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.
103 #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024)
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 fact = mul_u32_u32(fact, lw->inv_weight);
239 return mul_u64_u32_shr(delta_exec, fact, shift);
243 const struct sched_class fair_sched_class;
245 /**************************************************************
246 * CFS operations on generic schedulable entities:
249 #ifdef CONFIG_FAIR_GROUP_SCHED
250 static inline struct task_struct *task_of(struct sched_entity *se)
252 SCHED_WARN_ON(!entity_is_task(se));
253 return container_of(se, struct task_struct, se);
256 /* Walk up scheduling entities hierarchy */
257 #define for_each_sched_entity(se) \
258 for (; se; se = se->parent)
260 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
265 /* runqueue on which this entity is (to be) queued */
266 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
271 /* runqueue "owned" by this group */
272 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
277 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
282 if (cfs_rq && task_group_is_autogroup(cfs_rq->tg))
283 autogroup_path(cfs_rq->tg, path, len);
284 else if (cfs_rq && cfs_rq->tg->css.cgroup)
285 cgroup_path(cfs_rq->tg->css.cgroup, path, len);
287 strlcpy(path, "(null)", len);
290 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
292 struct rq *rq = rq_of(cfs_rq);
293 int cpu = cpu_of(rq);
296 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
301 * Ensure we either appear before our parent (if already
302 * enqueued) or force our parent to appear after us when it is
303 * enqueued. The fact that we always enqueue bottom-up
304 * reduces this to two cases and a special case for the root
305 * cfs_rq. Furthermore, it also means that we will always reset
306 * tmp_alone_branch either when the branch is connected
307 * to a tree or when we reach the top of the tree
309 if (cfs_rq->tg->parent &&
310 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
312 * If parent is already on the list, we add the child
313 * just before. Thanks to circular linked property of
314 * the list, this means to put the child at the tail
315 * of the list that starts by parent.
317 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
318 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
320 * The branch is now connected to its tree so we can
321 * reset tmp_alone_branch to the beginning of the
324 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
328 if (!cfs_rq->tg->parent) {
330 * cfs rq without parent should be put
331 * at the tail of the list.
333 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
334 &rq->leaf_cfs_rq_list);
336 * We have reach the top of a tree so we can reset
337 * tmp_alone_branch to the beginning of the list.
339 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
344 * The parent has not already been added so we want to
345 * make sure that it will be put after us.
346 * tmp_alone_branch points to the begin of the branch
347 * where we will add parent.
349 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
351 * update tmp_alone_branch to points to the new begin
354 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
358 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
360 if (cfs_rq->on_list) {
361 struct rq *rq = rq_of(cfs_rq);
364 * With cfs_rq being unthrottled/throttled during an enqueue,
365 * it can happen the tmp_alone_branch points the a leaf that
366 * we finally want to del. In this case, tmp_alone_branch moves
367 * to the prev element but it will point to rq->leaf_cfs_rq_list
368 * at the end of the enqueue.
370 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
371 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
373 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
378 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
380 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
383 /* Iterate thr' all leaf cfs_rq's on a runqueue */
384 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
385 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
388 /* Do the two (enqueued) entities belong to the same group ? */
389 static inline struct cfs_rq *
390 is_same_group(struct sched_entity *se, struct sched_entity *pse)
392 if (se->cfs_rq == pse->cfs_rq)
398 static inline struct sched_entity *parent_entity(struct sched_entity *se)
404 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
406 int se_depth, pse_depth;
409 * preemption test can be made between sibling entities who are in the
410 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
411 * both tasks until we find their ancestors who are siblings of common
415 /* First walk up until both entities are at same depth */
416 se_depth = (*se)->depth;
417 pse_depth = (*pse)->depth;
419 while (se_depth > pse_depth) {
421 *se = parent_entity(*se);
424 while (pse_depth > se_depth) {
426 *pse = parent_entity(*pse);
429 while (!is_same_group(*se, *pse)) {
430 *se = parent_entity(*se);
431 *pse = parent_entity(*pse);
435 #else /* !CONFIG_FAIR_GROUP_SCHED */
437 static inline struct task_struct *task_of(struct sched_entity *se)
439 return container_of(se, struct task_struct, se);
442 #define for_each_sched_entity(se) \
443 for (; se; se = NULL)
445 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
447 return &task_rq(p)->cfs;
450 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
452 struct task_struct *p = task_of(se);
453 struct rq *rq = task_rq(p);
458 /* runqueue "owned" by this group */
459 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
464 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
467 strlcpy(path, "(null)", len);
470 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
475 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
479 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
483 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
484 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
486 static inline struct sched_entity *parent_entity(struct sched_entity *se)
492 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
496 #endif /* CONFIG_FAIR_GROUP_SCHED */
498 static __always_inline
499 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
501 /**************************************************************
502 * Scheduling class tree data structure manipulation methods:
505 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
507 s64 delta = (s64)(vruntime - max_vruntime);
509 max_vruntime = vruntime;
514 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
516 s64 delta = (s64)(vruntime - min_vruntime);
518 min_vruntime = vruntime;
523 static inline int entity_before(struct sched_entity *a,
524 struct sched_entity *b)
526 return (s64)(a->vruntime - b->vruntime) < 0;
529 static void update_min_vruntime(struct cfs_rq *cfs_rq)
531 struct sched_entity *curr = cfs_rq->curr;
532 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
534 u64 vruntime = cfs_rq->min_vruntime;
538 vruntime = curr->vruntime;
543 if (leftmost) { /* non-empty tree */
544 struct sched_entity *se;
545 se = rb_entry(leftmost, struct sched_entity, run_node);
548 vruntime = se->vruntime;
550 vruntime = min_vruntime(vruntime, se->vruntime);
553 /* ensure we never gain time by being placed backwards. */
554 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
557 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
562 * Enqueue an entity into the rb-tree:
564 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
566 struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node;
567 struct rb_node *parent = NULL;
568 struct sched_entity *entry;
569 bool leftmost = true;
572 * Find the right place in the rbtree:
576 entry = rb_entry(parent, struct sched_entity, run_node);
578 * We dont care about collisions. Nodes with
579 * the same key stay together.
581 if (entity_before(se, entry)) {
582 link = &parent->rb_left;
584 link = &parent->rb_right;
589 rb_link_node(&se->run_node, parent, link);
590 rb_insert_color_cached(&se->run_node,
591 &cfs_rq->tasks_timeline, leftmost);
594 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
596 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
599 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
601 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
606 return rb_entry(left, struct sched_entity, run_node);
609 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
611 struct rb_node *next = rb_next(&se->run_node);
616 return rb_entry(next, struct sched_entity, run_node);
619 #ifdef CONFIG_SCHED_DEBUG
620 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
622 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
627 return rb_entry(last, struct sched_entity, run_node);
630 /**************************************************************
631 * Scheduling class statistics methods:
634 int sched_proc_update_handler(struct ctl_table *table, int write,
635 void __user *buffer, size_t *lenp,
638 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
639 unsigned int factor = get_update_sysctl_factor();
644 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
645 sysctl_sched_min_granularity);
647 #define WRT_SYSCTL(name) \
648 (normalized_sysctl_##name = sysctl_##name / (factor))
649 WRT_SYSCTL(sched_min_granularity);
650 WRT_SYSCTL(sched_latency);
651 WRT_SYSCTL(sched_wakeup_granularity);
661 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
663 if (unlikely(se->load.weight != NICE_0_LOAD))
664 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
670 * The idea is to set a period in which each task runs once.
672 * When there are too many tasks (sched_nr_latency) we have to stretch
673 * this period because otherwise the slices get too small.
675 * p = (nr <= nl) ? l : l*nr/nl
677 static u64 __sched_period(unsigned long nr_running)
679 if (unlikely(nr_running > sched_nr_latency))
680 return nr_running * sysctl_sched_min_granularity;
682 return sysctl_sched_latency;
686 * We calculate the wall-time slice from the period by taking a part
687 * proportional to the weight.
691 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
693 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
695 for_each_sched_entity(se) {
696 struct load_weight *load;
697 struct load_weight lw;
699 cfs_rq = cfs_rq_of(se);
700 load = &cfs_rq->load;
702 if (unlikely(!se->on_rq)) {
705 update_load_add(&lw, se->load.weight);
708 slice = __calc_delta(slice, se->load.weight, load);
714 * We calculate the vruntime slice of a to-be-inserted task.
718 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
720 return calc_delta_fair(sched_slice(cfs_rq, se), se);
726 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
727 static unsigned long task_h_load(struct task_struct *p);
728 static unsigned long capacity_of(int cpu);
730 /* Give new sched_entity start runnable values to heavy its load in infant time */
731 void init_entity_runnable_average(struct sched_entity *se)
733 struct sched_avg *sa = &se->avg;
735 memset(sa, 0, sizeof(*sa));
738 * Tasks are initialized with full load to be seen as heavy tasks until
739 * they get a chance to stabilize to their real load level.
740 * Group entities are initialized with zero load to reflect the fact that
741 * nothing has been attached to the task group yet.
743 if (entity_is_task(se))
744 sa->runnable_load_avg = sa->load_avg = scale_load_down(se->load.weight);
746 se->runnable_weight = se->load.weight;
748 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
751 static void attach_entity_cfs_rq(struct sched_entity *se);
754 * With new tasks being created, their initial util_avgs are extrapolated
755 * based on the cfs_rq's current util_avg:
757 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
759 * However, in many cases, the above util_avg does not give a desired
760 * value. Moreover, the sum of the util_avgs may be divergent, such
761 * as when the series is a harmonic series.
763 * To solve this problem, we also cap the util_avg of successive tasks to
764 * only 1/2 of the left utilization budget:
766 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
768 * where n denotes the nth task and cpu_scale the CPU capacity.
770 * For example, for a CPU with 1024 of capacity, a simplest series from
771 * the beginning would be like:
773 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
774 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
776 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
777 * if util_avg > util_avg_cap.
779 void post_init_entity_util_avg(struct task_struct *p)
781 struct sched_entity *se = &p->se;
782 struct cfs_rq *cfs_rq = cfs_rq_of(se);
783 struct sched_avg *sa = &se->avg;
784 long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
785 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
788 if (cfs_rq->avg.util_avg != 0) {
789 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
790 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
792 if (sa->util_avg > cap)
799 if (p->sched_class != &fair_sched_class) {
801 * For !fair tasks do:
803 update_cfs_rq_load_avg(now, cfs_rq);
804 attach_entity_load_avg(cfs_rq, se, 0);
805 switched_from_fair(rq, p);
807 * such that the next switched_to_fair() has the
810 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
814 attach_entity_cfs_rq(se);
817 #else /* !CONFIG_SMP */
818 void init_entity_runnable_average(struct sched_entity *se)
821 void post_init_entity_util_avg(struct task_struct *p)
824 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
827 #endif /* CONFIG_SMP */
830 * Update the current task's runtime statistics.
832 static void update_curr(struct cfs_rq *cfs_rq)
834 struct sched_entity *curr = cfs_rq->curr;
835 u64 now = rq_clock_task(rq_of(cfs_rq));
841 delta_exec = now - curr->exec_start;
842 if (unlikely((s64)delta_exec <= 0))
845 curr->exec_start = now;
847 schedstat_set(curr->statistics.exec_max,
848 max(delta_exec, curr->statistics.exec_max));
850 curr->sum_exec_runtime += delta_exec;
851 schedstat_add(cfs_rq->exec_clock, delta_exec);
853 curr->vruntime += calc_delta_fair(delta_exec, curr);
854 update_min_vruntime(cfs_rq);
856 if (entity_is_task(curr)) {
857 struct task_struct *curtask = task_of(curr);
859 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
860 cgroup_account_cputime(curtask, delta_exec);
861 account_group_exec_runtime(curtask, delta_exec);
864 account_cfs_rq_runtime(cfs_rq, delta_exec);
867 static void update_curr_fair(struct rq *rq)
869 update_curr(cfs_rq_of(&rq->curr->se));
873 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
875 u64 wait_start, prev_wait_start;
877 if (!schedstat_enabled())
880 wait_start = rq_clock(rq_of(cfs_rq));
881 prev_wait_start = schedstat_val(se->statistics.wait_start);
883 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
884 likely(wait_start > prev_wait_start))
885 wait_start -= prev_wait_start;
887 __schedstat_set(se->statistics.wait_start, wait_start);
891 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
893 struct task_struct *p;
896 if (!schedstat_enabled())
899 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
901 if (entity_is_task(se)) {
903 if (task_on_rq_migrating(p)) {
905 * Preserve migrating task's wait time so wait_start
906 * time stamp can be adjusted to accumulate wait time
907 * prior to migration.
909 __schedstat_set(se->statistics.wait_start, delta);
912 trace_sched_stat_wait(p, delta);
915 __schedstat_set(se->statistics.wait_max,
916 max(schedstat_val(se->statistics.wait_max), delta));
917 __schedstat_inc(se->statistics.wait_count);
918 __schedstat_add(se->statistics.wait_sum, delta);
919 __schedstat_set(se->statistics.wait_start, 0);
923 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
925 struct task_struct *tsk = NULL;
926 u64 sleep_start, block_start;
928 if (!schedstat_enabled())
931 sleep_start = schedstat_val(se->statistics.sleep_start);
932 block_start = schedstat_val(se->statistics.block_start);
934 if (entity_is_task(se))
938 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
943 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
944 __schedstat_set(se->statistics.sleep_max, delta);
946 __schedstat_set(se->statistics.sleep_start, 0);
947 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
950 account_scheduler_latency(tsk, delta >> 10, 1);
951 trace_sched_stat_sleep(tsk, delta);
955 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
960 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
961 __schedstat_set(se->statistics.block_max, delta);
963 __schedstat_set(se->statistics.block_start, 0);
964 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
967 if (tsk->in_iowait) {
968 __schedstat_add(se->statistics.iowait_sum, delta);
969 __schedstat_inc(se->statistics.iowait_count);
970 trace_sched_stat_iowait(tsk, delta);
973 trace_sched_stat_blocked(tsk, delta);
976 * Blocking time is in units of nanosecs, so shift by
977 * 20 to get a milliseconds-range estimation of the
978 * amount of time that the task spent sleeping:
980 if (unlikely(prof_on == SLEEP_PROFILING)) {
981 profile_hits(SLEEP_PROFILING,
982 (void *)get_wchan(tsk),
985 account_scheduler_latency(tsk, delta >> 10, 0);
991 * Task is being enqueued - update stats:
994 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
996 if (!schedstat_enabled())
1000 * Are we enqueueing a waiting task? (for current tasks
1001 * a dequeue/enqueue event is a NOP)
1003 if (se != cfs_rq->curr)
1004 update_stats_wait_start(cfs_rq, se);
1006 if (flags & ENQUEUE_WAKEUP)
1007 update_stats_enqueue_sleeper(cfs_rq, se);
1011 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1014 if (!schedstat_enabled())
1018 * Mark the end of the wait period if dequeueing a
1021 if (se != cfs_rq->curr)
1022 update_stats_wait_end(cfs_rq, se);
1024 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1025 struct task_struct *tsk = task_of(se);
1027 if (tsk->state & TASK_INTERRUPTIBLE)
1028 __schedstat_set(se->statistics.sleep_start,
1029 rq_clock(rq_of(cfs_rq)));
1030 if (tsk->state & TASK_UNINTERRUPTIBLE)
1031 __schedstat_set(se->statistics.block_start,
1032 rq_clock(rq_of(cfs_rq)));
1037 * We are picking a new current task - update its stats:
1040 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1043 * We are starting a new run period:
1045 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1048 /**************************************************
1049 * Scheduling class queueing methods:
1052 #ifdef CONFIG_NUMA_BALANCING
1054 * Approximate time to scan a full NUMA task in ms. The task scan period is
1055 * calculated based on the tasks virtual memory size and
1056 * numa_balancing_scan_size.
1058 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1059 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1061 /* Portion of address space to scan in MB */
1062 unsigned int sysctl_numa_balancing_scan_size = 256;
1064 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1065 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1068 refcount_t refcount;
1070 spinlock_t lock; /* nr_tasks, tasks */
1075 struct rcu_head rcu;
1076 unsigned long total_faults;
1077 unsigned long max_faults_cpu;
1079 * Faults_cpu is used to decide whether memory should move
1080 * towards the CPU. As a consequence, these stats are weighted
1081 * more by CPU use than by memory faults.
1083 unsigned long *faults_cpu;
1084 unsigned long faults[0];
1088 * For functions that can be called in multiple contexts that permit reading
1089 * ->numa_group (see struct task_struct for locking rules).
1091 static struct numa_group *deref_task_numa_group(struct task_struct *p)
1093 return rcu_dereference_check(p->numa_group, p == current ||
1094 (lockdep_is_held(&task_rq(p)->lock) && !READ_ONCE(p->on_cpu)));
1097 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1099 return rcu_dereference_protected(p->numa_group, p == current);
1102 static inline unsigned long group_faults_priv(struct numa_group *ng);
1103 static inline unsigned long group_faults_shared(struct numa_group *ng);
1105 static unsigned int task_nr_scan_windows(struct task_struct *p)
1107 unsigned long rss = 0;
1108 unsigned long nr_scan_pages;
1111 * Calculations based on RSS as non-present and empty pages are skipped
1112 * by the PTE scanner and NUMA hinting faults should be trapped based
1115 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1116 rss = get_mm_rss(p->mm);
1118 rss = nr_scan_pages;
1120 rss = round_up(rss, nr_scan_pages);
1121 return rss / nr_scan_pages;
1124 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1125 #define MAX_SCAN_WINDOW 2560
1127 static unsigned int task_scan_min(struct task_struct *p)
1129 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1130 unsigned int scan, floor;
1131 unsigned int windows = 1;
1133 if (scan_size < MAX_SCAN_WINDOW)
1134 windows = MAX_SCAN_WINDOW / scan_size;
1135 floor = 1000 / windows;
1137 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1138 return max_t(unsigned int, floor, scan);
1141 static unsigned int task_scan_start(struct task_struct *p)
1143 unsigned long smin = task_scan_min(p);
1144 unsigned long period = smin;
1145 struct numa_group *ng;
1147 /* Scale the maximum scan period with the amount of shared memory. */
1149 ng = rcu_dereference(p->numa_group);
1151 unsigned long shared = group_faults_shared(ng);
1152 unsigned long private = group_faults_priv(ng);
1154 period *= refcount_read(&ng->refcount);
1155 period *= shared + 1;
1156 period /= private + shared + 1;
1160 return max(smin, period);
1163 static unsigned int task_scan_max(struct task_struct *p)
1165 unsigned long smin = task_scan_min(p);
1167 struct numa_group *ng;
1169 /* Watch for min being lower than max due to floor calculations */
1170 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1172 /* Scale the maximum scan period with the amount of shared memory. */
1173 ng = deref_curr_numa_group(p);
1175 unsigned long shared = group_faults_shared(ng);
1176 unsigned long private = group_faults_priv(ng);
1177 unsigned long period = smax;
1179 period *= refcount_read(&ng->refcount);
1180 period *= shared + 1;
1181 period /= private + shared + 1;
1183 smax = max(smax, period);
1186 return max(smin, smax);
1189 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1191 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1192 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1195 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1197 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1198 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1201 /* Shared or private faults. */
1202 #define NR_NUMA_HINT_FAULT_TYPES 2
1204 /* Memory and CPU locality */
1205 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1207 /* Averaged statistics, and temporary buffers. */
1208 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1210 pid_t task_numa_group_id(struct task_struct *p)
1212 struct numa_group *ng;
1216 ng = rcu_dereference(p->numa_group);
1225 * The averaged statistics, shared & private, memory & CPU,
1226 * occupy the first half of the array. The second half of the
1227 * array is for current counters, which are averaged into the
1228 * first set by task_numa_placement.
1230 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1232 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1235 static inline unsigned long task_faults(struct task_struct *p, int nid)
1237 if (!p->numa_faults)
1240 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1241 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1244 static inline unsigned long group_faults(struct task_struct *p, int nid)
1246 struct numa_group *ng = deref_task_numa_group(p);
1251 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1252 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1255 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1257 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1258 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1261 static inline unsigned long group_faults_priv(struct numa_group *ng)
1263 unsigned long faults = 0;
1266 for_each_online_node(node) {
1267 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1273 static inline unsigned long group_faults_shared(struct numa_group *ng)
1275 unsigned long faults = 0;
1278 for_each_online_node(node) {
1279 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1286 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1287 * considered part of a numa group's pseudo-interleaving set. Migrations
1288 * between these nodes are slowed down, to allow things to settle down.
1290 #define ACTIVE_NODE_FRACTION 3
1292 static bool numa_is_active_node(int nid, struct numa_group *ng)
1294 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1297 /* Handle placement on systems where not all nodes are directly connected. */
1298 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1299 int maxdist, bool task)
1301 unsigned long score = 0;
1305 * All nodes are directly connected, and the same distance
1306 * from each other. No need for fancy placement algorithms.
1308 if (sched_numa_topology_type == NUMA_DIRECT)
1312 * This code is called for each node, introducing N^2 complexity,
1313 * which should be ok given the number of nodes rarely exceeds 8.
1315 for_each_online_node(node) {
1316 unsigned long faults;
1317 int dist = node_distance(nid, node);
1320 * The furthest away nodes in the system are not interesting
1321 * for placement; nid was already counted.
1323 if (dist == sched_max_numa_distance || node == nid)
1327 * On systems with a backplane NUMA topology, compare groups
1328 * of nodes, and move tasks towards the group with the most
1329 * memory accesses. When comparing two nodes at distance
1330 * "hoplimit", only nodes closer by than "hoplimit" are part
1331 * of each group. Skip other nodes.
1333 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1337 /* Add up the faults from nearby nodes. */
1339 faults = task_faults(p, node);
1341 faults = group_faults(p, node);
1344 * On systems with a glueless mesh NUMA topology, there are
1345 * no fixed "groups of nodes". Instead, nodes that are not
1346 * directly connected bounce traffic through intermediate
1347 * nodes; a numa_group can occupy any set of nodes.
1348 * The further away a node is, the less the faults count.
1349 * This seems to result in good task placement.
1351 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1352 faults *= (sched_max_numa_distance - dist);
1353 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1363 * These return the fraction of accesses done by a particular task, or
1364 * task group, on a particular numa node. The group weight is given a
1365 * larger multiplier, in order to group tasks together that are almost
1366 * evenly spread out between numa nodes.
1368 static inline unsigned long task_weight(struct task_struct *p, int nid,
1371 unsigned long faults, total_faults;
1373 if (!p->numa_faults)
1376 total_faults = p->total_numa_faults;
1381 faults = task_faults(p, nid);
1382 faults += score_nearby_nodes(p, nid, dist, true);
1384 return 1000 * faults / total_faults;
1387 static inline unsigned long group_weight(struct task_struct *p, int nid,
1390 struct numa_group *ng = deref_task_numa_group(p);
1391 unsigned long faults, total_faults;
1396 total_faults = ng->total_faults;
1401 faults = group_faults(p, nid);
1402 faults += score_nearby_nodes(p, nid, dist, false);
1404 return 1000 * faults / total_faults;
1407 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1408 int src_nid, int dst_cpu)
1410 struct numa_group *ng = deref_curr_numa_group(p);
1411 int dst_nid = cpu_to_node(dst_cpu);
1412 int last_cpupid, this_cpupid;
1414 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1415 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1418 * Allow first faults or private faults to migrate immediately early in
1419 * the lifetime of a task. The magic number 4 is based on waiting for
1420 * two full passes of the "multi-stage node selection" test that is
1423 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1424 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1428 * Multi-stage node selection is used in conjunction with a periodic
1429 * migration fault to build a temporal task<->page relation. By using
1430 * a two-stage filter we remove short/unlikely relations.
1432 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1433 * a task's usage of a particular page (n_p) per total usage of this
1434 * page (n_t) (in a given time-span) to a probability.
1436 * Our periodic faults will sample this probability and getting the
1437 * same result twice in a row, given these samples are fully
1438 * independent, is then given by P(n)^2, provided our sample period
1439 * is sufficiently short compared to the usage pattern.
1441 * This quadric squishes small probabilities, making it less likely we
1442 * act on an unlikely task<->page relation.
1444 if (!cpupid_pid_unset(last_cpupid) &&
1445 cpupid_to_nid(last_cpupid) != dst_nid)
1448 /* Always allow migrate on private faults */
1449 if (cpupid_match_pid(p, last_cpupid))
1452 /* A shared fault, but p->numa_group has not been set up yet. */
1457 * Destination node is much more heavily used than the source
1458 * node? Allow migration.
1460 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1461 ACTIVE_NODE_FRACTION)
1465 * Distribute memory according to CPU & memory use on each node,
1466 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1468 * faults_cpu(dst) 3 faults_cpu(src)
1469 * --------------- * - > ---------------
1470 * faults_mem(dst) 4 faults_mem(src)
1472 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1473 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1476 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq);
1478 static unsigned long cpu_runnable_load(struct rq *rq)
1480 return cfs_rq_runnable_load_avg(&rq->cfs);
1483 /* Cached statistics for all CPUs within a node */
1487 /* Total compute capacity of CPUs on a node */
1488 unsigned long compute_capacity;
1492 * XXX borrowed from update_sg_lb_stats
1494 static void update_numa_stats(struct numa_stats *ns, int nid)
1498 memset(ns, 0, sizeof(*ns));
1499 for_each_cpu(cpu, cpumask_of_node(nid)) {
1500 struct rq *rq = cpu_rq(cpu);
1502 ns->load += cpu_runnable_load(rq);
1503 ns->compute_capacity += capacity_of(cpu);
1508 struct task_numa_env {
1509 struct task_struct *p;
1511 int src_cpu, src_nid;
1512 int dst_cpu, dst_nid;
1514 struct numa_stats src_stats, dst_stats;
1519 struct task_struct *best_task;
1524 static void task_numa_assign(struct task_numa_env *env,
1525 struct task_struct *p, long imp)
1527 struct rq *rq = cpu_rq(env->dst_cpu);
1529 /* Bail out if run-queue part of active NUMA balance. */
1530 if (xchg(&rq->numa_migrate_on, 1))
1534 * Clear previous best_cpu/rq numa-migrate flag, since task now
1535 * found a better CPU to move/swap.
1537 if (env->best_cpu != -1) {
1538 rq = cpu_rq(env->best_cpu);
1539 WRITE_ONCE(rq->numa_migrate_on, 0);
1543 put_task_struct(env->best_task);
1548 env->best_imp = imp;
1549 env->best_cpu = env->dst_cpu;
1552 static bool load_too_imbalanced(long src_load, long dst_load,
1553 struct task_numa_env *env)
1556 long orig_src_load, orig_dst_load;
1557 long src_capacity, dst_capacity;
1560 * The load is corrected for the CPU capacity available on each node.
1563 * ------------ vs ---------
1564 * src_capacity dst_capacity
1566 src_capacity = env->src_stats.compute_capacity;
1567 dst_capacity = env->dst_stats.compute_capacity;
1569 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1571 orig_src_load = env->src_stats.load;
1572 orig_dst_load = env->dst_stats.load;
1574 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1576 /* Would this change make things worse? */
1577 return (imb > old_imb);
1581 * Maximum NUMA importance can be 1998 (2*999);
1582 * SMALLIMP @ 30 would be close to 1998/64.
1583 * Used to deter task migration.
1588 * This checks if the overall compute and NUMA accesses of the system would
1589 * be improved if the source tasks was migrated to the target dst_cpu taking
1590 * into account that it might be best if task running on the dst_cpu should
1591 * be exchanged with the source task
1593 static void task_numa_compare(struct task_numa_env *env,
1594 long taskimp, long groupimp, bool maymove)
1596 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
1597 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1598 long imp = p_ng ? groupimp : taskimp;
1599 struct task_struct *cur;
1600 long src_load, dst_load;
1601 int dist = env->dist;
1605 if (READ_ONCE(dst_rq->numa_migrate_on))
1609 cur = rcu_dereference(dst_rq->curr);
1610 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1614 * Because we have preemption enabled we can get migrated around and
1615 * end try selecting ourselves (current == env->p) as a swap candidate.
1621 if (maymove && moveimp >= env->best_imp)
1628 * "imp" is the fault differential for the source task between the
1629 * source and destination node. Calculate the total differential for
1630 * the source task and potential destination task. The more negative
1631 * the value is, the more remote accesses that would be expected to
1632 * be incurred if the tasks were swapped.
1634 /* Skip this swap candidate if cannot move to the source cpu */
1635 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
1639 * If dst and source tasks are in the same NUMA group, or not
1640 * in any group then look only at task weights.
