. The average
2861 * moves slower and therefore the approximation is cheaper and more stable.
2863 * So instead of the above, we substitute:
2865 * grq->load.weight -> grq->avg.load_avg (2)
2867 * which yields the following:
2869 * tg->weight * grq->avg.load_avg
2870 * ge->load.weight = ------------------------------ (3)
2873 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2875 * That is shares_avg, and it is right (given the approximation (2)).
2877 * The problem with it is that because the average is slow -- it was designed
2878 * to be exactly that of course -- this leads to transients in boundary
2879 * conditions. In specific, the case where the group was idle and we start the
2880 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2881 * yielding bad latency etc..
2883 * Now, in that special case (1) reduces to:
2885 * tg->weight * grq->load.weight
2886 * ge->load.weight = ----------------------------- = tg->weight (4)
2889 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2891 * So what we do is modify our approximation (3) to approach (4) in the (near)
2896 * tg->weight * grq->load.weight
2897 * --------------------------------------------------- (5)
2898 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2900 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2901 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2904 * tg->weight * grq->load.weight
2905 * ge->load.weight = ----------------------------- (6)
2910 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2911 * max(grq->load.weight, grq->avg.load_avg)
2913 * And that is shares_weight and is icky. In the (near) UP case it approaches
2914 * (4) while in the normal case it approaches (3). It consistently
2915 * overestimates the ge->load.weight and therefore:
2917 * \Sum ge->load.weight >= tg->weight
2921 static long calc_group_shares(struct cfs_rq *cfs_rq)
2923 long tg_weight, tg_shares, load, shares;
2924 struct task_group *tg = cfs_rq->tg;
2926 tg_shares = READ_ONCE(tg->shares);
2928 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
2930 tg_weight = atomic_long_read(&tg->load_avg);
2932 /* Ensure tg_weight >= load */
2933 tg_weight -= cfs_rq->tg_load_avg_contrib;
2936 shares = (tg_shares * load);
2938 shares /= tg_weight;
2941 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2942 * of a group with small tg->shares value. It is a floor value which is
2943 * assigned as a minimum load.weight to the sched_entity representing
2944 * the group on a CPU.
2946 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2947 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2948 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2949 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2952 return clamp_t(long, shares, MIN_SHARES, tg_shares);
2956 * This calculates the effective runnable weight for a group entity based on
2957 * the group entity weight calculated above.
2959 * Because of the above approximation (2), our group entity weight is
2960 * an load_avg based ratio (3). This means that it includes blocked load and
2961 * does not represent the runnable weight.
2963 * Approximate the group entity's runnable weight per ratio from the group
2966 * grq->avg.runnable_load_avg
2967 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
2970 * However, analogous to above, since the avg numbers are slow, this leads to
2971 * transients in the from-idle case. Instead we use:
2973 * ge->runnable_weight = ge->load.weight *
2975 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
2976 * ----------------------------------------------------- (8)
2977 * max(grq->avg.load_avg, grq->load.weight)
2979 * Where these max() serve both to use the 'instant' values to fix the slow
2980 * from-idle and avoid the /0 on to-idle, similar to (6).
2982 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
2984 long runnable, load_avg;
2986 load_avg = max(cfs_rq->avg.load_avg,
2987 scale_load_down(cfs_rq->load.weight));
2989 runnable = max(cfs_rq->avg.runnable_load_avg,
2990 scale_load_down(cfs_rq->runnable_weight));
2994 runnable /= load_avg;
2996 return clamp_t(long, runnable, MIN_SHARES, shares);
2998 #endif /* CONFIG_SMP */
3000 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3003 * Recomputes the group entity based on the current state of its group
3006 static void update_cfs_group(struct sched_entity *se)
3008 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3009 long shares, runnable;
3014 if (throttled_hierarchy(gcfs_rq))
3018 runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
3020 if (likely(se->load.weight == shares))
3023 shares = calc_group_shares(gcfs_rq);
3024 runnable = calc_group_runnable(gcfs_rq, shares);
3027 reweight_entity(cfs_rq_of(se), se, shares, runnable);
3030 #else /* CONFIG_FAIR_GROUP_SCHED */
3031 static inline void update_cfs_group(struct sched_entity *se)
3034 #endif /* CONFIG_FAIR_GROUP_SCHED */
3036 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3038 struct rq *rq = rq_of(cfs_rq);
3040 if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) {
3042 * There are a few boundary cases this might miss but it should
3043 * get called often enough that that should (hopefully) not be
3046 * It will not get called when we go idle, because the idle
3047 * thread is a different class (!fair), nor will the utilization
3048 * number include things like RT tasks.
3050 * As is, the util number is not freq-invariant (we'd have to
3051 * implement arch_scale_freq_capacity() for that).
3055 cpufreq_update_util(rq, flags);
3060 #ifdef CONFIG_FAIR_GROUP_SCHED
3062 * update_tg_load_avg - update the tg's load avg
3063 * @cfs_rq: the cfs_rq whose avg changed
3064 * @force: update regardless of how small the difference
3066 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3067 * However, because tg->load_avg is a global value there are performance
3070 * In order to avoid having to look at the other cfs_rq's, we use a
3071 * differential update where we store the last value we propagated. This in
3072 * turn allows skipping updates if the differential is 'small'.
3074 * Updating tg's load_avg is necessary before update_cfs_share().
3076 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3078 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3081 * No need to update load_avg for root_task_group as it is not used.
3083 if (cfs_rq->tg == &root_task_group)
3086 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3087 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3088 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3093 * Called within set_task_rq() right before setting a task's CPU. The
3094 * caller only guarantees p->pi_lock is held; no other assumptions,
3095 * including the state of rq->lock, should be made.
3097 void set_task_rq_fair(struct sched_entity *se,
3098 struct cfs_rq *prev, struct cfs_rq *next)
3100 u64 p_last_update_time;
3101 u64 n_last_update_time;
3103 if (!sched_feat(ATTACH_AGE_LOAD))
3107 * We are supposed to update the task to "current" time, then its up to
3108 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3109 * getting what current time is, so simply throw away the out-of-date
3110 * time. This will result in the wakee task is less decayed, but giving
3111 * the wakee more load sounds not bad.
3113 if (!(se->avg.last_update_time && prev))
3116 #ifndef CONFIG_64BIT
3118 u64 p_last_update_time_copy;
3119 u64 n_last_update_time_copy;
3122 p_last_update_time_copy = prev->load_last_update_time_copy;
3123 n_last_update_time_copy = next->load_last_update_time_copy;
3127 p_last_update_time = prev->avg.last_update_time;
3128 n_last_update_time = next->avg.last_update_time;
3130 } while (p_last_update_time != p_last_update_time_copy ||
3131 n_last_update_time != n_last_update_time_copy);
3134 p_last_update_time = prev->avg.last_update_time;
3135 n_last_update_time = next->avg.last_update_time;
3137 __update_load_avg_blocked_se(p_last_update_time, se);
3138 se->avg.last_update_time = n_last_update_time;
3143 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3144 * propagate its contribution. The key to this propagation is the invariant
3145 * that for each group:
3147 * ge->avg == grq->avg (1)
3149 * _IFF_ we look at the pure running and runnable sums. Because they
3150 * represent the very same entity, just at different points in the hierarchy.
3152 * Per the above update_tg_cfs_util() is trivial and simply copies the running
3153 * sum over (but still wrong, because the group entity and group rq do not have
3154 * their PELT windows aligned).
3156 * However, update_tg_cfs_runnable() is more complex. So we have:
3158 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3160 * And since, like util, the runnable part should be directly transferable,
3161 * the following would _appear_ to be the straight forward approach:
3163 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3165 * And per (1) we have:
3167 * ge->avg.runnable_avg == grq->avg.runnable_avg
3171 * ge->load.weight * grq->avg.load_avg
3172 * ge->avg.load_avg = ----------------------------------- (4)
3175 * Except that is wrong!
3177 * Because while for entities historical weight is not important and we
3178 * really only care about our future and therefore can consider a pure
3179 * runnable sum, runqueues can NOT do this.
3181 * We specifically want runqueues to have a load_avg that includes
3182 * historical weights. Those represent the blocked load, the load we expect
3183 * to (shortly) return to us. This only works by keeping the weights as
3184 * integral part of the sum. We therefore cannot decompose as per (3).
3186 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3187 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3188 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3189 * runnable section of these tasks overlap (or not). If they were to perfectly
3190 * align the rq as a whole would be runnable 2/3 of the time. If however we
3191 * always have at least 1 runnable task, the rq as a whole is always runnable.
3193 * So we'll have to approximate.. :/
3195 * Given the constraint:
3197 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3199 * We can construct a rule that adds runnable to a rq by assuming minimal
3202 * On removal, we'll assume each task is equally runnable; which yields:
3204 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3206 * XXX: only do this for the part of runnable > running ?
3211 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3213 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3215 /* Nothing to update */
3220 * The relation between sum and avg is:
3222 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3224 * however, the PELT windows are not aligned between grq and gse.
3227 /* Set new sched_entity's utilization */
3228 se->avg.util_avg = gcfs_rq->avg.util_avg;
3229 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3231 /* Update parent cfs_rq utilization */
3232 add_positive(&cfs_rq->avg.util_avg, delta);
3233 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3237 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3239 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3240 unsigned long runnable_load_avg, load_avg;
3241 u64 runnable_load_sum, load_sum = 0;
3247 gcfs_rq->prop_runnable_sum = 0;
3249 if (runnable_sum >= 0) {
3251 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3252 * the CPU is saturated running == runnable.
3254 runnable_sum += se->avg.load_sum;
3255 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3258 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3259 * assuming all tasks are equally runnable.
3261 if (scale_load_down(gcfs_rq->load.weight)) {
3262 load_sum = div_s64(gcfs_rq->avg.load_sum,
3263 scale_load_down(gcfs_rq->load.weight));
3266 /* But make sure to not inflate se's runnable */
3267 runnable_sum = min(se->avg.load_sum, load_sum);
3271 * runnable_sum can't be lower than running_sum
3272 * Rescale running sum to be in the same range as runnable sum
3273 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3274 * runnable_sum is in [0 : LOAD_AVG_MAX]
3276 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3277 runnable_sum = max(runnable_sum, running_sum);
3279 load_sum = (s64)se_weight(se) * runnable_sum;
3280 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3282 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3283 delta_avg = load_avg - se->avg.load_avg;
3285 se->avg.load_sum = runnable_sum;
3286 se->avg.load_avg = load_avg;
3287 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3288 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3290 runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3291 runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3292 delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3293 delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3295 se->avg.runnable_load_sum = runnable_sum;
3296 se->avg.runnable_load_avg = runnable_load_avg;
3299 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3300 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3304 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3306 cfs_rq->propagate = 1;
3307 cfs_rq->prop_runnable_sum += runnable_sum;
3310 /* Update task and its cfs_rq load average */
3311 static inline int propagate_entity_load_avg(struct sched_entity *se)
3313 struct cfs_rq *cfs_rq, *gcfs_rq;
3315 if (entity_is_task(se))
3318 gcfs_rq = group_cfs_rq(se);
3319 if (!gcfs_rq->propagate)
3322 gcfs_rq->propagate = 0;
3324 cfs_rq = cfs_rq_of(se);
3326 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3328 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3329 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3335 * Check if we need to update the load and the utilization of a blocked
3338 static inline bool skip_blocked_update(struct sched_entity *se)
3340 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3343 * If sched_entity still have not zero load or utilization, we have to
3346 if (se->avg.load_avg || se->avg.util_avg)
3350 * If there is a pending propagation, we have to update the load and
3351 * the utilization of the sched_entity:
3353 if (gcfs_rq->propagate)
3357 * Otherwise, the load and the utilization of the sched_entity is
3358 * already zero and there is no pending propagation, so it will be a
3359 * waste of time to try to decay it:
3364 #else /* CONFIG_FAIR_GROUP_SCHED */
3366 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3368 static inline int propagate_entity_load_avg(struct sched_entity *se)
3373 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3375 #endif /* CONFIG_FAIR_GROUP_SCHED */
3378 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3379 * @now: current time, as per cfs_rq_clock_pelt()
3380 * @cfs_rq: cfs_rq to update
3382 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3383 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3384 * post_init_entity_util_avg().
3386 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3388 * Returns true if the load decayed or we removed load.
3390 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3391 * call update_tg_load_avg() when this function returns true.
3394 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3396 unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3397 struct sched_avg *sa = &cfs_rq->avg;
3400 if (cfs_rq->removed.nr) {
3402 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3404 raw_spin_lock(&cfs_rq->removed.lock);
3405 swap(cfs_rq->removed.util_avg, removed_util);
3406 swap(cfs_rq->removed.load_avg, removed_load);
3407 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3408 cfs_rq->removed.nr = 0;
3409 raw_spin_unlock(&cfs_rq->removed.lock);
3412 sub_positive(&sa->load_avg, r);
3413 sub_positive(&sa->load_sum, r * divider);
3416 sub_positive(&sa->util_avg, r);
3417 sub_positive(&sa->util_sum, r * divider);
3419 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3424 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
3426 #ifndef CONFIG_64BIT
3428 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3432 cfs_rq_util_change(cfs_rq, 0);
3438 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3439 * @cfs_rq: cfs_rq to attach to
3440 * @se: sched_entity to attach
3441 * @flags: migration hints
3443 * Must call update_cfs_rq_load_avg() before this, since we rely on
3444 * cfs_rq->avg.last_update_time being current.
3446 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3448 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3451 * When we attach the @se to the @cfs_rq, we must align the decay
3452 * window because without that, really weird and wonderful things can
3457 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3458 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3461 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3462 * period_contrib. This isn't strictly correct, but since we're
3463 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3466 se->avg.util_sum = se->avg.util_avg * divider;
3468 se->avg.load_sum = divider;
3469 if (se_weight(se)) {
3471 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3474 se->avg.runnable_load_sum = se->avg.load_sum;
3476 enqueue_load_avg(cfs_rq, se);
3477 cfs_rq->avg.util_avg += se->avg.util_avg;
3478 cfs_rq->avg.util_sum += se->avg.util_sum;
3480 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3482 cfs_rq_util_change(cfs_rq, flags);
3486 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3487 * @cfs_rq: cfs_rq to detach from
3488 * @se: sched_entity to detach
3490 * Must call update_cfs_rq_load_avg() before this, since we rely on
3491 * cfs_rq->avg.last_update_time being current.
3493 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3495 dequeue_load_avg(cfs_rq, se);
3496 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3497 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3499 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3501 cfs_rq_util_change(cfs_rq, 0);
3505 * Optional action to be done while updating the load average
3507 #define UPDATE_TG 0x1
3508 #define SKIP_AGE_LOAD 0x2
3509 #define DO_ATTACH 0x4
3511 /* Update task and its cfs_rq load average */
3512 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3514 u64 now = cfs_rq_clock_pelt(cfs_rq);
3518 * Track task load average for carrying it to new CPU after migrated, and
3519 * track group sched_entity load average for task_h_load calc in migration
3521 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3522 __update_load_avg_se(now, cfs_rq, se);
3524 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3525 decayed |= propagate_entity_load_avg(se);
3527 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3530 * DO_ATTACH means we're here from enqueue_entity().
3531 * !last_update_time means we've passed through
3532 * migrate_task_rq_fair() indicating we migrated.
3534 * IOW we're enqueueing a task on a new CPU.
3536 attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
3537 update_tg_load_avg(cfs_rq, 0);
3539 } else if (decayed && (flags & UPDATE_TG))
3540 update_tg_load_avg(cfs_rq, 0);
3543 #ifndef CONFIG_64BIT
3544 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3546 u64 last_update_time_copy;
3547 u64 last_update_time;
3550 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3552 last_update_time = cfs_rq->avg.last_update_time;
3553 } while (last_update_time != last_update_time_copy);
3555 return last_update_time;
3558 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3560 return cfs_rq->avg.last_update_time;
3565 * Synchronize entity load avg of dequeued entity without locking
3568 static void sync_entity_load_avg(struct sched_entity *se)
3570 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3571 u64 last_update_time;
3573 last_update_time = cfs_rq_last_update_time(cfs_rq);
3574 __update_load_avg_blocked_se(last_update_time, se);
3578 * Task first catches up with cfs_rq, and then subtract
3579 * itself from the cfs_rq (task must be off the queue now).
3581 static void remove_entity_load_avg(struct sched_entity *se)
3583 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3584 unsigned long flags;
3587 * tasks cannot exit without having gone through wake_up_new_task() ->
3588 * post_init_entity_util_avg() which will have added things to the
3589 * cfs_rq, so we can remove unconditionally.
3592 sync_entity_load_avg(se);
3594 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3595 ++cfs_rq->removed.nr;
3596 cfs_rq->removed.util_avg += se->avg.util_avg;
3597 cfs_rq->removed.load_avg += se->avg.load_avg;
3598 cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */
3599 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3602 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3604 return cfs_rq->avg.runnable_load_avg;
3607 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3609 return cfs_rq->avg.load_avg;
3612 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3614 static inline unsigned long task_util(struct task_struct *p)
3616 return READ_ONCE(p->se.avg.util_avg);
3619 static inline unsigned long _task_util_est(struct task_struct *p)
3621 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3623 return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED);
3626 static inline unsigned long task_util_est(struct task_struct *p)
3628 return max(task_util(p), _task_util_est(p));
3631 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3632 struct task_struct *p)
3634 unsigned int enqueued;
3636 if (!sched_feat(UTIL_EST))
3639 /* Update root cfs_rq's estimated utilization */
3640 enqueued = cfs_rq->avg.util_est.enqueued;
3641 enqueued += _task_util_est(p);
3642 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3646 * Check if a (signed) value is within a specified (unsigned) margin,
3647 * based on the observation that:
3649 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3651 * NOTE: this only works when value + maring < INT_MAX.
3653 static inline bool within_margin(int value, int margin)
3655 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3659 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3661 long last_ewma_diff;
3665 if (!sched_feat(UTIL_EST))
3668 /* Update root cfs_rq's estimated utilization */
3669 ue.enqueued = cfs_rq->avg.util_est.enqueued;
3670 ue.enqueued -= min_t(unsigned int, ue.enqueued, _task_util_est(p));
3671 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3674 * Skip update of task's estimated utilization when the task has not
3675 * yet completed an activation, e.g. being migrated.
3681 * If the PELT values haven't changed since enqueue time,
3682 * skip the util_est update.
3684 ue = p->se.avg.util_est;
3685 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3689 * Skip update of task's estimated utilization when its EWMA is
3690 * already ~1% close to its last activation value.
3692 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3693 last_ewma_diff = ue.enqueued - ue.ewma;
3694 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3698 * To avoid overestimation of actual task utilization, skip updates if
3699 * we cannot grant there is idle time in this CPU.
3701 cpu = cpu_of(rq_of(cfs_rq));
3702 if (task_util(p) > capacity_orig_of(cpu))
3706 * Update Task's estimated utilization
3708 * When *p completes an activation we can consolidate another sample
3709 * of the task size. This is done by storing the current PELT value
3710 * as ue.enqueued and by using this value to update the Exponential
3711 * Weighted Moving Average (EWMA):
3713 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
3714 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
3715 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
3716 * = w * ( last_ewma_diff ) + ewma(t-1)
3717 * = w * (last_ewma_diff + ewma(t-1) / w)
3719 * Where 'w' is the weight of new samples, which is configured to be
3720 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3722 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
3723 ue.ewma += last_ewma_diff;
3724 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
3725 WRITE_ONCE(p->se.avg.util_est, ue);
3728 static inline int task_fits_capacity(struct task_struct *p, long capacity)
3730 return capacity * 1024 > task_util_est(p) * capacity_margin;
3733 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
3735 if (!static_branch_unlikely(&sched_asym_cpucapacity))
3739 rq->misfit_task_load = 0;
3743 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
3744 rq->misfit_task_load = 0;
3748 rq->misfit_task_load = task_h_load(p);
3751 #else /* CONFIG_SMP */
3753 #define UPDATE_TG 0x0
3754 #define SKIP_AGE_LOAD 0x0
3755 #define DO_ATTACH 0x0
3757 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3759 cfs_rq_util_change(cfs_rq, 0);
3762 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3765 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
3767 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3769 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3775 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
3778 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
3780 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
3782 #endif /* CONFIG_SMP */
3784 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3786 #ifdef CONFIG_SCHED_DEBUG
3787 s64 d = se->vruntime - cfs_rq->min_vruntime;
3792 if (d > 3*sysctl_sched_latency)
3793 schedstat_inc(cfs_rq->nr_spread_over);
3798 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3800 u64 vruntime = cfs_rq->min_vruntime;
3803 * The 'current' period is already promised to the current tasks,
3804 * however the extra weight of the new task will slow them down a
3805 * little, place the new task so that it fits in the slot that
3806 * stays open at the end.
3808 if (initial && sched_feat(START_DEBIT))
3809 vruntime += sched_vslice(cfs_rq, se);
3811 /* sleeps up to a single latency don't count. */
3813 unsigned long thresh = sysctl_sched_latency;
3816 * Halve their sleep time's effect, to allow
3817 * for a gentler effect of sleepers:
3819 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3825 /* ensure we never gain time by being placed backwards. */
3826 se->vruntime = max_vruntime(se->vruntime, vruntime);
3829 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3831 static inline void check_schedstat_required(void)
3833 #ifdef CONFIG_SCHEDSTATS
3834 if (schedstat_enabled())
3837 /* Force schedstat enabled if a dependent tracepoint is active */
3838 if (trace_sched_stat_wait_enabled() ||
3839 trace_sched_stat_sleep_enabled() ||
3840 trace_sched_stat_iowait_enabled() ||
3841 trace_sched_stat_blocked_enabled() ||
3842 trace_sched_stat_runtime_enabled()) {
3843 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3844 "stat_blocked and stat_runtime require the "
3845 "kernel parameter schedstats=enable or "
3846 "kernel.sched_schedstats=1\n");
3857 * update_min_vruntime()
3858 * vruntime -= min_vruntime
3862 * update_min_vruntime()
3863 * vruntime += min_vruntime
3865 * this way the vruntime transition between RQs is done when both
3866 * min_vruntime are up-to-date.
3870 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3871 * vruntime -= min_vruntime
3875 * update_min_vruntime()
3876 * vruntime += min_vruntime
3878 * this way we don't have the most up-to-date min_vruntime on the originating
3879 * CPU and an up-to-date min_vruntime on the destination CPU.
3883 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3885 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3886 bool curr = cfs_rq->curr == se;
3889 * If we're the current task, we must renormalise before calling
3893 se->vruntime += cfs_rq->min_vruntime;
3895 update_curr(cfs_rq);
3898 * Otherwise, renormalise after, such that we're placed at the current
3899 * moment in time, instead of some random moment in the past. Being
3900 * placed in the past could significantly boost this task to the
3901 * fairness detriment of existing tasks.
3903 if (renorm && !curr)
3904 se->vruntime += cfs_rq->min_vruntime;
3907 * When enqueuing a sched_entity, we must:
3908 * - Update loads to have both entity and cfs_rq synced with now.
3909 * - Add its load to cfs_rq->runnable_avg
3910 * - For group_entity, update its weight to reflect the new share of
3912 * - Add its new weight to cfs_rq->load.weight
3914 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
3915 update_cfs_group(se);
3916 enqueue_runnable_load_avg(cfs_rq, se);
3917 account_entity_enqueue(cfs_rq, se);
3919 if (flags & ENQUEUE_WAKEUP)
3920 place_entity(cfs_rq, se, 0);
3922 check_schedstat_required();
3923 update_stats_enqueue(cfs_rq, se, flags);
3924 check_spread(cfs_rq, se);
3926 __enqueue_entity(cfs_rq, se);
3929 if (cfs_rq->nr_running == 1) {
3930 list_add_leaf_cfs_rq(cfs_rq);
3931 check_enqueue_throttle(cfs_rq);
3935 static void __clear_buddies_last(struct sched_entity *se)
3937 for_each_sched_entity(se) {
3938 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3939 if (cfs_rq->last != se)
3942 cfs_rq->last = NULL;
3946 static void __clear_buddies_next(struct sched_entity *se)
3948 for_each_sched_entity(se) {
3949 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3950 if (cfs_rq->next != se)
3953 cfs_rq->next = NULL;
3957 static void __clear_buddies_skip(struct sched_entity *se)
3959 for_each_sched_entity(se) {
3960 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3961 if (cfs_rq->skip != se)
3964 cfs_rq->skip = NULL;
3968 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3970 if (cfs_rq->last == se)
3971 __clear_buddies_last(se);
3973 if (cfs_rq->next == se)
3974 __clear_buddies_next(se);
3976 if (cfs_rq->skip == se)
3977 __clear_buddies_skip(se);
3980 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3983 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3986 * Update run-time statistics of the 'current'.
3988 update_curr(cfs_rq);
3991 * When dequeuing a sched_entity, we must:
3992 * - Update loads to have both entity and cfs_rq synced with now.
3993 * - Subtract its load from the cfs_rq->runnable_avg.
3994 * - Subtract its previous weight from cfs_rq->load.weight.
3995 * - For group entity, update its weight to reflect the new share
3996 * of its group cfs_rq.
3998 update_load_avg(cfs_rq, se, UPDATE_TG);
3999 dequeue_runnable_load_avg(cfs_rq, se);
4001 update_stats_dequeue(cfs_rq, se, flags);
4003 clear_buddies(cfs_rq, se);
4005 if (se != cfs_rq->curr)
4006 __dequeue_entity(cfs_rq, se);
4008 account_entity_dequeue(cfs_rq, se);
4011 * Normalize after update_curr(); which will also have moved
4012 * min_vruntime if @se is the one holding it back. But before doing
4013 * update_min_vruntime() again, which will discount @se's position and
4014 * can move min_vruntime forward still more.
4016 if (!(flags & DEQUEUE_SLEEP))
4017 se->vruntime -= cfs_rq->min_vruntime;
4019 /* return excess runtime on last dequeue */
4020 return_cfs_rq_runtime(cfs_rq);
4022 update_cfs_group(se);
4025 * Now advance min_vruntime if @se was the entity holding it back,
4026 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4027 * put back on, and if we advance min_vruntime, we'll be placed back
4028 * further than we started -- ie. we'll be penalized.
4030 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4031 update_min_vruntime(cfs_rq);
4035 * Preempt the current task with a newly woken task if needed:
4038 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4040 unsigned long ideal_runtime, delta_exec;
4041 struct sched_entity *se;
4044 ideal_runtime = sched_slice(cfs_rq, curr);
4045 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4046 if (delta_exec > ideal_runtime) {
4047 resched_curr(rq_of(cfs_rq));
4049 * The current task ran long enough, ensure it doesn't get
4050 * re-elected due to buddy favours.
4052 clear_buddies(cfs_rq, curr);
4057 * Ensure that a task that missed wakeup preemption by a
4058 * narrow margin doesn't have to wait for a full slice.
4059 * This also mitigates buddy induced latencies under load.
4061 if (delta_exec < sysctl_sched_min_granularity)
4064 se = __pick_first_entity(cfs_rq);
4065 delta = curr->vruntime - se->vruntime;
4070 if (delta > ideal_runtime)
4071 resched_curr(rq_of(cfs_rq));
4075 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4077 /* 'current' is not kept within the tree. */
4080 * Any task has to be enqueued before it get to execute on
4081 * a CPU. So account for the time it spent waiting on the
4084 update_stats_wait_end(cfs_rq, se);
4085 __dequeue_entity(cfs_rq, se);
4086 update_load_avg(cfs_rq, se, UPDATE_TG);
4089 update_stats_curr_start(cfs_rq, se);
4093 * Track our maximum slice length, if the CPU's load is at
4094 * least twice that of our own weight (i.e. dont track it
4095 * when there are only lesser-weight tasks around):
4097 if (schedstat_enabled() &&
4098 rq_of(cfs_rq)->cfs.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 /* don't push forwards an existing deferred unthrottle */
4733 if (cfs_b->slack_started)
4735 cfs_b->slack_started = true;
4737 hrtimer_start(&cfs_b->slack_timer,
4738 ns_to_ktime(cfs_bandwidth_slack_period),
4742 /* we know any runtime found here is valid as update_curr() precedes return */
4743 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4745 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4746 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4748 if (slack_runtime <= 0)
4751 raw_spin_lock(&cfs_b->lock);
4752 if (cfs_b->quota != RUNTIME_INF &&
4753 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4754 cfs_b->runtime += slack_runtime;
4756 /* we are under rq->lock, defer unthrottling using a timer */
4757 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4758 !list_empty(&cfs_b->throttled_cfs_rq))
4759 start_cfs_slack_bandwidth(cfs_b);
4761 raw_spin_unlock(&cfs_b->lock);
4763 /* even if it's not valid for return we don't want to try again */
4764 cfs_rq->runtime_remaining -= slack_runtime;
4767 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4769 if (!cfs_bandwidth_used())
4772 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4775 __return_cfs_rq_runtime(cfs_rq);
4779 * This is done with a timer (instead of inline with bandwidth return) since
4780 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4782 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4784 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4785 unsigned long flags;
4788 /* confirm we're still not at a refresh boundary */
4789 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4790 cfs_b->slack_started = false;
4791 if (cfs_b->distribute_running) {
4792 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4796 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4797 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4801 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4802 runtime = cfs_b->runtime;
4804 expires = cfs_b->runtime_expires;
4806 cfs_b->distribute_running = 1;
4808 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4813 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4815 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4816 if (expires == cfs_b->runtime_expires)
4817 lsub_positive(&cfs_b->runtime, runtime);
4818 cfs_b->distribute_running = 0;
4819 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4823 * When a group wakes up we want to make sure that its quota is not already
4824 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4825 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4827 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4829 if (!cfs_bandwidth_used())
4832 /* an active group must be handled by the update_curr()->put() path */
4833 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4836 /* ensure the group is not already throttled */
4837 if (cfs_rq_throttled(cfs_rq))
4840 /* update runtime allocation */
4841 account_cfs_rq_runtime(cfs_rq, 0);
4842 if (cfs_rq->runtime_remaining <= 0)
4843 throttle_cfs_rq(cfs_rq);
4846 static void sync_throttle(struct task_group *tg, int cpu)
4848 struct cfs_rq *pcfs_rq, *cfs_rq;
4850 if (!cfs_bandwidth_used())
4856 cfs_rq = tg->cfs_rq[cpu];
4857 pcfs_rq = tg->parent->cfs_rq[cpu];
4859 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4860 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4863 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4864 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4866 if (!cfs_bandwidth_used())
4869 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4873 * it's possible for a throttled entity to be forced into a running
4874 * state (e.g. set_curr_task), in this case we're finished.
4876 if (cfs_rq_throttled(cfs_rq))
4879 throttle_cfs_rq(cfs_rq);
4883 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4885 struct cfs_bandwidth *cfs_b =
4886 container_of(timer, struct cfs_bandwidth, slack_timer);
4888 do_sched_cfs_slack_timer(cfs_b);
4890 return HRTIMER_NORESTART;
4893 extern const u64 max_cfs_quota_period;
4895 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4897 struct cfs_bandwidth *cfs_b =
4898 container_of(timer, struct cfs_bandwidth, period_timer);
4899 unsigned long flags;
4904 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4906 overrun = hrtimer_forward_now(timer, cfs_b->period);
4911 u64 new, old = ktime_to_ns(cfs_b->period);
4913 new = (old * 147) / 128; /* ~115% */
4914 new = min(new, max_cfs_quota_period);
4916 cfs_b->period = ns_to_ktime(new);
4918 /* since max is 1s, this is limited to 1e9^2, which fits in u64 */
4919 cfs_b->quota *= new;
4920 cfs_b->quota = div64_u64(cfs_b->quota, old);
4922 pr_warn_ratelimited(
4923 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us %lld, cfs_quota_us = %lld)\n",
4925 div_u64(new, NSEC_PER_USEC),
4926 div_u64(cfs_b->quota, NSEC_PER_USEC));
4928 /* reset count so we don't come right back in here */
4932 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
4935 cfs_b->period_active = 0;
4936 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4938 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4941 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4943 raw_spin_lock_init(&cfs_b->lock);
4945 cfs_b->quota = RUNTIME_INF;
4946 cfs_b->period = ns_to_ktime(default_cfs_period());
4948 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4949 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4950 cfs_b->period_timer.function = sched_cfs_period_timer;
4951 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4952 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4953 cfs_b->distribute_running = 0;
4954 cfs_b->slack_started = false;
4957 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4959 cfs_rq->runtime_enabled = 0;
4960 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4963 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4967 lockdep_assert_held(&cfs_b->lock);
4969 if (cfs_b->period_active)
4972 cfs_b->period_active = 1;
4973 overrun = hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4974 cfs_b->runtime_expires += (overrun + 1) * ktime_to_ns(cfs_b->period);
4975 cfs_b->expires_seq++;
4976 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4979 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4981 /* init_cfs_bandwidth() was not called */
4982 if (!cfs_b->throttled_cfs_rq.next)
4985 hrtimer_cancel(&cfs_b->period_timer);
4986 hrtimer_cancel(&cfs_b->slack_timer);
4990 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
4992 * The race is harmless, since modifying bandwidth settings of unhooked group
4993 * bits doesn't do much.
4996 /* cpu online calback */
4997 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4999 struct task_group *tg;
5001 lockdep_assert_held(&rq->lock);
5004 list_for_each_entry_rcu(tg, &task_groups, list) {
5005 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5006 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5008 raw_spin_lock(&cfs_b->lock);
5009 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5010 raw_spin_unlock(&cfs_b->lock);
5015 /* cpu offline callback */
5016 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5018 struct task_group *tg;
5020 lockdep_assert_held(&rq->lock);
5023 list_for_each_entry_rcu(tg, &task_groups, list) {
5024 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5026 if (!cfs_rq->runtime_enabled)
5030 * clock_task is not advancing so we just need to make sure
5031 * there's some valid quota amount
5033 cfs_rq->runtime_remaining = 1;
5035 * Offline rq is schedulable till CPU is completely disabled
5036 * in take_cpu_down(), so we prevent new cfs throttling here.
5038 cfs_rq->runtime_enabled = 0;
5040 if (cfs_rq_throttled(cfs_rq))
5041 unthrottle_cfs_rq(cfs_rq);
5046 #else /* CONFIG_CFS_BANDWIDTH */
5048 static inline bool cfs_bandwidth_used(void)
5053 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
5055 return rq_clock_task(rq_of(cfs_rq));
5058 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5059 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5060 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5061 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5062 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5064 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5069 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5074 static inline int throttled_lb_pair(struct task_group *tg,
5075 int src_cpu, int dest_cpu)
5080 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5082 #ifdef CONFIG_FAIR_GROUP_SCHED
5083 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5086 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5090 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5091 static inline void update_runtime_enabled(struct rq *rq) {}
5092 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5094 #endif /* CONFIG_CFS_BANDWIDTH */
5096 /**************************************************
5097 * CFS operations on tasks:
5100 #ifdef CONFIG_SCHED_HRTICK
5101 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5103 struct sched_entity *se = &p->se;
5104 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5106 SCHED_WARN_ON(task_rq(p) != rq);
5108 if (rq->cfs.h_nr_running > 1) {
5109 u64 slice = sched_slice(cfs_rq, se);
5110 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5111 s64 delta = slice - ran;
5118 hrtick_start(rq, delta);
5123 * called from enqueue/dequeue and updates the hrtick when the
5124 * current task is from our class and nr_running is low enough
5127 static void hrtick_update(struct rq *rq)
5129 struct task_struct *curr = rq->curr;
5131 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5134 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5135 hrtick_start_fair(rq, curr);
5137 #else /* !CONFIG_SCHED_HRTICK */
5139 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5143 static inline void hrtick_update(struct rq *rq)
5149 static inline unsigned long cpu_util(int cpu);
5151 static inline bool cpu_overutilized(int cpu)
5153 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5156 static inline void update_overutilized_status(struct rq *rq)
5158 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu))
5159 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5162 static inline void update_overutilized_status(struct rq *rq) { }
5166 * The enqueue_task method is called before nr_running is
5167 * increased. Here we update the fair scheduling stats and
5168 * then put the task into the rbtree:
5171 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5173 struct cfs_rq *cfs_rq;
5174 struct sched_entity *se = &p->se;
5177 * The code below (indirectly) updates schedutil which looks at
5178 * the cfs_rq utilization to select a frequency.
5179 * Let's add the task's estimated utilization to the cfs_rq's
5180 * estimated utilization, before we update schedutil.
5182 util_est_enqueue(&rq->cfs, p);
5185 * If in_iowait is set, the code below may not trigger any cpufreq
5186 * utilization updates, so do it here explicitly with the IOWAIT flag
5190 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5192 for_each_sched_entity(se) {
5195 cfs_rq = cfs_rq_of(se);
5196 enqueue_entity(cfs_rq, se, flags);
5199 * end evaluation on encountering a throttled cfs_rq
5201 * note: in the case of encountering a throttled cfs_rq we will
5202 * post the final h_nr_running increment below.
5204 if (cfs_rq_throttled(cfs_rq))
5206 cfs_rq->h_nr_running++;
5208 flags = ENQUEUE_WAKEUP;
5211 for_each_sched_entity(se) {
5212 cfs_rq = cfs_rq_of(se);
5213 cfs_rq->h_nr_running++;
5215 if (cfs_rq_throttled(cfs_rq))
5218 update_load_avg(cfs_rq, se, UPDATE_TG);
5219 update_cfs_group(se);
5223 add_nr_running(rq, 1);
5225 * Since new tasks are assigned an initial util_avg equal to
5226 * half of the spare capacity of their CPU, tiny tasks have the
5227 * ability to cross the overutilized threshold, which will
5228 * result in the load balancer ruining all the task placement
5229 * done by EAS. As a way to mitigate that effect, do not account
5230 * for the first enqueue operation of new tasks during the
5231 * overutilized flag detection.
5233 * A better way of solving this problem would be to wait for
5234 * the PELT signals of tasks to converge before taking them
5235 * into account, but that is not straightforward to implement,
5236 * and the following generally works well enough in practice.
5238 if (flags & ENQUEUE_WAKEUP)
5239 update_overutilized_status(rq);
5243 if (cfs_bandwidth_used()) {
5245 * When bandwidth control is enabled; the cfs_rq_throttled()
5246 * breaks in the above iteration can result in incomplete
5247 * leaf list maintenance, resulting in triggering the assertion
5250 for_each_sched_entity(se) {
5251 cfs_rq = cfs_rq_of(se);
5253 if (list_add_leaf_cfs_rq(cfs_rq))
5258 assert_list_leaf_cfs_rq(rq);
5263 static void set_next_buddy(struct sched_entity *se);
5266 * The dequeue_task method is called before nr_running is
5267 * decreased. We remove the task from the rbtree and
5268 * update the fair scheduling stats:
5270 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5272 struct cfs_rq *cfs_rq;
5273 struct sched_entity *se = &p->se;
5274 int task_sleep = flags & DEQUEUE_SLEEP;
5276 for_each_sched_entity(se) {
5277 cfs_rq = cfs_rq_of(se);
5278 dequeue_entity(cfs_rq, se, flags);
5281 * end evaluation on encountering a throttled cfs_rq
5283 * note: in the case of encountering a throttled cfs_rq we will
5284 * post the final h_nr_running decrement below.
5286 if (cfs_rq_throttled(cfs_rq))
5288 cfs_rq->h_nr_running--;
5290 /* Don't dequeue parent if it has other entities besides us */
5291 if (cfs_rq->load.weight) {
5292 /* Avoid re-evaluating load for this entity: */
5293 se = parent_entity(se);
5295 * Bias pick_next to pick a task from this cfs_rq, as
5296 * p is sleeping when it is within its sched_slice.