1642 cur_ng = rcu_dereference(cur->numa_group);
1643 if (cur_ng == p_ng) {
1644 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1645 task_weight(cur, env->dst_nid, dist);
1647 * Add some hysteresis to prevent swapping the
1648 * tasks within a group over tiny differences.
1654 * Compare the group weights. If a task is all by itself
1655 * (not part of a group), use the task weight instead.
1658 imp += group_weight(cur, env->src_nid, dist) -
1659 group_weight(cur, env->dst_nid, dist);
1661 imp += task_weight(cur, env->src_nid, dist) -
1662 task_weight(cur, env->dst_nid, dist);
1665 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1672 * If the NUMA importance is less than SMALLIMP,
1673 * task migration might only result in ping pong
1674 * of tasks and also hurt performance due to cache
1677 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1681 * In the overloaded case, try and keep the load balanced.
1683 load = task_h_load(env->p) - task_h_load(cur);
1687 dst_load = env->dst_stats.load + load;
1688 src_load = env->src_stats.load - load;
1690 if (load_too_imbalanced(src_load, dst_load, env))
1695 * One idle CPU per node is evaluated for a task numa move.
1696 * Call select_idle_sibling to maybe find a better one.
1700 * select_idle_siblings() uses an per-CPU cpumask that
1701 * can be used from IRQ context.
1703 local_irq_disable();
1704 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1709 task_numa_assign(env, cur, imp);
1714 static void task_numa_find_cpu(struct task_numa_env *env,
1715 long taskimp, long groupimp)
1717 long src_load, dst_load, load;
1718 bool maymove = false;
1721 load = task_h_load(env->p);
1722 dst_load = env->dst_stats.load + load;
1723 src_load = env->src_stats.load - load;
1726 * If the improvement from just moving env->p direction is better
1727 * than swapping tasks around, check if a move is possible.
1729 maymove = !load_too_imbalanced(src_load, dst_load, env);
1731 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1732 /* Skip this CPU if the source task cannot migrate */
1733 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
1737 task_numa_compare(env, taskimp, groupimp, maymove);
1741 static int task_numa_migrate(struct task_struct *p)
1743 struct task_numa_env env = {
1746 .src_cpu = task_cpu(p),
1747 .src_nid = task_node(p),
1749 .imbalance_pct = 112,
1755 unsigned long taskweight, groupweight;
1756 struct sched_domain *sd;
1757 long taskimp, groupimp;
1758 struct numa_group *ng;
1763 * Pick the lowest SD_NUMA domain, as that would have the smallest
1764 * imbalance and would be the first to start moving tasks about.
1766 * And we want to avoid any moving of tasks about, as that would create
1767 * random movement of tasks -- counter the numa conditions we're trying
1771 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1773 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1777 * Cpusets can break the scheduler domain tree into smaller
1778 * balance domains, some of which do not cross NUMA boundaries.
1779 * Tasks that are "trapped" in such domains cannot be migrated
1780 * elsewhere, so there is no point in (re)trying.
1782 if (unlikely(!sd)) {
1783 sched_setnuma(p, task_node(p));
1787 env.dst_nid = p->numa_preferred_nid;
1788 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1789 taskweight = task_weight(p, env.src_nid, dist);
1790 groupweight = group_weight(p, env.src_nid, dist);
1791 update_numa_stats(&env.src_stats, env.src_nid);
1792 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1793 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1794 update_numa_stats(&env.dst_stats, env.dst_nid);
1796 /* Try to find a spot on the preferred nid. */
1797 task_numa_find_cpu(&env, taskimp, groupimp);
1800 * Look at other nodes in these cases:
1801 * - there is no space available on the preferred_nid
1802 * - the task is part of a numa_group that is interleaved across
1803 * multiple NUMA nodes; in order to better consolidate the group,
1804 * we need to check other locations.
1806 ng = deref_curr_numa_group(p);
1807 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
1808 for_each_online_node(nid) {
1809 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1812 dist = node_distance(env.src_nid, env.dst_nid);
1813 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1815 taskweight = task_weight(p, env.src_nid, dist);
1816 groupweight = group_weight(p, env.src_nid, dist);
1819 /* Only consider nodes where both task and groups benefit */
1820 taskimp = task_weight(p, nid, dist) - taskweight;
1821 groupimp = group_weight(p, nid, dist) - groupweight;
1822 if (taskimp < 0 && groupimp < 0)
1827 update_numa_stats(&env.dst_stats, env.dst_nid);
1828 task_numa_find_cpu(&env, taskimp, groupimp);
1833 * If the task is part of a workload that spans multiple NUMA nodes,
1834 * and is migrating into one of the workload's active nodes, remember
1835 * this node as the task's preferred numa node, so the workload can
1837 * A task that migrated to a second choice node will be better off
1838 * trying for a better one later. Do not set the preferred node here.
1841 if (env.best_cpu == -1)
1844 nid = cpu_to_node(env.best_cpu);
1846 if (nid != p->numa_preferred_nid)
1847 sched_setnuma(p, nid);
1850 /* No better CPU than the current one was found. */
1851 if (env.best_cpu == -1)
1854 best_rq = cpu_rq(env.best_cpu);
1855 if (env.best_task == NULL) {
1856 ret = migrate_task_to(p, env.best_cpu);
1857 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1859 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1863 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
1864 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1867 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1868 put_task_struct(env.best_task);
1872 /* Attempt to migrate a task to a CPU on the preferred node. */
1873 static void numa_migrate_preferred(struct task_struct *p)
1875 unsigned long interval = HZ;
1877 /* This task has no NUMA fault statistics yet */
1878 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
1881 /* Periodically retry migrating the task to the preferred node */
1882 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1883 p->numa_migrate_retry = jiffies + interval;
1885 /* Success if task is already running on preferred CPU */
1886 if (task_node(p) == p->numa_preferred_nid)
1889 /* Otherwise, try migrate to a CPU on the preferred node */
1890 task_numa_migrate(p);
1894 * Find out how many nodes on the workload is actively running on. Do this by
1895 * tracking the nodes from which NUMA hinting faults are triggered. This can
1896 * be different from the set of nodes where the workload's memory is currently
1899 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1901 unsigned long faults, max_faults = 0;
1902 int nid, active_nodes = 0;
1904 for_each_online_node(nid) {
1905 faults = group_faults_cpu(numa_group, nid);
1906 if (faults > max_faults)
1907 max_faults = faults;
1910 for_each_online_node(nid) {
1911 faults = group_faults_cpu(numa_group, nid);
1912 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1916 numa_group->max_faults_cpu = max_faults;
1917 numa_group->active_nodes = active_nodes;
1921 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1922 * increments. The more local the fault statistics are, the higher the scan
1923 * period will be for the next scan window. If local/(local+remote) ratio is
1924 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1925 * the scan period will decrease. Aim for 70% local accesses.
1927 #define NUMA_PERIOD_SLOTS 10
1928 #define NUMA_PERIOD_THRESHOLD 7
1931 * Increase the scan period (slow down scanning) if the majority of
1932 * our memory is already on our local node, or if the majority of
1933 * the page accesses are shared with other processes.
1934 * Otherwise, decrease the scan period.
1936 static void update_task_scan_period(struct task_struct *p,
1937 unsigned long shared, unsigned long private)
1939 unsigned int period_slot;
1940 int lr_ratio, ps_ratio;
1943 unsigned long remote = p->numa_faults_locality[0];
1944 unsigned long local = p->numa_faults_locality[1];
1947 * If there were no record hinting faults then either the task is
1948 * completely idle or all activity is areas that are not of interest
1949 * to automatic numa balancing. Related to that, if there were failed
1950 * migration then it implies we are migrating too quickly or the local
1951 * node is overloaded. In either case, scan slower
1953 if (local + shared == 0 || p->numa_faults_locality[2]) {
1954 p->numa_scan_period = min(p->numa_scan_period_max,
1955 p->numa_scan_period << 1);
1957 p->mm->numa_next_scan = jiffies +
1958 msecs_to_jiffies(p->numa_scan_period);
1964 * Prepare to scale scan period relative to the current period.
1965 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1966 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1967 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1969 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1970 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1971 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1973 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1975 * Most memory accesses are local. There is no need to
1976 * do fast NUMA scanning, since memory is already local.
1978 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
1981 diff = slot * period_slot;
1982 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
1984 * Most memory accesses are shared with other tasks.
1985 * There is no point in continuing fast NUMA scanning,
1986 * since other tasks may just move the memory elsewhere.
1988 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
1991 diff = slot * period_slot;
1994 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
1995 * yet they are not on the local NUMA node. Speed up
1996 * NUMA scanning to get the memory moved over.
1998 int ratio = max(lr_ratio, ps_ratio);
1999 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2002 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2003 task_scan_min(p), task_scan_max(p));
2004 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2008 * Get the fraction of time the task has been running since the last
2009 * NUMA placement cycle. The scheduler keeps similar statistics, but
2010 * decays those on a 32ms period, which is orders of magnitude off
2011 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2012 * stats only if the task is so new there are no NUMA statistics yet.
2014 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2016 u64 runtime, delta, now;
2017 /* Use the start of this time slice to avoid calculations. */
2018 now = p->se.exec_start;
2019 runtime = p->se.sum_exec_runtime;
2021 if (p->last_task_numa_placement) {
2022 delta = runtime - p->last_sum_exec_runtime;
2023 *period = now - p->last_task_numa_placement;
2025 /* Avoid time going backwards, prevent potential divide error: */
2026 if (unlikely((s64)*period < 0))
2029 delta = p->se.avg.load_sum;
2030 *period = LOAD_AVG_MAX;
2033 p->last_sum_exec_runtime = runtime;
2034 p->last_task_numa_placement = now;
2040 * Determine the preferred nid for a task in a numa_group. This needs to
2041 * be done in a way that produces consistent results with group_weight,
2042 * otherwise workloads might not converge.
2044 static int preferred_group_nid(struct task_struct *p, int nid)
2049 /* Direct connections between all NUMA nodes. */
2050 if (sched_numa_topology_type == NUMA_DIRECT)
2054 * On a system with glueless mesh NUMA topology, group_weight
2055 * scores nodes according to the number of NUMA hinting faults on
2056 * both the node itself, and on nearby nodes.
2058 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2059 unsigned long score, max_score = 0;
2060 int node, max_node = nid;
2062 dist = sched_max_numa_distance;
2064 for_each_online_node(node) {
2065 score = group_weight(p, node, dist);
2066 if (score > max_score) {
2075 * Finding the preferred nid in a system with NUMA backplane
2076 * interconnect topology is more involved. The goal is to locate
2077 * tasks from numa_groups near each other in the system, and
2078 * untangle workloads from different sides of the system. This requires
2079 * searching down the hierarchy of node groups, recursively searching
2080 * inside the highest scoring group of nodes. The nodemask tricks
2081 * keep the complexity of the search down.
2083 nodes = node_online_map;
2084 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2085 unsigned long max_faults = 0;
2086 nodemask_t max_group = NODE_MASK_NONE;
2089 /* Are there nodes at this distance from each other? */
2090 if (!find_numa_distance(dist))
2093 for_each_node_mask(a, nodes) {
2094 unsigned long faults = 0;
2095 nodemask_t this_group;
2096 nodes_clear(this_group);
2098 /* Sum group's NUMA faults; includes a==b case. */
2099 for_each_node_mask(b, nodes) {
2100 if (node_distance(a, b) < dist) {
2101 faults += group_faults(p, b);
2102 node_set(b, this_group);
2103 node_clear(b, nodes);
2107 /* Remember the top group. */
2108 if (faults > max_faults) {
2109 max_faults = faults;
2110 max_group = this_group;
2112 * subtle: at the smallest distance there is
2113 * just one node left in each "group", the
2114 * winner is the preferred nid.
2119 /* Next round, evaluate the nodes within max_group. */
2127 static void task_numa_placement(struct task_struct *p)
2129 int seq, nid, max_nid = NUMA_NO_NODE;
2130 unsigned long max_faults = 0;
2131 unsigned long fault_types[2] = { 0, 0 };
2132 unsigned long total_faults;
2133 u64 runtime, period;
2134 spinlock_t *group_lock = NULL;
2135 struct numa_group *ng;
2138 * The p->mm->numa_scan_seq field gets updated without
2139 * exclusive access. Use READ_ONCE() here to ensure
2140 * that the field is read in a single access:
2142 seq = READ_ONCE(p->mm->numa_scan_seq);
2143 if (p->numa_scan_seq == seq)
2145 p->numa_scan_seq = seq;
2146 p->numa_scan_period_max = task_scan_max(p);
2148 total_faults = p->numa_faults_locality[0] +
2149 p->numa_faults_locality[1];
2150 runtime = numa_get_avg_runtime(p, &period);
2152 /* If the task is part of a group prevent parallel updates to group stats */
2153 ng = deref_curr_numa_group(p);
2155 group_lock = &ng->lock;
2156 spin_lock_irq(group_lock);
2159 /* Find the node with the highest number of faults */
2160 for_each_online_node(nid) {
2161 /* Keep track of the offsets in numa_faults array */
2162 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2163 unsigned long faults = 0, group_faults = 0;
2166 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2167 long diff, f_diff, f_weight;
2169 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2170 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2171 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2172 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2174 /* Decay existing window, copy faults since last scan */
2175 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2176 fault_types[priv] += p->numa_faults[membuf_idx];
2177 p->numa_faults[membuf_idx] = 0;
2180 * Normalize the faults_from, so all tasks in a group
2181 * count according to CPU use, instead of by the raw
2182 * number of faults. Tasks with little runtime have
2183 * little over-all impact on throughput, and thus their
2184 * faults are less important.
2186 f_weight = div64_u64(runtime << 16, period + 1);
2187 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2189 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2190 p->numa_faults[cpubuf_idx] = 0;
2192 p->numa_faults[mem_idx] += diff;
2193 p->numa_faults[cpu_idx] += f_diff;
2194 faults += p->numa_faults[mem_idx];
2195 p->total_numa_faults += diff;
2198 * safe because we can only change our own group
2200 * mem_idx represents the offset for a given
2201 * nid and priv in a specific region because it
2202 * is at the beginning of the numa_faults array.
2204 ng->faults[mem_idx] += diff;
2205 ng->faults_cpu[mem_idx] += f_diff;
2206 ng->total_faults += diff;
2207 group_faults += ng->faults[mem_idx];
2212 if (faults > max_faults) {
2213 max_faults = faults;
2216 } else if (group_faults > max_faults) {
2217 max_faults = group_faults;
2223 numa_group_count_active_nodes(ng);
2224 spin_unlock_irq(group_lock);
2225 max_nid = preferred_group_nid(p, max_nid);
2229 /* Set the new preferred node */
2230 if (max_nid != p->numa_preferred_nid)
2231 sched_setnuma(p, max_nid);
2234 update_task_scan_period(p, fault_types[0], fault_types[1]);
2237 static inline int get_numa_group(struct numa_group *grp)
2239 return refcount_inc_not_zero(&grp->refcount);
2242 static inline void put_numa_group(struct numa_group *grp)
2244 if (refcount_dec_and_test(&grp->refcount))
2245 kfree_rcu(grp, rcu);
2248 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2251 struct numa_group *grp, *my_grp;
2252 struct task_struct *tsk;
2254 int cpu = cpupid_to_cpu(cpupid);
2257 if (unlikely(!deref_curr_numa_group(p))) {
2258 unsigned int size = sizeof(struct numa_group) +
2259 4*nr_node_ids*sizeof(unsigned long);
2261 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2265 refcount_set(&grp->refcount, 1);
2266 grp->active_nodes = 1;
2267 grp->max_faults_cpu = 0;
2268 spin_lock_init(&grp->lock);
2270 /* Second half of the array tracks nids where faults happen */
2271 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2274 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2275 grp->faults[i] = p->numa_faults[i];
2277 grp->total_faults = p->total_numa_faults;
2280 rcu_assign_pointer(p->numa_group, grp);
2284 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2286 if (!cpupid_match_pid(tsk, cpupid))
2289 grp = rcu_dereference(tsk->numa_group);
2293 my_grp = deref_curr_numa_group(p);
2298 * Only join the other group if its bigger; if we're the bigger group,
2299 * the other task will join us.
2301 if (my_grp->nr_tasks > grp->nr_tasks)
2305 * Tie-break on the grp address.
2307 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2310 /* Always join threads in the same process. */
2311 if (tsk->mm == current->mm)
2314 /* Simple filter to avoid false positives due to PID collisions */
2315 if (flags & TNF_SHARED)
2318 /* Update priv based on whether false sharing was detected */
2321 if (join && !get_numa_group(grp))
2329 BUG_ON(irqs_disabled());
2330 double_lock_irq(&my_grp->lock, &grp->lock);
2332 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2333 my_grp->faults[i] -= p->numa_faults[i];
2334 grp->faults[i] += p->numa_faults[i];
2336 my_grp->total_faults -= p->total_numa_faults;
2337 grp->total_faults += p->total_numa_faults;
2342 spin_unlock(&my_grp->lock);
2343 spin_unlock_irq(&grp->lock);
2345 rcu_assign_pointer(p->numa_group, grp);
2347 put_numa_group(my_grp);
2356 * Get rid of NUMA staticstics associated with a task (either current or dead).
2357 * If @final is set, the task is dead and has reached refcount zero, so we can
2358 * safely free all relevant data structures. Otherwise, there might be
2359 * concurrent reads from places like load balancing and procfs, and we should
2360 * reset the data back to default state without freeing ->numa_faults.
2362 void task_numa_free(struct task_struct *p, bool final)
2364 /* safe: p either is current or is being freed by current */
2365 struct numa_group *grp = rcu_dereference_raw(p->numa_group);
2366 unsigned long *numa_faults = p->numa_faults;
2367 unsigned long flags;
2374 spin_lock_irqsave(&grp->lock, flags);
2375 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2376 grp->faults[i] -= p->numa_faults[i];
2377 grp->total_faults -= p->total_numa_faults;
2380 spin_unlock_irqrestore(&grp->lock, flags);
2381 RCU_INIT_POINTER(p->numa_group, NULL);
2382 put_numa_group(grp);
2386 p->numa_faults = NULL;
2389 p->total_numa_faults = 0;
2390 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2396 * Got a PROT_NONE fault for a page on @node.
2398 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2400 struct task_struct *p = current;
2401 bool migrated = flags & TNF_MIGRATED;
2402 int cpu_node = task_node(current);
2403 int local = !!(flags & TNF_FAULT_LOCAL);
2404 struct numa_group *ng;
2407 if (!static_branch_likely(&sched_numa_balancing))
2410 /* for example, ksmd faulting in a user's mm */
2414 /* Allocate buffer to track faults on a per-node basis */
2415 if (unlikely(!p->numa_faults)) {
2416 int size = sizeof(*p->numa_faults) *
2417 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2419 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2420 if (!p->numa_faults)
2423 p->total_numa_faults = 0;
2424 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2428 * First accesses are treated as private, otherwise consider accesses
2429 * to be private if the accessing pid has not changed
2431 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2434 priv = cpupid_match_pid(p, last_cpupid);
2435 if (!priv && !(flags & TNF_NO_GROUP))
2436 task_numa_group(p, last_cpupid, flags, &priv);
2440 * If a workload spans multiple NUMA nodes, a shared fault that
2441 * occurs wholly within the set of nodes that the workload is
2442 * actively using should be counted as local. This allows the
2443 * scan rate to slow down when a workload has settled down.
2445 ng = deref_curr_numa_group(p);
2446 if (!priv && !local && ng && ng->active_nodes > 1 &&
2447 numa_is_active_node(cpu_node, ng) &&
2448 numa_is_active_node(mem_node, ng))
2452 * Retry to migrate task to preferred node periodically, in case it
2453 * previously failed, or the scheduler moved us.
2455 if (time_after(jiffies, p->numa_migrate_retry)) {
2456 task_numa_placement(p);
2457 numa_migrate_preferred(p);
2461 p->numa_pages_migrated += pages;
2462 if (flags & TNF_MIGRATE_FAIL)
2463 p->numa_faults_locality[2] += pages;
2465 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2466 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2467 p->numa_faults_locality[local] += pages;
2470 static void reset_ptenuma_scan(struct task_struct *p)
2473 * We only did a read acquisition of the mmap sem, so
2474 * p->mm->numa_scan_seq is written to without exclusive access
2475 * and the update is not guaranteed to be atomic. That's not
2476 * much of an issue though, since this is just used for
2477 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2478 * expensive, to avoid any form of compiler optimizations:
2480 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2481 p->mm->numa_scan_offset = 0;
2485 * The expensive part of numa migration is done from task_work context.
2486 * Triggered from task_tick_numa().
2488 static void task_numa_work(struct callback_head *work)
2490 unsigned long migrate, next_scan, now = jiffies;
2491 struct task_struct *p = current;
2492 struct mm_struct *mm = p->mm;
2493 u64 runtime = p->se.sum_exec_runtime;
2494 struct vm_area_struct *vma;
2495 unsigned long start, end;
2496 unsigned long nr_pte_updates = 0;
2497 long pages, virtpages;
2499 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2503 * Who cares about NUMA placement when they're dying.
2505 * NOTE: make sure not to dereference p->mm before this check,
2506 * exit_task_work() happens _after_ exit_mm() so we could be called
2507 * without p->mm even though we still had it when we enqueued this
2510 if (p->flags & PF_EXITING)
2513 if (!mm->numa_next_scan) {
2514 mm->numa_next_scan = now +
2515 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2519 * Enforce maximal scan/migration frequency..
2521 migrate = mm->numa_next_scan;
2522 if (time_before(now, migrate))
2525 if (p->numa_scan_period == 0) {
2526 p->numa_scan_period_max = task_scan_max(p);
2527 p->numa_scan_period = task_scan_start(p);
2530 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2531 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2535 * Delay this task enough that another task of this mm will likely win
2536 * the next time around.
2538 p->node_stamp += 2 * TICK_NSEC;
2540 start = mm->numa_scan_offset;
2541 pages = sysctl_numa_balancing_scan_size;
2542 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2543 virtpages = pages * 8; /* Scan up to this much virtual space */
2548 if (!down_read_trylock(&mm->mmap_sem))
2550 vma = find_vma(mm, start);
2552 reset_ptenuma_scan(p);
2556 for (; vma; vma = vma->vm_next) {
2557 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2558 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2563 * Shared library pages mapped by multiple processes are not
2564 * migrated as it is expected they are cache replicated. Avoid
2565 * hinting faults in read-only file-backed mappings or the vdso
2566 * as migrating the pages will be of marginal benefit.
2569 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2573 * Skip inaccessible VMAs to avoid any confusion between
2574 * PROT_NONE and NUMA hinting ptes
2576 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2580 start = max(start, vma->vm_start);
2581 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2582 end = min(end, vma->vm_end);
2583 nr_pte_updates = change_prot_numa(vma, start, end);
2586 * Try to scan sysctl_numa_balancing_size worth of
2587 * hpages that have at least one present PTE that
2588 * is not already pte-numa. If the VMA contains
2589 * areas that are unused or already full of prot_numa
2590 * PTEs, scan up to virtpages, to skip through those
2594 pages -= (end - start) >> PAGE_SHIFT;
2595 virtpages -= (end - start) >> PAGE_SHIFT;
2598 if (pages <= 0 || virtpages <= 0)
2602 } while (end != vma->vm_end);
2607 * It is possible to reach the end of the VMA list but the last few
2608 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2609 * would find the !migratable VMA on the next scan but not reset the
2610 * scanner to the start so check it now.
2613 mm->numa_scan_offset = start;
2615 reset_ptenuma_scan(p);
2616 up_read(&mm->mmap_sem);
2619 * Make sure tasks use at least 32x as much time to run other code
2620 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2621 * Usually update_task_scan_period slows down scanning enough; on an
2622 * overloaded system we need to limit overhead on a per task basis.
2624 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2625 u64 diff = p->se.sum_exec_runtime - runtime;
2626 p->node_stamp += 32 * diff;
2630 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
2633 struct mm_struct *mm = p->mm;
2636 mm_users = atomic_read(&mm->mm_users);
2637 if (mm_users == 1) {
2638 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2639 mm->numa_scan_seq = 0;
2643 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
2644 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2645 /* Protect against double add, see task_tick_numa and task_numa_work */
2646 p->numa_work.next = &p->numa_work;
2647 p->numa_faults = NULL;
2648 RCU_INIT_POINTER(p->numa_group, NULL);
2649 p->last_task_numa_placement = 0;
2650 p->last_sum_exec_runtime = 0;
2652 init_task_work(&p->numa_work, task_numa_work);
2654 /* New address space, reset the preferred nid */
2655 if (!(clone_flags & CLONE_VM)) {
2656 p->numa_preferred_nid = NUMA_NO_NODE;
2661 * New thread, keep existing numa_preferred_nid which should be copied
2662 * already by arch_dup_task_struct but stagger when scans start.
2667 delay = min_t(unsigned int, task_scan_max(current),
2668 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
2669 delay += 2 * TICK_NSEC;
2670 p->node_stamp = delay;
2675 * Drive the periodic memory faults..
2677 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2679 struct callback_head *work = &curr->numa_work;
2683 * We don't care about NUMA placement if we don't have memory.
2685 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2689 * Using runtime rather than walltime has the dual advantage that
2690 * we (mostly) drive the selection from busy threads and that the
2691 * task needs to have done some actual work before we bother with
2694 now = curr->se.sum_exec_runtime;
2695 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2697 if (now > curr->node_stamp + period) {
2698 if (!curr->node_stamp)
2699 curr->numa_scan_period = task_scan_start(curr);
2700 curr->node_stamp += period;
2702 if (!time_before(jiffies, curr->mm->numa_next_scan))
2703 task_work_add(curr, work, true);
2707 static void update_scan_period(struct task_struct *p, int new_cpu)
2709 int src_nid = cpu_to_node(task_cpu(p));
2710 int dst_nid = cpu_to_node(new_cpu);
2712 if (!static_branch_likely(&sched_numa_balancing))
2715 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2718 if (src_nid == dst_nid)
2722 * Allow resets if faults have been trapped before one scan
2723 * has completed. This is most likely due to a new task that
2724 * is pulled cross-node due to wakeups or load balancing.
2726 if (p->numa_scan_seq) {
2728 * Avoid scan adjustments if moving to the preferred
2729 * node or if the task was not previously running on
2730 * the preferred node.
2732 if (dst_nid == p->numa_preferred_nid ||
2733 (p->numa_preferred_nid != NUMA_NO_NODE &&
2734 src_nid != p->numa_preferred_nid))
2738 p->numa_scan_period = task_scan_start(p);
2742 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2746 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2750 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2754 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2758 #endif /* CONFIG_NUMA_BALANCING */
2761 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2763 update_load_add(&cfs_rq->load, se->load.weight);
2765 if (entity_is_task(se)) {
2766 struct rq *rq = rq_of(cfs_rq);
2768 account_numa_enqueue(rq, task_of(se));
2769 list_add(&se->group_node, &rq->cfs_tasks);
2772 cfs_rq->nr_running++;
2776 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2778 update_load_sub(&cfs_rq->load, se->load.weight);
2780 if (entity_is_task(se)) {
2781 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2782 list_del_init(&se->group_node);
2785 cfs_rq->nr_running--;
2789 * Signed add and clamp on underflow.
2791 * Explicitly do a load-store to ensure the intermediate value never hits
2792 * memory. This allows lockless observations without ever seeing the negative
2795 #define add_positive(_ptr, _val) do { \
2796 typeof(_ptr) ptr = (_ptr); \
2797 typeof(_val) val = (_val); \
2798 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2802 if (val < 0 && res > var) \
2805 WRITE_ONCE(*ptr, res); \
2809 * Unsigned subtract and clamp on underflow.
2811 * Explicitly do a load-store to ensure the intermediate value never hits
2812 * memory. This allows lockless observations without ever seeing the negative
2815 #define sub_positive(_ptr, _val) do { \
2816 typeof(_ptr) ptr = (_ptr); \
2817 typeof(*ptr) val = (_val); \
2818 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2822 WRITE_ONCE(*ptr, res); \
2826 * Remove and clamp on negative, from a local variable.
2828 * A variant of sub_positive(), which does not use explicit load-store
2829 * and is thus optimized for local variable updates.