5298 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5302 flags |= DEQUEUE_SLEEP;
5305 for_each_sched_entity(se) {
5306 cfs_rq = cfs_rq_of(se);
5307 cfs_rq->h_nr_running--;
5309 if (cfs_rq_throttled(cfs_rq))
5312 update_load_avg(cfs_rq, se, UPDATE_TG);
5313 update_cfs_group(se);
5317 sub_nr_running(rq, 1);
5319 util_est_dequeue(&rq->cfs, p, task_sleep);
5325 /* Working cpumask for: load_balance, load_balance_newidle. */
5326 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5327 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5329 #ifdef CONFIG_NO_HZ_COMMON
5332 cpumask_var_t idle_cpus_mask;
5334 int has_blocked; /* Idle CPUS has blocked load */
5335 unsigned long next_balance; /* in jiffy units */
5336 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5337 } nohz ____cacheline_aligned;
5339 #endif /* CONFIG_NO_HZ_COMMON */
5341 static unsigned long weighted_cpuload(struct rq *rq)
5343 return cfs_rq_runnable_load_avg(&rq->cfs);
5346 static unsigned long capacity_of(int cpu)
5348 return cpu_rq(cpu)->cpu_capacity;
5351 static unsigned long cpu_avg_load_per_task(int cpu)
5353 struct rq *rq = cpu_rq(cpu);
5354 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5355 unsigned long load_avg = weighted_cpuload(rq);
5358 return load_avg / nr_running;
5363 static void record_wakee(struct task_struct *p)
5366 * Only decay a single time; tasks that have less then 1 wakeup per
5367 * jiffy will not have built up many flips.
5369 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5370 current->wakee_flips >>= 1;
5371 current->wakee_flip_decay_ts = jiffies;
5374 if (current->last_wakee != p) {
5375 current->last_wakee = p;
5376 current->wakee_flips++;
5381 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5383 * A waker of many should wake a different task than the one last awakened
5384 * at a frequency roughly N times higher than one of its wakees.
5386 * In order to determine whether we should let the load spread vs consolidating
5387 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5388 * partner, and a factor of lls_size higher frequency in the other.
5390 * With both conditions met, we can be relatively sure that the relationship is
5391 * non-monogamous, with partner count exceeding socket size.
5393 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5394 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5397 static int wake_wide(struct task_struct *p)
5399 unsigned int master = current->wakee_flips;
5400 unsigned int slave = p->wakee_flips;
5401 int factor = this_cpu_read(sd_llc_size);
5404 swap(master, slave);
5405 if (slave < factor || master < slave * factor)
5411 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5412 * soonest. For the purpose of speed we only consider the waking and previous
5415 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5416 * cache-affine and is (or will be) idle.
5418 * wake_affine_weight() - considers the weight to reflect the average
5419 * scheduling latency of the CPUs. This seems to work
5420 * for the overloaded case.
5423 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5426 * If this_cpu is idle, it implies the wakeup is from interrupt
5427 * context. Only allow the move if cache is shared. Otherwise an
5428 * interrupt intensive workload could force all tasks onto one
5429 * node depending on the IO topology or IRQ affinity settings.
5431 * If the prev_cpu is idle and cache affine then avoid a migration.
5432 * There is no guarantee that the cache hot data from an interrupt
5433 * is more important than cache hot data on the prev_cpu and from
5434 * a cpufreq perspective, it's better to have higher utilisation
5437 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5438 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5440 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5443 return nr_cpumask_bits;
5447 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5448 int this_cpu, int prev_cpu, int sync)
5450 s64 this_eff_load, prev_eff_load;
5451 unsigned long task_load;
5453 this_eff_load = weighted_cpuload(cpu_rq(this_cpu));
5456 unsigned long current_load = task_h_load(current);
5458 if (current_load > this_eff_load)
5461 this_eff_load -= current_load;
5464 task_load = task_h_load(p);
5466 this_eff_load += task_load;
5467 if (sched_feat(WA_BIAS))
5468 this_eff_load *= 100;
5469 this_eff_load *= capacity_of(prev_cpu);
5471 prev_eff_load = weighted_cpuload(cpu_rq(prev_cpu));
5472 prev_eff_load -= task_load;
5473 if (sched_feat(WA_BIAS))
5474 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5475 prev_eff_load *= capacity_of(this_cpu);
5478 * If sync, adjust the weight of prev_eff_load such that if
5479 * prev_eff == this_eff that select_idle_sibling() will consider
5480 * stacking the wakee on top of the waker if no other CPU is
5486 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5489 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5490 int this_cpu, int prev_cpu, int sync)
5492 int target = nr_cpumask_bits;
5494 if (sched_feat(WA_IDLE))
5495 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5497 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5498 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5500 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5501 if (target == nr_cpumask_bits)
5504 schedstat_inc(sd->ttwu_move_affine);
5505 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5509 static unsigned long cpu_util_without(int cpu, struct task_struct *p);
5511 static unsigned long capacity_spare_without(int cpu, struct task_struct *p)
5513 return max_t(long, capacity_of(cpu) - cpu_util_without(cpu, p), 0);
5517 * find_idlest_group finds and returns the least busy CPU group within the
5520 * Assumes p is allowed on at least one CPU in sd.
5522 static struct sched_group *
5523 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5524 int this_cpu, int sd_flag)
5526 struct sched_group *idlest = NULL, *group = sd->groups;
5527 struct sched_group *most_spare_sg = NULL;
5528 unsigned long min_runnable_load = ULONG_MAX;
5529 unsigned long this_runnable_load = ULONG_MAX;
5530 unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5531 unsigned long most_spare = 0, this_spare = 0;
5532 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5533 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5534 (sd->imbalance_pct-100) / 100;
5537 unsigned long load, avg_load, runnable_load;
5538 unsigned long spare_cap, max_spare_cap;
5542 /* Skip over this group if it has no CPUs allowed */
5543 if (!cpumask_intersects(sched_group_span(group),
5547 local_group = cpumask_test_cpu(this_cpu,
5548 sched_group_span(group));
5551 * Tally up the load of all CPUs in the group and find
5552 * the group containing the CPU with most spare capacity.
5558 for_each_cpu(i, sched_group_span(group)) {
5559 load = weighted_cpuload(cpu_rq(i));
5560 runnable_load += load;
5562 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5564 spare_cap = capacity_spare_without(i, p);
5566 if (spare_cap > max_spare_cap)
5567 max_spare_cap = spare_cap;
5570 /* Adjust by relative CPU capacity of the group */
5571 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5572 group->sgc->capacity;
5573 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5574 group->sgc->capacity;
5577 this_runnable_load = runnable_load;
5578 this_avg_load = avg_load;
5579 this_spare = max_spare_cap;
5581 if (min_runnable_load > (runnable_load + imbalance)) {
5583 * The runnable load is significantly smaller
5584 * so we can pick this new CPU:
5586 min_runnable_load = runnable_load;
5587 min_avg_load = avg_load;
5589 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5590 (100*min_avg_load > imbalance_scale*avg_load)) {
5592 * The runnable loads are close so take the
5593 * blocked load into account through avg_load:
5595 min_avg_load = avg_load;
5599 if (most_spare < max_spare_cap) {
5600 most_spare = max_spare_cap;
5601 most_spare_sg = group;
5604 } while (group = group->next, group != sd->groups);
5607 * The cross-over point between using spare capacity or least load
5608 * is too conservative for high utilization tasks on partially
5609 * utilized systems if we require spare_capacity > task_util(p),
5610 * so we allow for some task stuffing by using
5611 * spare_capacity > task_util(p)/2.
5613 * Spare capacity can't be used for fork because the utilization has
5614 * not been set yet, we must first select a rq to compute the initial
5617 if (sd_flag & SD_BALANCE_FORK)
5620 if (this_spare > task_util(p) / 2 &&
5621 imbalance_scale*this_spare > 100*most_spare)
5624 if (most_spare > task_util(p) / 2)
5625 return most_spare_sg;
5632 * When comparing groups across NUMA domains, it's possible for the
5633 * local domain to be very lightly loaded relative to the remote
5634 * domains but "imbalance" skews the comparison making remote CPUs
5635 * look much more favourable. When considering cross-domain, add
5636 * imbalance to the runnable load on the remote node and consider
5639 if ((sd->flags & SD_NUMA) &&
5640 min_runnable_load + imbalance >= this_runnable_load)
5643 if (min_runnable_load > (this_runnable_load + imbalance))
5646 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5647 (100*this_avg_load < imbalance_scale*min_avg_load))
5654 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5657 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5659 unsigned long load, min_load = ULONG_MAX;
5660 unsigned int min_exit_latency = UINT_MAX;
5661 u64 latest_idle_timestamp = 0;
5662 int least_loaded_cpu = this_cpu;
5663 int shallowest_idle_cpu = -1;
5666 /* Check if we have any choice: */
5667 if (group->group_weight == 1)
5668 return cpumask_first(sched_group_span(group));
5670 /* Traverse only the allowed CPUs */
5671 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
5672 if (available_idle_cpu(i)) {
5673 struct rq *rq = cpu_rq(i);
5674 struct cpuidle_state *idle = idle_get_state(rq);
5675 if (idle && idle->exit_latency < min_exit_latency) {
5677 * We give priority to a CPU whose idle state
5678 * has the smallest exit latency irrespective
5679 * of any idle timestamp.
5681 min_exit_latency = idle->exit_latency;
5682 latest_idle_timestamp = rq->idle_stamp;
5683 shallowest_idle_cpu = i;
5684 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5685 rq->idle_stamp > latest_idle_timestamp) {
5687 * If equal or no active idle state, then
5688 * the most recently idled CPU might have
5691 latest_idle_timestamp = rq->idle_stamp;
5692 shallowest_idle_cpu = i;
5694 } else if (shallowest_idle_cpu == -1) {
5695 load = weighted_cpuload(cpu_rq(i));
5696 if (load < min_load) {
5698 least_loaded_cpu = i;
5703 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5706 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5707 int cpu, int prev_cpu, int sd_flag)
5711 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
5715 * We need task's util for capacity_spare_without, sync it up to
5716 * prev_cpu's last_update_time.
5718 if (!(sd_flag & SD_BALANCE_FORK))
5719 sync_entity_load_avg(&p->se);
5722 struct sched_group *group;
5723 struct sched_domain *tmp;
5726 if (!(sd->flags & sd_flag)) {
5731 group = find_idlest_group(sd, p, cpu, sd_flag);
5737 new_cpu = find_idlest_group_cpu(group, p, cpu);
5738 if (new_cpu == cpu) {
5739 /* Now try balancing at a lower domain level of 'cpu': */
5744 /* Now try balancing at a lower domain level of 'new_cpu': */
5746 weight = sd->span_weight;
5748 for_each_domain(cpu, tmp) {
5749 if (weight <= tmp->span_weight)
5751 if (tmp->flags & sd_flag)
5759 #ifdef CONFIG_SCHED_SMT
5760 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5761 EXPORT_SYMBOL_GPL(sched_smt_present);
5763 static inline void set_idle_cores(int cpu, int val)
5765 struct sched_domain_shared *sds;
5767 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5769 WRITE_ONCE(sds->has_idle_cores, val);
5772 static inline bool test_idle_cores(int cpu, bool def)
5774 struct sched_domain_shared *sds;
5776 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5778 return READ_ONCE(sds->has_idle_cores);
5784 * Scans the local SMT mask to see if the entire core is idle, and records this
5785 * information in sd_llc_shared->has_idle_cores.
5787 * Since SMT siblings share all cache levels, inspecting this limited remote
5788 * state should be fairly cheap.
5790 void __update_idle_core(struct rq *rq)
5792 int core = cpu_of(rq);
5796 if (test_idle_cores(core, true))
5799 for_each_cpu(cpu, cpu_smt_mask(core)) {
5803 if (!available_idle_cpu(cpu))
5807 set_idle_cores(core, 1);
5813 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5814 * there are no idle cores left in the system; tracked through
5815 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5817 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5819 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5822 if (!static_branch_likely(&sched_smt_present))
5825 if (!test_idle_cores(target, false))
5828 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
5830 for_each_cpu_wrap(core, cpus, target) {
5833 for_each_cpu(cpu, cpu_smt_mask(core)) {
5834 __cpumask_clear_cpu(cpu, cpus);
5835 if (!available_idle_cpu(cpu))
5844 * Failed to find an idle core; stop looking for one.
5846 set_idle_cores(target, 0);
5852 * Scan the local SMT mask for idle CPUs.
5854 static int select_idle_smt(struct task_struct *p, int target)
5858 if (!static_branch_likely(&sched_smt_present))
5861 for_each_cpu(cpu, cpu_smt_mask(target)) {
5862 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
5864 if (available_idle_cpu(cpu))
5871 #else /* CONFIG_SCHED_SMT */
5873 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5878 static inline int select_idle_smt(struct task_struct *p, int target)
5883 #endif /* CONFIG_SCHED_SMT */
5886 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5887 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5888 * average idle time for this rq (as found in rq->avg_idle).
5890 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
5892 struct sched_domain *this_sd;
5893 u64 avg_cost, avg_idle;
5896 int cpu, nr = INT_MAX;
5898 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
5903 * Due to large variance we need a large fuzz factor; hackbench in
5904 * particularly is sensitive here.
5906 avg_idle = this_rq()->avg_idle / 512;
5907 avg_cost = this_sd->avg_scan_cost + 1;
5909 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
5912 if (sched_feat(SIS_PROP)) {
5913 u64 span_avg = sd->span_weight * avg_idle;
5914 if (span_avg > 4*avg_cost)
5915 nr = div_u64(span_avg, avg_cost);
5920 time = local_clock();
5922 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
5925 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
5927 if (available_idle_cpu(cpu))
5931 time = local_clock() - time;
5932 cost = this_sd->avg_scan_cost;
5933 delta = (s64)(time - cost) / 8;
5934 this_sd->avg_scan_cost += delta;
5940 * Try and locate an idle core/thread in the LLC cache domain.
5942 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5944 struct sched_domain *sd;
5945 int i, recent_used_cpu;
5947 if (available_idle_cpu(target))
5951 * If the previous CPU is cache affine and idle, don't be stupid:
5953 if (prev != target && cpus_share_cache(prev, target) && available_idle_cpu(prev))
5956 /* Check a recently used CPU as a potential idle candidate: */
5957 recent_used_cpu = p->recent_used_cpu;
5958 if (recent_used_cpu != prev &&
5959 recent_used_cpu != target &&
5960 cpus_share_cache(recent_used_cpu, target) &&
5961 available_idle_cpu(recent_used_cpu) &&
5962 cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr)) {
5964 * Replace recent_used_cpu with prev as it is a potential
5965 * candidate for the next wake:
5967 p->recent_used_cpu = prev;
5968 return recent_used_cpu;
5971 sd = rcu_dereference(per_cpu(sd_llc, target));
5975 i = select_idle_core(p, sd, target);
5976 if ((unsigned)i < nr_cpumask_bits)
5979 i = select_idle_cpu(p, sd, target);
5980 if ((unsigned)i < nr_cpumask_bits)
5983 i = select_idle_smt(p, target);
5984 if ((unsigned)i < nr_cpumask_bits)
5991 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
5992 * @cpu: the CPU to get the utilization of
5994 * The unit of the return value must be the one of capacity so we can compare
5995 * the utilization with the capacity of the CPU that is available for CFS task
5996 * (ie cpu_capacity).
5998 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5999 * recent utilization of currently non-runnable tasks on a CPU. It represents
6000 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6001 * capacity_orig is the cpu_capacity available at the highest frequency
6002 * (arch_scale_freq_capacity()).
6003 * The utilization of a CPU converges towards a sum equal to or less than the
6004 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6005 * the running time on this CPU scaled by capacity_curr.
6007 * The estimated utilization of a CPU is defined to be the maximum between its
6008 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6009 * currently RUNNABLE on that CPU.
6010 * This allows to properly represent the expected utilization of a CPU which
6011 * has just got a big task running since a long sleep period. At the same time
6012 * however it preserves the benefits of the "blocked utilization" in
6013 * describing the potential for other tasks waking up on the same CPU.
6015 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6016 * higher than capacity_orig because of unfortunate rounding in
6017 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6018 * the average stabilizes with the new running time. We need to check that the
6019 * utilization stays within the range of [0..capacity_orig] and cap it if
6020 * necessary. Without utilization capping, a group could be seen as overloaded
6021 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6022 * available capacity. We allow utilization to overshoot capacity_curr (but not
6023 * capacity_orig) as it useful for predicting the capacity required after task
6024 * migrations (scheduler-driven DVFS).
6026 * Return: the (estimated) utilization for the specified CPU
6028 static inline unsigned long cpu_util(int cpu)
6030 struct cfs_rq *cfs_rq;
6033 cfs_rq = &cpu_rq(cpu)->cfs;
6034 util = READ_ONCE(cfs_rq->avg.util_avg);
6036 if (sched_feat(UTIL_EST))
6037 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6039 return min_t(unsigned long, util, capacity_orig_of(cpu));
6043 * cpu_util_without: compute cpu utilization without any contributions from *p
6044 * @cpu: the CPU which utilization is requested
6045 * @p: the task which utilization should be discounted
6047 * The utilization of a CPU is defined by the utilization of tasks currently
6048 * enqueued on that CPU as well as tasks which are currently sleeping after an
6049 * execution on that CPU.
6051 * This method returns the utilization of the specified CPU by discounting the
6052 * utilization of the specified task, whenever the task is currently
6053 * contributing to the CPU utilization.
6055 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6057 struct cfs_rq *cfs_rq;
6060 /* Task has no contribution or is new */
6061 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6062 return cpu_util(cpu);
6064 cfs_rq = &cpu_rq(cpu)->cfs;
6065 util = READ_ONCE(cfs_rq->avg.util_avg);
6067 /* Discount task's util from CPU's util */
6068 lsub_positive(&util, task_util(p));
6073 * a) if *p is the only task sleeping on this CPU, then:
6074 * cpu_util (== task_util) > util_est (== 0)
6075 * and thus we return:
6076 * cpu_util_without = (cpu_util - task_util) = 0
6078 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6080 * cpu_util >= task_util
6081 * cpu_util > util_est (== 0)
6082 * and thus we discount *p's blocked utilization to return:
6083 * cpu_util_without = (cpu_util - task_util) >= 0
6085 * c) if other tasks are RUNNABLE on that CPU and
6086 * util_est > cpu_util
6087 * then we use util_est since it returns a more restrictive
6088 * estimation of the spare capacity on that CPU, by just
6089 * considering the expected utilization of tasks already
6090 * runnable on that CPU.
6092 * Cases a) and b) are covered by the above code, while case c) is
6093 * covered by the following code when estimated utilization is
6096 if (sched_feat(UTIL_EST)) {
6097 unsigned int estimated =
6098 READ_ONCE(cfs_rq->avg.util_est.enqueued);
6101 * Despite the following checks we still have a small window
6102 * for a possible race, when an execl's select_task_rq_fair()
6103 * races with LB's detach_task():
6106 * p->on_rq = TASK_ON_RQ_MIGRATING;
6107 * ---------------------------------- A
6108 * deactivate_task() \
6109 * dequeue_task() + RaceTime
6110 * util_est_dequeue() /
6111 * ---------------------------------- B
6113 * The additional check on "current == p" it's required to
6114 * properly fix the execl regression and it helps in further
6115 * reducing the chances for the above race.
6117 if (unlikely(task_on_rq_queued(p) || current == p))
6118 lsub_positive(&estimated, _task_util_est(p));
6120 util = max(util, estimated);
6124 * Utilization (estimated) can exceed the CPU capacity, thus let's
6125 * clamp to the maximum CPU capacity to ensure consistency with
6126 * the cpu_util call.
6128 return min_t(unsigned long, util, capacity_orig_of(cpu));
6132 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6133 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6135 * In that case WAKE_AFFINE doesn't make sense and we'll let
6136 * BALANCE_WAKE sort things out.
6138 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6140 long min_cap, max_cap;
6142 if (!static_branch_unlikely(&sched_asym_cpucapacity))
6145 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6146 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6148 /* Minimum capacity is close to max, no need to abort wake_affine */
6149 if (max_cap - min_cap < max_cap >> 3)
6152 /* Bring task utilization in sync with prev_cpu */
6153 sync_entity_load_avg(&p->se);
6155 return !task_fits_capacity(p, min_cap);
6159 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6162 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6164 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6165 unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
6168 * If @p migrates from @cpu to another, remove its contribution. Or,
6169 * if @p migrates from another CPU to @cpu, add its contribution. In
6170 * the other cases, @cpu is not impacted by the migration, so the
6171 * util_avg should already be correct.
6173 if (task_cpu(p) == cpu && dst_cpu != cpu)
6174 sub_positive(&util, task_util(p));
6175 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6176 util += task_util(p);
6178 if (sched_feat(UTIL_EST)) {
6179 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6182 * During wake-up, the task isn't enqueued yet and doesn't
6183 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6184 * so just add it (if needed) to "simulate" what will be
6185 * cpu_util() after the task has been enqueued.
6188 util_est += _task_util_est(p);
6190 util = max(util, util_est);
6193 return min(util, capacity_orig_of(cpu));
6197 * compute_energy(): Estimates the energy that would be consumed if @p was
6198 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6199 * landscape of the * CPUs after the task migration, and uses the Energy Model
6200 * to compute what would be the energy if we decided to actually migrate that
6204 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
6206 long util, max_util, sum_util, energy = 0;
6209 for (; pd; pd = pd->next) {
6210 max_util = sum_util = 0;
6212 * The capacity state of CPUs of the current rd can be driven by
6213 * CPUs of another rd if they belong to the same performance
6214 * domain. So, account for the utilization of these CPUs too
6215 * by masking pd with cpu_online_mask instead of the rd span.
6217 * If an entire performance domain is outside of the current rd,
6218 * it will not appear in its pd list and will not be accounted
6219 * by compute_energy().
6221 for_each_cpu_and(cpu, perf_domain_span(pd), cpu_online_mask) {
6222 util = cpu_util_next(cpu, p, dst_cpu);
6223 util = schedutil_energy_util(cpu, util);
6224 max_util = max(util, max_util);
6228 energy += em_pd_energy(pd->em_pd, max_util, sum_util);
6235 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6236 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6237 * spare capacity in each performance domain and uses it as a potential
6238 * candidate to execute the task. Then, it uses the Energy Model to figure
6239 * out which of the CPU candidates is the most energy-efficient.
6241 * The rationale for this heuristic is as follows. In a performance domain,
6242 * all the most energy efficient CPU candidates (according to the Energy
6243 * Model) are those for which we'll request a low frequency. When there are
6244 * several CPUs for which the frequency request will be the same, we don't
6245 * have enough data to break the tie between them, because the Energy Model
6246 * only includes active power costs. With this model, if we assume that
6247 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6248 * the maximum spare capacity in a performance domain is guaranteed to be among
6249 * the best candidates of the performance domain.
6251 * In practice, it could be preferable from an energy standpoint to pack
6252 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6253 * but that could also hurt our chances to go cluster idle, and we have no
6254 * ways to tell with the current Energy Model if this is actually a good
6255 * idea or not. So, find_energy_efficient_cpu() basically favors
6256 * cluster-packing, and spreading inside a cluster. That should at least be
6257 * a good thing for latency, and this is consistent with the idea that most
6258 * of the energy savings of EAS come from the asymmetry of the system, and
6259 * not so much from breaking the tie between identical CPUs. That's also the
6260 * reason why EAS is enabled in the topology code only for systems where
6261 * SD_ASYM_CPUCAPACITY is set.
6263 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6264 * they don't have any useful utilization data yet and it's not possible to
6265 * forecast their impact on energy consumption. Consequently, they will be
6266 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6267 * to be energy-inefficient in some use-cases. The alternative would be to
6268 * bias new tasks towards specific types of CPUs first, or to try to infer
6269 * their util_avg from the parent task, but those heuristics could hurt
6270 * other use-cases too. So, until someone finds a better way to solve this,
6271 * let's keep things simple by re-using the existing slow path.
6274 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6276 unsigned long prev_energy = ULONG_MAX, best_energy = ULONG_MAX;
6277 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6278 int cpu, best_energy_cpu = prev_cpu;
6279 struct perf_domain *head, *pd;
6280 unsigned long cpu_cap, util;
6281 struct sched_domain *sd;
6284 pd = rcu_dereference(rd->pd);
6285 if (!pd || READ_ONCE(rd->overutilized))
6290 * Energy-aware wake-up happens on the lowest sched_domain starting
6291 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6293 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6294 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6299 sync_entity_load_avg(&p->se);
6300 if (!task_util_est(p))
6303 for (; pd; pd = pd->next) {
6304 unsigned long cur_energy, spare_cap, max_spare_cap = 0;
6305 int max_spare_cap_cpu = -1;
6307 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
6308 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6311 /* Skip CPUs that will be overutilized. */
6312 util = cpu_util_next(cpu, p, cpu);
6313 cpu_cap = capacity_of(cpu);
6314 if (cpu_cap * 1024 < util * capacity_margin)
6317 /* Always use prev_cpu as a candidate. */
6318 if (cpu == prev_cpu) {
6319 prev_energy = compute_energy(p, prev_cpu, head);
6320 best_energy = min(best_energy, prev_energy);
6325 * Find the CPU with the maximum spare capacity in
6326 * the performance domain
6328 spare_cap = cpu_cap - util;
6329 if (spare_cap > max_spare_cap) {
6330 max_spare_cap = spare_cap;
6331 max_spare_cap_cpu = cpu;
6335 /* Evaluate the energy impact of using this CPU. */
6336 if (max_spare_cap_cpu >= 0) {
6337 cur_energy = compute_energy(p, max_spare_cap_cpu, head);
6338 if (cur_energy < best_energy) {
6339 best_energy = cur_energy;
6340 best_energy_cpu = max_spare_cap_cpu;
6348 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6349 * least 6% of the energy used by prev_cpu.
6351 if (prev_energy == ULONG_MAX)
6352 return best_energy_cpu;
6354 if ((prev_energy - best_energy) > (prev_energy >> 4))
6355 return best_energy_cpu;
6366 * select_task_rq_fair: Select target runqueue for the waking task in domains
6367 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6368 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6370 * Balances load by selecting the idlest CPU in the idlest group, or under
6371 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6373 * Returns the target CPU number.
6375 * preempt must be disabled.
6378 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6380 struct sched_domain *tmp, *sd = NULL;
6381 int cpu = smp_processor_id();
6382 int new_cpu = prev_cpu;
6383 int want_affine = 0;
6384 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6386 if (sd_flag & SD_BALANCE_WAKE) {
6389 if (sched_energy_enabled()) {
6390 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
6396 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu) &&
6397 cpumask_test_cpu(cpu, p->cpus_ptr);
6401 for_each_domain(cpu, tmp) {
6402 if (!(tmp->flags & SD_LOAD_BALANCE))
6406 * If both 'cpu' and 'prev_cpu' are part of this domain,
6407 * cpu is a valid SD_WAKE_AFFINE target.
6409 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6410 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6411 if (cpu != prev_cpu)
6412 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6414 sd = NULL; /* Prefer wake_affine over balance flags */
6418 if (tmp->flags & sd_flag)
6420 else if (!want_affine)
6426 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6427 } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6430 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6433 current->recent_used_cpu = cpu;
6440 static void detach_entity_cfs_rq(struct sched_entity *se);
6443 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6444 * cfs_rq_of(p) references at time of call are still valid and identify the
6445 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6447 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6450 * As blocked tasks retain absolute vruntime the migration needs to
6451 * deal with this by subtracting the old and adding the new
6452 * min_vruntime -- the latter is done by enqueue_entity() when placing
6453 * the task on the new runqueue.
6455 if (p->state == TASK_WAKING) {
6456 struct sched_entity *se = &p->se;
6457 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6460 #ifndef CONFIG_64BIT
6461 u64 min_vruntime_copy;
6464 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6466 min_vruntime = cfs_rq->min_vruntime;
6467 } while (min_vruntime != min_vruntime_copy);
6469 min_vruntime = cfs_rq->min_vruntime;
6472 se->vruntime -= min_vruntime;
6475 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6477 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6478 * rq->lock and can modify state directly.
6480 lockdep_assert_held(&task_rq(p)->lock);
6481 detach_entity_cfs_rq(&p->se);
6485 * We are supposed to update the task to "current" time, then
6486 * its up to date and ready to go to new CPU/cfs_rq. But we
6487 * have difficulty in getting what current time is, so simply
6488 * throw away the out-of-date time. This will result in the
6489 * wakee task is less decayed, but giving the wakee more load
6492 remove_entity_load_avg(&p->se);
6495 /* Tell new CPU we are migrated */
6496 p->se.avg.last_update_time = 0;
6498 /* We have migrated, no longer consider this task hot */
6499 p->se.exec_start = 0;
6501 update_scan_period(p, new_cpu);
6504 static void task_dead_fair(struct task_struct *p)
6506 remove_entity_load_avg(&p->se);
6508 #endif /* CONFIG_SMP */
6510 static unsigned long wakeup_gran(struct sched_entity *se)
6512 unsigned long gran = sysctl_sched_wakeup_granularity;
6515 * Since its curr running now, convert the gran from real-time
6516 * to virtual-time in his units.
6518 * By using 'se' instead of 'curr' we penalize light tasks, so
6519 * they get preempted easier. That is, if 'se' < 'curr' then
6520 * the resulting gran will be larger, therefore penalizing the
6521 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6522 * be smaller, again penalizing the lighter task.
6524 * This is especially important for buddies when the leftmost
6525 * task is higher priority than the buddy.
6527 return calc_delta_fair(gran, se);
6531 * Should 'se' preempt 'curr'.
6545 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6547 s64 gran, vdiff = curr->vruntime - se->vruntime;
6552 gran = wakeup_gran(se);
6559 static void set_last_buddy(struct sched_entity *se)
6561 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6564 for_each_sched_entity(se) {
6565 if (SCHED_WARN_ON(!se->on_rq))
6567 cfs_rq_of(se)->last = se;
6571 static void set_next_buddy(struct sched_entity *se)
6573 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6576 for_each_sched_entity(se) {
6577 if (SCHED_WARN_ON(!se->on_rq))
6579 cfs_rq_of(se)->next = se;
6583 static void set_skip_buddy(struct sched_entity *se)
6585 for_each_sched_entity(se)
6586 cfs_rq_of(se)->skip = se;
6590 * Preempt the current task with a newly woken task if needed:
6592 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6594 struct task_struct *curr = rq->curr;
6595 struct sched_entity *se = &curr->se, *pse = &p->se;
6596 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6597 int scale = cfs_rq->nr_running >= sched_nr_latency;
6598 int next_buddy_marked = 0;
6600 if (unlikely(se == pse))
6604 * This is possible from callers such as attach_tasks(), in which we
6605 * unconditionally check_prempt_curr() after an enqueue (which may have
6606 * lead to a throttle). This both saves work and prevents false
6607 * next-buddy nomination below.
6609 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6612 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6613 set_next_buddy(pse);
6614 next_buddy_marked = 1;
6618 * We can come here with TIF_NEED_RESCHED already set from new task
6621 * Note: this also catches the edge-case of curr being in a throttled
6622 * group (e.g. via set_curr_task), since update_curr() (in the
6623 * enqueue of curr) will have resulted in resched being set. This
6624 * prevents us from potentially nominating it as a false LAST_BUDDY
6627 if (test_tsk_need_resched(curr))
6630 /* Idle tasks are by definition preempted by non-idle tasks. */
6631 if (unlikely(task_has_idle_policy(curr)) &&
6632 likely(!task_has_idle_policy(p)))
6636 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6637 * is driven by the tick):
6639 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6642 find_matching_se(&se, &pse);
6643 update_curr(cfs_rq_of(se));
6645 if (wakeup_preempt_entity(se, pse) == 1) {
6647 * Bias pick_next to pick the sched entity that is
6648 * triggering this preemption.
6650 if (!next_buddy_marked)
6651 set_next_buddy(pse);
6660 * Only set the backward buddy when the current task is still
6661 * on the rq. This can happen when a wakeup gets interleaved
6662 * with schedule on the ->pre_schedule() or idle_balance()
6663 * point, either of which can * drop the rq lock.
6665 * Also, during early boot the idle thread is in the fair class,
6666 * for obvious reasons its a bad idea to schedule back to it.
6668 if (unlikely(!se->on_rq || curr == rq->idle))
6671 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6675 static struct task_struct *
6676 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6678 struct cfs_rq *cfs_rq = &rq->cfs;
6679 struct sched_entity *se;
6680 struct task_struct *p;
6684 if (!cfs_rq->nr_running)
6687 #ifdef CONFIG_FAIR_GROUP_SCHED
6688 if (prev->sched_class != &fair_sched_class)
6692 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6693 * likely that a next task is from the same cgroup as the current.
6695 * Therefore attempt to avoid putting and setting the entire cgroup
6696 * hierarchy, only change the part that actually changes.
6700 struct sched_entity *curr = cfs_rq->curr;
6703 * Since we got here without doing put_prev_entity() we also
6704 * have to consider cfs_rq->curr. If it is still a runnable
6705 * entity, update_curr() will update its vruntime, otherwise
6706 * forget we've ever seen it.
6710 update_curr(cfs_rq);
6715 * This call to check_cfs_rq_runtime() will do the
6716 * throttle and dequeue its entity in the parent(s).
6717 * Therefore the nr_running test will indeed
6720 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6723 if (!cfs_rq->nr_running)
6730 se = pick_next_entity(cfs_rq, curr);
6731 cfs_rq = group_cfs_rq(se);
6737 * Since we haven't yet done put_prev_entity and if the selected task
6738 * is a different task than we started out with, try and touch the
6739 * least amount of cfs_rqs.
6742 struct sched_entity *pse = &prev->se;
6744 while (!(cfs_rq = is_same_group(se, pse))) {
6745 int se_depth = se->depth;
6746 int pse_depth = pse->depth;
6748 if (se_depth <= pse_depth) {
6749 put_prev_entity(cfs_rq_of(pse), pse);
6750 pse = parent_entity(pse);
6752 if (se_depth >= pse_depth) {
6753 set_next_entity(cfs_rq_of(se), se);
6754 se = parent_entity(se);
6758 put_prev_entity(cfs_rq, pse);
6759 set_next_entity(cfs_rq, se);
6766 put_prev_task(rq, prev);
6769 se = pick_next_entity(cfs_rq, NULL);
6770 set_next_entity(cfs_rq, se);
6771 cfs_rq = group_cfs_rq(se);
6776 done: __maybe_unused;
6779 * Move the next running task to the front of
6780 * the list, so our cfs_tasks list becomes MRU
6783 list_move(&p->se.group_node, &rq->cfs_tasks);
6786 if (hrtick_enabled(rq))
6787 hrtick_start_fair(rq, p);
6789 update_misfit_status(p, rq);
6794 update_misfit_status(NULL, rq);
6795 new_tasks = idle_balance(rq, rf);
6798 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6799 * possible for any higher priority task to appear. In that case we
6800 * must re-start the pick_next_entity() loop.
6809 * rq is about to be idle, check if we need to update the
6810 * lost_idle_time of clock_pelt
6812 update_idle_rq_clock_pelt(rq);
6818 * Account for a descheduled task:
6820 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6822 struct sched_entity *se = &prev->se;
6823 struct cfs_rq *cfs_rq;
6825 for_each_sched_entity(se) {
6826 cfs_rq = cfs_rq_of(se);
6827 put_prev_entity(cfs_rq, se);
6832 * sched_yield() is very simple
6834 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6836 static void yield_task_fair(struct rq *rq)
6838 struct task_struct *curr = rq->curr;
6839 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6840 struct sched_entity *se = &curr->se;
6843 * Are we the only task in the tree?
6845 if (unlikely(rq->nr_running == 1))
6848 clear_buddies(cfs_rq, se);
6850 if (curr->policy != SCHED_BATCH) {
6851 update_rq_clock(rq);
6853 * Update run-time statistics of the 'current'.
6855 update_curr(cfs_rq);
6857 * Tell update_rq_clock() that we've just updated,
6858 * so we don't do microscopic update in schedule()
6859 * and double the fastpath cost.
6861 rq_clock_skip_update(rq);
6867 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6869 struct sched_entity *se = &p->se;
6871 /* throttled hierarchies are not runnable */
6872 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6875 /* Tell the scheduler that we'd really like pse to run next. */
6878 yield_task_fair(rq);
6884 /**************************************************
6885 * Fair scheduling class load-balancing methods.
6889 * The purpose of load-balancing is to achieve the same basic fairness the
6890 * per-CPU scheduler provides, namely provide a proportional amount of compute
6891 * time to each task. This is expressed in the following equation:
6893 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6895 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
6896 * W_i,0 is defined as:
6898 * W_i,0 = \Sum_j w_i,j (2)
6900 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
6901 * is derived from the nice value as per sched_prio_to_weight[].
6903 * The weight average is an exponential decay average of the instantaneous
6906 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6908 * C_i is the compute capacity of CPU i, typically it is the
6909 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6910 * can also include other factors [XXX].
6912 * To achieve this balance we define a measure of imbalance which follows
6913 * directly from (1):
6915 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6917 * We them move tasks around to minimize the imbalance. In the continuous
6918 * function space it is obvious this converges, in the discrete case we get
6919 * a few fun cases generally called infeasible weight scenarios.
6922 * - infeasible weights;
6923 * - local vs global optima in the discrete case. ]
6928 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6929 * for all i,j solution, we create a tree of CPUs that follows the hardware
6930 * topology where each level pairs two lower groups (or better). This results
6931 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
6932 * tree to only the first of the previous level and we decrease the frequency
6933 * of load-balance at each level inv. proportional to the number of CPUs in
6939 * \Sum { --- * --- * 2^i } = O(n) (5)
6941 * `- size of each group
6942 * | | `- number of CPUs doing load-balance
6944 * `- sum over all levels
6946 * Coupled with a limit on how many tasks we can migrate every balance pass,
6947 * this makes (5) the runtime complexity of the balancer.
6949 * An important property here is that each CPU is still (indirectly) connected
6950 * to every other CPU in at most O(log n) steps:
6952 * The adjacency matrix of the resulting graph is given by:
6955 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6958 * And you'll find that:
6960 * A^(log_2 n)_i,j != 0 for all i,j (7)
6962 * Showing there's indeed a path between every CPU in at most O(log n) steps.
6963 * The task movement gives a factor of O(m), giving a convergence complexity
6966 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6971 * In order to avoid CPUs going idle while there's still work to do, new idle
6972 * balancing is more aggressive and has the newly idle CPU iterate up the domain
6973 * tree itself instead of relying on other CPUs to bring it work.
6975 * This adds some complexity to both (5) and (8) but it reduces the total idle
6983 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6986 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6991 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6993 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
6995 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6998 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6999 * rewrite all of this once again.]
7002 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7004 enum fbq_type { regular, remote, all };
7013 #define LBF_ALL_PINNED 0x01
7014 #define LBF_NEED_BREAK 0x02
7015 #define LBF_DST_PINNED 0x04
7016 #define LBF_SOME_PINNED 0x08
7017 #define LBF_NOHZ_STATS 0x10
7018 #define LBF_NOHZ_AGAIN 0x20
7021 struct sched_domain *sd;
7029 struct cpumask *dst_grpmask;
7031 enum cpu_idle_type idle;
7033 /* The set of CPUs under consideration for load-balancing */
7034 struct cpumask *cpus;
7039 unsigned int loop_break;
7040 unsigned int loop_max;
7042 enum fbq_type fbq_type;
7043 enum group_type src_grp_type;
7044 struct list_head tasks;
7048 * Is this task likely cache-hot:
7050 static int task_hot(struct task_struct *p, struct lb_env *env)
7054 lockdep_assert_held(&env->src_rq->lock);
7056 if (p->sched_class != &fair_sched_class)
7059 if (unlikely(task_has_idle_policy(p)))
7063 * Buddy candidates are cache hot:
7065 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7066 (&p->se == cfs_rq_of(&p->se)->next ||
7067 &p->se == cfs_rq_of(&p->se)->last))
7070 if (sysctl_sched_migration_cost == -1)
7072 if (sysctl_sched_migration_cost == 0)
7075 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7077 return delta < (s64)sysctl_sched_migration_cost;
7080 #ifdef CONFIG_NUMA_BALANCING
7082 * Returns 1, if task migration degrades locality
7083 * Returns 0, if task migration improves locality i.e migration preferred.