2831 #define lsub_positive(_ptr, _val) do { \
2832 typeof(_ptr) ptr = (_ptr); \
2833 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
2838 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2840 cfs_rq->runnable_weight += se->runnable_weight;
2842 cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2843 cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2847 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2849 cfs_rq->runnable_weight -= se->runnable_weight;
2851 sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2852 sub_positive(&cfs_rq->avg.runnable_load_sum,
2853 se_runnable(se) * se->avg.runnable_load_sum);
2857 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2859 cfs_rq->avg.load_avg += se->avg.load_avg;
2860 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2864 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2866 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2867 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2871 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2873 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2875 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2877 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2880 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2881 unsigned long weight, unsigned long runnable)
2884 /* commit outstanding execution time */
2885 if (cfs_rq->curr == se)
2886 update_curr(cfs_rq);
2887 account_entity_dequeue(cfs_rq, se);
2888 dequeue_runnable_load_avg(cfs_rq, se);
2890 dequeue_load_avg(cfs_rq, se);
2892 se->runnable_weight = runnable;
2893 update_load_set(&se->load, weight);
2897 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2899 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2900 se->avg.runnable_load_avg =
2901 div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2905 enqueue_load_avg(cfs_rq, se);
2907 account_entity_enqueue(cfs_rq, se);
2908 enqueue_runnable_load_avg(cfs_rq, se);
2912 void reweight_task(struct task_struct *p, int prio)
2914 struct sched_entity *se = &p->se;
2915 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2916 struct load_weight *load = &se->load;
2917 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2919 reweight_entity(cfs_rq, se, weight, weight);
2920 load->inv_weight = sched_prio_to_wmult[prio];
2923 #ifdef CONFIG_FAIR_GROUP_SCHED
2926 * All this does is approximate the hierarchical proportion which includes that
2927 * global sum we all love to hate.
2929 * That is, the weight of a group entity, is the proportional share of the
2930 * group weight based on the group runqueue weights. That is:
2932 * tg->weight * grq->load.weight
2933 * ge->load.weight = ----------------------------- (1)
2934 * \Sum grq->load.weight
2936 * Now, because computing that sum is prohibitively expensive to compute (been
2937 * there, done that) we approximate it with this average stuff. The average
2938 * moves slower and therefore the approximation is cheaper and more stable.
2940 * So instead of the above, we substitute:
2942 * grq->load.weight -> grq->avg.load_avg (2)
2944 * which yields the following:
2946 * tg->weight * grq->avg.load_avg
2947 * ge->load.weight = ------------------------------ (3)
2950 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2952 * That is shares_avg, and it is right (given the approximation (2)).
2954 * The problem with it is that because the average is slow -- it was designed
2955 * to be exactly that of course -- this leads to transients in boundary
2956 * conditions. In specific, the case where the group was idle and we start the
2957 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2958 * yielding bad latency etc..
2960 * Now, in that special case (1) reduces to:
2962 * tg->weight * grq->load.weight
2963 * ge->load.weight = ----------------------------- = tg->weight (4)
2966 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2968 * So what we do is modify our approximation (3) to approach (4) in the (near)
2973 * tg->weight * grq->load.weight
2974 * --------------------------------------------------- (5)
2975 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2977 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2978 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2981 * tg->weight * grq->load.weight
2982 * ge->load.weight = ----------------------------- (6)
2987 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2988 * max(grq->load.weight, grq->avg.load_avg)
2990 * And that is shares_weight and is icky. In the (near) UP case it approaches
2991 * (4) while in the normal case it approaches (3). It consistently
2992 * overestimates the ge->load.weight and therefore:
2994 * \Sum ge->load.weight >= tg->weight
2998 static long calc_group_shares(struct cfs_rq *cfs_rq)
3000 long tg_weight, tg_shares, load, shares;
3001 struct task_group *tg = cfs_rq->tg;
3003 tg_shares = READ_ONCE(tg->shares);
3005 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3007 tg_weight = atomic_long_read(&tg->load_avg);
3009 /* Ensure tg_weight >= load */
3010 tg_weight -= cfs_rq->tg_load_avg_contrib;
3013 shares = (tg_shares * load);
3015 shares /= tg_weight;
3018 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3019 * of a group with small tg->shares value. It is a floor value which is
3020 * assigned as a minimum load.weight to the sched_entity representing
3021 * the group on a CPU.
3023 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3024 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3025 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3026 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3029 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3033 * This calculates the effective runnable weight for a group entity based on
3034 * the group entity weight calculated above.
3036 * Because of the above approximation (2), our group entity weight is
3037 * an load_avg based ratio (3). This means that it includes blocked load and
3038 * does not represent the runnable weight.
3040 * Approximate the group entity's runnable weight per ratio from the group
3043 * grq->avg.runnable_load_avg
3044 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
3047 * However, analogous to above, since the avg numbers are slow, this leads to
3048 * transients in the from-idle case. Instead we use:
3050 * ge->runnable_weight = ge->load.weight *
3052 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
3053 * ----------------------------------------------------- (8)
3054 * max(grq->avg.load_avg, grq->load.weight)
3056 * Where these max() serve both to use the 'instant' values to fix the slow
3057 * from-idle and avoid the /0 on to-idle, similar to (6).
3059 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
3061 long runnable, load_avg;
3063 load_avg = max(cfs_rq->avg.load_avg,
3064 scale_load_down(cfs_rq->load.weight));
3066 runnable = max(cfs_rq->avg.runnable_load_avg,
3067 scale_load_down(cfs_rq->runnable_weight));
3071 runnable /= load_avg;
3073 return clamp_t(long, runnable, MIN_SHARES, shares);
3075 #endif /* CONFIG_SMP */
3077 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3080 * Recomputes the group entity based on the current state of its group
3083 static void update_cfs_group(struct sched_entity *se)
3085 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3086 long shares, runnable;
3091 if (throttled_hierarchy(gcfs_rq))
3095 runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
3097 if (likely(se->load.weight == shares))
3100 shares = calc_group_shares(gcfs_rq);
3101 runnable = calc_group_runnable(gcfs_rq, shares);
3104 reweight_entity(cfs_rq_of(se), se, shares, runnable);
3107 #else /* CONFIG_FAIR_GROUP_SCHED */
3108 static inline void update_cfs_group(struct sched_entity *se)
3111 #endif /* CONFIG_FAIR_GROUP_SCHED */
3113 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3115 struct rq *rq = rq_of(cfs_rq);
3117 if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) {
3119 * There are a few boundary cases this might miss but it should
3120 * get called often enough that that should (hopefully) not be
3123 * It will not get called when we go idle, because the idle
3124 * thread is a different class (!fair), nor will the utilization
3125 * number include things like RT tasks.
3127 * As is, the util number is not freq-invariant (we'd have to
3128 * implement arch_scale_freq_capacity() for that).
3132 cpufreq_update_util(rq, flags);
3137 #ifdef CONFIG_FAIR_GROUP_SCHED
3139 * update_tg_load_avg - update the tg's load avg
3140 * @cfs_rq: the cfs_rq whose avg changed
3141 * @force: update regardless of how small the difference
3143 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3144 * However, because tg->load_avg is a global value there are performance
3147 * In order to avoid having to look at the other cfs_rq's, we use a
3148 * differential update where we store the last value we propagated. This in
3149 * turn allows skipping updates if the differential is 'small'.
3151 * Updating tg's load_avg is necessary before update_cfs_share().
3153 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3155 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3158 * No need to update load_avg for root_task_group as it is not used.
3160 if (cfs_rq->tg == &root_task_group)
3163 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3164 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3165 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3170 * Called within set_task_rq() right before setting a task's CPU. The
3171 * caller only guarantees p->pi_lock is held; no other assumptions,
3172 * including the state of rq->lock, should be made.
3174 void set_task_rq_fair(struct sched_entity *se,
3175 struct cfs_rq *prev, struct cfs_rq *next)
3177 u64 p_last_update_time;
3178 u64 n_last_update_time;
3180 if (!sched_feat(ATTACH_AGE_LOAD))
3184 * We are supposed to update the task to "current" time, then its up to
3185 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3186 * getting what current time is, so simply throw away the out-of-date
3187 * time. This will result in the wakee task is less decayed, but giving
3188 * the wakee more load sounds not bad.
3190 if (!(se->avg.last_update_time && prev))
3193 #ifndef CONFIG_64BIT
3195 u64 p_last_update_time_copy;
3196 u64 n_last_update_time_copy;
3199 p_last_update_time_copy = prev->load_last_update_time_copy;
3200 n_last_update_time_copy = next->load_last_update_time_copy;
3204 p_last_update_time = prev->avg.last_update_time;
3205 n_last_update_time = next->avg.last_update_time;
3207 } while (p_last_update_time != p_last_update_time_copy ||
3208 n_last_update_time != n_last_update_time_copy);
3211 p_last_update_time = prev->avg.last_update_time;
3212 n_last_update_time = next->avg.last_update_time;
3214 __update_load_avg_blocked_se(p_last_update_time, se);
3215 se->avg.last_update_time = n_last_update_time;
3220 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3221 * propagate its contribution. The key to this propagation is the invariant
3222 * that for each group:
3224 * ge->avg == grq->avg (1)
3226 * _IFF_ we look at the pure running and runnable sums. Because they
3227 * represent the very same entity, just at different points in the hierarchy.
3229 * Per the above update_tg_cfs_util() is trivial and simply copies the running
3230 * sum over (but still wrong, because the group entity and group rq do not have
3231 * their PELT windows aligned).
3233 * However, update_tg_cfs_runnable() is more complex. So we have:
3235 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3237 * And since, like util, the runnable part should be directly transferable,
3238 * the following would _appear_ to be the straight forward approach:
3240 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3242 * And per (1) we have:
3244 * ge->avg.runnable_avg == grq->avg.runnable_avg
3248 * ge->load.weight * grq->avg.load_avg
3249 * ge->avg.load_avg = ----------------------------------- (4)
3252 * Except that is wrong!
3254 * Because while for entities historical weight is not important and we
3255 * really only care about our future and therefore can consider a pure
3256 * runnable sum, runqueues can NOT do this.
3258 * We specifically want runqueues to have a load_avg that includes
3259 * historical weights. Those represent the blocked load, the load we expect
3260 * to (shortly) return to us. This only works by keeping the weights as
3261 * integral part of the sum. We therefore cannot decompose as per (3).
3263 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3264 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3265 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3266 * runnable section of these tasks overlap (or not). If they were to perfectly
3267 * align the rq as a whole would be runnable 2/3 of the time. If however we
3268 * always have at least 1 runnable task, the rq as a whole is always runnable.
3270 * So we'll have to approximate.. :/
3272 * Given the constraint:
3274 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3276 * We can construct a rule that adds runnable to a rq by assuming minimal
3279 * On removal, we'll assume each task is equally runnable; which yields:
3281 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3283 * XXX: only do this for the part of runnable > running ?
3288 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3290 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3292 /* Nothing to update */
3297 * The relation between sum and avg is:
3299 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3301 * however, the PELT windows are not aligned between grq and gse.
3304 /* Set new sched_entity's utilization */
3305 se->avg.util_avg = gcfs_rq->avg.util_avg;
3306 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3308 /* Update parent cfs_rq utilization */
3309 add_positive(&cfs_rq->avg.util_avg, delta);
3310 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3314 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3316 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3317 unsigned long runnable_load_avg, load_avg;
3318 u64 runnable_load_sum, load_sum = 0;
3324 gcfs_rq->prop_runnable_sum = 0;
3326 if (runnable_sum >= 0) {
3328 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3329 * the CPU is saturated running == runnable.
3331 runnable_sum += se->avg.load_sum;
3332 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3335 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3336 * assuming all tasks are equally runnable.
3338 if (scale_load_down(gcfs_rq->load.weight)) {
3339 load_sum = div_s64(gcfs_rq->avg.load_sum,
3340 scale_load_down(gcfs_rq->load.weight));
3343 /* But make sure to not inflate se's runnable */
3344 runnable_sum = min(se->avg.load_sum, load_sum);
3348 * runnable_sum can't be lower than running_sum
3349 * Rescale running sum to be in the same range as runnable sum
3350 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3351 * runnable_sum is in [0 : LOAD_AVG_MAX]
3353 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3354 runnable_sum = max(runnable_sum, running_sum);
3356 load_sum = (s64)se_weight(se) * runnable_sum;
3357 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3359 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3360 delta_avg = load_avg - se->avg.load_avg;
3362 se->avg.load_sum = runnable_sum;
3363 se->avg.load_avg = load_avg;
3364 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3365 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3367 runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3368 runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3369 delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3370 delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3372 se->avg.runnable_load_sum = runnable_sum;
3373 se->avg.runnable_load_avg = runnable_load_avg;
3376 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3377 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3381 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3383 cfs_rq->propagate = 1;
3384 cfs_rq->prop_runnable_sum += runnable_sum;
3387 /* Update task and its cfs_rq load average */
3388 static inline int propagate_entity_load_avg(struct sched_entity *se)
3390 struct cfs_rq *cfs_rq, *gcfs_rq;
3392 if (entity_is_task(se))
3395 gcfs_rq = group_cfs_rq(se);
3396 if (!gcfs_rq->propagate)
3399 gcfs_rq->propagate = 0;
3401 cfs_rq = cfs_rq_of(se);
3403 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3405 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3406 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3408 trace_pelt_cfs_tp(cfs_rq);
3409 trace_pelt_se_tp(se);
3415 * Check if we need to update the load and the utilization of a blocked
3418 static inline bool skip_blocked_update(struct sched_entity *se)
3420 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3423 * If sched_entity still have not zero load or utilization, we have to
3426 if (se->avg.load_avg || se->avg.util_avg)
3430 * If there is a pending propagation, we have to update the load and
3431 * the utilization of the sched_entity:
3433 if (gcfs_rq->propagate)
3437 * Otherwise, the load and the utilization of the sched_entity is
3438 * already zero and there is no pending propagation, so it will be a
3439 * waste of time to try to decay it:
3444 #else /* CONFIG_FAIR_GROUP_SCHED */
3446 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3448 static inline int propagate_entity_load_avg(struct sched_entity *se)
3453 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3455 #endif /* CONFIG_FAIR_GROUP_SCHED */
3458 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3459 * @now: current time, as per cfs_rq_clock_pelt()
3460 * @cfs_rq: cfs_rq to update
3462 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3463 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3464 * post_init_entity_util_avg().
3466 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3468 * Returns true if the load decayed or we removed load.
3470 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3471 * call update_tg_load_avg() when this function returns true.
3474 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3476 unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3477 struct sched_avg *sa = &cfs_rq->avg;
3480 if (cfs_rq->removed.nr) {
3482 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3484 raw_spin_lock(&cfs_rq->removed.lock);
3485 swap(cfs_rq->removed.util_avg, removed_util);
3486 swap(cfs_rq->removed.load_avg, removed_load);
3487 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3488 cfs_rq->removed.nr = 0;
3489 raw_spin_unlock(&cfs_rq->removed.lock);
3492 sub_positive(&sa->load_avg, r);
3493 sub_positive(&sa->load_sum, r * divider);
3496 sub_positive(&sa->util_avg, r);
3497 sub_positive(&sa->util_sum, r * divider);
3499 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3504 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
3506 #ifndef CONFIG_64BIT
3508 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3515 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3516 * @cfs_rq: cfs_rq to attach to
3517 * @se: sched_entity to attach
3518 * @flags: migration hints
3520 * Must call update_cfs_rq_load_avg() before this, since we rely on
3521 * cfs_rq->avg.last_update_time being current.
3523 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3525 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3528 * When we attach the @se to the @cfs_rq, we must align the decay
3529 * window because without that, really weird and wonderful things can
3534 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3535 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3538 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3539 * period_contrib. This isn't strictly correct, but since we're
3540 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3543 se->avg.util_sum = se->avg.util_avg * divider;
3545 se->avg.load_sum = divider;
3546 if (se_weight(se)) {
3548 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3551 se->avg.runnable_load_sum = se->avg.load_sum;
3553 enqueue_load_avg(cfs_rq, se);
3554 cfs_rq->avg.util_avg += se->avg.util_avg;
3555 cfs_rq->avg.util_sum += se->avg.util_sum;
3557 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3559 cfs_rq_util_change(cfs_rq, flags);
3561 trace_pelt_cfs_tp(cfs_rq);
3565 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3566 * @cfs_rq: cfs_rq to detach from
3567 * @se: sched_entity to detach
3569 * Must call update_cfs_rq_load_avg() before this, since we rely on
3570 * cfs_rq->avg.last_update_time being current.
3572 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3574 dequeue_load_avg(cfs_rq, se);
3575 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3576 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3578 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3580 cfs_rq_util_change(cfs_rq, 0);
3582 trace_pelt_cfs_tp(cfs_rq);
3586 * Optional action to be done while updating the load average
3588 #define UPDATE_TG 0x1
3589 #define SKIP_AGE_LOAD 0x2
3590 #define DO_ATTACH 0x4
3592 /* Update task and its cfs_rq load average */
3593 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3595 u64 now = cfs_rq_clock_pelt(cfs_rq);
3599 * Track task load average for carrying it to new CPU after migrated, and
3600 * track group sched_entity load average for task_h_load calc in migration
3602 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3603 __update_load_avg_se(now, cfs_rq, se);
3605 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3606 decayed |= propagate_entity_load_avg(se);
3608 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3611 * DO_ATTACH means we're here from enqueue_entity().
3612 * !last_update_time means we've passed through
3613 * migrate_task_rq_fair() indicating we migrated.
3615 * IOW we're enqueueing a task on a new CPU.
3617 attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
3618 update_tg_load_avg(cfs_rq, 0);
3620 } else if (decayed) {
3621 cfs_rq_util_change(cfs_rq, 0);
3623 if (flags & UPDATE_TG)
3624 update_tg_load_avg(cfs_rq, 0);
3628 #ifndef CONFIG_64BIT
3629 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3631 u64 last_update_time_copy;
3632 u64 last_update_time;
3635 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3637 last_update_time = cfs_rq->avg.last_update_time;
3638 } while (last_update_time != last_update_time_copy);
3640 return last_update_time;
3643 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3645 return cfs_rq->avg.last_update_time;
3650 * Synchronize entity load avg of dequeued entity without locking
3653 static void sync_entity_load_avg(struct sched_entity *se)
3655 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3656 u64 last_update_time;
3658 last_update_time = cfs_rq_last_update_time(cfs_rq);
3659 __update_load_avg_blocked_se(last_update_time, se);
3663 * Task first catches up with cfs_rq, and then subtract
3664 * itself from the cfs_rq (task must be off the queue now).
3666 static void remove_entity_load_avg(struct sched_entity *se)
3668 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3669 unsigned long flags;
3672 * tasks cannot exit without having gone through wake_up_new_task() ->
3673 * post_init_entity_util_avg() which will have added things to the
3674 * cfs_rq, so we can remove unconditionally.
3677 sync_entity_load_avg(se);
3679 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3680 ++cfs_rq->removed.nr;
3681 cfs_rq->removed.util_avg += se->avg.util_avg;
3682 cfs_rq->removed.load_avg += se->avg.load_avg;
3683 cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */
3684 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3687 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3689 return cfs_rq->avg.runnable_load_avg;
3692 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3694 return cfs_rq->avg.load_avg;
3697 static inline unsigned long task_util(struct task_struct *p)
3699 return READ_ONCE(p->se.avg.util_avg);
3702 static inline unsigned long _task_util_est(struct task_struct *p)
3704 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3706 return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED);
3709 static inline unsigned long task_util_est(struct task_struct *p)
3711 return max(task_util(p), _task_util_est(p));
3714 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3715 struct task_struct *p)
3717 unsigned int enqueued;
3719 if (!sched_feat(UTIL_EST))
3722 /* Update root cfs_rq's estimated utilization */
3723 enqueued = cfs_rq->avg.util_est.enqueued;
3724 enqueued += _task_util_est(p);
3725 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3729 * Check if a (signed) value is within a specified (unsigned) margin,
3730 * based on the observation that:
3732 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3734 * NOTE: this only works when value + maring < INT_MAX.
3736 static inline bool within_margin(int value, int margin)
3738 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3742 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3744 long last_ewma_diff;
3748 if (!sched_feat(UTIL_EST))
3751 /* Update root cfs_rq's estimated utilization */
3752 ue.enqueued = cfs_rq->avg.util_est.enqueued;
3753 ue.enqueued -= min_t(unsigned int, ue.enqueued, _task_util_est(p));
3754 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3757 * Skip update of task's estimated utilization when the task has not
3758 * yet completed an activation, e.g. being migrated.
3764 * If the PELT values haven't changed since enqueue time,
3765 * skip the util_est update.
3767 ue = p->se.avg.util_est;
3768 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3772 * Reset EWMA on utilization increases, the moving average is used only
3773 * to smooth utilization decreases.
3775 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3776 if (sched_feat(UTIL_EST_FASTUP)) {
3777 if (ue.ewma < ue.enqueued) {
3778 ue.ewma = ue.enqueued;
3784 * Skip update of task's estimated utilization when its EWMA is
3785 * already ~1% close to its last activation value.
3787 last_ewma_diff = ue.enqueued - ue.ewma;
3788 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3792 * To avoid overestimation of actual task utilization, skip updates if
3793 * we cannot grant there is idle time in this CPU.
3795 cpu = cpu_of(rq_of(cfs_rq));
3796 if (task_util(p) > capacity_orig_of(cpu))
3800 * Update Task's estimated utilization
3802 * When *p completes an activation we can consolidate another sample
3803 * of the task size. This is done by storing the current PELT value
3804 * as ue.enqueued and by using this value to update the Exponential
3805 * Weighted Moving Average (EWMA):
3807 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
3808 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
3809 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
3810 * = w * ( last_ewma_diff ) + ewma(t-1)
3811 * = w * (last_ewma_diff + ewma(t-1) / w)
3813 * Where 'w' is the weight of new samples, which is configured to be
3814 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3816 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
3817 ue.ewma += last_ewma_diff;
3818 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
3820 WRITE_ONCE(p->se.avg.util_est, ue);
3823 static inline int task_fits_capacity(struct task_struct *p, long capacity)
3825 return fits_capacity(task_util_est(p), capacity);
3828 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
3830 if (!static_branch_unlikely(&sched_asym_cpucapacity))
3834 rq->misfit_task_load = 0;
3838 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
3839 rq->misfit_task_load = 0;
3843 rq->misfit_task_load = task_h_load(p);
3846 #else /* CONFIG_SMP */
3848 #define UPDATE_TG 0x0
3849 #define SKIP_AGE_LOAD 0x0
3850 #define DO_ATTACH 0x0
3852 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3854 cfs_rq_util_change(cfs_rq, 0);
3857 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3860 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
3862 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3864 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3870 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
3873 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
3875 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
3877 #endif /* CONFIG_SMP */
3879 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3881 #ifdef CONFIG_SCHED_DEBUG
3882 s64 d = se->vruntime - cfs_rq->min_vruntime;
3887 if (d > 3*sysctl_sched_latency)
3888 schedstat_inc(cfs_rq->nr_spread_over);
3893 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3895 u64 vruntime = cfs_rq->min_vruntime;
3898 * The 'current' period is already promised to the current tasks,
3899 * however the extra weight of the new task will slow them down a
3900 * little, place the new task so that it fits in the slot that
3901 * stays open at the end.
3903 if (initial && sched_feat(START_DEBIT))
3904 vruntime += sched_vslice(cfs_rq, se);
3906 /* sleeps up to a single latency don't count. */
3908 unsigned long thresh = sysctl_sched_latency;
3911 * Halve their sleep time's effect, to allow
3912 * for a gentler effect of sleepers:
3914 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3920 /* ensure we never gain time by being placed backwards. */
3921 se->vruntime = max_vruntime(se->vruntime, vruntime);
3924 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3926 static inline void check_schedstat_required(void)
3928 #ifdef CONFIG_SCHEDSTATS
3929 if (schedstat_enabled())
3932 /* Force schedstat enabled if a dependent tracepoint is active */
3933 if (trace_sched_stat_wait_enabled() ||
3934 trace_sched_stat_sleep_enabled() ||
3935 trace_sched_stat_iowait_enabled() ||
3936 trace_sched_stat_blocked_enabled() ||
3937 trace_sched_stat_runtime_enabled()) {
3938 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3939 "stat_blocked and stat_runtime require the "
3940 "kernel parameter schedstats=enable or "
3941 "kernel.sched_schedstats=1\n");
3952 * update_min_vruntime()
3953 * vruntime -= min_vruntime
3957 * update_min_vruntime()
3958 * vruntime += min_vruntime
3960 * this way the vruntime transition between RQs is done when both
3961 * min_vruntime are up-to-date.
3965 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3966 * vruntime -= min_vruntime
3970 * update_min_vruntime()
3971 * vruntime += min_vruntime
3973 * this way we don't have the most up-to-date min_vruntime on the originating
3974 * CPU and an up-to-date min_vruntime on the destination CPU.
3978 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3980 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3981 bool curr = cfs_rq->curr == se;
3984 * If we're the current task, we must renormalise before calling
3988 se->vruntime += cfs_rq->min_vruntime;
3990 update_curr(cfs_rq);
3993 * Otherwise, renormalise after, such that we're placed at the current
3994 * moment in time, instead of some random moment in the past. Being
3995 * placed in the past could significantly boost this task to the
3996 * fairness detriment of existing tasks.
3998 if (renorm && !curr)
3999 se->vruntime += cfs_rq->min_vruntime;
4002 * When enqueuing a sched_entity, we must:
4003 * - Update loads to have both entity and cfs_rq synced with now.
4004 * - Add its load to cfs_rq->runnable_avg
4005 * - For group_entity, update its weight to reflect the new share of
4007 * - Add its new weight to cfs_rq->load.weight
4009 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
4010 update_cfs_group(se);
4011 enqueue_runnable_load_avg(cfs_rq, se);
4012 account_entity_enqueue(cfs_rq, se);
4014 if (flags & ENQUEUE_WAKEUP)
4015 place_entity(cfs_rq, se, 0);
4017 check_schedstat_required();
4018 update_stats_enqueue(cfs_rq, se, flags);
4019 check_spread(cfs_rq, se);
4021 __enqueue_entity(cfs_rq, se);
4024 if (cfs_rq->nr_running == 1) {
4025 list_add_leaf_cfs_rq(cfs_rq);
4026 check_enqueue_throttle(cfs_rq);
4030 static void __clear_buddies_last(struct sched_entity *se)
4032 for_each_sched_entity(se) {
4033 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4034 if (cfs_rq->last != se)
4037 cfs_rq->last = NULL;
4041 static void __clear_buddies_next(struct sched_entity *se)
4043 for_each_sched_entity(se) {
4044 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4045 if (cfs_rq->next != se)
4048 cfs_rq->next = NULL;
4052 static void __clear_buddies_skip(struct sched_entity *se)
4054 for_each_sched_entity(se) {
4055 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4056 if (cfs_rq->skip != se)
4059 cfs_rq->skip = NULL;
4063 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4065 if (cfs_rq->last == se)
4066 __clear_buddies_last(se);
4068 if (cfs_rq->next == se)
4069 __clear_buddies_next(se);
4071 if (cfs_rq->skip == se)
4072 __clear_buddies_skip(se);
4075 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4078 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4081 * Update run-time statistics of the 'current'.
4083 update_curr(cfs_rq);
4086 * When dequeuing a sched_entity, we must:
4087 * - Update loads to have both entity and cfs_rq synced with now.
4088 * - Subtract its load from the cfs_rq->runnable_avg.
4089 * - Subtract its previous weight from cfs_rq->load.weight.
4090 * - For group entity, update its weight to reflect the new share
4091 * of its group cfs_rq.
4093 update_load_avg(cfs_rq, se, UPDATE_TG);
4094 dequeue_runnable_load_avg(cfs_rq, se);
4096 update_stats_dequeue(cfs_rq, se, flags);
4098 clear_buddies(cfs_rq, se);
4100 if (se != cfs_rq->curr)
4101 __dequeue_entity(cfs_rq, se);
4103 account_entity_dequeue(cfs_rq, se);
4106 * Normalize after update_curr(); which will also have moved
4107 * min_vruntime if @se is the one holding it back. But before doing
4108 * update_min_vruntime() again, which will discount @se's position and
4109 * can move min_vruntime forward still more.
4111 if (!(flags & DEQUEUE_SLEEP))
4112 se->vruntime -= cfs_rq->min_vruntime;
4114 /* return excess runtime on last dequeue */
4115 return_cfs_rq_runtime(cfs_rq);
4117 update_cfs_group(se);
4120 * Now advance min_vruntime if @se was the entity holding it back,
4121 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4122 * put back on, and if we advance min_vruntime, we'll be placed back
4123 * further than we started -- ie. we'll be penalized.