7084 * Returns -1, if task migration is not affected by locality.
7086 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7088 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7089 unsigned long src_weight, dst_weight;
7090 int src_nid, dst_nid, dist;
7092 if (!static_branch_likely(&sched_numa_balancing))
7095 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7098 src_nid = cpu_to_node(env->src_cpu);
7099 dst_nid = cpu_to_node(env->dst_cpu);
7101 if (src_nid == dst_nid)
7104 /* Migrating away from the preferred node is always bad. */
7105 if (src_nid == p->numa_preferred_nid) {
7106 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7112 /* Encourage migration to the preferred node. */
7113 if (dst_nid == p->numa_preferred_nid)
7116 /* Leaving a core idle is often worse than degrading locality. */
7117 if (env->idle == CPU_IDLE)
7120 dist = node_distance(src_nid, dst_nid);
7122 src_weight = group_weight(p, src_nid, dist);
7123 dst_weight = group_weight(p, dst_nid, dist);
7125 src_weight = task_weight(p, src_nid, dist);
7126 dst_weight = task_weight(p, dst_nid, dist);
7129 return dst_weight < src_weight;
7133 static inline int migrate_degrades_locality(struct task_struct *p,
7141 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7144 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7148 lockdep_assert_held(&env->src_rq->lock);
7151 * We do not migrate tasks that are:
7152 * 1) throttled_lb_pair, or
7153 * 2) cannot be migrated to this CPU due to cpus_ptr, or
7154 * 3) running (obviously), or
7155 * 4) are cache-hot on their current CPU.
7157 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7160 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
7163 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7165 env->flags |= LBF_SOME_PINNED;
7168 * Remember if this task can be migrated to any other CPU in
7169 * our sched_group. We may want to revisit it if we couldn't
7170 * meet load balance goals by pulling other tasks on src_cpu.
7172 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7173 * already computed one in current iteration.
7175 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7178 /* Prevent to re-select dst_cpu via env's CPUs: */
7179 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7180 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
7181 env->flags |= LBF_DST_PINNED;
7182 env->new_dst_cpu = cpu;
7190 /* Record that we found atleast one task that could run on dst_cpu */
7191 env->flags &= ~LBF_ALL_PINNED;
7193 if (task_running(env->src_rq, p)) {
7194 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7199 * Aggressive migration if:
7200 * 1) destination numa is preferred
7201 * 2) task is cache cold, or
7202 * 3) too many balance attempts have failed.
7204 tsk_cache_hot = migrate_degrades_locality(p, env);
7205 if (tsk_cache_hot == -1)
7206 tsk_cache_hot = task_hot(p, env);
7208 if (tsk_cache_hot <= 0 ||
7209 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7210 if (tsk_cache_hot == 1) {
7211 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7212 schedstat_inc(p->se.statistics.nr_forced_migrations);
7217 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7222 * detach_task() -- detach the task for the migration specified in env
7224 static void detach_task(struct task_struct *p, struct lb_env *env)
7226 lockdep_assert_held(&env->src_rq->lock);
7228 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7229 set_task_cpu(p, env->dst_cpu);
7233 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7234 * part of active balancing operations within "domain".
7236 * Returns a task if successful and NULL otherwise.
7238 static struct task_struct *detach_one_task(struct lb_env *env)
7240 struct task_struct *p;
7242 lockdep_assert_held(&env->src_rq->lock);
7244 list_for_each_entry_reverse(p,
7245 &env->src_rq->cfs_tasks, se.group_node) {
7246 if (!can_migrate_task(p, env))
7249 detach_task(p, env);
7252 * Right now, this is only the second place where
7253 * lb_gained[env->idle] is updated (other is detach_tasks)
7254 * so we can safely collect stats here rather than
7255 * inside detach_tasks().
7257 schedstat_inc(env->sd->lb_gained[env->idle]);
7263 static const unsigned int sched_nr_migrate_break = 32;
7266 * detach_tasks() -- tries to detach up to imbalance weighted load from
7267 * busiest_rq, as part of a balancing operation within domain "sd".
7269 * Returns number of detached tasks if successful and 0 otherwise.
7271 static int detach_tasks(struct lb_env *env)
7273 struct list_head *tasks = &env->src_rq->cfs_tasks;
7274 struct task_struct *p;
7278 lockdep_assert_held(&env->src_rq->lock);
7280 if (env->imbalance <= 0)
7283 while (!list_empty(tasks)) {
7285 * We don't want to steal all, otherwise we may be treated likewise,
7286 * which could at worst lead to a livelock crash.
7288 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7291 p = list_last_entry(tasks, struct task_struct, se.group_node);
7294 /* We've more or less seen every task there is, call it quits */
7295 if (env->loop > env->loop_max)
7298 /* take a breather every nr_migrate tasks */
7299 if (env->loop > env->loop_break) {
7300 env->loop_break += sched_nr_migrate_break;
7301 env->flags |= LBF_NEED_BREAK;
7305 if (!can_migrate_task(p, env))
7308 load = task_h_load(p);
7310 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7313 if ((load / 2) > env->imbalance)
7316 detach_task(p, env);
7317 list_add(&p->se.group_node, &env->tasks);
7320 env->imbalance -= load;
7322 #ifdef CONFIG_PREEMPT
7324 * NEWIDLE balancing is a source of latency, so preemptible
7325 * kernels will stop after the first task is detached to minimize
7326 * the critical section.
7328 if (env->idle == CPU_NEWLY_IDLE)
7333 * We only want to steal up to the prescribed amount of
7336 if (env->imbalance <= 0)
7341 list_move(&p->se.group_node, tasks);
7345 * Right now, this is one of only two places we collect this stat
7346 * so we can safely collect detach_one_task() stats here rather
7347 * than inside detach_one_task().
7349 schedstat_add(env->sd->lb_gained[env->idle], detached);
7355 * attach_task() -- attach the task detached by detach_task() to its new rq.
7357 static void attach_task(struct rq *rq, struct task_struct *p)
7359 lockdep_assert_held(&rq->lock);
7361 BUG_ON(task_rq(p) != rq);
7362 activate_task(rq, p, ENQUEUE_NOCLOCK);
7363 check_preempt_curr(rq, p, 0);
7367 * attach_one_task() -- attaches the task returned from detach_one_task() to
7370 static void attach_one_task(struct rq *rq, struct task_struct *p)
7375 update_rq_clock(rq);
7381 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7384 static void attach_tasks(struct lb_env *env)
7386 struct list_head *tasks = &env->tasks;
7387 struct task_struct *p;
7390 rq_lock(env->dst_rq, &rf);
7391 update_rq_clock(env->dst_rq);
7393 while (!list_empty(tasks)) {
7394 p = list_first_entry(tasks, struct task_struct, se.group_node);
7395 list_del_init(&p->se.group_node);
7397 attach_task(env->dst_rq, p);
7400 rq_unlock(env->dst_rq, &rf);
7403 #ifdef CONFIG_NO_HZ_COMMON
7404 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7406 if (cfs_rq->avg.load_avg)
7409 if (cfs_rq->avg.util_avg)
7415 static inline bool others_have_blocked(struct rq *rq)
7417 if (READ_ONCE(rq->avg_rt.util_avg))
7420 if (READ_ONCE(rq->avg_dl.util_avg))
7423 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7424 if (READ_ONCE(rq->avg_irq.util_avg))
7431 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
7433 rq->last_blocked_load_update_tick = jiffies;
7436 rq->has_blocked_load = 0;
7439 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
7440 static inline bool others_have_blocked(struct rq *rq) { return false; }
7441 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
7444 #ifdef CONFIG_FAIR_GROUP_SCHED
7446 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7448 if (cfs_rq->load.weight)
7451 if (cfs_rq->avg.load_sum)
7454 if (cfs_rq->avg.util_sum)
7457 if (cfs_rq->avg.runnable_load_sum)
7463 static void update_blocked_averages(int cpu)
7465 struct rq *rq = cpu_rq(cpu);
7466 struct cfs_rq *cfs_rq, *pos;
7467 const struct sched_class *curr_class;
7471 rq_lock_irqsave(rq, &rf);
7472 update_rq_clock(rq);
7475 * Iterates the task_group tree in a bottom up fashion, see
7476 * list_add_leaf_cfs_rq() for details.
7478 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7479 struct sched_entity *se;
7481 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq))
7482 update_tg_load_avg(cfs_rq, 0);
7484 /* Propagate pending load changes to the parent, if any: */
7485 se = cfs_rq->tg->se[cpu];
7486 if (se && !skip_blocked_update(se))
7487 update_load_avg(cfs_rq_of(se), se, 0);
7490 * There can be a lot of idle CPU cgroups. Don't let fully
7491 * decayed cfs_rqs linger on the list.
7493 if (cfs_rq_is_decayed(cfs_rq))
7494 list_del_leaf_cfs_rq(cfs_rq);
7496 /* Don't need periodic decay once load/util_avg are null */
7497 if (cfs_rq_has_blocked(cfs_rq))
7501 curr_class = rq->curr->sched_class;
7502 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class);
7503 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class);
7504 update_irq_load_avg(rq, 0);
7505 /* Don't need periodic decay once load/util_avg are null */
7506 if (others_have_blocked(rq))
7509 update_blocked_load_status(rq, !done);
7510 rq_unlock_irqrestore(rq, &rf);
7514 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7515 * This needs to be done in a top-down fashion because the load of a child
7516 * group is a fraction of its parents load.
7518 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7520 struct rq *rq = rq_of(cfs_rq);
7521 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7522 unsigned long now = jiffies;
7525 if (cfs_rq->last_h_load_update == now)
7528 WRITE_ONCE(cfs_rq->h_load_next, NULL);
7529 for_each_sched_entity(se) {
7530 cfs_rq = cfs_rq_of(se);
7531 WRITE_ONCE(cfs_rq->h_load_next, se);
7532 if (cfs_rq->last_h_load_update == now)
7537 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7538 cfs_rq->last_h_load_update = now;
7541 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
7542 load = cfs_rq->h_load;
7543 load = div64_ul(load * se->avg.load_avg,
7544 cfs_rq_load_avg(cfs_rq) + 1);
7545 cfs_rq = group_cfs_rq(se);
7546 cfs_rq->h_load = load;
7547 cfs_rq->last_h_load_update = now;
7551 static unsigned long task_h_load(struct task_struct *p)
7553 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7555 update_cfs_rq_h_load(cfs_rq);
7556 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7557 cfs_rq_load_avg(cfs_rq) + 1);
7560 static inline void update_blocked_averages(int cpu)
7562 struct rq *rq = cpu_rq(cpu);
7563 struct cfs_rq *cfs_rq = &rq->cfs;
7564 const struct sched_class *curr_class;
7567 rq_lock_irqsave(rq, &rf);
7568 update_rq_clock(rq);
7569 update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
7571 curr_class = rq->curr->sched_class;
7572 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class);
7573 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class);
7574 update_irq_load_avg(rq, 0);
7575 update_blocked_load_status(rq, cfs_rq_has_blocked(cfs_rq) || others_have_blocked(rq));
7576 rq_unlock_irqrestore(rq, &rf);
7579 static unsigned long task_h_load(struct task_struct *p)
7581 return p->se.avg.load_avg;
7585 /********** Helpers for find_busiest_group ************************/
7588 * sg_lb_stats - stats of a sched_group required for load_balancing
7590 struct sg_lb_stats {
7591 unsigned long avg_load; /*Avg load across the CPUs of the group */
7592 unsigned long group_load; /* Total load over the CPUs of the group */
7593 unsigned long load_per_task;
7594 unsigned long group_capacity;
7595 unsigned long group_util; /* Total utilization of the group */
7596 unsigned int sum_nr_running; /* Nr tasks running in the group */
7597 unsigned int idle_cpus;
7598 unsigned int group_weight;
7599 enum group_type group_type;
7600 int group_no_capacity;
7601 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
7602 #ifdef CONFIG_NUMA_BALANCING
7603 unsigned int nr_numa_running;
7604 unsigned int nr_preferred_running;
7609 * sd_lb_stats - Structure to store the statistics of a sched_domain
7610 * during load balancing.
7612 struct sd_lb_stats {
7613 struct sched_group *busiest; /* Busiest group in this sd */
7614 struct sched_group *local; /* Local group in this sd */
7615 unsigned long total_running;
7616 unsigned long total_load; /* Total load of all groups in sd */
7617 unsigned long total_capacity; /* Total capacity of all groups in sd */
7618 unsigned long avg_load; /* Average load across all groups in sd */
7620 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7621 struct sg_lb_stats local_stat; /* Statistics of the local group */
7624 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7627 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7628 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7629 * We must however clear busiest_stat::avg_load because
7630 * update_sd_pick_busiest() reads this before assignment.
7632 *sds = (struct sd_lb_stats){
7635 .total_running = 0UL,
7637 .total_capacity = 0UL,
7640 .sum_nr_running = 0,
7641 .group_type = group_other,
7646 static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu)
7648 struct rq *rq = cpu_rq(cpu);
7649 unsigned long max = arch_scale_cpu_capacity(cpu);
7650 unsigned long used, free;
7653 irq = cpu_util_irq(rq);
7655 if (unlikely(irq >= max))
7658 used = READ_ONCE(rq->avg_rt.util_avg);
7659 used += READ_ONCE(rq->avg_dl.util_avg);
7661 if (unlikely(used >= max))
7666 return scale_irq_capacity(free, irq, max);
7669 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7671 unsigned long capacity = scale_rt_capacity(sd, cpu);
7672 struct sched_group *sdg = sd->groups;
7674 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
7679 cpu_rq(cpu)->cpu_capacity = capacity;
7680 sdg->sgc->capacity = capacity;
7681 sdg->sgc->min_capacity = capacity;
7682 sdg->sgc->max_capacity = capacity;
7685 void update_group_capacity(struct sched_domain *sd, int cpu)
7687 struct sched_domain *child = sd->child;
7688 struct sched_group *group, *sdg = sd->groups;
7689 unsigned long capacity, min_capacity, max_capacity;
7690 unsigned long interval;
7692 interval = msecs_to_jiffies(sd->balance_interval);
7693 interval = clamp(interval, 1UL, max_load_balance_interval);
7694 sdg->sgc->next_update = jiffies + interval;
7697 update_cpu_capacity(sd, cpu);
7702 min_capacity = ULONG_MAX;
7705 if (child->flags & SD_OVERLAP) {
7707 * SD_OVERLAP domains cannot assume that child groups
7708 * span the current group.
7711 for_each_cpu(cpu, sched_group_span(sdg)) {
7712 struct sched_group_capacity *sgc;
7713 struct rq *rq = cpu_rq(cpu);
7716 * build_sched_domains() -> init_sched_groups_capacity()
7717 * gets here before we've attached the domains to the
7720 * Use capacity_of(), which is set irrespective of domains
7721 * in update_cpu_capacity().
7723 * This avoids capacity from being 0 and
7724 * causing divide-by-zero issues on boot.
7726 if (unlikely(!rq->sd)) {
7727 capacity += capacity_of(cpu);
7729 sgc = rq->sd->groups->sgc;
7730 capacity += sgc->capacity;
7733 min_capacity = min(capacity, min_capacity);
7734 max_capacity = max(capacity, max_capacity);
7738 * !SD_OVERLAP domains can assume that child groups
7739 * span the current group.
7742 group = child->groups;
7744 struct sched_group_capacity *sgc = group->sgc;
7746 capacity += sgc->capacity;
7747 min_capacity = min(sgc->min_capacity, min_capacity);
7748 max_capacity = max(sgc->max_capacity, max_capacity);
7749 group = group->next;
7750 } while (group != child->groups);
7753 sdg->sgc->capacity = capacity;
7754 sdg->sgc->min_capacity = min_capacity;
7755 sdg->sgc->max_capacity = max_capacity;
7759 * Check whether the capacity of the rq has been noticeably reduced by side
7760 * activity. The imbalance_pct is used for the threshold.
7761 * Return true is the capacity is reduced
7764 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7766 return ((rq->cpu_capacity * sd->imbalance_pct) <
7767 (rq->cpu_capacity_orig * 100));
7771 * Check whether a rq has a misfit task and if it looks like we can actually
7772 * help that task: we can migrate the task to a CPU of higher capacity, or
7773 * the task's current CPU is heavily pressured.
7775 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
7777 return rq->misfit_task_load &&
7778 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
7779 check_cpu_capacity(rq, sd));
7783 * Group imbalance indicates (and tries to solve) the problem where balancing
7784 * groups is inadequate due to ->cpus_ptr constraints.
7786 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
7787 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
7790 * { 0 1 2 3 } { 4 5 6 7 }
7793 * If we were to balance group-wise we'd place two tasks in the first group and
7794 * two tasks in the second group. Clearly this is undesired as it will overload
7795 * cpu 3 and leave one of the CPUs in the second group unused.
7797 * The current solution to this issue is detecting the skew in the first group
7798 * by noticing the lower domain failed to reach balance and had difficulty
7799 * moving tasks due to affinity constraints.
7801 * When this is so detected; this group becomes a candidate for busiest; see
7802 * update_sd_pick_busiest(). And calculate_imbalance() and
7803 * find_busiest_group() avoid some of the usual balance conditions to allow it
7804 * to create an effective group imbalance.
7806 * This is a somewhat tricky proposition since the next run might not find the
7807 * group imbalance and decide the groups need to be balanced again. A most
7808 * subtle and fragile situation.
7811 static inline int sg_imbalanced(struct sched_group *group)
7813 return group->sgc->imbalance;
7817 * group_has_capacity returns true if the group has spare capacity that could
7818 * be used by some tasks.
7819 * We consider that a group has spare capacity if the * number of task is
7820 * smaller than the number of CPUs or if the utilization is lower than the
7821 * available capacity for CFS tasks.
7822 * For the latter, we use a threshold to stabilize the state, to take into
7823 * account the variance of the tasks' load and to return true if the available
7824 * capacity in meaningful for the load balancer.
7825 * As an example, an available capacity of 1% can appear but it doesn't make
7826 * any benefit for the load balance.
7829 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7831 if (sgs->sum_nr_running < sgs->group_weight)
7834 if ((sgs->group_capacity * 100) >
7835 (sgs->group_util * env->sd->imbalance_pct))
7842 * group_is_overloaded returns true if the group has more tasks than it can
7844 * group_is_overloaded is not equals to !group_has_capacity because a group
7845 * with the exact right number of tasks, has no more spare capacity but is not
7846 * overloaded so both group_has_capacity and group_is_overloaded return
7850 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7852 if (sgs->sum_nr_running <= sgs->group_weight)
7855 if ((sgs->group_capacity * 100) <
7856 (sgs->group_util * env->sd->imbalance_pct))
7863 * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller
7864 * per-CPU capacity than sched_group ref.
7867 group_smaller_min_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7869 return sg->sgc->min_capacity * capacity_margin <
7870 ref->sgc->min_capacity * 1024;
7874 * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller
7875 * per-CPU capacity_orig than sched_group ref.
7878 group_smaller_max_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7880 return sg->sgc->max_capacity * capacity_margin <
7881 ref->sgc->max_capacity * 1024;
7885 group_type group_classify(struct sched_group *group,
7886 struct sg_lb_stats *sgs)
7888 if (sgs->group_no_capacity)
7889 return group_overloaded;
7891 if (sg_imbalanced(group))
7892 return group_imbalanced;
7894 if (sgs->group_misfit_task_load)
7895 return group_misfit_task;
7900 static bool update_nohz_stats(struct rq *rq, bool force)
7902 #ifdef CONFIG_NO_HZ_COMMON
7903 unsigned int cpu = rq->cpu;
7905 if (!rq->has_blocked_load)
7908 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
7911 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
7914 update_blocked_averages(cpu);
7916 return rq->has_blocked_load;
7923 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7924 * @env: The load balancing environment.
7925 * @group: sched_group whose statistics are to be updated.
7926 * @sgs: variable to hold the statistics for this group.
7927 * @sg_status: Holds flag indicating the status of the sched_group
7929 static inline void update_sg_lb_stats(struct lb_env *env,
7930 struct sched_group *group,
7931 struct sg_lb_stats *sgs,
7936 memset(sgs, 0, sizeof(*sgs));
7938 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
7939 struct rq *rq = cpu_rq(i);
7941 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
7942 env->flags |= LBF_NOHZ_AGAIN;
7944 sgs->group_load += weighted_cpuload(rq);
7945 sgs->group_util += cpu_util(i);
7946 sgs->sum_nr_running += rq->cfs.h_nr_running;
7948 nr_running = rq->nr_running;
7950 *sg_status |= SG_OVERLOAD;
7952 if (cpu_overutilized(i))
7953 *sg_status |= SG_OVERUTILIZED;
7955 #ifdef CONFIG_NUMA_BALANCING
7956 sgs->nr_numa_running += rq->nr_numa_running;
7957 sgs->nr_preferred_running += rq->nr_preferred_running;
7960 * No need to call idle_cpu() if nr_running is not 0
7962 if (!nr_running && idle_cpu(i))
7965 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
7966 sgs->group_misfit_task_load < rq->misfit_task_load) {
7967 sgs->group_misfit_task_load = rq->misfit_task_load;
7968 *sg_status |= SG_OVERLOAD;
7972 /* Adjust by relative CPU capacity of the group */
7973 sgs->group_capacity = group->sgc->capacity;
7974 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7976 if (sgs->sum_nr_running)
7977 sgs->load_per_task = sgs->group_load / sgs->sum_nr_running;
7979 sgs->group_weight = group->group_weight;
7981 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7982 sgs->group_type = group_classify(group, sgs);
7986 * update_sd_pick_busiest - return 1 on busiest group
7987 * @env: The load balancing environment.
7988 * @sds: sched_domain statistics
7989 * @sg: sched_group candidate to be checked for being the busiest
7990 * @sgs: sched_group statistics
7992 * Determine if @sg is a busier group than the previously selected
7995 * Return: %true if @sg is a busier group than the previously selected
7996 * busiest group. %false otherwise.
7998 static bool update_sd_pick_busiest(struct lb_env *env,
7999 struct sd_lb_stats *sds,
8000 struct sched_group *sg,
8001 struct sg_lb_stats *sgs)
8003 struct sg_lb_stats *busiest = &sds->busiest_stat;
8006 * Don't try to pull misfit tasks we can't help.
8007 * We can use max_capacity here as reduction in capacity on some
8008 * CPUs in the group should either be possible to resolve
8009 * internally or be covered by avg_load imbalance (eventually).
8011 if (sgs->group_type == group_misfit_task &&
8012 (!group_smaller_max_cpu_capacity(sg, sds->local) ||
8013 !group_has_capacity(env, &sds->local_stat)))
8016 if (sgs->group_type > busiest->group_type)
8019 if (sgs->group_type < busiest->group_type)
8022 if (sgs->avg_load <= busiest->avg_load)
8025 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
8029 * Candidate sg has no more than one task per CPU and
8030 * has higher per-CPU capacity. Migrating tasks to less
8031 * capable CPUs may harm throughput. Maximize throughput,
8032 * power/energy consequences are not considered.
8034 if (sgs->sum_nr_running <= sgs->group_weight &&
8035 group_smaller_min_cpu_capacity(sds->local, sg))
8039 * If we have more than one misfit sg go with the biggest misfit.
8041 if (sgs->group_type == group_misfit_task &&
8042 sgs->group_misfit_task_load < busiest->group_misfit_task_load)
8046 /* This is the busiest node in its class. */
8047 if (!(env->sd->flags & SD_ASYM_PACKING))
8050 /* No ASYM_PACKING if target CPU is already busy */
8051 if (env->idle == CPU_NOT_IDLE)
8054 * ASYM_PACKING needs to move all the work to the highest
8055 * prority CPUs in the group, therefore mark all groups
8056 * of lower priority than ourself as busy.
8058 if (sgs->sum_nr_running &&
8059 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
8063 /* Prefer to move from lowest priority CPU's work */
8064 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
8065 sg->asym_prefer_cpu))
8072 #ifdef CONFIG_NUMA_BALANCING
8073 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8075 if (sgs->sum_nr_running > sgs->nr_numa_running)
8077 if (sgs->sum_nr_running > sgs->nr_preferred_running)
8082 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8084 if (rq->nr_running > rq->nr_numa_running)
8086 if (rq->nr_running > rq->nr_preferred_running)
8091 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8096 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8100 #endif /* CONFIG_NUMA_BALANCING */
8103 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8104 * @env: The load balancing environment.
8105 * @sds: variable to hold the statistics for this sched_domain.
8107 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8109 struct sched_domain *child = env->sd->child;
8110 struct sched_group *sg = env->sd->groups;
8111 struct sg_lb_stats *local = &sds->local_stat;
8112 struct sg_lb_stats tmp_sgs;
8113 bool prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
8116 #ifdef CONFIG_NO_HZ_COMMON
8117 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
8118 env->flags |= LBF_NOHZ_STATS;
8122 struct sg_lb_stats *sgs = &tmp_sgs;
8125 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
8130 if (env->idle != CPU_NEWLY_IDLE ||
8131 time_after_eq(jiffies, sg->sgc->next_update))
8132 update_group_capacity(env->sd, env->dst_cpu);
8135 update_sg_lb_stats(env, sg, sgs, &sg_status);
8141 * In case the child domain prefers tasks go to siblings
8142 * first, lower the sg capacity so that we'll try
8143 * and move all the excess tasks away. We lower the capacity
8144 * of a group only if the local group has the capacity to fit
8145 * these excess tasks. The extra check prevents the case where
8146 * you always pull from the heaviest group when it is already
8147 * under-utilized (possible with a large weight task outweighs
8148 * the tasks on the system).
8150 if (prefer_sibling && sds->local &&
8151 group_has_capacity(env, local) &&
8152 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
8153 sgs->group_no_capacity = 1;
8154 sgs->group_type = group_classify(sg, sgs);
8157 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8159 sds->busiest_stat = *sgs;
8163 /* Now, start updating sd_lb_stats */
8164 sds->total_running += sgs->sum_nr_running;
8165 sds->total_load += sgs->group_load;
8166 sds->total_capacity += sgs->group_capacity;
8169 } while (sg != env->sd->groups);
8171 #ifdef CONFIG_NO_HZ_COMMON
8172 if ((env->flags & LBF_NOHZ_AGAIN) &&
8173 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8175 WRITE_ONCE(nohz.next_blocked,
8176 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8180 if (env->sd->flags & SD_NUMA)
8181 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8183 if (!env->sd->parent) {
8184 struct root_domain *rd = env->dst_rq->rd;
8186 /* update overload indicator if we are at root domain */
8187 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
8189 /* Update over-utilization (tipping point, U >= 0) indicator */
8190 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
8191 } else if (sg_status & SG_OVERUTILIZED) {
8192 WRITE_ONCE(env->dst_rq->rd->overutilized, SG_OVERUTILIZED);
8197 * check_asym_packing - Check to see if the group is packed into the
8200 * This is primarily intended to used at the sibling level. Some
8201 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8202 * case of POWER7, it can move to lower SMT modes only when higher
8203 * threads are idle. When in lower SMT modes, the threads will
8204 * perform better since they share less core resources. Hence when we
8205 * have idle threads, we want them to be the higher ones.
8207 * This packing function is run on idle threads. It checks to see if
8208 * the busiest CPU in this domain (core in the P7 case) has a higher
8209 * CPU number than the packing function is being run on. Here we are
8210 * assuming lower CPU number will be equivalent to lower a SMT thread
8213 * Return: 1 when packing is required and a task should be moved to
8214 * this CPU. The amount of the imbalance is returned in env->imbalance.
8216 * @env: The load balancing environment.
8217 * @sds: Statistics of the sched_domain which is to be packed
8219 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8223 if (!(env->sd->flags & SD_ASYM_PACKING))
8226 if (env->idle == CPU_NOT_IDLE)
8232 busiest_cpu = sds->busiest->asym_prefer_cpu;
8233 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
8236 env->imbalance = sds->busiest_stat.group_load;
8242 * fix_small_imbalance - Calculate the minor imbalance that exists
8243 * amongst the groups of a sched_domain, during
8245 * @env: The load balancing environment.
8246 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8249 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8251 unsigned long tmp, capa_now = 0, capa_move = 0;
8252 unsigned int imbn = 2;
8253 unsigned long scaled_busy_load_per_task;
8254 struct sg_lb_stats *local, *busiest;
8256 local = &sds->local_stat;
8257 busiest = &sds->busiest_stat;
8259 if (!local->sum_nr_running)
8260 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8261 else if (busiest->load_per_task > local->load_per_task)
8264 scaled_busy_load_per_task =
8265 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8266 busiest->group_capacity;
8268 if (busiest->avg_load + scaled_busy_load_per_task >=
8269 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8270 env->imbalance = busiest->load_per_task;
8275 * OK, we don't have enough imbalance to justify moving tasks,
8276 * however we may be able to increase total CPU capacity used by
8280 capa_now += busiest->group_capacity *
8281 min(busiest->load_per_task, busiest->avg_load);
8282 capa_now += local->group_capacity *
8283 min(local->load_per_task, local->avg_load);
8284 capa_now /= SCHED_CAPACITY_SCALE;
8286 /* Amount of load we'd subtract */
8287 if (busiest->avg_load > scaled_busy_load_per_task) {
8288 capa_move += busiest->group_capacity *
8289 min(busiest->load_per_task,
8290 busiest->avg_load - scaled_busy_load_per_task);
8293 /* Amount of load we'd add */
8294 if (busiest->avg_load * busiest->group_capacity <
8295 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8296 tmp = (busiest->avg_load * busiest->group_capacity) /
8297 local->group_capacity;
8299 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8300 local->group_capacity;
8302 capa_move += local->group_capacity *
8303 min(local->load_per_task, local->avg_load + tmp);
8304 capa_move /= SCHED_CAPACITY_SCALE;
8306 /* Move if we gain throughput */
8307 if (capa_move > capa_now)
8308 env->imbalance = busiest->load_per_task;
8312 * calculate_imbalance - Calculate the amount of imbalance present within the
8313 * groups of a given sched_domain during load balance.
8314 * @env: load balance environment
8315 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8317 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8319 unsigned long max_pull, load_above_capacity = ~0UL;
8320 struct sg_lb_stats *local, *busiest;
8322 local = &sds->local_stat;
8323 busiest = &sds->busiest_stat;
8325 if (busiest->group_type == group_imbalanced) {
8327 * In the group_imb case we cannot rely on group-wide averages
8328 * to ensure CPU-load equilibrium, look at wider averages. XXX
8330 busiest->load_per_task =
8331 min(busiest->load_per_task, sds->avg_load);
8335 * Avg load of busiest sg can be less and avg load of local sg can
8336 * be greater than avg load across all sgs of sd because avg load
8337 * factors in sg capacity and sgs with smaller group_type are
8338 * skipped when updating the busiest sg:
8340 if (busiest->group_type != group_misfit_task &&
8341 (busiest->avg_load <= sds->avg_load ||
8342 local->avg_load >= sds->avg_load)) {
8344 return fix_small_imbalance(env, sds);
8348 * If there aren't any idle CPUs, avoid creating some.
8350 if (busiest->group_type == group_overloaded &&
8351 local->group_type == group_overloaded) {
8352 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
8353 if (load_above_capacity > busiest->group_capacity) {
8354 load_above_capacity -= busiest->group_capacity;
8355 load_above_capacity *= scale_load_down(NICE_0_LOAD);
8356 load_above_capacity /= busiest->group_capacity;
8358 load_above_capacity = ~0UL;
8362 * We're trying to get all the CPUs to the average_load, so we don't
8363 * want to push ourselves above the average load, nor do we wish to
8364 * reduce the max loaded CPU below the average load. At the same time,
8365 * we also don't want to reduce the group load below the group
8366 * capacity. Thus we look for the minimum possible imbalance.
8368 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8370 /* How much load to actually move to equalise the imbalance */
8371 env->imbalance = min(
8372 max_pull * busiest->group_capacity,
8373 (sds->avg_load - local->avg_load) * local->group_capacity
8374 ) / SCHED_CAPACITY_SCALE;
8376 /* Boost imbalance to allow misfit task to be balanced. */
8377 if (busiest->group_type == group_misfit_task) {
8378 env->imbalance = max_t(long, env->imbalance,
8379 busiest->group_misfit_task_load);
8383 * if *imbalance is less than the average load per runnable task
8384 * there is no guarantee that any tasks will be moved so we'll have
8385 * a think about bumping its value to force at least one task to be
8388 if (env->imbalance < busiest->load_per_task)
8389 return fix_small_imbalance(env, sds);
8392 /******* find_busiest_group() helpers end here *********************/
8395 * find_busiest_group - Returns the busiest group within the sched_domain
8396 * if there is an imbalance.
8398 * Also calculates the amount of weighted load which should be moved
8399 * to restore balance.
8401 * @env: The load balancing environment.
8403 * Return: - The busiest group if imbalance exists.
8405 static struct sched_group *find_busiest_group(struct lb_env *env)
8407 struct sg_lb_stats *local, *busiest;
8408 struct sd_lb_stats sds;
8410 init_sd_lb_stats(&sds);
8413 * Compute the various statistics relavent for load balancing at
8416 update_sd_lb_stats(env, &sds);
8418 if (sched_energy_enabled()) {
8419 struct root_domain *rd = env->dst_rq->rd;
8421 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
8425 local = &sds.local_stat;
8426 busiest = &sds.busiest_stat;
8428 /* ASYM feature bypasses nice load balance check */
8429 if (check_asym_packing(env, &sds))
8432 /* There is no busy sibling group to pull tasks from */
8433 if (!sds.busiest || busiest->sum_nr_running == 0)
8436 /* XXX broken for overlapping NUMA groups */
8437 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8438 / sds.total_capacity;
8441 * If the busiest group is imbalanced the below checks don't
8442 * work because they assume all things are equal, which typically
8443 * isn't true due to cpus_ptr constraints and the like.
8445 if (busiest->group_type == group_imbalanced)
8449 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8450 * capacities from resulting in underutilization due to avg_load.
8452 if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8453 busiest->group_no_capacity)
8456 /* Misfit tasks should be dealt with regardless of the avg load */
8457 if (busiest->group_type == group_misfit_task)
8461 * If the local group is busier than the selected busiest group
8462 * don't try and pull any tasks.
8464 if (local->avg_load >= busiest->avg_load)
8468 * Don't pull any tasks if this group is already above the domain
8471 if (local->avg_load >= sds.avg_load)
8474 if (env->idle == CPU_IDLE) {
8476 * This CPU is idle. If the busiest group is not overloaded
8477 * and there is no imbalance between this and busiest group
8478 * wrt idle CPUs, it is balanced. The imbalance becomes
8479 * significant if the diff is greater than 1 otherwise we
8480 * might end up to just move the imbalance on another group
8482 if ((busiest->group_type != group_overloaded) &&
8483 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8487 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8488 * imbalance_pct to be conservative.
8490 if (100 * busiest->avg_load <=
8491 env->sd->imbalance_pct * local->avg_load)
8496 /* Looks like there is an imbalance. Compute it */
8497 env->src_grp_type = busiest->group_type;
8498 calculate_imbalance(env, &sds);
8499 return env->imbalance ? sds.busiest : NULL;
8507 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
8509 static struct rq *find_busiest_queue(struct lb_env *env,
8510 struct sched_group *group)
8512 struct rq *busiest = NULL, *rq;
8513 unsigned long busiest_load = 0, busiest_capacity = 1;
8516 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8517 unsigned long capacity, wl;
8521 rt = fbq_classify_rq(rq);
8524 * We classify groups/runqueues into three groups:
8525 * - regular: there are !numa tasks
8526 * - remote: there are numa tasks that run on the 'wrong' node
8527 * - all: there is no distinction
8529 * In order to avoid migrating ideally placed numa tasks,
8530 * ignore those when there's better options.
8532 * If we ignore the actual busiest queue to migrate another
8533 * task, the next balance pass can still reduce the busiest
8534 * queue by moving tasks around inside the node.
8536 * If we cannot move enough load due to this classification
8537 * the next pass will adjust the group classification and
8538 * allow migration of more tasks.
8540 * Both cases only affect the total convergence complexity.
8542 if (rt > env->fbq_type)
8546 * For ASYM_CPUCAPACITY domains with misfit tasks we simply
8547 * seek the "biggest" misfit task.
8549 if (env->src_grp_type == group_misfit_task) {
8550 if (rq->misfit_task_load > busiest_load) {
8551 busiest_load = rq->misfit_task_load;
8558 capacity = capacity_of(i);
8561 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
8562 * eventually lead to active_balancing high->low capacity.
8563 * Higher per-CPU capacity is considered better than balancing
8566 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8567 capacity_of(env->dst_cpu) < capacity &&
8568 rq->nr_running == 1)
8571 wl = weighted_cpuload(rq);
8574 * When comparing with imbalance, use weighted_cpuload()
8575 * which is not scaled with the CPU capacity.
8578 if (rq->nr_running == 1 && wl > env->imbalance &&
8579 !check_cpu_capacity(rq, env->sd))
8583 * For the load comparisons with the other CPU's, consider
8584 * the weighted_cpuload() scaled with the CPU capacity, so
8585 * that the load can be moved away from the CPU that is
8586 * potentially running at a lower capacity.
8588 * Thus we're looking for max(wl_i / capacity_i), crosswise
8589 * multiplication to rid ourselves of the division works out
8590 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8591 * our previous maximum.
8593 if (wl * busiest_capacity > busiest_load * capacity) {
8595 busiest_capacity = capacity;
8604 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8605 * so long as it is large enough.
8607 #define MAX_PINNED_INTERVAL 512
8610 asym_active_balance(struct lb_env *env)
8613 * ASYM_PACKING needs to force migrate tasks from busy but
8614 * lower priority CPUs in order to pack all tasks in the
8615 * highest priority CPUs.
8617 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
8618 sched_asym_prefer(env->dst_cpu, env->src_cpu);
8622 voluntary_active_balance(struct lb_env *env)
8624 struct sched_domain *sd = env->sd;
8626 if (asym_active_balance(env))
8630 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8631 * It's worth migrating the task if the src_cpu's capacity is reduced
8632 * because of other sched_class or IRQs if more capacity stays
8633 * available on dst_cpu.
8635 if ((env->idle != CPU_NOT_IDLE) &&
8636 (env->src_rq->cfs.h_nr_running == 1)) {
8637 if ((check_cpu_capacity(env->src_rq, sd)) &&
8638 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8642 if (env->src_grp_type == group_misfit_task)
8648 static int need_active_balance(struct lb_env *env)
8650 struct sched_domain *sd = env->sd;
8652 if (voluntary_active_balance(env))
8655 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8658 static int active_load_balance_cpu_stop(void *data);
8660 static int should_we_balance(struct lb_env *env)
8662 struct sched_group *sg = env->sd->groups;
8663 int cpu, balance_cpu = -1;
8666 * Ensure the balancing environment is consistent; can happen
8667 * when the softirq triggers 'during' hotplug.
8669 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8673 * In the newly idle case, we will allow all the CPUs
8674 * to do the newly idle load balance.
8676 if (env->idle == CPU_NEWLY_IDLE)
8679 /* Try to find first idle CPU */
8680 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8688 if (balance_cpu == -1)
8689 balance_cpu = group_balance_cpu(sg);
8692 * First idle CPU or the first CPU(busiest) in this sched group
8693 * is eligible for doing load balancing at this and above domains.