4125 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4126 update_min_vruntime(cfs_rq);
4130 * Preempt the current task with a newly woken task if needed:
4133 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4135 unsigned long ideal_runtime, delta_exec;
4136 struct sched_entity *se;
4139 ideal_runtime = sched_slice(cfs_rq, curr);
4140 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4141 if (delta_exec > ideal_runtime) {
4142 resched_curr(rq_of(cfs_rq));
4144 * The current task ran long enough, ensure it doesn't get
4145 * re-elected due to buddy favours.
4147 clear_buddies(cfs_rq, curr);
4152 * Ensure that a task that missed wakeup preemption by a
4153 * narrow margin doesn't have to wait for a full slice.
4154 * This also mitigates buddy induced latencies under load.
4156 if (delta_exec < sysctl_sched_min_granularity)
4159 se = __pick_first_entity(cfs_rq);
4160 delta = curr->vruntime - se->vruntime;
4165 if (delta > ideal_runtime)
4166 resched_curr(rq_of(cfs_rq));
4170 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4172 /* 'current' is not kept within the tree. */
4175 * Any task has to be enqueued before it get to execute on
4176 * a CPU. So account for the time it spent waiting on the
4179 update_stats_wait_end(cfs_rq, se);
4180 __dequeue_entity(cfs_rq, se);
4181 update_load_avg(cfs_rq, se, UPDATE_TG);
4184 update_stats_curr_start(cfs_rq, se);
4188 * Track our maximum slice length, if the CPU's load is at
4189 * least twice that of our own weight (i.e. dont track it
4190 * when there are only lesser-weight tasks around):
4192 if (schedstat_enabled() &&
4193 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
4194 schedstat_set(se->statistics.slice_max,
4195 max((u64)schedstat_val(se->statistics.slice_max),
4196 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4199 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4203 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4206 * Pick the next process, keeping these things in mind, in this order:
4207 * 1) keep things fair between processes/task groups
4208 * 2) pick the "next" process, since someone really wants that to run
4209 * 3) pick the "last" process, for cache locality
4210 * 4) do not run the "skip" process, if something else is available
4212 static struct sched_entity *
4213 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4215 struct sched_entity *left = __pick_first_entity(cfs_rq);
4216 struct sched_entity *se;
4219 * If curr is set we have to see if its left of the leftmost entity
4220 * still in the tree, provided there was anything in the tree at all.
4222 if (!left || (curr && entity_before(curr, left)))
4225 se = left; /* ideally we run the leftmost entity */
4228 * Avoid running the skip buddy, if running something else can
4229 * be done without getting too unfair.
4231 if (cfs_rq->skip == se) {
4232 struct sched_entity *second;
4235 second = __pick_first_entity(cfs_rq);
4237 second = __pick_next_entity(se);
4238 if (!second || (curr && entity_before(curr, second)))
4242 if (second && wakeup_preempt_entity(second, left) < 1)
4247 * Prefer last buddy, try to return the CPU to a preempted task.
4249 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4253 * Someone really wants this to run. If it's not unfair, run it.
4255 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4258 clear_buddies(cfs_rq, se);
4263 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4265 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4268 * If still on the runqueue then deactivate_task()
4269 * was not called and update_curr() has to be done:
4272 update_curr(cfs_rq);
4274 /* throttle cfs_rqs exceeding runtime */
4275 check_cfs_rq_runtime(cfs_rq);
4277 check_spread(cfs_rq, prev);
4280 update_stats_wait_start(cfs_rq, prev);
4281 /* Put 'current' back into the tree. */
4282 __enqueue_entity(cfs_rq, prev);
4283 /* in !on_rq case, update occurred at dequeue */
4284 update_load_avg(cfs_rq, prev, 0);
4286 cfs_rq->curr = NULL;
4290 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4293 * Update run-time statistics of the 'current'.
4295 update_curr(cfs_rq);
4298 * Ensure that runnable average is periodically updated.
4300 update_load_avg(cfs_rq, curr, UPDATE_TG);
4301 update_cfs_group(curr);
4303 #ifdef CONFIG_SCHED_HRTICK
4305 * queued ticks are scheduled to match the slice, so don't bother
4306 * validating it and just reschedule.
4309 resched_curr(rq_of(cfs_rq));
4313 * don't let the period tick interfere with the hrtick preemption
4315 if (!sched_feat(DOUBLE_TICK) &&
4316 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4320 if (cfs_rq->nr_running > 1)
4321 check_preempt_tick(cfs_rq, curr);
4325 /**************************************************
4326 * CFS bandwidth control machinery
4329 #ifdef CONFIG_CFS_BANDWIDTH
4331 #ifdef CONFIG_JUMP_LABEL
4332 static struct static_key __cfs_bandwidth_used;
4334 static inline bool cfs_bandwidth_used(void)
4336 return static_key_false(&__cfs_bandwidth_used);
4339 void cfs_bandwidth_usage_inc(void)
4341 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4344 void cfs_bandwidth_usage_dec(void)
4346 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4348 #else /* CONFIG_JUMP_LABEL */
4349 static bool cfs_bandwidth_used(void)
4354 void cfs_bandwidth_usage_inc(void) {}
4355 void cfs_bandwidth_usage_dec(void) {}
4356 #endif /* CONFIG_JUMP_LABEL */
4359 * default period for cfs group bandwidth.
4360 * default: 0.1s, units: nanoseconds
4362 static inline u64 default_cfs_period(void)
4364 return 100000000ULL;
4367 static inline u64 sched_cfs_bandwidth_slice(void)
4369 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4373 * Replenish runtime according to assigned quota. We use sched_clock_cpu
4374 * directly instead of rq->clock to avoid adding additional synchronization
4377 * requires cfs_b->lock
4379 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4381 if (cfs_b->quota != RUNTIME_INF)
4382 cfs_b->runtime = cfs_b->quota;
4385 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4387 return &tg->cfs_bandwidth;
4390 /* returns 0 on failure to allocate runtime */
4391 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4393 struct task_group *tg = cfs_rq->tg;
4394 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4395 u64 amount = 0, min_amount;
4397 /* note: this is a positive sum as runtime_remaining <= 0 */
4398 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4400 raw_spin_lock(&cfs_b->lock);
4401 if (cfs_b->quota == RUNTIME_INF)
4402 amount = min_amount;
4404 start_cfs_bandwidth(cfs_b);
4406 if (cfs_b->runtime > 0) {
4407 amount = min(cfs_b->runtime, min_amount);
4408 cfs_b->runtime -= amount;
4412 raw_spin_unlock(&cfs_b->lock);
4414 cfs_rq->runtime_remaining += amount;
4416 return cfs_rq->runtime_remaining > 0;
4419 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4421 /* dock delta_exec before expiring quota (as it could span periods) */
4422 cfs_rq->runtime_remaining -= delta_exec;
4424 if (likely(cfs_rq->runtime_remaining > 0))
4427 if (cfs_rq->throttled)
4430 * if we're unable to extend our runtime we resched so that the active
4431 * hierarchy can be throttled
4433 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4434 resched_curr(rq_of(cfs_rq));
4437 static __always_inline
4438 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4440 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4443 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4446 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4448 return cfs_bandwidth_used() && cfs_rq->throttled;
4451 /* check whether cfs_rq, or any parent, is throttled */
4452 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4454 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4458 * Ensure that neither of the group entities corresponding to src_cpu or
4459 * dest_cpu are members of a throttled hierarchy when performing group
4460 * load-balance operations.
4462 static inline int throttled_lb_pair(struct task_group *tg,
4463 int src_cpu, int dest_cpu)
4465 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4467 src_cfs_rq = tg->cfs_rq[src_cpu];
4468 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4470 return throttled_hierarchy(src_cfs_rq) ||
4471 throttled_hierarchy(dest_cfs_rq);
4474 static int tg_unthrottle_up(struct task_group *tg, void *data)
4476 struct rq *rq = data;
4477 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4479 cfs_rq->throttle_count--;
4480 if (!cfs_rq->throttle_count) {
4481 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4482 cfs_rq->throttled_clock_task;
4484 /* Add cfs_rq with already running entity in the list */
4485 if (cfs_rq->nr_running >= 1)
4486 list_add_leaf_cfs_rq(cfs_rq);
4492 static int tg_throttle_down(struct task_group *tg, void *data)
4494 struct rq *rq = data;
4495 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4497 /* group is entering throttled state, stop time */
4498 if (!cfs_rq->throttle_count) {
4499 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4500 list_del_leaf_cfs_rq(cfs_rq);
4502 cfs_rq->throttle_count++;
4507 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4509 struct rq *rq = rq_of(cfs_rq);
4510 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4511 struct sched_entity *se;
4512 long task_delta, idle_task_delta, dequeue = 1;
4515 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4517 /* freeze hierarchy runnable averages while throttled */
4519 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4522 task_delta = cfs_rq->h_nr_running;
4523 idle_task_delta = cfs_rq->idle_h_nr_running;
4524 for_each_sched_entity(se) {
4525 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4526 /* throttled entity or throttle-on-deactivate */
4531 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4532 qcfs_rq->h_nr_running -= task_delta;
4533 qcfs_rq->idle_h_nr_running -= idle_task_delta;
4535 if (qcfs_rq->load.weight)
4540 sub_nr_running(rq, task_delta);
4542 cfs_rq->throttled = 1;
4543 cfs_rq->throttled_clock = rq_clock(rq);
4544 raw_spin_lock(&cfs_b->lock);
4545 empty = list_empty(&cfs_b->throttled_cfs_rq);
4548 * Add to the _head_ of the list, so that an already-started
4549 * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
4550 * not running add to the tail so that later runqueues don't get starved.
4552 if (cfs_b->distribute_running)
4553 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4555 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4558 * If we're the first throttled task, make sure the bandwidth
4562 start_cfs_bandwidth(cfs_b);
4564 raw_spin_unlock(&cfs_b->lock);
4567 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4569 struct rq *rq = rq_of(cfs_rq);
4570 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4571 struct sched_entity *se;
4573 long task_delta, idle_task_delta;
4575 se = cfs_rq->tg->se[cpu_of(rq)];
4577 cfs_rq->throttled = 0;
4579 update_rq_clock(rq);
4581 raw_spin_lock(&cfs_b->lock);
4582 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4583 list_del_rcu(&cfs_rq->throttled_list);
4584 raw_spin_unlock(&cfs_b->lock);
4586 /* update hierarchical throttle state */
4587 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4589 if (!cfs_rq->load.weight)
4592 task_delta = cfs_rq->h_nr_running;
4593 idle_task_delta = cfs_rq->idle_h_nr_running;
4594 for_each_sched_entity(se) {
4598 cfs_rq = cfs_rq_of(se);
4600 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4601 cfs_rq->h_nr_running += task_delta;
4602 cfs_rq->idle_h_nr_running += idle_task_delta;
4604 if (cfs_rq_throttled(cfs_rq))
4608 assert_list_leaf_cfs_rq(rq);
4611 add_nr_running(rq, task_delta);
4613 /* Determine whether we need to wake up potentially idle CPU: */
4614 if (rq->curr == rq->idle && rq->cfs.nr_running)
4618 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, u64 remaining)
4620 struct cfs_rq *cfs_rq;
4622 u64 starting_runtime = remaining;
4625 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4627 struct rq *rq = rq_of(cfs_rq);
4630 rq_lock_irqsave(rq, &rf);
4631 if (!cfs_rq_throttled(cfs_rq))
4634 /* By the above check, this should never be true */
4635 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
4637 runtime = -cfs_rq->runtime_remaining + 1;
4638 if (runtime > remaining)
4639 runtime = remaining;
4640 remaining -= runtime;
4642 cfs_rq->runtime_remaining += runtime;
4644 /* we check whether we're throttled above */
4645 if (cfs_rq->runtime_remaining > 0)
4646 unthrottle_cfs_rq(cfs_rq);
4649 rq_unlock_irqrestore(rq, &rf);
4656 return starting_runtime - remaining;
4660 * Responsible for refilling a task_group's bandwidth and unthrottling its
4661 * cfs_rqs as appropriate. If there has been no activity within the last
4662 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4663 * used to track this state.
4665 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
4670 /* no need to continue the timer with no bandwidth constraint */
4671 if (cfs_b->quota == RUNTIME_INF)
4672 goto out_deactivate;
4674 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4675 cfs_b->nr_periods += overrun;
4678 * idle depends on !throttled (for the case of a large deficit), and if
4679 * we're going inactive then everything else can be deferred
4681 if (cfs_b->idle && !throttled)
4682 goto out_deactivate;
4684 __refill_cfs_bandwidth_runtime(cfs_b);
4687 /* mark as potentially idle for the upcoming period */
4692 /* account preceding periods in which throttling occurred */
4693 cfs_b->nr_throttled += overrun;
4696 * This check is repeated as we are holding onto the new bandwidth while
4697 * we unthrottle. This can potentially race with an unthrottled group
4698 * trying to acquire new bandwidth from the global pool. This can result
4699 * in us over-using our runtime if it is all used during this loop, but
4700 * only by limited amounts in that extreme case.
4702 while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
4703 runtime = cfs_b->runtime;
4704 cfs_b->distribute_running = 1;
4705 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4706 /* we can't nest cfs_b->lock while distributing bandwidth */
4707 runtime = distribute_cfs_runtime(cfs_b, runtime);
4708 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4710 cfs_b->distribute_running = 0;
4711 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4713 lsub_positive(&cfs_b->runtime, runtime);
4717 * While we are ensured activity in the period following an
4718 * unthrottle, this also covers the case in which the new bandwidth is
4719 * insufficient to cover the existing bandwidth deficit. (Forcing the
4720 * timer to remain active while there are any throttled entities.)
4730 /* a cfs_rq won't donate quota below this amount */
4731 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4732 /* minimum remaining period time to redistribute slack quota */
4733 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4734 /* how long we wait to gather additional slack before distributing */
4735 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4738 * Are we near the end of the current quota period?
4740 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4741 * hrtimer base being cleared by hrtimer_start. In the case of
4742 * migrate_hrtimers, base is never cleared, so we are fine.
4744 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4746 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4749 /* if the call-back is running a quota refresh is already occurring */
4750 if (hrtimer_callback_running(refresh_timer))
4753 /* is a quota refresh about to occur? */
4754 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4755 if (remaining < min_expire)
4761 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4763 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4765 /* if there's a quota refresh soon don't bother with slack */
4766 if (runtime_refresh_within(cfs_b, min_left))
4769 /* don't push forwards an existing deferred unthrottle */
4770 if (cfs_b->slack_started)
4772 cfs_b->slack_started = true;
4774 hrtimer_start(&cfs_b->slack_timer,
4775 ns_to_ktime(cfs_bandwidth_slack_period),
4779 /* we know any runtime found here is valid as update_curr() precedes return */
4780 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4782 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4783 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4785 if (slack_runtime <= 0)
4788 raw_spin_lock(&cfs_b->lock);
4789 if (cfs_b->quota != RUNTIME_INF) {
4790 cfs_b->runtime += slack_runtime;
4792 /* we are under rq->lock, defer unthrottling using a timer */
4793 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4794 !list_empty(&cfs_b->throttled_cfs_rq))
4795 start_cfs_slack_bandwidth(cfs_b);
4797 raw_spin_unlock(&cfs_b->lock);
4799 /* even if it's not valid for return we don't want to try again */
4800 cfs_rq->runtime_remaining -= slack_runtime;
4803 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4805 if (!cfs_bandwidth_used())
4808 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4811 __return_cfs_rq_runtime(cfs_rq);
4815 * This is done with a timer (instead of inline with bandwidth return) since
4816 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4818 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4820 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4821 unsigned long flags;
4823 /* confirm we're still not at a refresh boundary */
4824 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4825 cfs_b->slack_started = false;
4826 if (cfs_b->distribute_running) {
4827 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4831 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4832 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4836 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4837 runtime = cfs_b->runtime;
4840 cfs_b->distribute_running = 1;
4842 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4847 runtime = distribute_cfs_runtime(cfs_b, runtime);
4849 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4850 lsub_positive(&cfs_b->runtime, runtime);
4851 cfs_b->distribute_running = 0;
4852 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4856 * When a group wakes up we want to make sure that its quota is not already
4857 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4858 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4860 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4862 if (!cfs_bandwidth_used())
4865 /* an active group must be handled by the update_curr()->put() path */
4866 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4869 /* ensure the group is not already throttled */
4870 if (cfs_rq_throttled(cfs_rq))
4873 /* update runtime allocation */
4874 account_cfs_rq_runtime(cfs_rq, 0);
4875 if (cfs_rq->runtime_remaining <= 0)
4876 throttle_cfs_rq(cfs_rq);
4879 static void sync_throttle(struct task_group *tg, int cpu)
4881 struct cfs_rq *pcfs_rq, *cfs_rq;
4883 if (!cfs_bandwidth_used())
4889 cfs_rq = tg->cfs_rq[cpu];
4890 pcfs_rq = tg->parent->cfs_rq[cpu];
4892 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4893 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4896 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4897 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4899 if (!cfs_bandwidth_used())
4902 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4906 * it's possible for a throttled entity to be forced into a running
4907 * state (e.g. set_curr_task), in this case we're finished.
4909 if (cfs_rq_throttled(cfs_rq))
4912 throttle_cfs_rq(cfs_rq);
4916 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4918 struct cfs_bandwidth *cfs_b =
4919 container_of(timer, struct cfs_bandwidth, slack_timer);
4921 do_sched_cfs_slack_timer(cfs_b);
4923 return HRTIMER_NORESTART;
4926 extern const u64 max_cfs_quota_period;
4928 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4930 struct cfs_bandwidth *cfs_b =
4931 container_of(timer, struct cfs_bandwidth, period_timer);
4932 unsigned long flags;
4937 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4939 overrun = hrtimer_forward_now(timer, cfs_b->period);
4944 u64 new, old = ktime_to_ns(cfs_b->period);
4947 * Grow period by a factor of 2 to avoid losing precision.
4948 * Precision loss in the quota/period ratio can cause __cfs_schedulable
4952 if (new < max_cfs_quota_period) {
4953 cfs_b->period = ns_to_ktime(new);
4956 pr_warn_ratelimited(
4957 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
4959 div_u64(new, NSEC_PER_USEC),
4960 div_u64(cfs_b->quota, NSEC_PER_USEC));
4962 pr_warn_ratelimited(
4963 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
4965 div_u64(old, NSEC_PER_USEC),
4966 div_u64(cfs_b->quota, NSEC_PER_USEC));
4969 /* reset count so we don't come right back in here */
4973 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
4976 cfs_b->period_active = 0;
4977 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4979 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4982 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4984 raw_spin_lock_init(&cfs_b->lock);
4986 cfs_b->quota = RUNTIME_INF;
4987 cfs_b->period = ns_to_ktime(default_cfs_period());
4989 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4990 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4991 cfs_b->period_timer.function = sched_cfs_period_timer;
4992 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4993 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4994 cfs_b->distribute_running = 0;
4995 cfs_b->slack_started = false;
4998 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5000 cfs_rq->runtime_enabled = 0;
5001 INIT_LIST_HEAD(&cfs_rq->throttled_list);
5004 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5006 lockdep_assert_held(&cfs_b->lock);
5008 if (cfs_b->period_active)
5011 cfs_b->period_active = 1;
5012 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5013 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5016 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5018 /* init_cfs_bandwidth() was not called */
5019 if (!cfs_b->throttled_cfs_rq.next)
5022 hrtimer_cancel(&cfs_b->period_timer);
5023 hrtimer_cancel(&cfs_b->slack_timer);
5027 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5029 * The race is harmless, since modifying bandwidth settings of unhooked group
5030 * bits doesn't do much.
5033 /* cpu online calback */
5034 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5036 struct task_group *tg;
5038 lockdep_assert_held(&rq->lock);
5041 list_for_each_entry_rcu(tg, &task_groups, list) {
5042 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5043 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5045 raw_spin_lock(&cfs_b->lock);
5046 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5047 raw_spin_unlock(&cfs_b->lock);
5052 /* cpu offline callback */
5053 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5055 struct task_group *tg;
5057 lockdep_assert_held(&rq->lock);
5060 list_for_each_entry_rcu(tg, &task_groups, list) {
5061 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5063 if (!cfs_rq->runtime_enabled)
5067 * clock_task is not advancing so we just need to make sure
5068 * there's some valid quota amount
5070 cfs_rq->runtime_remaining = 1;
5072 * Offline rq is schedulable till CPU is completely disabled
5073 * in take_cpu_down(), so we prevent new cfs throttling here.
5075 cfs_rq->runtime_enabled = 0;
5077 if (cfs_rq_throttled(cfs_rq))
5078 unthrottle_cfs_rq(cfs_rq);
5083 #else /* CONFIG_CFS_BANDWIDTH */
5085 static inline bool cfs_bandwidth_used(void)
5090 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5091 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5092 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5093 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5094 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5096 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5101 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5106 static inline int throttled_lb_pair(struct task_group *tg,
5107 int src_cpu, int dest_cpu)
5112 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5114 #ifdef CONFIG_FAIR_GROUP_SCHED
5115 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5118 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5122 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5123 static inline void update_runtime_enabled(struct rq *rq) {}
5124 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5126 #endif /* CONFIG_CFS_BANDWIDTH */
5128 /**************************************************
5129 * CFS operations on tasks:
5132 #ifdef CONFIG_SCHED_HRTICK
5133 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5135 struct sched_entity *se = &p->se;
5136 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5138 SCHED_WARN_ON(task_rq(p) != rq);
5140 if (rq->cfs.h_nr_running > 1) {
5141 u64 slice = sched_slice(cfs_rq, se);
5142 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5143 s64 delta = slice - ran;
5150 hrtick_start(rq, delta);
5155 * called from enqueue/dequeue and updates the hrtick when the
5156 * current task is from our class and nr_running is low enough
5159 static void hrtick_update(struct rq *rq)
5161 struct task_struct *curr = rq->curr;
5163 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5166 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5167 hrtick_start_fair(rq, curr);
5169 #else /* !CONFIG_SCHED_HRTICK */
5171 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5175 static inline void hrtick_update(struct rq *rq)
5181 static inline unsigned long cpu_util(int cpu);
5183 static inline bool cpu_overutilized(int cpu)
5185 return !fits_capacity(cpu_util(cpu), capacity_of(cpu));
5188 static inline void update_overutilized_status(struct rq *rq)
5190 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
5191 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5192 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
5196 static inline void update_overutilized_status(struct rq *rq) { }
5200 * The enqueue_task method is called before nr_running is
5201 * increased. Here we update the fair scheduling stats and
5202 * then put the task into the rbtree:
5205 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5207 struct cfs_rq *cfs_rq;
5208 struct sched_entity *se = &p->se;
5209 int idle_h_nr_running = task_has_idle_policy(p);
5212 * The code below (indirectly) updates schedutil which looks at
5213 * the cfs_rq utilization to select a frequency.
5214 * Let's add the task's estimated utilization to the cfs_rq's
5215 * estimated utilization, before we update schedutil.
5217 util_est_enqueue(&rq->cfs, p);
5220 * If in_iowait is set, the code below may not trigger any cpufreq
5221 * utilization updates, so do it here explicitly with the IOWAIT flag
5225 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5227 for_each_sched_entity(se) {
5230 cfs_rq = cfs_rq_of(se);
5231 enqueue_entity(cfs_rq, se, flags);
5234 * end evaluation on encountering a throttled cfs_rq
5236 * note: in the case of encountering a throttled cfs_rq we will
5237 * post the final h_nr_running increment below.
5239 if (cfs_rq_throttled(cfs_rq))
5241 cfs_rq->h_nr_running++;
5242 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5244 flags = ENQUEUE_WAKEUP;
5247 for_each_sched_entity(se) {
5248 cfs_rq = cfs_rq_of(se);
5249 cfs_rq->h_nr_running++;
5250 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5252 if (cfs_rq_throttled(cfs_rq))
5255 update_load_avg(cfs_rq, se, UPDATE_TG);
5256 update_cfs_group(se);
5260 add_nr_running(rq, 1);
5262 * Since new tasks are assigned an initial util_avg equal to
5263 * half of the spare capacity of their CPU, tiny tasks have the
5264 * ability to cross the overutilized threshold, which will
5265 * result in the load balancer ruining all the task placement
5266 * done by EAS. As a way to mitigate that effect, do not account
5267 * for the first enqueue operation of new tasks during the
5268 * overutilized flag detection.
5270 * A better way of solving this problem would be to wait for
5271 * the PELT signals of tasks to converge before taking them
5272 * into account, but that is not straightforward to implement,
5273 * and the following generally works well enough in practice.
5275 if (flags & ENQUEUE_WAKEUP)
5276 update_overutilized_status(rq);
5280 if (cfs_bandwidth_used()) {
5282 * When bandwidth control is enabled; the cfs_rq_throttled()
5283 * breaks in the above iteration can result in incomplete
5284 * leaf list maintenance, resulting in triggering the assertion
5287 for_each_sched_entity(se) {
5288 cfs_rq = cfs_rq_of(se);
5290 if (list_add_leaf_cfs_rq(cfs_rq))
5295 assert_list_leaf_cfs_rq(rq);
5300 static void set_next_buddy(struct sched_entity *se);
5303 * The dequeue_task method is called before nr_running is
5304 * decreased. We remove the task from the rbtree and
5305 * update the fair scheduling stats:
5307 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5309 struct cfs_rq *cfs_rq;
5310 struct sched_entity *se = &p->se;
5311 int task_sleep = flags & DEQUEUE_SLEEP;
5312 int idle_h_nr_running = task_has_idle_policy(p);
5314 for_each_sched_entity(se) {
5315 cfs_rq = cfs_rq_of(se);
5316 dequeue_entity(cfs_rq, se, flags);
5319 * end evaluation on encountering a throttled cfs_rq
5321 * note: in the case of encountering a throttled cfs_rq we will
5322 * post the final h_nr_running decrement below.
5324 if (cfs_rq_throttled(cfs_rq))
5326 cfs_rq->h_nr_running--;
5327 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5329 /* Don't dequeue parent if it has other entities besides us */
5330 if (cfs_rq->load.weight) {
5331 /* Avoid re-evaluating load for this entity: */
5332 se = parent_entity(se);
5334 * Bias pick_next to pick a task from this cfs_rq, as
5335 * p is sleeping when it is within its sched_slice.
5337 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5341 flags |= DEQUEUE_SLEEP;
5344 for_each_sched_entity(se) {
5345 cfs_rq = cfs_rq_of(se);
5346 cfs_rq->h_nr_running--;
5347 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5349 if (cfs_rq_throttled(cfs_rq))
5352 update_load_avg(cfs_rq, se, UPDATE_TG);
5353 update_cfs_group(se);
5357 sub_nr_running(rq, 1);
5359 util_est_dequeue(&rq->cfs, p, task_sleep);
5365 /* Working cpumask for: load_balance, load_balance_newidle. */
5366 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5367 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5369 #ifdef CONFIG_NO_HZ_COMMON
5372 cpumask_var_t idle_cpus_mask;
5374 int has_blocked; /* Idle CPUS has blocked load */
5375 unsigned long next_balance; /* in jiffy units */
5376 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5377 } nohz ____cacheline_aligned;
5379 #endif /* CONFIG_NO_HZ_COMMON */
5381 /* CPU only has SCHED_IDLE tasks enqueued */
5382 static int sched_idle_cpu(int cpu)
5384 struct rq *rq = cpu_rq(cpu);
5386 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
5390 static unsigned long cpu_load(struct rq *rq)
5392 return cfs_rq_load_avg(&rq->cfs);
5396 * cpu_load_without - compute CPU load without any contributions from *p
5397 * @cpu: the CPU which load is requested
5398 * @p: the task which load should be discounted
5400 * The load of a CPU is defined by the load of tasks currently enqueued on that
5401 * CPU as well as tasks which are currently sleeping after an execution on that
5404 * This method returns the load of the specified CPU by discounting the load of
5405 * the specified task, whenever the task is currently contributing to the CPU
5408 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
5410 struct cfs_rq *cfs_rq;
5413 /* Task has no contribution or is new */
5414 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
5415 return cpu_load(rq);
5418 load = READ_ONCE(cfs_rq->avg.load_avg);
5420 /* Discount task's util from CPU's util */
5421 lsub_positive(&load, task_h_load(p));
5426 static unsigned long capacity_of(int cpu)
5428 return cpu_rq(cpu)->cpu_capacity;
5431 static void record_wakee(struct task_struct *p)
5434 * Only decay a single time; tasks that have less then 1 wakeup per
5435 * jiffy will not have built up many flips.
5437 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5438 current->wakee_flips >>= 1;
5439 current->wakee_flip_decay_ts = jiffies;
5442 if (current->last_wakee != p) {
5443 current->last_wakee = p;
5444 current->wakee_flips++;
5449 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5451 * A waker of many should wake a different task than the one last awakened
5452 * at a frequency roughly N times higher than one of its wakees.
5454 * In order to determine whether we should let the load spread vs consolidating
5455 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5456 * partner, and a factor of lls_size higher frequency in the other.
5458 * With both conditions met, we can be relatively sure that the relationship is
5459 * non-monogamous, with partner count exceeding socket size.
5461 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5462 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5465 static int wake_wide(struct task_struct *p)
5467 unsigned int master = current->wakee_flips;
5468 unsigned int slave = p->wakee_flips;
5469 int factor = this_cpu_read(sd_llc_size);
5472 swap(master, slave);
5473 if (slave < factor || master < slave * factor)
5479 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5480 * soonest. For the purpose of speed we only consider the waking and previous
5483 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5484 * cache-affine and is (or will be) idle.
5486 * wake_affine_weight() - considers the weight to reflect the average
5487 * scheduling latency of the CPUs. This seems to work
5488 * for the overloaded case.
5491 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5494 * If this_cpu is idle, it implies the wakeup is from interrupt
5495 * context. Only allow the move if cache is shared. Otherwise an
5496 * interrupt intensive workload could force all tasks onto one
5497 * node depending on the IO topology or IRQ affinity settings.
5499 * If the prev_cpu is idle and cache affine then avoid a migration.
5500 * There is no guarantee that the cache hot data from an interrupt
5501 * is more important than cache hot data on the prev_cpu and from
5502 * a cpufreq perspective, it's better to have higher utilisation
5505 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5506 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5508 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5511 return nr_cpumask_bits;
5515 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5516 int this_cpu, int prev_cpu, int sync)
5518 s64 this_eff_load, prev_eff_load;
5519 unsigned long task_load;
5521 this_eff_load = cpu_load(cpu_rq(this_cpu));
5524 unsigned long current_load = task_h_load(current);
5526 if (current_load > this_eff_load)
5529 this_eff_load -= current_load;
5532 task_load = task_h_load(p);
5534 this_eff_load += task_load;
5535 if (sched_feat(WA_BIAS))
5536 this_eff_load *= 100;
5537 this_eff_load *= capacity_of(prev_cpu);
5539 prev_eff_load = cpu_load(cpu_rq(prev_cpu));
5540 prev_eff_load -= task_load;
5541 if (sched_feat(WA_BIAS))
5542 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5543 prev_eff_load *= capacity_of(this_cpu);
5546 * If sync, adjust the weight of prev_eff_load such that if
5547 * prev_eff == this_eff that select_idle_sibling() will consider
5548 * stacking the wakee on top of the waker if no other CPU is
5554 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5557 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5558 int this_cpu, int prev_cpu, int sync)
5560 int target = nr_cpumask_bits;
5562 if (sched_feat(WA_IDLE))
5563 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5565 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5566 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5568 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5569 if (target == nr_cpumask_bits)
5572 schedstat_inc(sd->ttwu_move_affine);
5573 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5577 static struct sched_group *
5578 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5579 int this_cpu, int sd_flag);
5582 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5585 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5587 unsigned long load, min_load = ULONG_MAX;
5588 unsigned int min_exit_latency = UINT_MAX;
5589 u64 latest_idle_timestamp = 0;
5590 int least_loaded_cpu = this_cpu;
5591 int shallowest_idle_cpu = -1, si_cpu = -1;
5594 /* Check if we have any choice: */
5595 if (group->group_weight == 1)
5596 return cpumask_first(sched_group_span(group));
5598 /* Traverse only the allowed CPUs */
5599 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
5600 if (available_idle_cpu(i)) {
5601 struct rq *rq = cpu_rq(i);
5602 struct cpuidle_state *idle = idle_get_state(rq);
5603 if (idle && idle->exit_latency < min_exit_latency) {
5605 * We give priority to a CPU whose idle state
5606 * has the smallest exit latency irrespective
5607 * of any idle timestamp.
5609 min_exit_latency = idle->exit_latency;
5610 latest_idle_timestamp = rq->idle_stamp;
5611 shallowest_idle_cpu = i;
5612 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5613 rq->idle_stamp > latest_idle_timestamp) {
5615 * If equal or no active idle state, then
5616 * the most recently idled CPU might have
5619 latest_idle_timestamp = rq->idle_stamp;
5620 shallowest_idle_cpu = i;
5622 } else if (shallowest_idle_cpu == -1 && si_cpu == -1) {
5623 if (sched_idle_cpu(i)) {
5628 load = cpu_load(cpu_rq(i));
5629 if (load < min_load) {
5631 least_loaded_cpu = i;
5636 if (shallowest_idle_cpu != -1)
5637 return shallowest_idle_cpu;
5640 return least_loaded_cpu;
5643 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5644 int cpu, int prev_cpu, int sd_flag)
5648 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
5652 * We need task's util for cpu_util_without, sync it up to
5653 * prev_cpu's last_update_time.
5655 if (!(sd_flag & SD_BALANCE_FORK))
5656 sync_entity_load_avg(&p->se);
5659 struct sched_group *group;
5660 struct sched_domain *tmp;
5663 if (!(sd->flags & sd_flag)) {
5668 group = find_idlest_group(sd, p, cpu, sd_flag);
5674 new_cpu = find_idlest_group_cpu(group, p, cpu);
5675 if (new_cpu == cpu) {
5676 /* Now try balancing at a lower domain level of 'cpu': */
5681 /* Now try balancing at a lower domain level of 'new_cpu': */
5683 weight = sd->span_weight;
5685 for_each_domain(cpu, tmp) {
5686 if (weight <= tmp->span_weight)
5688 if (tmp->flags & sd_flag)
5696 #ifdef CONFIG_SCHED_SMT
5697 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5698 EXPORT_SYMBOL_GPL(sched_smt_present);
5700 static inline void set_idle_cores(int cpu, int val)
5702 struct sched_domain_shared *sds;
5704 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5706 WRITE_ONCE(sds->has_idle_cores, val);
5709 static inline bool test_idle_cores(int cpu, bool def)
5711 struct sched_domain_shared *sds;
5713 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5715 return READ_ONCE(sds->has_idle_cores);
5721 * Scans the local SMT mask to see if the entire core is idle, and records this
5722 * information in sd_llc_shared->has_idle_cores.
5724 * Since SMT siblings share all cache levels, inspecting this limited remote
5725 * state should be fairly cheap.
5727 void __update_idle_core(struct rq *rq)
5729 int core = cpu_of(rq);
5733 if (test_idle_cores(core, true))
5736 for_each_cpu(cpu, cpu_smt_mask(core)) {
5740 if (!available_idle_cpu(cpu))
5744 set_idle_cores(core, 1);
5750 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5751 * there are no idle cores left in the system; tracked through
5752 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5754 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5756 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5759 if (!static_branch_likely(&sched_smt_present))
5762 if (!test_idle_cores(target, false))
5765 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
5767 for_each_cpu_wrap(core, cpus, target) {
5770 for_each_cpu(cpu, cpu_smt_mask(core)) {
5771 __cpumask_clear_cpu(cpu, cpus);
5772 if (!available_idle_cpu(cpu))
5781 * Failed to find an idle core; stop looking for one.
5783 set_idle_cores(target, 0);
5789 * Scan the local SMT mask for idle CPUs.
5791 static int select_idle_smt(struct task_struct *p, int target)
5793 int cpu, si_cpu = -1;
5795 if (!static_branch_likely(&sched_smt_present))
5798 for_each_cpu(cpu, cpu_smt_mask(target)) {
5799 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
5801 if (available_idle_cpu(cpu))
5803 if (si_cpu == -1 && sched_idle_cpu(cpu))
5810 #else /* CONFIG_SCHED_SMT */
5812 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5817 static inline int select_idle_smt(struct task_struct *p, int target)
5822 #endif /* CONFIG_SCHED_SMT */
5825 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5826 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5827 * average idle time for this rq (as found in rq->avg_idle).
5829 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
5831 struct sched_domain *this_sd;
5832 u64 avg_cost, avg_idle;
5835 int this = smp_processor_id();
5836 int cpu, nr = INT_MAX, si_cpu = -1;
5838 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
5843 * Due to large variance we need a large fuzz factor; hackbench in
5844 * particularly is sensitive here.
5846 avg_idle = this_rq()->avg_idle / 512;
5847 avg_cost = this_sd->avg_scan_cost + 1;
5849 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
5852 if (sched_feat(SIS_PROP)) {
5853 u64 span_avg = sd->span_weight * avg_idle;
5854 if (span_avg > 4*avg_cost)
5855 nr = div_u64(span_avg, avg_cost);
5860 time = cpu_clock(this);
5862 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
5865 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
5867 if (available_idle_cpu(cpu))
5869 if (si_cpu == -1 && sched_idle_cpu(cpu))
5873 time = cpu_clock(this) - time;
5874 cost = this_sd->avg_scan_cost;
5875 delta = (s64)(time - cost) / 8;
5876 this_sd->avg_scan_cost += delta;
5882 * Try and locate an idle core/thread in the LLC cache domain.
5884 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5886 struct sched_domain *sd;
5887 int i, recent_used_cpu;
5889 if (available_idle_cpu(target) || sched_idle_cpu(target))
5893 * If the previous CPU is cache affine and idle, don't be stupid:
5895 if (prev != target && cpus_share_cache(prev, target) &&
5896 (available_idle_cpu(prev) || sched_idle_cpu(prev)))
5899 /* Check a recently used CPU as a potential idle candidate: */
5900 recent_used_cpu = p->recent_used_cpu;
5901 if (recent_used_cpu != prev &&
5902 recent_used_cpu != target &&
5903 cpus_share_cache(recent_used_cpu, target) &&
5904 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
5905 cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr)) {
5907 * Replace recent_used_cpu with prev as it is a potential
5908 * candidate for the next wake:
5910 p->recent_used_cpu = prev;
5911 return recent_used_cpu;
5914 sd = rcu_dereference(per_cpu(sd_llc, target));
5918 i = select_idle_core(p, sd, target);
5919 if ((unsigned)i < nr_cpumask_bits)
5922 i = select_idle_cpu(p, sd, target);
5923 if ((unsigned)i < nr_cpumask_bits)
5926 i = select_idle_smt(p, target);
5927 if ((unsigned)i < nr_cpumask_bits)
5934 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
5935 * @cpu: the CPU to get the utilization of
5937 * The unit of the return value must be the one of capacity so we can compare
5938 * the utilization with the capacity of the CPU that is available for CFS task
5939 * (ie cpu_capacity).
5941 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5942 * recent utilization of currently non-runnable tasks on a CPU. It represents
5943 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5944 * capacity_orig is the cpu_capacity available at the highest frequency
5945 * (arch_scale_freq_capacity()).
5946 * The utilization of a CPU converges towards a sum equal to or less than the
5947 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5948 * the running time on this CPU scaled by capacity_curr.
5950 * The estimated utilization of a CPU is defined to be the maximum between its
5951 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
5952 * currently RUNNABLE on that CPU.
5953 * This allows to properly represent the expected utilization of a CPU which
5954 * has just got a big task running since a long sleep period. At the same time
5955 * however it preserves the benefits of the "blocked utilization" in
5956 * describing the potential for other tasks waking up on the same CPU.
5958 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5959 * higher than capacity_orig because of unfortunate rounding in
5960 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5961 * the average stabilizes with the new running time. We need to check that the
5962 * utilization stays within the range of [0..capacity_orig] and cap it if
5963 * necessary. Without utilization capping, a group could be seen as overloaded
5964 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5965 * available capacity. We allow utilization to overshoot capacity_curr (but not
5966 * capacity_orig) as it useful for predicting the capacity required after task
5967 * migrations (scheduler-driven DVFS).
5969 * Return: the (estimated) utilization for the specified CPU
5971 static inline unsigned long cpu_util(int cpu)
5973 struct cfs_rq *cfs_rq;
5976 cfs_rq = &cpu_rq(cpu)->cfs;
5977 util = READ_ONCE(cfs_rq->avg.util_avg);
5979 if (sched_feat(UTIL_EST))
5980 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
5982 return min_t(unsigned long, util, capacity_orig_of(cpu));
5986 * cpu_util_without: compute cpu utilization without any contributions from *p
5987 * @cpu: the CPU which utilization is requested
5988 * @p: the task which utilization should be discounted
5990 * The utilization of a CPU is defined by the utilization of tasks currently
5991 * enqueued on that CPU as well as tasks which are currently sleeping after an
5992 * execution on that CPU.
5994 * This method returns the utilization of the specified CPU by discounting the
5995 * utilization of the specified task, whenever the task is currently
5996 * contributing to the CPU utilization.
5998 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6000 struct cfs_rq *cfs_rq;
6003 /* Task has no contribution or is new */
6004 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6005 return cpu_util(cpu);
6007 cfs_rq = &cpu_rq(cpu)->cfs;
6008 util = READ_ONCE(cfs_rq->avg.util_avg);
6010 /* Discount task's util from CPU's util */
6011 lsub_positive(&util, task_util(p));
6016 * a) if *p is the only task sleeping on this CPU, then:
6017 * cpu_util (== task_util) > util_est (== 0)
6018 * and thus we return:
6019 * cpu_util_without = (cpu_util - task_util) = 0
6021 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6023 * cpu_util >= task_util
6024 * cpu_util > util_est (== 0)
6025 * and thus we discount *p's blocked utilization to return:
6026 * cpu_util_without = (cpu_util - task_util) >= 0
6028 * c) if other tasks are RUNNABLE on that CPU and
6029 * util_est > cpu_util
6030 * then we use util_est since it returns a more restrictive
6031 * estimation of the spare capacity on that CPU, by just
6032 * considering the expected utilization of tasks already
6033 * runnable on that CPU.
6035 * Cases a) and b) are covered by the above code, while case c) is
6036 * covered by the following code when estimated utilization is
6039 if (sched_feat(UTIL_EST)) {
6040 unsigned int estimated =
6041 READ_ONCE(cfs_rq->avg.util_est.enqueued);
6044 * Despite the following checks we still have a small window
6045 * for a possible race, when an execl's select_task_rq_fair()
6046 * races with LB's detach_task():
6049 * p->on_rq = TASK_ON_RQ_MIGRATING;
6050 * ---------------------------------- A
6051 * deactivate_task() \
6052 * dequeue_task() + RaceTime
6053 * util_est_dequeue() /
6054 * ---------------------------------- B
6056 * The additional check on "current == p" it's required to
6057 * properly fix the execl regression and it helps in further
6058 * reducing the chances for the above race.
6060 if (unlikely(task_on_rq_queued(p) || current == p))
6061 lsub_positive(&estimated, _task_util_est(p));
6063 util = max(util, estimated);
6067 * Utilization (estimated) can exceed the CPU capacity, thus let's
6068 * clamp to the maximum CPU capacity to ensure consistency with
6069 * the cpu_util call.
6071 return min_t(unsigned long, util, capacity_orig_of(cpu));
6075 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6076 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6078 * In that case WAKE_AFFINE doesn't make sense and we'll let
6079 * BALANCE_WAKE sort things out.
6081 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6083 long min_cap, max_cap;
6085 if (!static_branch_unlikely(&sched_asym_cpucapacity))
6088 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6089 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6091 /* Minimum capacity is close to max, no need to abort wake_affine */
6092 if (max_cap - min_cap < max_cap >> 3)
6095 /* Bring task utilization in sync with prev_cpu */
6096 sync_entity_load_avg(&p->se);
6098 return !task_fits_capacity(p, min_cap);
6102 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6105 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6107 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6108 unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
6111 * If @p migrates from @cpu to another, remove its contribution. Or,
6112 * if @p migrates from another CPU to @cpu, add its contribution. In
6113 * the other cases, @cpu is not impacted by the migration, so the
6114 * util_avg should already be correct.
6116 if (task_cpu(p) == cpu && dst_cpu != cpu)
6117 sub_positive(&util, task_util(p));
6118 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6119 util += task_util(p);
6121 if (sched_feat(UTIL_EST)) {
6122 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6125 * During wake-up, the task isn't enqueued yet and doesn't
6126 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6127 * so just add it (if needed) to "simulate" what will be
6128 * cpu_util() after the task has been enqueued.
6131 util_est += _task_util_est(p);
6133 util = max(util, util_est);
6136 return min(util, capacity_orig_of(cpu));
6140 * compute_energy(): Estimates the energy that @pd would consume if @p was
6141 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6142 * landscape of @pd's CPUs after the task migration, and uses the Energy Model
6143 * to compute what would be the energy if we decided to actually migrate that
6147 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
6149 struct cpumask *pd_mask = perf_domain_span(pd);
6150 unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask));
6151 unsigned long max_util = 0, sum_util = 0;
6155 * The capacity state of CPUs of the current rd can be driven by CPUs
6156 * of another rd if they belong to the same pd. So, account for the
6157 * utilization of these CPUs too by masking pd with cpu_online_mask
6158 * instead of the rd span.
6160 * If an entire pd is outside of the current rd, it will not appear in
6161 * its pd list and will not be accounted by compute_energy().
6163 for_each_cpu_and(cpu, pd_mask, cpu_online_mask) {
6164 unsigned long cpu_util, util_cfs = cpu_util_next(cpu, p, dst_cpu);
6165 struct task_struct *tsk = cpu == dst_cpu ? p : NULL;
6168 * Busy time computation: utilization clamping is not
6169 * required since the ratio (sum_util / cpu_capacity)
6170 * is already enough to scale the EM reported power
6171 * consumption at the (eventually clamped) cpu_capacity.
6173 sum_util += schedutil_cpu_util(cpu, util_cfs, cpu_cap,
6177 * Performance domain frequency: utilization clamping
6178 * must be considered since it affects the selection
6179 * of the performance domain frequency.
6180 * NOTE: in case RT tasks are running, by default the
6181 * FREQUENCY_UTIL's utilization can be max OPP.
6183 cpu_util = schedutil_cpu_util(cpu, util_cfs, cpu_cap,
6184 FREQUENCY_UTIL, tsk);
6185 max_util = max(max_util, cpu_util);
6188 return em_pd_energy(pd->em_pd, max_util, sum_util);
6192 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6193 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6194 * spare capacity in each performance domain and uses it as a potential
6195 * candidate to execute the task. Then, it uses the Energy Model to figure
6196 * out which of the CPU candidates is the most energy-efficient.
6198 * The rationale for this heuristic is as follows. In a performance domain,
6199 * all the most energy efficient CPU candidates (according to the Energy
6200 * Model) are those for which we'll request a low frequency. When there are
6201 * several CPUs for which the frequency request will be the same, we don't
6202 * have enough data to break the tie between them, because the Energy Model
6203 * only includes active power costs. With this model, if we assume that
6204 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6205 * the maximum spare capacity in a performance domain is guaranteed to be among
6206 * the best candidates of the performance domain.
6208 * In practice, it could be preferable from an energy standpoint to pack
6209 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6210 * but that could also hurt our chances to go cluster idle, and we have no
6211 * ways to tell with the current Energy Model if this is actually a good
6212 * idea or not. So, find_energy_efficient_cpu() basically favors
6213 * cluster-packing, and spreading inside a cluster. That should at least be
6214 * a good thing for latency, and this is consistent with the idea that most
6215 * of the energy savings of EAS come from the asymmetry of the system, and
6216 * not so much from breaking the tie between identical CPUs. That's also the
6217 * reason why EAS is enabled in the topology code only for systems where
6218 * SD_ASYM_CPUCAPACITY is set.
6220 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6221 * they don't have any useful utilization data yet and it's not possible to
6222 * forecast their impact on energy consumption. Consequently, they will be
6223 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6224 * to be energy-inefficient in some use-cases. The alternative would be to
6225 * bias new tasks towards specific types of CPUs first, or to try to infer
6226 * their util_avg from the parent task, but those heuristics could hurt
6227 * other use-cases too. So, until someone finds a better way to solve this,
6228 * let's keep things simple by re-using the existing slow path.
6230 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6232 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
6233 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6234 unsigned long cpu_cap, util, base_energy = 0;
6235 int cpu, best_energy_cpu = prev_cpu;
6236 struct sched_domain *sd;
6237 struct perf_domain *pd;
6240 pd = rcu_dereference(rd->pd);
6241 if (!pd || READ_ONCE(rd->overutilized))
6245 * Energy-aware wake-up happens on the lowest sched_domain starting
6246 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6248 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6249 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6254 sync_entity_load_avg(&p->se);
6255 if (!task_util_est(p))
6258 for (; pd; pd = pd->next) {
6259 unsigned long cur_delta, spare_cap, max_spare_cap = 0;
6260 unsigned long base_energy_pd;
6261 int max_spare_cap_cpu = -1;
6263 /* Compute the 'base' energy of the pd, without @p */
6264 base_energy_pd = compute_energy(p, -1, pd);
6265 base_energy += base_energy_pd;
6267 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
6268 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6271 /* Skip CPUs that will be overutilized. */
6272 util = cpu_util_next(cpu, p, cpu);
6273 cpu_cap = capacity_of(cpu);
6274 if (!fits_capacity(util, cpu_cap))
6277 /* Always use prev_cpu as a candidate. */
6278 if (cpu == prev_cpu) {
6279 prev_delta = compute_energy(p, prev_cpu, pd);
6280 prev_delta -= base_energy_pd;
6281 best_delta = min(best_delta, prev_delta);
6285 * Find the CPU with the maximum spare capacity in
6286 * the performance domain
6288 spare_cap = cpu_cap - util;
6289 if (spare_cap > max_spare_cap) {
6290 max_spare_cap = spare_cap;
6291 max_spare_cap_cpu = cpu;
6295 /* Evaluate the energy impact of using this CPU. */
6296 if (max_spare_cap_cpu >= 0 && max_spare_cap_cpu != prev_cpu) {
6297 cur_delta = compute_energy(p, max_spare_cap_cpu, pd);
6298 cur_delta -= base_energy_pd;
6299 if (cur_delta < best_delta) {
6300 best_delta = cur_delta;
6301 best_energy_cpu = max_spare_cap_cpu;
6309 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6310 * least 6% of the energy used by prev_cpu.
6312 if (prev_delta == ULONG_MAX)
6313 return best_energy_cpu;
6315 if ((prev_delta - best_delta) > ((prev_delta + base_energy) >> 4))
6316 return best_energy_cpu;
6327 * select_task_rq_fair: Select target runqueue for the waking task in domains
6328 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6329 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6331 * Balances load by selecting the idlest CPU in the idlest group, or under
6332 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6334 * Returns the target CPU number.
6336 * preempt must be disabled.
6339 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6341 struct sched_domain *tmp, *sd = NULL;
6342 int cpu = smp_processor_id();
6343 int new_cpu = prev_cpu;
6344 int want_affine = 0;
6345 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6347 if (sd_flag & SD_BALANCE_WAKE) {
6350 if (sched_energy_enabled()) {
6351 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
6357 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu) &&
6358 cpumask_test_cpu(cpu, p->cpus_ptr);
6362 for_each_domain(cpu, tmp) {
6363 if (!(tmp->flags & SD_LOAD_BALANCE))
6367 * If both 'cpu' and 'prev_cpu' are part of this domain,
6368 * cpu is a valid SD_WAKE_AFFINE target.
6370 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6371 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6372 if (cpu != prev_cpu)
6373 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6375 sd = NULL; /* Prefer wake_affine over balance flags */
6379 if (tmp->flags & sd_flag)
6381 else if (!want_affine)
6387 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6388 } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6391 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6394 current->recent_used_cpu = cpu;
6401 static void detach_entity_cfs_rq(struct sched_entity *se);
6404 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6405 * cfs_rq_of(p) references at time of call are still valid and identify the
6406 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6408 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6411 * As blocked tasks retain absolute vruntime the migration needs to
6412 * deal with this by subtracting the old and adding the new
6413 * min_vruntime -- the latter is done by enqueue_entity() when placing
6414 * the task on the new runqueue.
6416 if (p->state == TASK_WAKING) {
6417 struct sched_entity *se = &p->se;
6418 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6421 #ifndef CONFIG_64BIT
6422 u64 min_vruntime_copy;
6425 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6427 min_vruntime = cfs_rq->min_vruntime;
6428 } while (min_vruntime != min_vruntime_copy);
6430 min_vruntime = cfs_rq->min_vruntime;
6433 se->vruntime -= min_vruntime;
6436 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6438 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6439 * rq->lock and can modify state directly.
6441 lockdep_assert_held(&task_rq(p)->lock);
6442 detach_entity_cfs_rq(&p->se);
6446 * We are supposed to update the task to "current" time, then
6447 * its up to date and ready to go to new CPU/cfs_rq. But we
6448 * have difficulty in getting what current time is, so simply
6449 * throw away the out-of-date time. This will result in the
6450 * wakee task is less decayed, but giving the wakee more load
6453 remove_entity_load_avg(&p->se);
6456 /* Tell new CPU we are migrated */
6457 p->se.avg.last_update_time = 0;
6459 /* We have migrated, no longer consider this task hot */
6460 p->se.exec_start = 0;
6462 update_scan_period(p, new_cpu);
6465 static void task_dead_fair(struct task_struct *p)
6467 remove_entity_load_avg(&p->se);
6471 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6476 return newidle_balance(rq, rf) != 0;
6478 #endif /* CONFIG_SMP */
6480 static unsigned long wakeup_gran(struct sched_entity *se)
6482 unsigned long gran = sysctl_sched_wakeup_granularity;
6485 * Since its curr running now, convert the gran from real-time
6486 * to virtual-time in his units.
6488 * By using 'se' instead of 'curr' we penalize light tasks, so
6489 * they get preempted easier. That is, if 'se' < 'curr' then
6490 * the resulting gran will be larger, therefore penalizing the
6491 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6492 * be smaller, again penalizing the lighter task.
6494 * This is especially important for buddies when the leftmost
6495 * task is higher priority than the buddy.
6497 return calc_delta_fair(gran, se);
6501 * Should 'se' preempt 'curr'.
6515 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6517 s64 gran, vdiff = curr->vruntime - se->vruntime;
6522 gran = wakeup_gran(se);
6529 static void set_last_buddy(struct sched_entity *se)
6531 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6534 for_each_sched_entity(se) {
6535 if (SCHED_WARN_ON(!se->on_rq))
6537 cfs_rq_of(se)->last = se;
6541 static void set_next_buddy(struct sched_entity *se)
6543 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6546 for_each_sched_entity(se) {
6547 if (SCHED_WARN_ON(!se->on_rq))
6549 cfs_rq_of(se)->next = se;
6553 static void set_skip_buddy(struct sched_entity *se)
6555 for_each_sched_entity(se)
6556 cfs_rq_of(se)->skip = se;
6560 * Preempt the current task with a newly woken task if needed:
6562 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6564 struct task_struct *curr = rq->curr;
6565 struct sched_entity *se = &curr->se, *pse = &p->se;
6566 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6567 int scale = cfs_rq->nr_running >= sched_nr_latency;
6568 int next_buddy_marked = 0;
6570 if (unlikely(se == pse))
6574 * This is possible from callers such as attach_tasks(), in which we
6575 * unconditionally check_prempt_curr() after an enqueue (which may have
6576 * lead to a throttle). This both saves work and prevents false
6577 * next-buddy nomination below.
6579 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6582 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6583 set_next_buddy(pse);
6584 next_buddy_marked = 1;
6588 * We can come here with TIF_NEED_RESCHED already set from new task
6591 * Note: this also catches the edge-case of curr being in a throttled
6592 * group (e.g. via set_curr_task), since update_curr() (in the
6593 * enqueue of curr) will have resulted in resched being set. This
6594 * prevents us from potentially nominating it as a false LAST_BUDDY
6597 if (test_tsk_need_resched(curr))
6600 /* Idle tasks are by definition preempted by non-idle tasks. */
6601 if (unlikely(task_has_idle_policy(curr)) &&
6602 likely(!task_has_idle_policy(p)))
6606 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6607 * is driven by the tick):
6609 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6612 find_matching_se(&se, &pse);
6613 update_curr(cfs_rq_of(se));
6615 if (wakeup_preempt_entity(se, pse) == 1) {
6617 * Bias pick_next to pick the sched entity that is
6618 * triggering this preemption.
6620 if (!next_buddy_marked)
6621 set_next_buddy(pse);
6630 * Only set the backward buddy when the current task is still
6631 * on the rq. This can happen when a wakeup gets interleaved
6632 * with schedule on the ->pre_schedule() or idle_balance()
6633 * point, either of which can * drop the rq lock.