8695 return balance_cpu == env->dst_cpu;
8699 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8700 * tasks if there is an imbalance.
8702 static int load_balance(int this_cpu, struct rq *this_rq,
8703 struct sched_domain *sd, enum cpu_idle_type idle,
8704 int *continue_balancing)
8706 int ld_moved, cur_ld_moved, active_balance = 0;
8707 struct sched_domain *sd_parent = sd->parent;
8708 struct sched_group *group;
8711 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8713 struct lb_env env = {
8715 .dst_cpu = this_cpu,
8717 .dst_grpmask = sched_group_span(sd->groups),
8719 .loop_break = sched_nr_migrate_break,
8722 .tasks = LIST_HEAD_INIT(env.tasks),
8725 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8727 schedstat_inc(sd->lb_count[idle]);
8730 if (!should_we_balance(&env)) {
8731 *continue_balancing = 0;
8735 group = find_busiest_group(&env);
8737 schedstat_inc(sd->lb_nobusyg[idle]);
8741 busiest = find_busiest_queue(&env, group);
8743 schedstat_inc(sd->lb_nobusyq[idle]);
8747 BUG_ON(busiest == env.dst_rq);
8749 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8751 env.src_cpu = busiest->cpu;
8752 env.src_rq = busiest;
8755 if (busiest->nr_running > 1) {
8757 * Attempt to move tasks. If find_busiest_group has found
8758 * an imbalance but busiest->nr_running <= 1, the group is
8759 * still unbalanced. ld_moved simply stays zero, so it is
8760 * correctly treated as an imbalance.
8762 env.flags |= LBF_ALL_PINNED;
8763 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8766 rq_lock_irqsave(busiest, &rf);
8767 update_rq_clock(busiest);
8770 * cur_ld_moved - load moved in current iteration
8771 * ld_moved - cumulative load moved across iterations
8773 cur_ld_moved = detach_tasks(&env);
8776 * We've detached some tasks from busiest_rq. Every
8777 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8778 * unlock busiest->lock, and we are able to be sure
8779 * that nobody can manipulate the tasks in parallel.
8780 * See task_rq_lock() family for the details.
8783 rq_unlock(busiest, &rf);
8787 ld_moved += cur_ld_moved;
8790 local_irq_restore(rf.flags);
8792 if (env.flags & LBF_NEED_BREAK) {
8793 env.flags &= ~LBF_NEED_BREAK;
8798 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8799 * us and move them to an alternate dst_cpu in our sched_group
8800 * where they can run. The upper limit on how many times we
8801 * iterate on same src_cpu is dependent on number of CPUs in our
8804 * This changes load balance semantics a bit on who can move
8805 * load to a given_cpu. In addition to the given_cpu itself
8806 * (or a ilb_cpu acting on its behalf where given_cpu is
8807 * nohz-idle), we now have balance_cpu in a position to move
8808 * load to given_cpu. In rare situations, this may cause
8809 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8810 * _independently_ and at _same_ time to move some load to
8811 * given_cpu) causing exceess load to be moved to given_cpu.
8812 * This however should not happen so much in practice and
8813 * moreover subsequent load balance cycles should correct the
8814 * excess load moved.
8816 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8818 /* Prevent to re-select dst_cpu via env's CPUs */
8819 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
8821 env.dst_rq = cpu_rq(env.new_dst_cpu);
8822 env.dst_cpu = env.new_dst_cpu;
8823 env.flags &= ~LBF_DST_PINNED;
8825 env.loop_break = sched_nr_migrate_break;
8828 * Go back to "more_balance" rather than "redo" since we
8829 * need to continue with same src_cpu.
8835 * We failed to reach balance because of affinity.
8838 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8840 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8841 *group_imbalance = 1;
8844 /* All tasks on this runqueue were pinned by CPU affinity */
8845 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8846 __cpumask_clear_cpu(cpu_of(busiest), cpus);
8848 * Attempting to continue load balancing at the current
8849 * sched_domain level only makes sense if there are
8850 * active CPUs remaining as possible busiest CPUs to
8851 * pull load from which are not contained within the
8852 * destination group that is receiving any migrated
8855 if (!cpumask_subset(cpus, env.dst_grpmask)) {
8857 env.loop_break = sched_nr_migrate_break;
8860 goto out_all_pinned;
8865 schedstat_inc(sd->lb_failed[idle]);
8867 * Increment the failure counter only on periodic balance.
8868 * We do not want newidle balance, which can be very
8869 * frequent, pollute the failure counter causing
8870 * excessive cache_hot migrations and active balances.
8872 if (idle != CPU_NEWLY_IDLE)
8873 sd->nr_balance_failed++;
8875 if (need_active_balance(&env)) {
8876 unsigned long flags;
8878 raw_spin_lock_irqsave(&busiest->lock, flags);
8881 * Don't kick the active_load_balance_cpu_stop,
8882 * if the curr task on busiest CPU can't be
8883 * moved to this_cpu:
8885 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
8886 raw_spin_unlock_irqrestore(&busiest->lock,
8888 env.flags |= LBF_ALL_PINNED;
8889 goto out_one_pinned;
8893 * ->active_balance synchronizes accesses to
8894 * ->active_balance_work. Once set, it's cleared
8895 * only after active load balance is finished.
8897 if (!busiest->active_balance) {
8898 busiest->active_balance = 1;
8899 busiest->push_cpu = this_cpu;
8902 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8904 if (active_balance) {
8905 stop_one_cpu_nowait(cpu_of(busiest),
8906 active_load_balance_cpu_stop, busiest,
8907 &busiest->active_balance_work);
8910 /* We've kicked active balancing, force task migration. */
8911 sd->nr_balance_failed = sd->cache_nice_tries+1;
8914 sd->nr_balance_failed = 0;
8916 if (likely(!active_balance) || voluntary_active_balance(&env)) {
8917 /* We were unbalanced, so reset the balancing interval */
8918 sd->balance_interval = sd->min_interval;
8921 * If we've begun active balancing, start to back off. This
8922 * case may not be covered by the all_pinned logic if there
8923 * is only 1 task on the busy runqueue (because we don't call
8926 if (sd->balance_interval < sd->max_interval)
8927 sd->balance_interval *= 2;
8934 * We reach balance although we may have faced some affinity
8935 * constraints. Clear the imbalance flag if it was set.
8938 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8940 if (*group_imbalance)
8941 *group_imbalance = 0;
8946 * We reach balance because all tasks are pinned at this level so
8947 * we can't migrate them. Let the imbalance flag set so parent level
8948 * can try to migrate them.
8950 schedstat_inc(sd->lb_balanced[idle]);
8952 sd->nr_balance_failed = 0;
8958 * idle_balance() disregards balance intervals, so we could repeatedly
8959 * reach this code, which would lead to balance_interval skyrocketting
8960 * in a short amount of time. Skip the balance_interval increase logic
8963 if (env.idle == CPU_NEWLY_IDLE)
8966 /* tune up the balancing interval */
8967 if ((env.flags & LBF_ALL_PINNED &&
8968 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8969 sd->balance_interval < sd->max_interval)
8970 sd->balance_interval *= 2;
8975 static inline unsigned long
8976 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8978 unsigned long interval = sd->balance_interval;
8981 interval *= sd->busy_factor;
8983 /* scale ms to jiffies */
8984 interval = msecs_to_jiffies(interval);
8985 interval = clamp(interval, 1UL, max_load_balance_interval);
8991 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
8993 unsigned long interval, next;
8995 /* used by idle balance, so cpu_busy = 0 */
8996 interval = get_sd_balance_interval(sd, 0);
8997 next = sd->last_balance + interval;
8999 if (time_after(*next_balance, next))
9000 *next_balance = next;
9004 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9005 * running tasks off the busiest CPU onto idle CPUs. It requires at
9006 * least 1 task to be running on each physical CPU where possible, and
9007 * avoids physical / logical imbalances.
9009 static int active_load_balance_cpu_stop(void *data)
9011 struct rq *busiest_rq = data;
9012 int busiest_cpu = cpu_of(busiest_rq);
9013 int target_cpu = busiest_rq->push_cpu;
9014 struct rq *target_rq = cpu_rq(target_cpu);
9015 struct sched_domain *sd;
9016 struct task_struct *p = NULL;
9019 rq_lock_irq(busiest_rq, &rf);
9021 * Between queueing the stop-work and running it is a hole in which
9022 * CPUs can become inactive. We should not move tasks from or to
9025 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
9028 /* Make sure the requested CPU hasn't gone down in the meantime: */
9029 if (unlikely(busiest_cpu != smp_processor_id() ||
9030 !busiest_rq->active_balance))
9033 /* Is there any task to move? */
9034 if (busiest_rq->nr_running <= 1)
9038 * This condition is "impossible", if it occurs
9039 * we need to fix it. Originally reported by
9040 * Bjorn Helgaas on a 128-CPU setup.
9042 BUG_ON(busiest_rq == target_rq);
9044 /* Search for an sd spanning us and the target CPU. */
9046 for_each_domain(target_cpu, sd) {
9047 if ((sd->flags & SD_LOAD_BALANCE) &&
9048 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9053 struct lb_env env = {
9055 .dst_cpu = target_cpu,
9056 .dst_rq = target_rq,
9057 .src_cpu = busiest_rq->cpu,
9058 .src_rq = busiest_rq,
9061 * can_migrate_task() doesn't need to compute new_dst_cpu
9062 * for active balancing. Since we have CPU_IDLE, but no
9063 * @dst_grpmask we need to make that test go away with lying
9066 .flags = LBF_DST_PINNED,
9069 schedstat_inc(sd->alb_count);
9070 update_rq_clock(busiest_rq);
9072 p = detach_one_task(&env);
9074 schedstat_inc(sd->alb_pushed);
9075 /* Active balancing done, reset the failure counter. */
9076 sd->nr_balance_failed = 0;
9078 schedstat_inc(sd->alb_failed);
9083 busiest_rq->active_balance = 0;
9084 rq_unlock(busiest_rq, &rf);
9087 attach_one_task(target_rq, p);
9094 static DEFINE_SPINLOCK(balancing);
9097 * Scale the max load_balance interval with the number of CPUs in the system.
9098 * This trades load-balance latency on larger machines for less cross talk.
9100 void update_max_interval(void)
9102 max_load_balance_interval = HZ*num_online_cpus()/10;
9106 * It checks each scheduling domain to see if it is due to be balanced,
9107 * and initiates a balancing operation if so.
9109 * Balancing parameters are set up in init_sched_domains.
9111 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9113 int continue_balancing = 1;
9115 unsigned long interval;
9116 struct sched_domain *sd;
9117 /* Earliest time when we have to do rebalance again */
9118 unsigned long next_balance = jiffies + 60*HZ;
9119 int update_next_balance = 0;
9120 int need_serialize, need_decay = 0;
9124 for_each_domain(cpu, sd) {
9126 * Decay the newidle max times here because this is a regular
9127 * visit to all the domains. Decay ~1% per second.
9129 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9130 sd->max_newidle_lb_cost =
9131 (sd->max_newidle_lb_cost * 253) / 256;
9132 sd->next_decay_max_lb_cost = jiffies + HZ;
9135 max_cost += sd->max_newidle_lb_cost;
9137 if (!(sd->flags & SD_LOAD_BALANCE))
9141 * Stop the load balance at this level. There is another
9142 * CPU in our sched group which is doing load balancing more
9145 if (!continue_balancing) {
9151 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9153 need_serialize = sd->flags & SD_SERIALIZE;
9154 if (need_serialize) {
9155 if (!spin_trylock(&balancing))
9159 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9160 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9162 * The LBF_DST_PINNED logic could have changed
9163 * env->dst_cpu, so we can't know our idle
9164 * state even if we migrated tasks. Update it.
9166 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9168 sd->last_balance = jiffies;
9169 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9172 spin_unlock(&balancing);
9174 if (time_after(next_balance, sd->last_balance + interval)) {
9175 next_balance = sd->last_balance + interval;
9176 update_next_balance = 1;
9181 * Ensure the rq-wide value also decays but keep it at a
9182 * reasonable floor to avoid funnies with rq->avg_idle.
9184 rq->max_idle_balance_cost =
9185 max((u64)sysctl_sched_migration_cost, max_cost);
9190 * next_balance will be updated only when there is a need.
9191 * When the cpu is attached to null domain for ex, it will not be
9194 if (likely(update_next_balance)) {
9195 rq->next_balance = next_balance;
9197 #ifdef CONFIG_NO_HZ_COMMON
9199 * If this CPU has been elected to perform the nohz idle
9200 * balance. Other idle CPUs have already rebalanced with
9201 * nohz_idle_balance() and nohz.next_balance has been
9202 * updated accordingly. This CPU is now running the idle load
9203 * balance for itself and we need to update the
9204 * nohz.next_balance accordingly.
9206 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9207 nohz.next_balance = rq->next_balance;
9212 static inline int on_null_domain(struct rq *rq)
9214 return unlikely(!rcu_dereference_sched(rq->sd));
9217 #ifdef CONFIG_NO_HZ_COMMON
9219 * idle load balancing details
9220 * - When one of the busy CPUs notice that there may be an idle rebalancing
9221 * needed, they will kick the idle load balancer, which then does idle
9222 * load balancing for all the idle CPUs.
9223 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
9227 static inline int find_new_ilb(void)
9231 for_each_cpu_and(ilb, nohz.idle_cpus_mask,
9232 housekeeping_cpumask(HK_FLAG_MISC)) {
9241 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
9242 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
9244 static void kick_ilb(unsigned int flags)
9248 nohz.next_balance++;
9250 ilb_cpu = find_new_ilb();
9252 if (ilb_cpu >= nr_cpu_ids)
9255 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
9256 if (flags & NOHZ_KICK_MASK)
9260 * Use smp_send_reschedule() instead of resched_cpu().
9261 * This way we generate a sched IPI on the target CPU which
9262 * is idle. And the softirq performing nohz idle load balance
9263 * will be run before returning from the IPI.
9265 smp_send_reschedule(ilb_cpu);
9269 * Current decision point for kicking the idle load balancer in the presence
9270 * of idle CPUs in the system.
9272 static void nohz_balancer_kick(struct rq *rq)
9274 unsigned long now = jiffies;
9275 struct sched_domain_shared *sds;
9276 struct sched_domain *sd;
9277 int nr_busy, i, cpu = rq->cpu;
9278 unsigned int flags = 0;
9280 if (unlikely(rq->idle_balance))
9284 * We may be recently in ticked or tickless idle mode. At the first
9285 * busy tick after returning from idle, we will update the busy stats.
9287 nohz_balance_exit_idle(rq);
9290 * None are in tickless mode and hence no need for NOHZ idle load
9293 if (likely(!atomic_read(&nohz.nr_cpus)))
9296 if (READ_ONCE(nohz.has_blocked) &&
9297 time_after(now, READ_ONCE(nohz.next_blocked)))
9298 flags = NOHZ_STATS_KICK;
9300 if (time_before(now, nohz.next_balance))
9303 if (rq->nr_running >= 2) {
9304 flags = NOHZ_KICK_MASK;
9310 sd = rcu_dereference(rq->sd);
9313 * If there's a CFS task and the current CPU has reduced
9314 * capacity; kick the ILB to see if there's a better CPU to run
9317 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
9318 flags = NOHZ_KICK_MASK;
9323 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
9326 * When ASYM_PACKING; see if there's a more preferred CPU
9327 * currently idle; in which case, kick the ILB to move tasks
9330 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
9331 if (sched_asym_prefer(i, cpu)) {
9332 flags = NOHZ_KICK_MASK;
9338 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
9341 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
9342 * to run the misfit task on.
9344 if (check_misfit_status(rq, sd)) {
9345 flags = NOHZ_KICK_MASK;
9350 * For asymmetric systems, we do not want to nicely balance
9351 * cache use, instead we want to embrace asymmetry and only
9352 * ensure tasks have enough CPU capacity.
9354 * Skip the LLC logic because it's not relevant in that case.
9359 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9362 * If there is an imbalance between LLC domains (IOW we could
9363 * increase the overall cache use), we need some less-loaded LLC
9364 * domain to pull some load. Likewise, we may need to spread
9365 * load within the current LLC domain (e.g. packed SMT cores but
9366 * other CPUs are idle). We can't really know from here how busy
9367 * the others are - so just get a nohz balance going if it looks
9368 * like this LLC domain has tasks we could move.
9370 nr_busy = atomic_read(&sds->nr_busy_cpus);
9372 flags = NOHZ_KICK_MASK;
9383 static void set_cpu_sd_state_busy(int cpu)
9385 struct sched_domain *sd;
9388 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9390 if (!sd || !sd->nohz_idle)
9394 atomic_inc(&sd->shared->nr_busy_cpus);
9399 void nohz_balance_exit_idle(struct rq *rq)
9401 SCHED_WARN_ON(rq != this_rq());
9403 if (likely(!rq->nohz_tick_stopped))
9406 rq->nohz_tick_stopped = 0;
9407 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
9408 atomic_dec(&nohz.nr_cpus);
9410 set_cpu_sd_state_busy(rq->cpu);
9413 static void set_cpu_sd_state_idle(int cpu)
9415 struct sched_domain *sd;
9418 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9420 if (!sd || sd->nohz_idle)
9424 atomic_dec(&sd->shared->nr_busy_cpus);
9430 * This routine will record that the CPU is going idle with tick stopped.
9431 * This info will be used in performing idle load balancing in the future.
9433 void nohz_balance_enter_idle(int cpu)
9435 struct rq *rq = cpu_rq(cpu);
9437 SCHED_WARN_ON(cpu != smp_processor_id());
9439 /* If this CPU is going down, then nothing needs to be done: */
9440 if (!cpu_active(cpu))
9443 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9444 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9448 * Can be set safely without rq->lock held
9449 * If a clear happens, it will have evaluated last additions because
9450 * rq->lock is held during the check and the clear
9452 rq->has_blocked_load = 1;
9455 * The tick is still stopped but load could have been added in the
9456 * meantime. We set the nohz.has_blocked flag to trig a check of the
9457 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
9458 * of nohz.has_blocked can only happen after checking the new load
9460 if (rq->nohz_tick_stopped)
9463 /* If we're a completely isolated CPU, we don't play: */
9464 if (on_null_domain(rq))
9467 rq->nohz_tick_stopped = 1;
9469 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9470 atomic_inc(&nohz.nr_cpus);
9473 * Ensures that if nohz_idle_balance() fails to observe our
9474 * @idle_cpus_mask store, it must observe the @has_blocked
9477 smp_mb__after_atomic();
9479 set_cpu_sd_state_idle(cpu);
9483 * Each time a cpu enter idle, we assume that it has blocked load and
9484 * enable the periodic update of the load of idle cpus
9486 WRITE_ONCE(nohz.has_blocked, 1);
9490 * Internal function that runs load balance for all idle cpus. The load balance
9491 * can be a simple update of blocked load or a complete load balance with
9492 * tasks movement depending of flags.
9493 * The function returns false if the loop has stopped before running
9494 * through all idle CPUs.
9496 static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
9497 enum cpu_idle_type idle)
9499 /* Earliest time when we have to do rebalance again */
9500 unsigned long now = jiffies;
9501 unsigned long next_balance = now + 60*HZ;
9502 bool has_blocked_load = false;
9503 int update_next_balance = 0;
9504 int this_cpu = this_rq->cpu;
9509 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
9512 * We assume there will be no idle load after this update and clear
9513 * the has_blocked flag. If a cpu enters idle in the mean time, it will
9514 * set the has_blocked flag and trig another update of idle load.
9515 * Because a cpu that becomes idle, is added to idle_cpus_mask before
9516 * setting the flag, we are sure to not clear the state and not
9517 * check the load of an idle cpu.
9519 WRITE_ONCE(nohz.has_blocked, 0);
9522 * Ensures that if we miss the CPU, we must see the has_blocked
9523 * store from nohz_balance_enter_idle().