6635 * Also, during early boot the idle thread is in the fair class,
6636 * for obvious reasons its a bad idea to schedule back to it.
6638 if (unlikely(!se->on_rq || curr == rq->idle))
6641 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6645 struct task_struct *
6646 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6648 struct cfs_rq *cfs_rq = &rq->cfs;
6649 struct sched_entity *se;
6650 struct task_struct *p;
6654 if (!sched_fair_runnable(rq))
6657 #ifdef CONFIG_FAIR_GROUP_SCHED
6658 if (!prev || prev->sched_class != &fair_sched_class)
6662 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6663 * likely that a next task is from the same cgroup as the current.
6665 * Therefore attempt to avoid putting and setting the entire cgroup
6666 * hierarchy, only change the part that actually changes.
6670 struct sched_entity *curr = cfs_rq->curr;
6673 * Since we got here without doing put_prev_entity() we also
6674 * have to consider cfs_rq->curr. If it is still a runnable
6675 * entity, update_curr() will update its vruntime, otherwise
6676 * forget we've ever seen it.
6680 update_curr(cfs_rq);
6685 * This call to check_cfs_rq_runtime() will do the
6686 * throttle and dequeue its entity in the parent(s).
6687 * Therefore the nr_running test will indeed
6690 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6693 if (!cfs_rq->nr_running)
6700 se = pick_next_entity(cfs_rq, curr);
6701 cfs_rq = group_cfs_rq(se);
6707 * Since we haven't yet done put_prev_entity and if the selected task
6708 * is a different task than we started out with, try and touch the
6709 * least amount of cfs_rqs.
6712 struct sched_entity *pse = &prev->se;
6714 while (!(cfs_rq = is_same_group(se, pse))) {
6715 int se_depth = se->depth;
6716 int pse_depth = pse->depth;
6718 if (se_depth <= pse_depth) {
6719 put_prev_entity(cfs_rq_of(pse), pse);
6720 pse = parent_entity(pse);
6722 if (se_depth >= pse_depth) {
6723 set_next_entity(cfs_rq_of(se), se);
6724 se = parent_entity(se);
6728 put_prev_entity(cfs_rq, pse);
6729 set_next_entity(cfs_rq, se);
6736 put_prev_task(rq, prev);
6739 se = pick_next_entity(cfs_rq, NULL);
6740 set_next_entity(cfs_rq, se);
6741 cfs_rq = group_cfs_rq(se);
6746 done: __maybe_unused;
6749 * Move the next running task to the front of
6750 * the list, so our cfs_tasks list becomes MRU
6753 list_move(&p->se.group_node, &rq->cfs_tasks);
6756 if (hrtick_enabled(rq))
6757 hrtick_start_fair(rq, p);
6759 update_misfit_status(p, rq);
6767 new_tasks = newidle_balance(rq, rf);
6770 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
6771 * possible for any higher priority task to appear. In that case we
6772 * must re-start the pick_next_entity() loop.
6781 * rq is about to be idle, check if we need to update the
6782 * lost_idle_time of clock_pelt
6784 update_idle_rq_clock_pelt(rq);
6789 static struct task_struct *__pick_next_task_fair(struct rq *rq)
6791 return pick_next_task_fair(rq, NULL, NULL);
6795 * Account for a descheduled task:
6797 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6799 struct sched_entity *se = &prev->se;
6800 struct cfs_rq *cfs_rq;
6802 for_each_sched_entity(se) {
6803 cfs_rq = cfs_rq_of(se);
6804 put_prev_entity(cfs_rq, se);
6809 * sched_yield() is very simple
6811 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6813 static void yield_task_fair(struct rq *rq)
6815 struct task_struct *curr = rq->curr;
6816 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6817 struct sched_entity *se = &curr->se;
6820 * Are we the only task in the tree?
6822 if (unlikely(rq->nr_running == 1))
6825 clear_buddies(cfs_rq, se);
6827 if (curr->policy != SCHED_BATCH) {
6828 update_rq_clock(rq);
6830 * Update run-time statistics of the 'current'.
6832 update_curr(cfs_rq);
6834 * Tell update_rq_clock() that we've just updated,
6835 * so we don't do microscopic update in schedule()
6836 * and double the fastpath cost.
6838 rq_clock_skip_update(rq);
6844 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6846 struct sched_entity *se = &p->se;
6848 /* throttled hierarchies are not runnable */
6849 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6852 /* Tell the scheduler that we'd really like pse to run next. */
6855 yield_task_fair(rq);
6861 /**************************************************
6862 * Fair scheduling class load-balancing methods.
6866 * The purpose of load-balancing is to achieve the same basic fairness the
6867 * per-CPU scheduler provides, namely provide a proportional amount of compute
6868 * time to each task. This is expressed in the following equation:
6870 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6872 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
6873 * W_i,0 is defined as:
6875 * W_i,0 = \Sum_j w_i,j (2)
6877 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
6878 * is derived from the nice value as per sched_prio_to_weight[].
6880 * The weight average is an exponential decay average of the instantaneous
6883 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6885 * C_i is the compute capacity of CPU i, typically it is the
6886 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6887 * can also include other factors [XXX].
6889 * To achieve this balance we define a measure of imbalance which follows
6890 * directly from (1):
6892 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6894 * We them move tasks around to minimize the imbalance. In the continuous
6895 * function space it is obvious this converges, in the discrete case we get
6896 * a few fun cases generally called infeasible weight scenarios.
6899 * - infeasible weights;
6900 * - local vs global optima in the discrete case. ]
6905 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6906 * for all i,j solution, we create a tree of CPUs that follows the hardware
6907 * topology where each level pairs two lower groups (or better). This results
6908 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
6909 * tree to only the first of the previous level and we decrease the frequency
6910 * of load-balance at each level inv. proportional to the number of CPUs in
6916 * \Sum { --- * --- * 2^i } = O(n) (5)
6918 * `- size of each group
6919 * | | `- number of CPUs doing load-balance
6921 * `- sum over all levels
6923 * Coupled with a limit on how many tasks we can migrate every balance pass,
6924 * this makes (5) the runtime complexity of the balancer.
6926 * An important property here is that each CPU is still (indirectly) connected
6927 * to every other CPU in at most O(log n) steps:
6929 * The adjacency matrix of the resulting graph is given by:
6932 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6935 * And you'll find that:
6937 * A^(log_2 n)_i,j != 0 for all i,j (7)
6939 * Showing there's indeed a path between every CPU in at most O(log n) steps.
6940 * The task movement gives a factor of O(m), giving a convergence complexity
6943 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6948 * In order to avoid CPUs going idle while there's still work to do, new idle
6949 * balancing is more aggressive and has the newly idle CPU iterate up the domain
6950 * tree itself instead of relying on other CPUs to bring it work.
6952 * This adds some complexity to both (5) and (8) but it reduces the total idle
6960 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6963 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6968 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6970 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
6972 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6975 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6976 * rewrite all of this once again.]
6979 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6981 enum fbq_type { regular, remote, all };
6984 * 'group_type' describes the group of CPUs at the moment of load balancing.
6986 * The enum is ordered by pulling priority, with the group with lowest priority
6987 * first so the group_type can simply be compared when selecting the busiest
6988 * group. See update_sd_pick_busiest().
6991 /* The group has spare capacity that can be used to run more tasks. */
6992 group_has_spare = 0,
6994 * The group is fully used and the tasks don't compete for more CPU
6995 * cycles. Nevertheless, some tasks might wait before running.
6999 * SD_ASYM_CPUCAPACITY only: One task doesn't fit with CPU's capacity
7000 * and must be migrated to a more powerful CPU.
7004 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
7005 * and the task should be migrated to it instead of running on the
7010 * The tasks' affinity constraints previously prevented the scheduler
7011 * from balancing the load across the system.
7015 * The CPU is overloaded and can't provide expected CPU cycles to all
7021 enum migration_type {
7028 #define LBF_ALL_PINNED 0x01
7029 #define LBF_NEED_BREAK 0x02
7030 #define LBF_DST_PINNED 0x04
7031 #define LBF_SOME_PINNED 0x08
7032 #define LBF_NOHZ_STATS 0x10
7033 #define LBF_NOHZ_AGAIN 0x20
7036 struct sched_domain *sd;
7044 struct cpumask *dst_grpmask;
7046 enum cpu_idle_type idle;
7048 /* The set of CPUs under consideration for load-balancing */
7049 struct cpumask *cpus;
7054 unsigned int loop_break;
7055 unsigned int loop_max;
7057 enum fbq_type fbq_type;
7058 enum migration_type migration_type;
7059 struct list_head tasks;
7063 * Is this task likely cache-hot:
7065 static int task_hot(struct task_struct *p, struct lb_env *env)
7069 lockdep_assert_held(&env->src_rq->lock);
7071 if (p->sched_class != &fair_sched_class)
7074 if (unlikely(task_has_idle_policy(p)))
7078 * Buddy candidates are cache hot:
7080 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7081 (&p->se == cfs_rq_of(&p->se)->next ||
7082 &p->se == cfs_rq_of(&p->se)->last))
7085 if (sysctl_sched_migration_cost == -1)
7087 if (sysctl_sched_migration_cost == 0)
7090 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7092 return delta < (s64)sysctl_sched_migration_cost;
7095 #ifdef CONFIG_NUMA_BALANCING
7097 * Returns 1, if task migration degrades locality
7098 * Returns 0, if task migration improves locality i.e migration preferred.
7099 * Returns -1, if task migration is not affected by locality.
7101 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7103 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7104 unsigned long src_weight, dst_weight;
7105 int src_nid, dst_nid, dist;
7107 if (!static_branch_likely(&sched_numa_balancing))
7110 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7113 src_nid = cpu_to_node(env->src_cpu);
7114 dst_nid = cpu_to_node(env->dst_cpu);
7116 if (src_nid == dst_nid)
7119 /* Migrating away from the preferred node is always bad. */
7120 if (src_nid == p->numa_preferred_nid) {
7121 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7127 /* Encourage migration to the preferred node. */
7128 if (dst_nid == p->numa_preferred_nid)
7131 /* Leaving a core idle is often worse than degrading locality. */
7132 if (env->idle == CPU_IDLE)
7135 dist = node_distance(src_nid, dst_nid);
7137 src_weight = group_weight(p, src_nid, dist);
7138 dst_weight = group_weight(p, dst_nid, dist);
7140 src_weight = task_weight(p, src_nid, dist);
7141 dst_weight = task_weight(p, dst_nid, dist);
7144 return dst_weight < src_weight;
7148 static inline int migrate_degrades_locality(struct task_struct *p,
7156 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7159 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7163 lockdep_assert_held(&env->src_rq->lock);
7166 * We do not migrate tasks that are:
7167 * 1) throttled_lb_pair, or
7168 * 2) cannot be migrated to this CPU due to cpus_ptr, or
7169 * 3) running (obviously), or
7170 * 4) are cache-hot on their current CPU.
7172 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7175 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
7178 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7180 env->flags |= LBF_SOME_PINNED;
7183 * Remember if this task can be migrated to any other CPU in
7184 * our sched_group. We may want to revisit it if we couldn't
7185 * meet load balance goals by pulling other tasks on src_cpu.
7187 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7188 * already computed one in current iteration.
7190 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7193 /* Prevent to re-select dst_cpu via env's CPUs: */
7194 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7195 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
7196 env->flags |= LBF_DST_PINNED;
7197 env->new_dst_cpu = cpu;
7205 /* Record that we found atleast one task that could run on dst_cpu */
7206 env->flags &= ~LBF_ALL_PINNED;
7208 if (task_running(env->src_rq, p)) {
7209 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7214 * Aggressive migration if:
7215 * 1) destination numa is preferred
7216 * 2) task is cache cold, or
7217 * 3) too many balance attempts have failed.
7219 tsk_cache_hot = migrate_degrades_locality(p, env);
7220 if (tsk_cache_hot == -1)
7221 tsk_cache_hot = task_hot(p, env);
7223 if (tsk_cache_hot <= 0 ||
7224 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7225 if (tsk_cache_hot == 1) {
7226 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7227 schedstat_inc(p->se.statistics.nr_forced_migrations);
7232 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7237 * detach_task() -- detach the task for the migration specified in env
7239 static void detach_task(struct task_struct *p, struct lb_env *env)
7241 lockdep_assert_held(&env->src_rq->lock);
7243 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7244 set_task_cpu(p, env->dst_cpu);
7248 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7249 * part of active balancing operations within "domain".
7251 * Returns a task if successful and NULL otherwise.
7253 static struct task_struct *detach_one_task(struct lb_env *env)
7255 struct task_struct *p;
7257 lockdep_assert_held(&env->src_rq->lock);
7259 list_for_each_entry_reverse(p,
7260 &env->src_rq->cfs_tasks, se.group_node) {
7261 if (!can_migrate_task(p, env))
7264 detach_task(p, env);
7267 * Right now, this is only the second place where
7268 * lb_gained[env->idle] is updated (other is detach_tasks)
7269 * so we can safely collect stats here rather than
7270 * inside detach_tasks().
7272 schedstat_inc(env->sd->lb_gained[env->idle]);
7278 static const unsigned int sched_nr_migrate_break = 32;
7281 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
7282 * busiest_rq, as part of a balancing operation within domain "sd".
7284 * Returns number of detached tasks if successful and 0 otherwise.
7286 static int detach_tasks(struct lb_env *env)
7288 struct list_head *tasks = &env->src_rq->cfs_tasks;
7289 unsigned long util, load;
7290 struct task_struct *p;
7293 lockdep_assert_held(&env->src_rq->lock);
7295 if (env->imbalance <= 0)
7298 while (!list_empty(tasks)) {
7300 * We don't want to steal all, otherwise we may be treated likewise,
7301 * which could at worst lead to a livelock crash.
7303 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7306 p = list_last_entry(tasks, struct task_struct, se.group_node);
7309 /* We've more or less seen every task there is, call it quits */
7310 if (env->loop > env->loop_max)
7313 /* take a breather every nr_migrate tasks */
7314 if (env->loop > env->loop_break) {
7315 env->loop_break += sched_nr_migrate_break;
7316 env->flags |= LBF_NEED_BREAK;
7320 if (!can_migrate_task(p, env))
7323 switch (env->migration_type) {
7325 load = task_h_load(p);
7327 if (sched_feat(LB_MIN) &&
7328 load < 16 && !env->sd->nr_balance_failed)
7331 if (load/2 > env->imbalance)
7334 env->imbalance -= load;
7338 util = task_util_est(p);
7340 if (util > env->imbalance)
7343 env->imbalance -= util;
7350 case migrate_misfit:
7351 /* This is not a misfit task */
7352 if (task_fits_capacity(p, capacity_of(env->src_cpu)))
7359 detach_task(p, env);
7360 list_add(&p->se.group_node, &env->tasks);
7364 #ifdef CONFIG_PREEMPTION
7366 * NEWIDLE balancing is a source of latency, so preemptible
7367 * kernels will stop after the first task is detached to minimize
7368 * the critical section.
7370 if (env->idle == CPU_NEWLY_IDLE)
7375 * We only want to steal up to the prescribed amount of
7378 if (env->imbalance <= 0)
7383 list_move(&p->se.group_node, tasks);
7387 * Right now, this is one of only two places we collect this stat
7388 * so we can safely collect detach_one_task() stats here rather
7389 * than inside detach_one_task().
7391 schedstat_add(env->sd->lb_gained[env->idle], detached);
7397 * attach_task() -- attach the task detached by detach_task() to its new rq.
7399 static void attach_task(struct rq *rq, struct task_struct *p)
7401 lockdep_assert_held(&rq->lock);
7403 BUG_ON(task_rq(p) != rq);
7404 activate_task(rq, p, ENQUEUE_NOCLOCK);
7405 check_preempt_curr(rq, p, 0);
7409 * attach_one_task() -- attaches the task returned from detach_one_task() to
7412 static void attach_one_task(struct rq *rq, struct task_struct *p)
7417 update_rq_clock(rq);
7423 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7426 static void attach_tasks(struct lb_env *env)
7428 struct list_head *tasks = &env->tasks;
7429 struct task_struct *p;
7432 rq_lock(env->dst_rq, &rf);
7433 update_rq_clock(env->dst_rq);
7435 while (!list_empty(tasks)) {
7436 p = list_first_entry(tasks, struct task_struct, se.group_node);
7437 list_del_init(&p->se.group_node);
7439 attach_task(env->dst_rq, p);
7442 rq_unlock(env->dst_rq, &rf);
7445 #ifdef CONFIG_NO_HZ_COMMON
7446 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7448 if (cfs_rq->avg.load_avg)
7451 if (cfs_rq->avg.util_avg)
7457 static inline bool others_have_blocked(struct rq *rq)
7459 if (READ_ONCE(rq->avg_rt.util_avg))
7462 if (READ_ONCE(rq->avg_dl.util_avg))
7465 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7466 if (READ_ONCE(rq->avg_irq.util_avg))
7473 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
7475 rq->last_blocked_load_update_tick = jiffies;
7478 rq->has_blocked_load = 0;
7481 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
7482 static inline bool others_have_blocked(struct rq *rq) { return false; }
7483 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
7486 static bool __update_blocked_others(struct rq *rq, bool *done)
7488 const struct sched_class *curr_class;
7489 u64 now = rq_clock_pelt(rq);
7493 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
7494 * DL and IRQ signals have been updated before updating CFS.
7496 curr_class = rq->curr->sched_class;
7498 decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
7499 update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
7500 update_irq_load_avg(rq, 0);
7502 if (others_have_blocked(rq))
7508 #ifdef CONFIG_FAIR_GROUP_SCHED
7510 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7512 if (cfs_rq->load.weight)
7515 if (cfs_rq->avg.load_sum)
7518 if (cfs_rq->avg.util_sum)
7521 if (cfs_rq->avg.runnable_load_sum)
7527 static bool __update_blocked_fair(struct rq *rq, bool *done)
7529 struct cfs_rq *cfs_rq, *pos;
7530 bool decayed = false;
7531 int cpu = cpu_of(rq);
7534 * Iterates the task_group tree in a bottom up fashion, see
7535 * list_add_leaf_cfs_rq() for details.
7537 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7538 struct sched_entity *se;
7540 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
7541 update_tg_load_avg(cfs_rq, 0);
7543 if (cfs_rq == &rq->cfs)
7547 /* Propagate pending load changes to the parent, if any: */
7548 se = cfs_rq->tg->se[cpu];
7549 if (se && !skip_blocked_update(se))
7550 update_load_avg(cfs_rq_of(se), se, 0);
7553 * There can be a lot of idle CPU cgroups. Don't let fully
7554 * decayed cfs_rqs linger on the list.
7556 if (cfs_rq_is_decayed(cfs_rq))
7557 list_del_leaf_cfs_rq(cfs_rq);
7559 /* Don't need periodic decay once load/util_avg are null */
7560 if (cfs_rq_has_blocked(cfs_rq))
7568 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7569 * This needs to be done in a top-down fashion because the load of a child
7570 * group is a fraction of its parents load.
7572 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7574 struct rq *rq = rq_of(cfs_rq);
7575 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7576 unsigned long now = jiffies;
7579 if (cfs_rq->last_h_load_update == now)
7582 WRITE_ONCE(cfs_rq->h_load_next, NULL);
7583 for_each_sched_entity(se) {
7584 cfs_rq = cfs_rq_of(se);
7585 WRITE_ONCE(cfs_rq->h_load_next, se);
7586 if (cfs_rq->last_h_load_update == now)
7591 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7592 cfs_rq->last_h_load_update = now;
7595 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
7596 load = cfs_rq->h_load;
7597 load = div64_ul(load * se->avg.load_avg,
7598 cfs_rq_load_avg(cfs_rq) + 1);
7599 cfs_rq = group_cfs_rq(se);
7600 cfs_rq->h_load = load;
7601 cfs_rq->last_h_load_update = now;
7605 static unsigned long task_h_load(struct task_struct *p)
7607 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7609 update_cfs_rq_h_load(cfs_rq);
7610 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7611 cfs_rq_load_avg(cfs_rq) + 1);
7614 static bool __update_blocked_fair(struct rq *rq, bool *done)
7616 struct cfs_rq *cfs_rq = &rq->cfs;
7619 decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
7620 if (cfs_rq_has_blocked(cfs_rq))
7626 static unsigned long task_h_load(struct task_struct *p)
7628 return p->se.avg.load_avg;
7632 static void update_blocked_averages(int cpu)
7634 bool decayed = false, done = true;
7635 struct rq *rq = cpu_rq(cpu);
7638 rq_lock_irqsave(rq, &rf);
7639 update_rq_clock(rq);
7641 decayed |= __update_blocked_others(rq, &done);
7642 decayed |= __update_blocked_fair(rq, &done);
7644 update_blocked_load_status(rq, !done);
7646 cpufreq_update_util(rq, 0);
7647 rq_unlock_irqrestore(rq, &rf);
7650 /********** Helpers for find_busiest_group ************************/
7653 * sg_lb_stats - stats of a sched_group required for load_balancing
7655 struct sg_lb_stats {
7656 unsigned long avg_load; /*Avg load across the CPUs of the group */
7657 unsigned long group_load; /* Total load over the CPUs of the group */
7658 unsigned long group_capacity;
7659 unsigned long group_util; /* Total utilization of the group */
7660 unsigned int sum_nr_running; /* Nr of tasks running in the group */
7661 unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
7662 unsigned int idle_cpus;
7663 unsigned int group_weight;
7664 enum group_type group_type;
7665 unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
7666 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
7667 #ifdef CONFIG_NUMA_BALANCING
7668 unsigned int nr_numa_running;
7669 unsigned int nr_preferred_running;
7674 * sd_lb_stats - Structure to store the statistics of a sched_domain
7675 * during load balancing.
7677 struct sd_lb_stats {
7678 struct sched_group *busiest; /* Busiest group in this sd */
7679 struct sched_group *local; /* Local group in this sd */
7680 unsigned long total_load; /* Total load of all groups in sd */
7681 unsigned long total_capacity; /* Total capacity of all groups in sd */
7682 unsigned long avg_load; /* Average load across all groups in sd */
7683 unsigned int prefer_sibling; /* tasks should go to sibling first */
7685 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7686 struct sg_lb_stats local_stat; /* Statistics of the local group */
7689 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7692 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7693 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7694 * We must however set busiest_stat::group_type and
7695 * busiest_stat::idle_cpus to the worst busiest group because
7696 * update_sd_pick_busiest() reads these before assignment.
7698 *sds = (struct sd_lb_stats){
7702 .total_capacity = 0UL,
7704 .idle_cpus = UINT_MAX,
7705 .group_type = group_has_spare,
7710 static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu)
7712 struct rq *rq = cpu_rq(cpu);
7713 unsigned long max = arch_scale_cpu_capacity(cpu);
7714 unsigned long used, free;
7717 irq = cpu_util_irq(rq);
7719 if (unlikely(irq >= max))
7722 used = READ_ONCE(rq->avg_rt.util_avg);
7723 used += READ_ONCE(rq->avg_dl.util_avg);
7725 if (unlikely(used >= max))
7730 return scale_irq_capacity(free, irq, max);
7733 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7735 unsigned long capacity = scale_rt_capacity(sd, cpu);
7736 struct sched_group *sdg = sd->groups;
7738 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
7743 cpu_rq(cpu)->cpu_capacity = capacity;
7744 sdg->sgc->capacity = capacity;
7745 sdg->sgc->min_capacity = capacity;
7746 sdg->sgc->max_capacity = capacity;
7749 void update_group_capacity(struct sched_domain *sd, int cpu)
7751 struct sched_domain *child = sd->child;
7752 struct sched_group *group, *sdg = sd->groups;
7753 unsigned long capacity, min_capacity, max_capacity;
7754 unsigned long interval;
7756 interval = msecs_to_jiffies(sd->balance_interval);
7757 interval = clamp(interval, 1UL, max_load_balance_interval);
7758 sdg->sgc->next_update = jiffies + interval;
7761 update_cpu_capacity(sd, cpu);
7766 min_capacity = ULONG_MAX;
7769 if (child->flags & SD_OVERLAP) {
7771 * SD_OVERLAP domains cannot assume that child groups
7772 * span the current group.
7775 for_each_cpu(cpu, sched_group_span(sdg)) {
7776 struct sched_group_capacity *sgc;
7777 struct rq *rq = cpu_rq(cpu);
7780 * build_sched_domains() -> init_sched_groups_capacity()
7781 * gets here before we've attached the domains to the
7784 * Use capacity_of(), which is set irrespective of domains
7785 * in update_cpu_capacity().
7787 * This avoids capacity from being 0 and
7788 * causing divide-by-zero issues on boot.
7790 if (unlikely(!rq->sd)) {
7791 capacity += capacity_of(cpu);
7793 sgc = rq->sd->groups->sgc;
7794 capacity += sgc->capacity;
7797 min_capacity = min(capacity, min_capacity);
7798 max_capacity = max(capacity, max_capacity);
7802 * !SD_OVERLAP domains can assume that child groups
7803 * span the current group.
7806 group = child->groups;
7808 struct sched_group_capacity *sgc = group->sgc;
7810 capacity += sgc->capacity;
7811 min_capacity = min(sgc->min_capacity, min_capacity);
7812 max_capacity = max(sgc->max_capacity, max_capacity);
7813 group = group->next;
7814 } while (group != child->groups);
7817 sdg->sgc->capacity = capacity;
7818 sdg->sgc->min_capacity = min_capacity;
7819 sdg->sgc->max_capacity = max_capacity;
7823 * Check whether the capacity of the rq has been noticeably reduced by side
7824 * activity. The imbalance_pct is used for the threshold.
7825 * Return true is the capacity is reduced
7828 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7830 return ((rq->cpu_capacity * sd->imbalance_pct) <
7831 (rq->cpu_capacity_orig * 100));
7835 * Check whether a rq has a misfit task and if it looks like we can actually
7836 * help that task: we can migrate the task to a CPU of higher capacity, or
7837 * the task's current CPU is heavily pressured.
7839 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
7841 return rq->misfit_task_load &&
7842 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
7843 check_cpu_capacity(rq, sd));
7847 * Group imbalance indicates (and tries to solve) the problem where balancing
7848 * groups is inadequate due to ->cpus_ptr constraints.
7850 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
7851 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
7854 * { 0 1 2 3 } { 4 5 6 7 }
7857 * If we were to balance group-wise we'd place two tasks in the first group and
7858 * two tasks in the second group. Clearly this is undesired as it will overload
7859 * cpu 3 and leave one of the CPUs in the second group unused.
7861 * The current solution to this issue is detecting the skew in the first group
7862 * by noticing the lower domain failed to reach balance and had difficulty
7863 * moving tasks due to affinity constraints.
7865 * When this is so detected; this group becomes a candidate for busiest; see
7866 * update_sd_pick_busiest(). And calculate_imbalance() and
7867 * find_busiest_group() avoid some of the usual balance conditions to allow it
7868 * to create an effective group imbalance.
7870 * This is a somewhat tricky proposition since the next run might not find the
7871 * group imbalance and decide the groups need to be balanced again. A most
7872 * subtle and fragile situation.
7875 static inline int sg_imbalanced(struct sched_group *group)
7877 return group->sgc->imbalance;
7881 * group_has_capacity returns true if the group has spare capacity that could
7882 * be used by some tasks.
7883 * We consider that a group has spare capacity if the * number of task is
7884 * smaller than the number of CPUs or if the utilization is lower than the
7885 * available capacity for CFS tasks.
7886 * For the latter, we use a threshold to stabilize the state, to take into
7887 * account the variance of the tasks' load and to return true if the available
7888 * capacity in meaningful for the load balancer.
7889 * As an example, an available capacity of 1% can appear but it doesn't make
7890 * any benefit for the load balance.
7893 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
7895 if (sgs->sum_nr_running < sgs->group_weight)
7898 if ((sgs->group_capacity * 100) >
7899 (sgs->group_util * imbalance_pct))
7906 * group_is_overloaded returns true if the group has more tasks than it can
7908 * group_is_overloaded is not equals to !group_has_capacity because a group
7909 * with the exact right number of tasks, has no more spare capacity but is not
7910 * overloaded so both group_has_capacity and group_is_overloaded return
7914 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
7916 if (sgs->sum_nr_running <= sgs->group_weight)
7919 if ((sgs->group_capacity * 100) <
7920 (sgs->group_util * imbalance_pct))
7927 * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller
7928 * per-CPU capacity than sched_group ref.
7931 group_smaller_min_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7933 return fits_capacity(sg->sgc->min_capacity, ref->sgc->min_capacity);
7937 * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller
7938 * per-CPU capacity_orig than sched_group ref.
7941 group_smaller_max_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7943 return fits_capacity(sg->sgc->max_capacity, ref->sgc->max_capacity);
7947 group_type group_classify(unsigned int imbalance_pct,
7948 struct sched_group *group,
7949 struct sg_lb_stats *sgs)
7951 if (group_is_overloaded(imbalance_pct, sgs))
7952 return group_overloaded;
7954 if (sg_imbalanced(group))
7955 return group_imbalanced;
7957 if (sgs->group_asym_packing)
7958 return group_asym_packing;
7960 if (sgs->group_misfit_task_load)
7961 return group_misfit_task;
7963 if (!group_has_capacity(imbalance_pct, sgs))
7964 return group_fully_busy;
7966 return group_has_spare;
7969 static bool update_nohz_stats(struct rq *rq, bool force)
7971 #ifdef CONFIG_NO_HZ_COMMON
7972 unsigned int cpu = rq->cpu;
7974 if (!rq->has_blocked_load)
7977 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
7980 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
7983 update_blocked_averages(cpu);
7985 return rq->has_blocked_load;
7992 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7993 * @env: The load balancing environment.
7994 * @group: sched_group whose statistics are to be updated.
7995 * @sgs: variable to hold the statistics for this group.
7996 * @sg_status: Holds flag indicating the status of the sched_group
7998 static inline void update_sg_lb_stats(struct lb_env *env,
7999 struct sched_group *group,
8000 struct sg_lb_stats *sgs,
8003 int i, nr_running, local_group;
8005 memset(sgs, 0, sizeof(*sgs));
8007 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(group));
8009 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8010 struct rq *rq = cpu_rq(i);
8012 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
8013 env->flags |= LBF_NOHZ_AGAIN;
8015 sgs->group_load += cpu_load(rq);
8016 sgs->group_util += cpu_util(i);
8017 sgs->sum_h_nr_running += rq->cfs.h_nr_running;
8019 nr_running = rq->nr_running;
8020 sgs->sum_nr_running += nr_running;
8023 *sg_status |= SG_OVERLOAD;
8025 if (cpu_overutilized(i))
8026 *sg_status |= SG_OVERUTILIZED;
8028 #ifdef CONFIG_NUMA_BALANCING
8029 sgs->nr_numa_running += rq->nr_numa_running;
8030 sgs->nr_preferred_running += rq->nr_preferred_running;
8033 * No need to call idle_cpu() if nr_running is not 0
8035 if (!nr_running && idle_cpu(i)) {
8037 /* Idle cpu can't have misfit task */
8044 /* Check for a misfit task on the cpu */
8045 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8046 sgs->group_misfit_task_load < rq->misfit_task_load) {
8047 sgs->group_misfit_task_load = rq->misfit_task_load;
8048 *sg_status |= SG_OVERLOAD;
8052 /* Check if dst CPU is idle and preferred to this group */
8053 if (env->sd->flags & SD_ASYM_PACKING &&
8054 env->idle != CPU_NOT_IDLE &&
8055 sgs->sum_h_nr_running &&
8056 sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu)) {
8057 sgs->group_asym_packing = 1;
8060 sgs->group_capacity = group->sgc->capacity;
8062 sgs->group_weight = group->group_weight;
8064 sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
8066 /* Computing avg_load makes sense only when group is overloaded */
8067 if (sgs->group_type == group_overloaded)
8068 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
8069 sgs->group_capacity;
8073 * update_sd_pick_busiest - return 1 on busiest group
8074 * @env: The load balancing environment.
8075 * @sds: sched_domain statistics
8076 * @sg: sched_group candidate to be checked for being the busiest
8077 * @sgs: sched_group statistics
8079 * Determine if @sg is a busier group than the previously selected
8082 * Return: %true if @sg is a busier group than the previously selected
8083 * busiest group. %false otherwise.
8085 static bool update_sd_pick_busiest(struct lb_env *env,
8086 struct sd_lb_stats *sds,
8087 struct sched_group *sg,
8088 struct sg_lb_stats *sgs)
8090 struct sg_lb_stats *busiest = &sds->busiest_stat;
8092 /* Make sure that there is at least one task to pull */
8093 if (!sgs->sum_h_nr_running)
8097 * Don't try to pull misfit tasks we can't help.
8098 * We can use max_capacity here as reduction in capacity on some
8099 * CPUs in the group should either be possible to resolve
8100 * internally or be covered by avg_load imbalance (eventually).
8102 if (sgs->group_type == group_misfit_task &&
8103 (!group_smaller_max_cpu_capacity(sg, sds->local) ||
8104 sds->local_stat.group_type != group_has_spare))
8107 if (sgs->group_type > busiest->group_type)
8110 if (sgs->group_type < busiest->group_type)
8114 * The candidate and the current busiest group are the same type of
8115 * group. Let check which one is the busiest according to the type.
8118 switch (sgs->group_type) {
8119 case group_overloaded:
8120 /* Select the overloaded group with highest avg_load. */
8121 if (sgs->avg_load <= busiest->avg_load)
8125 case group_imbalanced:
8127 * Select the 1st imbalanced group as we don't have any way to
8128 * choose one more than another.
8132 case group_asym_packing:
8133 /* Prefer to move from lowest priority CPU's work */
8134 if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu))
8138 case group_misfit_task:
8140 * If we have more than one misfit sg go with the biggest
8143 if (sgs->group_misfit_task_load < busiest->group_misfit_task_load)
8147 case group_fully_busy:
8149 * Select the fully busy group with highest avg_load. In
8150 * theory, there is no need to pull task from such kind of
8151 * group because tasks have all compute capacity that they need
8152 * but we can still improve the overall throughput by reducing
8153 * contention when accessing shared HW resources.
8155 * XXX for now avg_load is not computed and always 0 so we
8156 * select the 1st one.
8158 if (sgs->avg_load <= busiest->avg_load)
8162 case group_has_spare:
8164 * Select not overloaded group with lowest number of
8165 * idle cpus. We could also compare the spare capacity
8166 * which is more stable but it can end up that the
8167 * group has less spare capacity but finally more idle
8168 * CPUs which means less opportunity to pull tasks.
8170 if (sgs->idle_cpus >= busiest->idle_cpus)
8176 * Candidate sg has no more than one task per CPU and has higher
8177 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
8178 * throughput. Maximize throughput, power/energy consequences are not
8181 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
8182 (sgs->group_type <= group_fully_busy) &&
8183 (group_smaller_min_cpu_capacity(sds->local, sg)))
8189 #ifdef CONFIG_NUMA_BALANCING
8190 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8192 if (sgs->sum_h_nr_running > sgs->nr_numa_running)
8194 if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
8199 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8201 if (rq->nr_running > rq->nr_numa_running)
8203 if (rq->nr_running > rq->nr_preferred_running)
8208 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8213 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8217 #endif /* CONFIG_NUMA_BALANCING */
8223 * task_running_on_cpu - return 1 if @p is running on @cpu.
8226 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
8228 /* Task has no contribution or is new */
8229 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
8232 if (task_on_rq_queued(p))
8239 * idle_cpu_without - would a given CPU be idle without p ?
8240 * @cpu: the processor on which idleness is tested.
8241 * @p: task which should be ignored.
8243 * Return: 1 if the CPU would be idle. 0 otherwise.
8245 static int idle_cpu_without(int cpu, struct task_struct *p)
8247 struct rq *rq = cpu_rq(cpu);
8249 if (rq->curr != rq->idle && rq->curr != p)
8253 * rq->nr_running can't be used but an updated version without the
8254 * impact of p on cpu must be used instead. The updated nr_running
8255 * be computed and tested before calling idle_cpu_without().
8259 if (!llist_empty(&rq->wake_list))
8267 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
8268 * @sd: The sched_domain level to look for idlest group.
8269 * @group: sched_group whose statistics are to be updated.
8270 * @sgs: variable to hold the statistics for this group.
8271 * @p: The task for which we look for the idlest group/CPU.
8273 static inline void update_sg_wakeup_stats(struct sched_domain *sd,
8274 struct sched_group *group,
8275 struct sg_lb_stats *sgs,
8276 struct task_struct *p)
8280 memset(sgs, 0, sizeof(*sgs));
8282 for_each_cpu(i, sched_group_span(group)) {
8283 struct rq *rq = cpu_rq(i);
8286 sgs->group_load += cpu_load_without(rq, p);
8287 sgs->group_util += cpu_util_without(i, p);
8288 local = task_running_on_cpu(i, p);
8289 sgs->sum_h_nr_running += rq->cfs.h_nr_running - local;
8291 nr_running = rq->nr_running - local;
8292 sgs->sum_nr_running += nr_running;
8295 * No need to call idle_cpu_without() if nr_running is not 0
8297 if (!nr_running && idle_cpu_without(i, p))
8302 /* Check if task fits in the group */
8303 if (sd->flags & SD_ASYM_CPUCAPACITY &&
8304 !task_fits_capacity(p, group->sgc->max_capacity)) {
8305 sgs->group_misfit_task_load = 1;
8308 sgs->group_capacity = group->sgc->capacity;
8310 sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
8313 * Computing avg_load makes sense only when group is fully busy or
8316 if (sgs->group_type < group_fully_busy)
8317 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
8318 sgs->group_capacity;
8321 static bool update_pick_idlest(struct sched_group *idlest,
8322 struct sg_lb_stats *idlest_sgs,
8323 struct sched_group *group,
8324 struct sg_lb_stats *sgs)
8326 if (sgs->group_type < idlest_sgs->group_type)
8329 if (sgs->group_type > idlest_sgs->group_type)
8333 * The candidate and the current idlest group are the same type of
8334 * group. Let check which one is the idlest according to the type.
8337 switch (sgs->group_type) {
8338 case group_overloaded:
8339 case group_fully_busy:
8340 /* Select the group with lowest avg_load. */
8341 if (idlest_sgs->avg_load <= sgs->avg_load)
8345 case group_imbalanced:
8346 case group_asym_packing:
8347 /* Those types are not used in the slow wakeup path */
8350 case group_misfit_task:
8351 /* Select group with the highest max capacity */
8352 if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
8356 case group_has_spare:
8357 /* Select group with most idle CPUs */
8358 if (idlest_sgs->idle_cpus >= sgs->idle_cpus)
8367 * find_idlest_group() finds and returns the least busy CPU group within the
8370 * Assumes p is allowed on at least one CPU in sd.
8372 static struct sched_group *
8373 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
8374 int this_cpu, int sd_flag)
8376 struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
8377 struct sg_lb_stats local_sgs, tmp_sgs;
8378 struct sg_lb_stats *sgs;
8379 unsigned long imbalance;
8380 struct sg_lb_stats idlest_sgs = {
8381 .avg_load = UINT_MAX,
8382 .group_type = group_overloaded,
8385 imbalance = scale_load_down(NICE_0_LOAD) *
8386 (sd->imbalance_pct-100) / 100;
8391 /* Skip over this group if it has no CPUs allowed */
8392 if (!cpumask_intersects(sched_group_span(group),
8396 local_group = cpumask_test_cpu(this_cpu,
8397 sched_group_span(group));
8406 update_sg_wakeup_stats(sd, group, sgs, p);
8408 if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
8413 } while (group = group->next, group != sd->groups);
8416 /* There is no idlest group to push tasks to */
8421 * If the local group is idler than the selected idlest group
8422 * don't try and push the task.
8424 if (local_sgs.group_type < idlest_sgs.group_type)
8428 * If the local group is busier than the selected idlest group
8429 * try and push the task.
8431 if (local_sgs.group_type > idlest_sgs.group_type)
8434 switch (local_sgs.group_type) {
8435 case group_overloaded:
8436 case group_fully_busy:
8438 * When comparing groups across NUMA domains, it's possible for
8439 * the local domain to be very lightly loaded relative to the
8440 * remote domains but "imbalance" skews the comparison making
8441 * remote CPUs look much more favourable. When considering
8442 * cross-domain, add imbalance to the load on the remote node
8443 * and consider staying local.
8446 if ((sd->flags & SD_NUMA) &&
8447 ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
8451 * If the local group is less loaded than the selected
8452 * idlest group don't try and push any tasks.
8454 if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
8457 if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
8461 case group_imbalanced:
8462 case group_asym_packing:
8463 /* Those type are not used in the slow wakeup path */
8466 case group_misfit_task:
8467 /* Select group with the highest max capacity */
8468 if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
8472 case group_has_spare:
8473 if (sd->flags & SD_NUMA) {
8474 #ifdef CONFIG_NUMA_BALANCING
8477 * If there is spare capacity at NUMA, try to select
8478 * the preferred node
8480 if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
8483 idlest_cpu = cpumask_first(sched_group_span(idlest));
8484 if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
8488 * Otherwise, keep the task on this node to stay close
8489 * its wakeup source and improve locality. If there is
8490 * a real need of migration, periodic load balance will
8493 if (local_sgs.idle_cpus)
8498 * Select group with highest number of idle CPUs. We could also
8499 * compare the utilization which is more stable but it can end
8500 * up that the group has less spare capacity but finally more
8501 * idle CPUs which means more opportunity to run task.
8503 if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
8512 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8513 * @env: The load balancing environment.
8514 * @sds: variable to hold the statistics for this sched_domain.
8517 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8519 struct sched_domain *child = env->sd->child;
8520 struct sched_group *sg = env->sd->groups;
8521 struct sg_lb_stats *local = &sds->local_stat;
8522 struct sg_lb_stats tmp_sgs;
8525 #ifdef CONFIG_NO_HZ_COMMON
8526 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
8527 env->flags |= LBF_NOHZ_STATS;
8531 struct sg_lb_stats *sgs = &tmp_sgs;
8534 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
8539 if (env->idle != CPU_NEWLY_IDLE ||
8540 time_after_eq(jiffies, sg->sgc->next_update))
8541 update_group_capacity(env->sd, env->dst_cpu);
8544 update_sg_lb_stats(env, sg, sgs, &sg_status);
8550 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8552 sds->busiest_stat = *sgs;
8556 /* Now, start updating sd_lb_stats */
8557 sds->total_load += sgs->group_load;
8558 sds->total_capacity += sgs->group_capacity;
8561 } while (sg != env->sd->groups);
8563 /* Tag domain that child domain prefers tasks go to siblings first */
8564 sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
8566 #ifdef CONFIG_NO_HZ_COMMON
8567 if ((env->flags & LBF_NOHZ_AGAIN) &&
8568 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8570 WRITE_ONCE(nohz.next_blocked,
8571 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8575 if (env->sd->flags & SD_NUMA)
8576 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8578 if (!env->sd->parent) {
8579 struct root_domain *rd = env->dst_rq->rd;
8581 /* update overload indicator if we are at root domain */
8582 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
8584 /* Update over-utilization (tipping point, U >= 0) indicator */
8585 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
8586 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
8587 } else if (sg_status & SG_OVERUTILIZED) {
8588 struct root_domain *rd = env->dst_rq->rd;
8590 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
8591 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
8596 * calculate_imbalance - Calculate the amount of imbalance present within the
8597 * groups of a given sched_domain during load balance.
8598 * @env: load balance environment
8599 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8601 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8603 struct sg_lb_stats *local, *busiest;
8605 local = &sds->local_stat;
8606 busiest = &sds->busiest_stat;
8608 if (busiest->group_type == group_misfit_task) {
8609 /* Set imbalance to allow misfit tasks to be balanced. */
8610 env->migration_type = migrate_misfit;
8615 if (busiest->group_type == group_asym_packing) {
8617 * In case of asym capacity, we will try to migrate all load to
8618 * the preferred CPU.
8620 env->migration_type = migrate_task;
8621 env->imbalance = busiest->sum_h_nr_running;
8625 if (busiest->group_type == group_imbalanced) {
8627 * In the group_imb case we cannot rely on group-wide averages
8628 * to ensure CPU-load equilibrium, try to move any task to fix
8629 * the imbalance. The next load balance will take care of
8630 * balancing back the system.
8632 env->migration_type = migrate_task;
8638 * Try to use spare capacity of local group without overloading it or
8640 * XXX Spreading tasks across NUMA nodes is not always the best policy
8641 * and special care should be taken for SD_NUMA domain level before
8642 * spreading the tasks. For now, load_balance() fully relies on
8643 * NUMA_BALANCING and fbq_classify_group/rq to override the decision.
8645 if (local->group_type == group_has_spare) {
8646 if (busiest->group_type > group_fully_busy) {
8648 * If busiest is overloaded, try to fill spare
8649 * capacity. This might end up creating spare capacity
8650 * in busiest or busiest still being overloaded but
8651 * there is no simple way to directly compute the
8652 * amount of load to migrate in order to balance the
8655 env->migration_type = migrate_util;
8656 env->imbalance = max(local->group_capacity, local->group_util) -
8660 * In some cases, the group's utilization is max or even
8661 * higher than capacity because of migrations but the
8662 * local CPU is (newly) idle. There is at least one
8663 * waiting task in this overloaded busiest group. Let's
8666 if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) {
8667 env->migration_type = migrate_task;
8674 if (busiest->group_weight == 1 || sds->prefer_sibling) {
8675 unsigned int nr_diff = busiest->sum_nr_running;
8677 * When prefer sibling, evenly spread running tasks on
8680 env->migration_type = migrate_task;
8681 lsub_positive(&nr_diff, local->sum_nr_running);
8682 env->imbalance = nr_diff >> 1;
8687 * If there is no overload, we just want to even the number of
8690 env->migration_type = migrate_task;
8691 env->imbalance = max_t(long, 0, (local->idle_cpus -
8692 busiest->idle_cpus) >> 1);
8697 * Local is fully busy but has to take more load to relieve the
8700 if (local->group_type < group_overloaded) {
8702 * Local will become overloaded so the avg_load metrics are
8706 local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
8707 local->group_capacity;
8709 sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
8710 sds->total_capacity;
8714 * Both group are or will become overloaded and we're trying to get all
8715 * the CPUs to the average_load, so we don't want to push ourselves
8716 * above the average load, nor do we wish to reduce the max loaded CPU
8717 * below the average load. At the same time, we also don't want to
8718 * reduce the group load below the group capacity. Thus we look for
8719 * the minimum possible imbalance.
8721 env->migration_type = migrate_load;
8722 env->imbalance = min(
8723 (busiest->avg_load - sds->avg_load) * busiest->group_capacity,
8724 (sds->avg_load - local->avg_load) * local->group_capacity
8725 ) / SCHED_CAPACITY_SCALE;
8728 /******* find_busiest_group() helpers end here *********************/
8731 * Decision matrix according to the local and busiest group type:
8733 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
8734 * has_spare nr_idle balanced N/A N/A balanced balanced
8735 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
8736 * misfit_task force N/A N/A N/A force force
8737 * asym_packing force force N/A N/A force force
8738 * imbalanced force force N/A N/A force force
8739 * overloaded force force N/A N/A force avg_load
8741 * N/A : Not Applicable because already filtered while updating
8743 * balanced : The system is balanced for these 2 groups.
8744 * force : Calculate the imbalance as load migration is probably needed.
8745 * avg_load : Only if imbalance is significant enough.
8746 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
8747 * different in groups.
8751 * find_busiest_group - Returns the busiest group within the sched_domain
8752 * if there is an imbalance.
8754 * Also calculates the amount of runnable load which should be moved
8755 * to restore balance.
8757 * @env: The load balancing environment.
8759 * Return: - The busiest group if imbalance exists.
8761 static struct sched_group *find_busiest_group(struct lb_env *env)
8763 struct sg_lb_stats *local, *busiest;
8764 struct sd_lb_stats sds;
8766 init_sd_lb_stats(&sds);
8769 * Compute the various statistics relevant for load balancing at
8772 update_sd_lb_stats(env, &sds);
8774 if (sched_energy_enabled()) {
8775 struct root_domain *rd = env->dst_rq->rd;
8777 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
8781 local = &sds.local_stat;
8782 busiest = &sds.busiest_stat;
8784 /* There is no busy sibling group to pull tasks from */
8788 /* Misfit tasks should be dealt with regardless of the avg load */
8789 if (busiest->group_type == group_misfit_task)
8792 /* ASYM feature bypasses nice load balance check */
8793 if (busiest->group_type == group_asym_packing)
8797 * If the busiest group is imbalanced the below checks don't
8798 * work because they assume all things are equal, which typically
8799 * isn't true due to cpus_ptr constraints and the like.
8801 if (busiest->group_type == group_imbalanced)
8805 * If the local group is busier than the selected busiest group
8806 * don't try and pull any tasks.
8808 if (local->group_type > busiest->group_type)
8812 * When groups are overloaded, use the avg_load to ensure fairness
8815 if (local->group_type == group_overloaded) {
8817 * If the local group is more loaded than the selected
8818 * busiest group don't try to pull any tasks.
8820 if (local->avg_load >= busiest->avg_load)
8823 /* XXX broken for overlapping NUMA groups */
8824 sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
8828 * Don't pull any tasks if this group is already above the
8829 * domain average load.
8831 if (local->avg_load >= sds.avg_load)
8835 * If the busiest group is more loaded, use imbalance_pct to be
8838 if (100 * busiest->avg_load <=
8839 env->sd->imbalance_pct * local->avg_load)
8843 /* Try to move all excess tasks to child's sibling domain */
8844 if (sds.prefer_sibling && local->group_type == group_has_spare &&
8845 busiest->sum_nr_running > local->sum_nr_running + 1)
8848 if (busiest->group_type != group_overloaded) {
8849 if (env->idle == CPU_NOT_IDLE)
8851 * If the busiest group is not overloaded (and as a
8852 * result the local one too) but this CPU is already
8853 * busy, let another idle CPU try to pull task.
8857 if (busiest->group_weight > 1 &&
8858 local->idle_cpus <= (busiest->idle_cpus + 1))
8860 * If the busiest group is not overloaded
8861 * and there is no imbalance between this and busiest
8862 * group wrt idle CPUs, it is balanced. The imbalance
8863 * becomes significant if the diff is greater than 1
8864 * otherwise we might end up to just move the imbalance
8865 * on another group. Of course this applies only if
8866 * there is more than 1 CPU per group.
8870 if (busiest->sum_h_nr_running == 1)
8872 * busiest doesn't have any tasks waiting to run
8878 /* Looks like there is an imbalance. Compute it */
8879 calculate_imbalance(env, &sds);
8880 return env->imbalance ? sds.busiest : NULL;
8888 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
8890 static struct rq *find_busiest_queue(struct lb_env *env,
8891 struct sched_group *group)
8893 struct rq *busiest = NULL, *rq;
8894 unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
8895 unsigned int busiest_nr = 0;
8898 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8899 unsigned long capacity, load, util;
8900 unsigned int nr_running;
8904 rt = fbq_classify_rq(rq);
8907 * We classify groups/runqueues into three groups:
8908 * - regular: there are !numa tasks
8909 * - remote: there are numa tasks that run on the 'wrong' node
8910 * - all: there is no distinction
8912 * In order to avoid migrating ideally placed numa tasks,
8913 * ignore those when there's better options.
8915 * If we ignore the actual busiest queue to migrate another
8916 * task, the next balance pass can still reduce the busiest
8917 * queue by moving tasks around inside the node.
8919 * If we cannot move enough load due to this classification
8920 * the next pass will adjust the group classification and
8921 * allow migration of more tasks.
8923 * Both cases only affect the total convergence complexity.
8925 if (rt > env->fbq_type)
8928 capacity = capacity_of(i);
8929 nr_running = rq->cfs.h_nr_running;
8932 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
8933 * eventually lead to active_balancing high->low capacity.
8934 * Higher per-CPU capacity is considered better than balancing
8937 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8938 capacity_of(env->dst_cpu) < capacity &&
8942 switch (env->migration_type) {
8945 * When comparing with load imbalance, use cpu_load()
8946 * which is not scaled with the CPU capacity.
8948 load = cpu_load(rq);
8950 if (nr_running == 1 && load > env->imbalance &&
8951 !check_cpu_capacity(rq, env->sd))
8955 * For the load comparisons with the other CPUs,
8956 * consider the cpu_load() scaled with the CPU
8957 * capacity, so that the load can be moved away
8958 * from the CPU that is potentially running at a
8961 * Thus we're looking for max(load_i / capacity_i),
8962 * crosswise multiplication to rid ourselves of the
8963 * division works out to:
8964 * load_i * capacity_j > load_j * capacity_i;
8965 * where j is our previous maximum.
8967 if (load * busiest_capacity > busiest_load * capacity) {
8968 busiest_load = load;
8969 busiest_capacity = capacity;
8975 util = cpu_util(cpu_of(rq));
8977 if (busiest_util < util) {
8978 busiest_util = util;
8984 if (busiest_nr < nr_running) {
8985 busiest_nr = nr_running;
8990 case migrate_misfit:
8992 * For ASYM_CPUCAPACITY domains with misfit tasks we
8993 * simply seek the "biggest" misfit task.
8995 if (rq->misfit_task_load > busiest_load) {
8996 busiest_load = rq->misfit_task_load;
9009 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
9010 * so long as it is large enough.
9012 #define MAX_PINNED_INTERVAL 512
9015 asym_active_balance(struct lb_env *env)
9018 * ASYM_PACKING needs to force migrate tasks from busy but
9019 * lower priority CPUs in order to pack all tasks in the
9020 * highest priority CPUs.
9022 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
9023 sched_asym_prefer(env->dst_cpu, env->src_cpu);
9027 voluntary_active_balance(struct lb_env *env)
9029 struct sched_domain *sd = env->sd;
9031 if (asym_active_balance(env))
9035 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
9036 * It's worth migrating the task if the src_cpu's capacity is reduced
9037 * because of other sched_class or IRQs if more capacity stays
9038 * available on dst_cpu.
9040 if ((env->idle != CPU_NOT_IDLE) &&
9041 (env->src_rq->cfs.h_nr_running == 1)) {
9042 if ((check_cpu_capacity(env->src_rq, sd)) &&
9043 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
9047 if (env->migration_type == migrate_misfit)
9053 static int need_active_balance(struct lb_env *env)
9055 struct sched_domain *sd = env->sd;
9057 if (voluntary_active_balance(env))
9060 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
9063 static int active_load_balance_cpu_stop(void *data);
9065 static int should_we_balance(struct lb_env *env)
9067 struct sched_group *sg = env->sd->groups;
9068 int cpu, balance_cpu = -1;
9071 * Ensure the balancing environment is consistent; can happen
9072 * when the softirq triggers 'during' hotplug.
9074 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
9078 * In the newly idle case, we will allow all the CPUs
9079 * to do the newly idle load balance.
9081 if (env->idle == CPU_NEWLY_IDLE)
9084 /* Try to find first idle CPU */
9085 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
9093 if (balance_cpu == -1)
9094 balance_cpu = group_balance_cpu(sg);
9097 * First idle CPU or the first CPU(busiest) in this sched group
9098 * is eligible for doing load balancing at this and above domains.
9100 return balance_cpu == env->dst_cpu;
9104 * Check this_cpu to ensure it is balanced within domain. Attempt to move
9105 * tasks if there is an imbalance.
9107 static int load_balance(int this_cpu, struct rq *this_rq,
9108 struct sched_domain *sd, enum cpu_idle_type idle,
9109 int *continue_balancing)
9111 int ld_moved, cur_ld_moved, active_balance = 0;
9112 struct sched_domain *sd_parent = sd->parent;
9113 struct sched_group *group;
9116 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
9118 struct lb_env env = {
9120 .dst_cpu = this_cpu,
9122 .dst_grpmask = sched_group_span(sd->groups),
9124 .loop_break = sched_nr_migrate_break,
9127 .tasks = LIST_HEAD_INIT(env.tasks),
9130 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
9132 schedstat_inc(sd->lb_count[idle]);
9135 if (!should_we_balance(&env)) {
9136 *continue_balancing = 0;
9140 group = find_busiest_group(&env);
9142 schedstat_inc(sd->lb_nobusyg[idle]);
9146 busiest = find_busiest_queue(&env, group);
9148 schedstat_inc(sd->lb_nobusyq[idle]);
9152 BUG_ON(busiest == env.dst_rq);
9154 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
9156 env.src_cpu = busiest->cpu;
9157 env.src_rq = busiest;
9160 if (busiest->nr_running > 1) {
9162 * Attempt to move tasks. If find_busiest_group has found
9163 * an imbalance but busiest->nr_running <= 1, the group is
9164 * still unbalanced. ld_moved simply stays zero, so it is
9165 * correctly treated as an imbalance.