9527 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9528 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9532 * If this CPU gets work to do, stop the load balancing
9533 * work being done for other CPUs. Next load
9534 * balancing owner will pick it up.
9536 if (need_resched()) {
9537 has_blocked_load = true;
9541 rq = cpu_rq(balance_cpu);
9543 has_blocked_load |= update_nohz_stats(rq, true);
9546 * If time for next balance is due,
9549 if (time_after_eq(jiffies, rq->next_balance)) {
9552 rq_lock_irqsave(rq, &rf);
9553 update_rq_clock(rq);
9554 rq_unlock_irqrestore(rq, &rf);
9556 if (flags & NOHZ_BALANCE_KICK)
9557 rebalance_domains(rq, CPU_IDLE);
9560 if (time_after(next_balance, rq->next_balance)) {
9561 next_balance = rq->next_balance;
9562 update_next_balance = 1;
9566 /* Newly idle CPU doesn't need an update */
9567 if (idle != CPU_NEWLY_IDLE) {
9568 update_blocked_averages(this_cpu);
9569 has_blocked_load |= this_rq->has_blocked_load;
9572 if (flags & NOHZ_BALANCE_KICK)
9573 rebalance_domains(this_rq, CPU_IDLE);
9575 WRITE_ONCE(nohz.next_blocked,
9576 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
9578 /* The full idle balance loop has been done */
9582 /* There is still blocked load, enable periodic update */
9583 if (has_blocked_load)
9584 WRITE_ONCE(nohz.has_blocked, 1);
9587 * next_balance will be updated only when there is a need.
9588 * When the CPU is attached to null domain for ex, it will not be
9591 if (likely(update_next_balance))
9592 nohz.next_balance = next_balance;
9598 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9599 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9601 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9603 int this_cpu = this_rq->cpu;
9606 if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
9609 if (idle != CPU_IDLE) {
9610 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9614 /* could be _relaxed() */
9615 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9616 if (!(flags & NOHZ_KICK_MASK))
9619 _nohz_idle_balance(this_rq, flags, idle);
9624 static void nohz_newidle_balance(struct rq *this_rq)
9626 int this_cpu = this_rq->cpu;
9629 * This CPU doesn't want to be disturbed by scheduler
9632 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
9635 /* Will wake up very soon. No time for doing anything else*/
9636 if (this_rq->avg_idle < sysctl_sched_migration_cost)
9639 /* Don't need to update blocked load of idle CPUs*/
9640 if (!READ_ONCE(nohz.has_blocked) ||
9641 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
9644 raw_spin_unlock(&this_rq->lock);
9646 * This CPU is going to be idle and blocked load of idle CPUs
9647 * need to be updated. Run the ilb locally as it is a good
9648 * candidate for ilb instead of waking up another idle CPU.
9649 * Kick an normal ilb if we failed to do the update.
9651 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
9652 kick_ilb(NOHZ_STATS_KICK);
9653 raw_spin_lock(&this_rq->lock);
9656 #else /* !CONFIG_NO_HZ_COMMON */
9657 static inline void nohz_balancer_kick(struct rq *rq) { }
9659 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9664 static inline void nohz_newidle_balance(struct rq *this_rq) { }
9665 #endif /* CONFIG_NO_HZ_COMMON */
9668 * idle_balance is called by schedule() if this_cpu is about to become
9669 * idle. Attempts to pull tasks from other CPUs.
9671 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
9673 unsigned long next_balance = jiffies + HZ;
9674 int this_cpu = this_rq->cpu;
9675 struct sched_domain *sd;
9676 int pulled_task = 0;
9680 * We must set idle_stamp _before_ calling idle_balance(), such that we
9681 * measure the duration of idle_balance() as idle time.
9683 this_rq->idle_stamp = rq_clock(this_rq);
9686 * Do not pull tasks towards !active CPUs...
9688 if (!cpu_active(this_cpu))
9692 * This is OK, because current is on_cpu, which avoids it being picked
9693 * for load-balance and preemption/IRQs are still disabled avoiding
9694 * further scheduler activity on it and we're being very careful to
9695 * re-start the picking loop.
9697 rq_unpin_lock(this_rq, rf);
9699 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
9700 !READ_ONCE(this_rq->rd->overload)) {
9703 sd = rcu_dereference_check_sched_domain(this_rq->sd);
9705 update_next_balance(sd, &next_balance);
9708 nohz_newidle_balance(this_rq);
9713 raw_spin_unlock(&this_rq->lock);
9715 update_blocked_averages(this_cpu);
9717 for_each_domain(this_cpu, sd) {
9718 int continue_balancing = 1;
9719 u64 t0, domain_cost;
9721 if (!(sd->flags & SD_LOAD_BALANCE))
9724 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
9725 update_next_balance(sd, &next_balance);
9729 if (sd->flags & SD_BALANCE_NEWIDLE) {
9730 t0 = sched_clock_cpu(this_cpu);
9732 pulled_task = load_balance(this_cpu, this_rq,
9734 &continue_balancing);
9736 domain_cost = sched_clock_cpu(this_cpu) - t0;
9737 if (domain_cost > sd->max_newidle_lb_cost)
9738 sd->max_newidle_lb_cost = domain_cost;
9740 curr_cost += domain_cost;
9743 update_next_balance(sd, &next_balance);
9746 * Stop searching for tasks to pull if there are
9747 * now runnable tasks on this rq.
9749 if (pulled_task || this_rq->nr_running > 0)
9754 raw_spin_lock(&this_rq->lock);
9756 if (curr_cost > this_rq->max_idle_balance_cost)
9757 this_rq->max_idle_balance_cost = curr_cost;
9761 * While browsing the domains, we released the rq lock, a task could
9762 * have been enqueued in the meantime. Since we're not going idle,
9763 * pretend we pulled a task.
9765 if (this_rq->cfs.h_nr_running && !pulled_task)
9768 /* Move the next balance forward */
9769 if (time_after(this_rq->next_balance, next_balance))
9770 this_rq->next_balance = next_balance;
9772 /* Is there a task of a high priority class? */
9773 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
9777 this_rq->idle_stamp = 0;
9779 rq_repin_lock(this_rq, rf);
9785 * run_rebalance_domains is triggered when needed from the scheduler tick.
9786 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9788 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9790 struct rq *this_rq = this_rq();
9791 enum cpu_idle_type idle = this_rq->idle_balance ?
9792 CPU_IDLE : CPU_NOT_IDLE;
9795 * If this CPU has a pending nohz_balance_kick, then do the
9796 * balancing on behalf of the other idle CPUs whose ticks are
9797 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9798 * give the idle CPUs a chance to load balance. Else we may
9799 * load balance only within the local sched_domain hierarchy
9800 * and abort nohz_idle_balance altogether if we pull some load.
9802 if (nohz_idle_balance(this_rq, idle))
9805 /* normal load balance */
9806 update_blocked_averages(this_rq->cpu);
9807 rebalance_domains(this_rq, idle);
9811 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9813 void trigger_load_balance(struct rq *rq)
9815 /* Don't need to rebalance while attached to NULL domain */
9816 if (unlikely(on_null_domain(rq)))
9819 if (time_after_eq(jiffies, rq->next_balance))
9820 raise_softirq(SCHED_SOFTIRQ);
9822 nohz_balancer_kick(rq);
9825 static void rq_online_fair(struct rq *rq)
9829 update_runtime_enabled(rq);
9832 static void rq_offline_fair(struct rq *rq)
9836 /* Ensure any throttled groups are reachable by pick_next_task */
9837 unthrottle_offline_cfs_rqs(rq);
9840 #endif /* CONFIG_SMP */
9843 * scheduler tick hitting a task of our scheduling class.
9845 * NOTE: This function can be called remotely by the tick offload that
9846 * goes along full dynticks. Therefore no local assumption can be made
9847 * and everything must be accessed through the @rq and @curr passed in
9850 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9852 struct cfs_rq *cfs_rq;
9853 struct sched_entity *se = &curr->se;
9855 for_each_sched_entity(se) {
9856 cfs_rq = cfs_rq_of(se);
9857 entity_tick(cfs_rq, se, queued);
9860 if (static_branch_unlikely(&sched_numa_balancing))
9861 task_tick_numa(rq, curr);
9863 update_misfit_status(curr, rq);
9864 update_overutilized_status(task_rq(curr));
9868 * called on fork with the child task as argument from the parent's context
9869 * - child not yet on the tasklist
9870 * - preemption disabled
9872 static void task_fork_fair(struct task_struct *p)
9874 struct cfs_rq *cfs_rq;
9875 struct sched_entity *se = &p->se, *curr;
9876 struct rq *rq = this_rq();
9880 update_rq_clock(rq);
9882 cfs_rq = task_cfs_rq(current);
9883 curr = cfs_rq->curr;
9885 update_curr(cfs_rq);
9886 se->vruntime = curr->vruntime;
9888 place_entity(cfs_rq, se, 1);
9890 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9892 * Upon rescheduling, sched_class::put_prev_task() will place
9893 * 'current' within the tree based on its new key value.
9895 swap(curr->vruntime, se->vruntime);
9899 se->vruntime -= cfs_rq->min_vruntime;
9904 * Priority of the task has changed. Check to see if we preempt
9908 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9910 if (!task_on_rq_queued(p))
9914 * Reschedule if we are currently running on this runqueue and
9915 * our priority decreased, or if we are not currently running on
9916 * this runqueue and our priority is higher than the current's
9918 if (rq->curr == p) {
9919 if (p->prio > oldprio)
9922 check_preempt_curr(rq, p, 0);
9925 static inline bool vruntime_normalized(struct task_struct *p)
9927 struct sched_entity *se = &p->se;
9930 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9931 * the dequeue_entity(.flags=0) will already have normalized the
9938 * When !on_rq, vruntime of the task has usually NOT been normalized.
9939 * But there are some cases where it has already been normalized:
9941 * - A forked child which is waiting for being woken up by
9942 * wake_up_new_task().
9943 * - A task which has been woken up by try_to_wake_up() and
9944 * waiting for actually being woken up by sched_ttwu_pending().
9946 if (!se->sum_exec_runtime ||
9947 (p->state == TASK_WAKING && p->sched_remote_wakeup))
9953 #ifdef CONFIG_FAIR_GROUP_SCHED
9955 * Propagate the changes of the sched_entity across the tg tree to make it
9956 * visible to the root
9958 static void propagate_entity_cfs_rq(struct sched_entity *se)
9960 struct cfs_rq *cfs_rq;
9962 /* Start to propagate at parent */
9965 for_each_sched_entity(se) {
9966 cfs_rq = cfs_rq_of(se);
9968 if (cfs_rq_throttled(cfs_rq))
9971 update_load_avg(cfs_rq, se, UPDATE_TG);
9975 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
9978 static void detach_entity_cfs_rq(struct sched_entity *se)
9980 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9982 /* Catch up with the cfs_rq and remove our load when we leave */
9983 update_load_avg(cfs_rq, se, 0);
9984 detach_entity_load_avg(cfs_rq, se);
9985 update_tg_load_avg(cfs_rq, false);
9986 propagate_entity_cfs_rq(se);
9989 static void attach_entity_cfs_rq(struct sched_entity *se)
9991 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9993 #ifdef CONFIG_FAIR_GROUP_SCHED
9995 * Since the real-depth could have been changed (only FAIR
9996 * class maintain depth value), reset depth properly.
9998 se->depth = se->parent ? se->parent->depth + 1 : 0;
10001 /* Synchronize entity with its cfs_rq */
10002 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
10003 attach_entity_load_avg(cfs_rq, se, 0);
10004 update_tg_load_avg(cfs_rq, false);
10005 propagate_entity_cfs_rq(se);
10008 static void detach_task_cfs_rq(struct task_struct *p)
10010 struct sched_entity *se = &p->se;
10011 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10013 if (!vruntime_normalized(p)) {
10015 * Fix up our vruntime so that the current sleep doesn't
10016 * cause 'unlimited' sleep bonus.
10018 place_entity(cfs_rq, se, 0);
10019 se->vruntime -= cfs_rq->min_vruntime;
10022 detach_entity_cfs_rq(se);
10025 static void attach_task_cfs_rq(struct task_struct *p)
10027 struct sched_entity *se = &p->se;
10028 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10030 attach_entity_cfs_rq(se);
10032 if (!vruntime_normalized(p))
10033 se->vruntime += cfs_rq->min_vruntime;
10036 static void switched_from_fair(struct rq *rq, struct task_struct *p)
10038 detach_task_cfs_rq(p);
10041 static void switched_to_fair(struct rq *rq, struct task_struct *p)
10043 attach_task_cfs_rq(p);
10045 if (task_on_rq_queued(p)) {
10047 * We were most likely switched from sched_rt, so
10048 * kick off the schedule if running, otherwise just see
10049 * if we can still preempt the current task.
10054 check_preempt_curr(rq, p, 0);
10058 /* Account for a task changing its policy or group.
10060 * This routine is mostly called to set cfs_rq->curr field when a task
10061 * migrates between groups/classes.
10063 static void set_curr_task_fair(struct rq *rq)
10065 struct sched_entity *se = &rq->curr->se;
10067 for_each_sched_entity(se) {
10068 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10070 set_next_entity(cfs_rq, se);
10071 /* ensure bandwidth has been allocated on our new cfs_rq */
10072 account_cfs_rq_runtime(cfs_rq, 0);
10076 void init_cfs_rq(struct cfs_rq *cfs_rq)
10078 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
10079 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
10080 #ifndef CONFIG_64BIT
10081 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
10084 raw_spin_lock_init(&cfs_rq->removed.lock);
10088 #ifdef CONFIG_FAIR_GROUP_SCHED
10089 static void task_set_group_fair(struct task_struct *p)
10091 struct sched_entity *se = &p->se;
10093 set_task_rq(p, task_cpu(p));
10094 se->depth = se->parent ? se->parent->depth + 1 : 0;
10097 static void task_move_group_fair(struct task_struct *p)
10099 detach_task_cfs_rq(p);
10100 set_task_rq(p, task_cpu(p));
10103 /* Tell se's cfs_rq has been changed -- migrated */
10104 p->se.avg.last_update_time = 0;
10106 attach_task_cfs_rq(p);
10109 static void task_change_group_fair(struct task_struct *p, int type)
10112 case TASK_SET_GROUP:
10113 task_set_group_fair(p);
10116 case TASK_MOVE_GROUP:
10117 task_move_group_fair(p);
10122 void free_fair_sched_group(struct task_group *tg)
10126 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
10128 for_each_possible_cpu(i) {
10130 kfree(tg->cfs_rq[i]);
10139 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10141 struct sched_entity *se;
10142 struct cfs_rq *cfs_rq;
10145 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
10148 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
10152 tg->shares = NICE_0_LOAD;
10154 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
10156 for_each_possible_cpu(i) {
10157 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
10158 GFP_KERNEL, cpu_to_node(i));
10162 se = kzalloc_node(sizeof(struct sched_entity),
10163 GFP_KERNEL, cpu_to_node(i));
10167 init_cfs_rq(cfs_rq);
10168 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
10169 init_entity_runnable_average(se);
10180 void online_fair_sched_group(struct task_group *tg)
10182 struct sched_entity *se;
10186 for_each_possible_cpu(i) {
10190 raw_spin_lock_irq(&rq->lock);
10191 update_rq_clock(rq);
10192 attach_entity_cfs_rq(se);
10193 sync_throttle(tg, i);
10194 raw_spin_unlock_irq(&rq->lock);
10198 void unregister_fair_sched_group(struct task_group *tg)
10200 unsigned long flags;
10204 for_each_possible_cpu(cpu) {
10206 remove_entity_load_avg(tg->se[cpu]);
10209 * Only empty task groups can be destroyed; so we can speculatively
10210 * check on_list without danger of it being re-added.
10212 if (!tg->cfs_rq[cpu]->on_list)
10217 raw_spin_lock_irqsave(&rq->lock, flags);
10218 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
10219 raw_spin_unlock_irqrestore(&rq->lock, flags);
10223 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
10224 struct sched_entity *se, int cpu,
10225 struct sched_entity *parent)
10227 struct rq *rq = cpu_rq(cpu);
10231 init_cfs_rq_runtime(cfs_rq);
10233 tg->cfs_rq[cpu] = cfs_rq;
10236 /* se could be NULL for root_task_group */
10241 se->cfs_rq = &rq->cfs;
10244 se->cfs_rq = parent->my_q;
10245 se->depth = parent->depth + 1;
10249 /* guarantee group entities always have weight */
10250 update_load_set(&se->load, NICE_0_LOAD);
10251 se->parent = parent;
10254 static DEFINE_MUTEX(shares_mutex);
10256 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10261 * We can't change the weight of the root cgroup.
10266 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10268 mutex_lock(&shares_mutex);
10269 if (tg->shares == shares)
10272 tg->shares = shares;
10273 for_each_possible_cpu(i) {
10274 struct rq *rq = cpu_rq(i);
10275 struct sched_entity *se = tg->se[i];
10276 struct rq_flags rf;
10278 /* Propagate contribution to hierarchy */
10279 rq_lock_irqsave(rq, &rf);
10280 update_rq_clock(rq);
10281 for_each_sched_entity(se) {
10282 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
10283 update_cfs_group(se);
10285 rq_unlock_irqrestore(rq, &rf);
10289 mutex_unlock(&shares_mutex);
10292 #else /* CONFIG_FAIR_GROUP_SCHED */
10294 void free_fair_sched_group(struct task_group *tg) { }
10296 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10301 void online_fair_sched_group(struct task_group *tg) { }
10303 void unregister_fair_sched_group(struct task_group *tg) { }
10305 #endif /* CONFIG_FAIR_GROUP_SCHED */
10308 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10310 struct sched_entity *se = &task->se;
10311 unsigned int rr_interval = 0;
10314 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10317 if (rq->cfs.load.weight)
10318 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10320 return rr_interval;
10324 * All the scheduling class methods:
10326 const struct sched_class fair_sched_class = {
10327 .next = &idle_sched_class,
10328 .enqueue_task = enqueue_task_fair,
10329 .dequeue_task = dequeue_task_fair,
10330 .yield_task = yield_task_fair,
10331 .yield_to_task = yield_to_task_fair,
10333 .check_preempt_curr = check_preempt_wakeup,
10335 .pick_next_task = pick_next_task_fair,
10336 .put_prev_task = put_prev_task_fair,
10339 .select_task_rq = select_task_rq_fair,
10340 .migrate_task_rq = migrate_task_rq_fair,
10342 .rq_online = rq_online_fair,
10343 .rq_offline = rq_offline_fair,
10345 .task_dead = task_dead_fair,
10346 .set_cpus_allowed = set_cpus_allowed_common,
10349 .set_curr_task = set_curr_task_fair,
10350 .task_tick = task_tick_fair,
10351 .task_fork = task_fork_fair,
10353 .prio_changed = prio_changed_fair,
10354 .switched_from = switched_from_fair,
10355 .switched_to = switched_to_fair,
10357 .get_rr_interval = get_rr_interval_fair,
10359 .update_curr = update_curr_fair,
10361 #ifdef CONFIG_FAIR_GROUP_SCHED
10362 .task_change_group = task_change_group_fair,
10366 #ifdef CONFIG_SCHED_DEBUG
10367 void print_cfs_stats(struct seq_file *m, int cpu)
10369 struct cfs_rq *cfs_rq, *pos;
10372 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
10373 print_cfs_rq(m, cpu, cfs_rq);
10377 #ifdef CONFIG_NUMA_BALANCING
10378 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10381 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10383 for_each_online_node(node) {
10384 if (p->numa_faults) {
10385 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10386 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10388 if (p->numa_group) {
10389 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
10390 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
10392 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10395 #endif /* CONFIG_NUMA_BALANCING */
10396 #endif /* CONFIG_SCHED_DEBUG */
10398 __init void init_sched_fair_class(void)
10401 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10403 #ifdef CONFIG_NO_HZ_COMMON
10404 nohz.next_balance = jiffies;
10405 nohz.next_blocked = jiffies;
10406 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);