9167 env.flags |= LBF_ALL_PINNED;
9168 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
9171 rq_lock_irqsave(busiest, &rf);
9172 update_rq_clock(busiest);
9175 * cur_ld_moved - load moved in current iteration
9176 * ld_moved - cumulative load moved across iterations
9178 cur_ld_moved = detach_tasks(&env);
9181 * We've detached some tasks from busiest_rq. Every
9182 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
9183 * unlock busiest->lock, and we are able to be sure
9184 * that nobody can manipulate the tasks in parallel.
9185 * See task_rq_lock() family for the details.
9188 rq_unlock(busiest, &rf);
9192 ld_moved += cur_ld_moved;
9195 local_irq_restore(rf.flags);
9197 if (env.flags & LBF_NEED_BREAK) {
9198 env.flags &= ~LBF_NEED_BREAK;
9203 * Revisit (affine) tasks on src_cpu that couldn't be moved to
9204 * us and move them to an alternate dst_cpu in our sched_group
9205 * where they can run. The upper limit on how many times we
9206 * iterate on same src_cpu is dependent on number of CPUs in our
9209 * This changes load balance semantics a bit on who can move
9210 * load to a given_cpu. In addition to the given_cpu itself
9211 * (or a ilb_cpu acting on its behalf where given_cpu is
9212 * nohz-idle), we now have balance_cpu in a position to move
9213 * load to given_cpu. In rare situations, this may cause
9214 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
9215 * _independently_ and at _same_ time to move some load to
9216 * given_cpu) causing exceess load to be moved to given_cpu.
9217 * This however should not happen so much in practice and
9218 * moreover subsequent load balance cycles should correct the
9219 * excess load moved.
9221 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
9223 /* Prevent to re-select dst_cpu via env's CPUs */
9224 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
9226 env.dst_rq = cpu_rq(env.new_dst_cpu);
9227 env.dst_cpu = env.new_dst_cpu;
9228 env.flags &= ~LBF_DST_PINNED;
9230 env.loop_break = sched_nr_migrate_break;
9233 * Go back to "more_balance" rather than "redo" since we
9234 * need to continue with same src_cpu.
9240 * We failed to reach balance because of affinity.
9243 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9245 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
9246 *group_imbalance = 1;
9249 /* All tasks on this runqueue were pinned by CPU affinity */
9250 if (unlikely(env.flags & LBF_ALL_PINNED)) {
9251 __cpumask_clear_cpu(cpu_of(busiest), cpus);
9253 * Attempting to continue load balancing at the current
9254 * sched_domain level only makes sense if there are
9255 * active CPUs remaining as possible busiest CPUs to
9256 * pull load from which are not contained within the
9257 * destination group that is receiving any migrated
9260 if (!cpumask_subset(cpus, env.dst_grpmask)) {
9262 env.loop_break = sched_nr_migrate_break;
9265 goto out_all_pinned;
9270 schedstat_inc(sd->lb_failed[idle]);
9272 * Increment the failure counter only on periodic balance.
9273 * We do not want newidle balance, which can be very
9274 * frequent, pollute the failure counter causing
9275 * excessive cache_hot migrations and active balances.
9277 if (idle != CPU_NEWLY_IDLE)
9278 sd->nr_balance_failed++;
9280 if (need_active_balance(&env)) {
9281 unsigned long flags;
9283 raw_spin_lock_irqsave(&busiest->lock, flags);
9286 * Don't kick the active_load_balance_cpu_stop,
9287 * if the curr task on busiest CPU can't be
9288 * moved to this_cpu:
9290 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
9291 raw_spin_unlock_irqrestore(&busiest->lock,
9293 env.flags |= LBF_ALL_PINNED;
9294 goto out_one_pinned;
9298 * ->active_balance synchronizes accesses to
9299 * ->active_balance_work. Once set, it's cleared
9300 * only after active load balance is finished.
9302 if (!busiest->active_balance) {
9303 busiest->active_balance = 1;
9304 busiest->push_cpu = this_cpu;
9307 raw_spin_unlock_irqrestore(&busiest->lock, flags);
9309 if (active_balance) {
9310 stop_one_cpu_nowait(cpu_of(busiest),
9311 active_load_balance_cpu_stop, busiest,
9312 &busiest->active_balance_work);
9315 /* We've kicked active balancing, force task migration. */
9316 sd->nr_balance_failed = sd->cache_nice_tries+1;
9319 sd->nr_balance_failed = 0;
9321 if (likely(!active_balance) || voluntary_active_balance(&env)) {
9322 /* We were unbalanced, so reset the balancing interval */
9323 sd->balance_interval = sd->min_interval;
9326 * If we've begun active balancing, start to back off. This
9327 * case may not be covered by the all_pinned logic if there
9328 * is only 1 task on the busy runqueue (because we don't call
9331 if (sd->balance_interval < sd->max_interval)
9332 sd->balance_interval *= 2;
9339 * We reach balance although we may have faced some affinity
9340 * constraints. Clear the imbalance flag only if other tasks got
9341 * a chance to move and fix the imbalance.
9343 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
9344 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9346 if (*group_imbalance)
9347 *group_imbalance = 0;
9352 * We reach balance because all tasks are pinned at this level so
9353 * we can't migrate them. Let the imbalance flag set so parent level
9354 * can try to migrate them.
9356 schedstat_inc(sd->lb_balanced[idle]);
9358 sd->nr_balance_failed = 0;
9364 * newidle_balance() disregards balance intervals, so we could
9365 * repeatedly reach this code, which would lead to balance_interval
9366 * skyrocketting in a short amount of time. Skip the balance_interval
9367 * increase logic to avoid that.
9369 if (env.idle == CPU_NEWLY_IDLE)
9372 /* tune up the balancing interval */
9373 if ((env.flags & LBF_ALL_PINNED &&
9374 sd->balance_interval < MAX_PINNED_INTERVAL) ||
9375 sd->balance_interval < sd->max_interval)
9376 sd->balance_interval *= 2;
9381 static inline unsigned long
9382 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
9384 unsigned long interval = sd->balance_interval;
9387 interval *= sd->busy_factor;
9389 /* scale ms to jiffies */
9390 interval = msecs_to_jiffies(interval);
9391 interval = clamp(interval, 1UL, max_load_balance_interval);
9397 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
9399 unsigned long interval, next;
9401 /* used by idle balance, so cpu_busy = 0 */
9402 interval = get_sd_balance_interval(sd, 0);
9403 next = sd->last_balance + interval;
9405 if (time_after(*next_balance, next))
9406 *next_balance = next;
9410 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9411 * running tasks off the busiest CPU onto idle CPUs. It requires at
9412 * least 1 task to be running on each physical CPU where possible, and
9413 * avoids physical / logical imbalances.
9415 static int active_load_balance_cpu_stop(void *data)
9417 struct rq *busiest_rq = data;
9418 int busiest_cpu = cpu_of(busiest_rq);
9419 int target_cpu = busiest_rq->push_cpu;
9420 struct rq *target_rq = cpu_rq(target_cpu);
9421 struct sched_domain *sd;
9422 struct task_struct *p = NULL;
9425 rq_lock_irq(busiest_rq, &rf);
9427 * Between queueing the stop-work and running it is a hole in which
9428 * CPUs can become inactive. We should not move tasks from or to
9431 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
9434 /* Make sure the requested CPU hasn't gone down in the meantime: */
9435 if (unlikely(busiest_cpu != smp_processor_id() ||
9436 !busiest_rq->active_balance))
9439 /* Is there any task to move? */
9440 if (busiest_rq->nr_running <= 1)
9444 * This condition is "impossible", if it occurs
9445 * we need to fix it. Originally reported by
9446 * Bjorn Helgaas on a 128-CPU setup.
9448 BUG_ON(busiest_rq == target_rq);
9450 /* Search for an sd spanning us and the target CPU. */
9452 for_each_domain(target_cpu, sd) {
9453 if ((sd->flags & SD_LOAD_BALANCE) &&
9454 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9459 struct lb_env env = {
9461 .dst_cpu = target_cpu,
9462 .dst_rq = target_rq,
9463 .src_cpu = busiest_rq->cpu,
9464 .src_rq = busiest_rq,
9467 * can_migrate_task() doesn't need to compute new_dst_cpu
9468 * for active balancing. Since we have CPU_IDLE, but no
9469 * @dst_grpmask we need to make that test go away with lying
9472 .flags = LBF_DST_PINNED,
9475 schedstat_inc(sd->alb_count);
9476 update_rq_clock(busiest_rq);
9478 p = detach_one_task(&env);
9480 schedstat_inc(sd->alb_pushed);
9481 /* Active balancing done, reset the failure counter. */
9482 sd->nr_balance_failed = 0;
9484 schedstat_inc(sd->alb_failed);
9489 busiest_rq->active_balance = 0;
9490 rq_unlock(busiest_rq, &rf);
9493 attach_one_task(target_rq, p);
9500 static DEFINE_SPINLOCK(balancing);
9503 * Scale the max load_balance interval with the number of CPUs in the system.
9504 * This trades load-balance latency on larger machines for less cross talk.
9506 void update_max_interval(void)
9508 max_load_balance_interval = HZ*num_online_cpus()/10;
9512 * It checks each scheduling domain to see if it is due to be balanced,
9513 * and initiates a balancing operation if so.
9515 * Balancing parameters are set up in init_sched_domains.
9517 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9519 int continue_balancing = 1;
9521 unsigned long interval;
9522 struct sched_domain *sd;
9523 /* Earliest time when we have to do rebalance again */
9524 unsigned long next_balance = jiffies + 60*HZ;
9525 int update_next_balance = 0;
9526 int need_serialize, need_decay = 0;
9530 for_each_domain(cpu, sd) {
9532 * Decay the newidle max times here because this is a regular
9533 * visit to all the domains. Decay ~1% per second.
9535 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9536 sd->max_newidle_lb_cost =
9537 (sd->max_newidle_lb_cost * 253) / 256;
9538 sd->next_decay_max_lb_cost = jiffies + HZ;
9541 max_cost += sd->max_newidle_lb_cost;
9543 if (!(sd->flags & SD_LOAD_BALANCE))
9547 * Stop the load balance at this level. There is another
9548 * CPU in our sched group which is doing load balancing more
9551 if (!continue_balancing) {
9557 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9559 need_serialize = sd->flags & SD_SERIALIZE;
9560 if (need_serialize) {
9561 if (!spin_trylock(&balancing))
9565 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9566 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9568 * The LBF_DST_PINNED logic could have changed
9569 * env->dst_cpu, so we can't know our idle
9570 * state even if we migrated tasks. Update it.
9572 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9574 sd->last_balance = jiffies;
9575 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9578 spin_unlock(&balancing);
9580 if (time_after(next_balance, sd->last_balance + interval)) {
9581 next_balance = sd->last_balance + interval;
9582 update_next_balance = 1;
9587 * Ensure the rq-wide value also decays but keep it at a
9588 * reasonable floor to avoid funnies with rq->avg_idle.
9590 rq->max_idle_balance_cost =
9591 max((u64)sysctl_sched_migration_cost, max_cost);
9596 * next_balance will be updated only when there is a need.
9597 * When the cpu is attached to null domain for ex, it will not be
9600 if (likely(update_next_balance)) {
9601 rq->next_balance = next_balance;
9603 #ifdef CONFIG_NO_HZ_COMMON
9605 * If this CPU has been elected to perform the nohz idle
9606 * balance. Other idle CPUs have already rebalanced with
9607 * nohz_idle_balance() and nohz.next_balance has been
9608 * updated accordingly. This CPU is now running the idle load
9609 * balance for itself and we need to update the
9610 * nohz.next_balance accordingly.
9612 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9613 nohz.next_balance = rq->next_balance;
9618 static inline int on_null_domain(struct rq *rq)
9620 return unlikely(!rcu_dereference_sched(rq->sd));
9623 #ifdef CONFIG_NO_HZ_COMMON
9625 * idle load balancing details
9626 * - When one of the busy CPUs notice that there may be an idle rebalancing
9627 * needed, they will kick the idle load balancer, which then does idle
9628 * load balancing for all the idle CPUs.
9629 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
9633 static inline int find_new_ilb(void)
9637 for_each_cpu_and(ilb, nohz.idle_cpus_mask,
9638 housekeeping_cpumask(HK_FLAG_MISC)) {
9647 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
9648 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
9650 static void kick_ilb(unsigned int flags)
9654 nohz.next_balance++;
9656 ilb_cpu = find_new_ilb();
9658 if (ilb_cpu >= nr_cpu_ids)
9661 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
9662 if (flags & NOHZ_KICK_MASK)
9666 * Use smp_send_reschedule() instead of resched_cpu().
9667 * This way we generate a sched IPI on the target CPU which
9668 * is idle. And the softirq performing nohz idle load balance
9669 * will be run before returning from the IPI.
9671 smp_send_reschedule(ilb_cpu);
9675 * Current decision point for kicking the idle load balancer in the presence
9676 * of idle CPUs in the system.
9678 static void nohz_balancer_kick(struct rq *rq)
9680 unsigned long now = jiffies;
9681 struct sched_domain_shared *sds;
9682 struct sched_domain *sd;
9683 int nr_busy, i, cpu = rq->cpu;
9684 unsigned int flags = 0;
9686 if (unlikely(rq->idle_balance))
9690 * We may be recently in ticked or tickless idle mode. At the first
9691 * busy tick after returning from idle, we will update the busy stats.
9693 nohz_balance_exit_idle(rq);
9696 * None are in tickless mode and hence no need for NOHZ idle load
9699 if (likely(!atomic_read(&nohz.nr_cpus)))
9702 if (READ_ONCE(nohz.has_blocked) &&
9703 time_after(now, READ_ONCE(nohz.next_blocked)))
9704 flags = NOHZ_STATS_KICK;
9706 if (time_before(now, nohz.next_balance))
9709 if (rq->nr_running >= 2) {
9710 flags = NOHZ_KICK_MASK;
9716 sd = rcu_dereference(rq->sd);
9719 * If there's a CFS task and the current CPU has reduced
9720 * capacity; kick the ILB to see if there's a better CPU to run
9723 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
9724 flags = NOHZ_KICK_MASK;
9729 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
9732 * When ASYM_PACKING; see if there's a more preferred CPU
9733 * currently idle; in which case, kick the ILB to move tasks
9736 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
9737 if (sched_asym_prefer(i, cpu)) {
9738 flags = NOHZ_KICK_MASK;
9744 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
9747 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
9748 * to run the misfit task on.
9750 if (check_misfit_status(rq, sd)) {
9751 flags = NOHZ_KICK_MASK;
9756 * For asymmetric systems, we do not want to nicely balance
9757 * cache use, instead we want to embrace asymmetry and only
9758 * ensure tasks have enough CPU capacity.
9760 * Skip the LLC logic because it's not relevant in that case.
9765 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9768 * If there is an imbalance between LLC domains (IOW we could
9769 * increase the overall cache use), we need some less-loaded LLC
9770 * domain to pull some load. Likewise, we may need to spread
9771 * load within the current LLC domain (e.g. packed SMT cores but
9772 * other CPUs are idle). We can't really know from here how busy
9773 * the others are - so just get a nohz balance going if it looks
9774 * like this LLC domain has tasks we could move.
9776 nr_busy = atomic_read(&sds->nr_busy_cpus);
9778 flags = NOHZ_KICK_MASK;
9789 static void set_cpu_sd_state_busy(int cpu)
9791 struct sched_domain *sd;
9794 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9796 if (!sd || !sd->nohz_idle)
9800 atomic_inc(&sd->shared->nr_busy_cpus);
9805 void nohz_balance_exit_idle(struct rq *rq)
9807 SCHED_WARN_ON(rq != this_rq());
9809 if (likely(!rq->nohz_tick_stopped))
9812 rq->nohz_tick_stopped = 0;
9813 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
9814 atomic_dec(&nohz.nr_cpus);
9816 set_cpu_sd_state_busy(rq->cpu);
9819 static void set_cpu_sd_state_idle(int cpu)
9821 struct sched_domain *sd;
9824 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9826 if (!sd || sd->nohz_idle)
9830 atomic_dec(&sd->shared->nr_busy_cpus);
9836 * This routine will record that the CPU is going idle with tick stopped.
9837 * This info will be used in performing idle load balancing in the future.
9839 void nohz_balance_enter_idle(int cpu)
9841 struct rq *rq = cpu_rq(cpu);
9843 SCHED_WARN_ON(cpu != smp_processor_id());
9845 /* If this CPU is going down, then nothing needs to be done: */
9846 if (!cpu_active(cpu))
9849 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9850 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9854 * Can be set safely without rq->lock held
9855 * If a clear happens, it will have evaluated last additions because
9856 * rq->lock is held during the check and the clear
9858 rq->has_blocked_load = 1;
9861 * The tick is still stopped but load could have been added in the
9862 * meantime. We set the nohz.has_blocked flag to trig a check of the
9863 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
9864 * of nohz.has_blocked can only happen after checking the new load
9866 if (rq->nohz_tick_stopped)
9869 /* If we're a completely isolated CPU, we don't play: */
9870 if (on_null_domain(rq))
9873 rq->nohz_tick_stopped = 1;
9875 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9876 atomic_inc(&nohz.nr_cpus);
9879 * Ensures that if nohz_idle_balance() fails to observe our
9880 * @idle_cpus_mask store, it must observe the @has_blocked
9883 smp_mb__after_atomic();
9885 set_cpu_sd_state_idle(cpu);
9889 * Each time a cpu enter idle, we assume that it has blocked load and
9890 * enable the periodic update of the load of idle cpus
9892 WRITE_ONCE(nohz.has_blocked, 1);
9896 * Internal function that runs load balance for all idle cpus. The load balance
9897 * can be a simple update of blocked load or a complete load balance with
9898 * tasks movement depending of flags.
9899 * The function returns false if the loop has stopped before running
9900 * through all idle CPUs.
9902 static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
9903 enum cpu_idle_type idle)
9905 /* Earliest time when we have to do rebalance again */
9906 unsigned long now = jiffies;
9907 unsigned long next_balance = now + 60*HZ;
9908 bool has_blocked_load = false;
9909 int update_next_balance = 0;
9910 int this_cpu = this_rq->cpu;
9915 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
9918 * We assume there will be no idle load after this update and clear
9919 * the has_blocked flag. If a cpu enters idle in the mean time, it will
9920 * set the has_blocked flag and trig another update of idle load.
9921 * Because a cpu that becomes idle, is added to idle_cpus_mask before
9922 * setting the flag, we are sure to not clear the state and not
9923 * check the load of an idle cpu.
9925 WRITE_ONCE(nohz.has_blocked, 0);
9928 * Ensures that if we miss the CPU, we must see the has_blocked
9929 * store from nohz_balance_enter_idle().
9933 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9934 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9938 * If this CPU gets work to do, stop the load balancing
9939 * work being done for other CPUs. Next load
9940 * balancing owner will pick it up.
9942 if (need_resched()) {
9943 has_blocked_load = true;
9947 rq = cpu_rq(balance_cpu);
9949 has_blocked_load |= update_nohz_stats(rq, true);
9952 * If time for next balance is due,
9955 if (time_after_eq(jiffies, rq->next_balance)) {
9958 rq_lock_irqsave(rq, &rf);
9959 update_rq_clock(rq);
9960 rq_unlock_irqrestore(rq, &rf);
9962 if (flags & NOHZ_BALANCE_KICK)
9963 rebalance_domains(rq, CPU_IDLE);
9966 if (time_after(next_balance, rq->next_balance)) {
9967 next_balance = rq->next_balance;
9968 update_next_balance = 1;
9972 /* Newly idle CPU doesn't need an update */
9973 if (idle != CPU_NEWLY_IDLE) {
9974 update_blocked_averages(this_cpu);
9975 has_blocked_load |= this_rq->has_blocked_load;
9978 if (flags & NOHZ_BALANCE_KICK)
9979 rebalance_domains(this_rq, CPU_IDLE);
9981 WRITE_ONCE(nohz.next_blocked,
9982 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
9984 /* The full idle balance loop has been done */
9988 /* There is still blocked load, enable periodic update */
9989 if (has_blocked_load)
9990 WRITE_ONCE(nohz.has_blocked, 1);
9993 * next_balance will be updated only when there is a need.
9994 * When the CPU is attached to null domain for ex, it will not be
9997 if (likely(update_next_balance))
9998 nohz.next_balance = next_balance;
10004 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
10005 * rebalancing for all the cpus for whom scheduler ticks are stopped.
10007 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
10009 int this_cpu = this_rq->cpu;
10010 unsigned int flags;
10012 if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
10015 if (idle != CPU_IDLE) {
10016 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
10020 /* could be _relaxed() */
10021 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
10022 if (!(flags & NOHZ_KICK_MASK))
10025 _nohz_idle_balance(this_rq, flags, idle);
10030 static void nohz_newidle_balance(struct rq *this_rq)
10032 int this_cpu = this_rq->cpu;
10035 * This CPU doesn't want to be disturbed by scheduler
10038 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
10041 /* Will wake up very soon. No time for doing anything else*/
10042 if (this_rq->avg_idle < sysctl_sched_migration_cost)
10045 /* Don't need to update blocked load of idle CPUs*/
10046 if (!READ_ONCE(nohz.has_blocked) ||
10047 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
10050 raw_spin_unlock(&this_rq->lock);
10052 * This CPU is going to be idle and blocked load of idle CPUs
10053 * need to be updated. Run the ilb locally as it is a good
10054 * candidate for ilb instead of waking up another idle CPU.
10055 * Kick an normal ilb if we failed to do the update.
10057 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
10058 kick_ilb(NOHZ_STATS_KICK);
10059 raw_spin_lock(&this_rq->lock);
10062 #else /* !CONFIG_NO_HZ_COMMON */
10063 static inline void nohz_balancer_kick(struct rq *rq) { }
10065 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
10070 static inline void nohz_newidle_balance(struct rq *this_rq) { }
10071 #endif /* CONFIG_NO_HZ_COMMON */
10074 * idle_balance is called by schedule() if this_cpu is about to become
10075 * idle. Attempts to pull tasks from other CPUs.
10078 * < 0 - we released the lock and there are !fair tasks present
10079 * 0 - failed, no new tasks
10080 * > 0 - success, new (fair) tasks present
10082 int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
10084 unsigned long next_balance = jiffies + HZ;
10085 int this_cpu = this_rq->cpu;
10086 struct sched_domain *sd;
10087 int pulled_task = 0;
10090 update_misfit_status(NULL, this_rq);
10092 * We must set idle_stamp _before_ calling idle_balance(), such that we
10093 * measure the duration of idle_balance() as idle time.
10095 this_rq->idle_stamp = rq_clock(this_rq);
10098 * Do not pull tasks towards !active CPUs...
10100 if (!cpu_active(this_cpu))
10104 * This is OK, because current is on_cpu, which avoids it being picked
10105 * for load-balance and preemption/IRQs are still disabled avoiding
10106 * further scheduler activity on it and we're being very careful to
10107 * re-start the picking loop.
10109 rq_unpin_lock(this_rq, rf);
10111 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
10112 !READ_ONCE(this_rq->rd->overload)) {
10115 sd = rcu_dereference_check_sched_domain(this_rq->sd);
10117 update_next_balance(sd, &next_balance);
10120 nohz_newidle_balance(this_rq);
10125 raw_spin_unlock(&this_rq->lock);
10127 update_blocked_averages(this_cpu);
10129 for_each_domain(this_cpu, sd) {
10130 int continue_balancing = 1;
10131 u64 t0, domain_cost;
10133 if (!(sd->flags & SD_LOAD_BALANCE))
10136 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
10137 update_next_balance(sd, &next_balance);
10141 if (sd->flags & SD_BALANCE_NEWIDLE) {
10142 t0 = sched_clock_cpu(this_cpu);
10144 pulled_task = load_balance(this_cpu, this_rq,
10145 sd, CPU_NEWLY_IDLE,
10146 &continue_balancing);
10148 domain_cost = sched_clock_cpu(this_cpu) - t0;
10149 if (domain_cost > sd->max_newidle_lb_cost)
10150 sd->max_newidle_lb_cost = domain_cost;
10152 curr_cost += domain_cost;
10155 update_next_balance(sd, &next_balance);
10158 * Stop searching for tasks to pull if there are
10159 * now runnable tasks on this rq.
10161 if (pulled_task || this_rq->nr_running > 0)
10166 raw_spin_lock(&this_rq->lock);
10168 if (curr_cost > this_rq->max_idle_balance_cost)
10169 this_rq->max_idle_balance_cost = curr_cost;
10173 * While browsing the domains, we released the rq lock, a task could
10174 * have been enqueued in the meantime. Since we're not going idle,
10175 * pretend we pulled a task.
10177 if (this_rq->cfs.h_nr_running && !pulled_task)
10180 /* Move the next balance forward */
10181 if (time_after(this_rq->next_balance, next_balance))
10182 this_rq->next_balance = next_balance;
10184 /* Is there a task of a high priority class? */
10185 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
10189 this_rq->idle_stamp = 0;
10191 rq_repin_lock(this_rq, rf);
10193 return pulled_task;
10197 * run_rebalance_domains is triggered when needed from the scheduler tick.
10198 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
10200 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
10202 struct rq *this_rq = this_rq();
10203 enum cpu_idle_type idle = this_rq->idle_balance ?
10204 CPU_IDLE : CPU_NOT_IDLE;
10207 * If this CPU has a pending nohz_balance_kick, then do the
10208 * balancing on behalf of the other idle CPUs whose ticks are
10209 * stopped. Do nohz_idle_balance *before* rebalance_domains to
10210 * give the idle CPUs a chance to load balance. Else we may
10211 * load balance only within the local sched_domain hierarchy
10212 * and abort nohz_idle_balance altogether if we pull some load.
10214 if (nohz_idle_balance(this_rq, idle))
10217 /* normal load balance */
10218 update_blocked_averages(this_rq->cpu);
10219 rebalance_domains(this_rq, idle);
10223 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
10225 void trigger_load_balance(struct rq *rq)
10227 /* Don't need to rebalance while attached to NULL domain */
10228 if (unlikely(on_null_domain(rq)))
10231 if (time_after_eq(jiffies, rq->next_balance))
10232 raise_softirq(SCHED_SOFTIRQ);
10234 nohz_balancer_kick(rq);
10237 static void rq_online_fair(struct rq *rq)
10241 update_runtime_enabled(rq);
10244 static void rq_offline_fair(struct rq *rq)
10248 /* Ensure any throttled groups are reachable by pick_next_task */
10249 unthrottle_offline_cfs_rqs(rq);
10252 #endif /* CONFIG_SMP */
10255 * scheduler tick hitting a task of our scheduling class.
10257 * NOTE: This function can be called remotely by the tick offload that
10258 * goes along full dynticks. Therefore no local assumption can be made
10259 * and everything must be accessed through the @rq and @curr passed in
10262 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
10264 struct cfs_rq *cfs_rq;
10265 struct sched_entity *se = &curr->se;
10267 for_each_sched_entity(se) {
10268 cfs_rq = cfs_rq_of(se);
10269 entity_tick(cfs_rq, se, queued);
10272 if (static_branch_unlikely(&sched_numa_balancing))
10273 task_tick_numa(rq, curr);
10275 update_misfit_status(curr, rq);
10276 update_overutilized_status(task_rq(curr));
10280 * called on fork with the child task as argument from the parent's context
10281 * - child not yet on the tasklist
10282 * - preemption disabled
10284 static void task_fork_fair(struct task_struct *p)
10286 struct cfs_rq *cfs_rq;
10287 struct sched_entity *se = &p->se, *curr;
10288 struct rq *rq = this_rq();
10289 struct rq_flags rf;
10292 update_rq_clock(rq);
10294 cfs_rq = task_cfs_rq(current);
10295 curr = cfs_rq->curr;
10297 update_curr(cfs_rq);
10298 se->vruntime = curr->vruntime;
10300 place_entity(cfs_rq, se, 1);
10302 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
10304 * Upon rescheduling, sched_class::put_prev_task() will place
10305 * 'current' within the tree based on its new key value.
10307 swap(curr->vruntime, se->vruntime);
10311 se->vruntime -= cfs_rq->min_vruntime;
10312 rq_unlock(rq, &rf);
10316 * Priority of the task has changed. Check to see if we preempt
10317 * the current task.
10320 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
10322 if (!task_on_rq_queued(p))
10326 * Reschedule if we are currently running on this runqueue and
10327 * our priority decreased, or if we are not currently running on
10328 * this runqueue and our priority is higher than the current's
10330 if (rq->curr == p) {
10331 if (p->prio > oldprio)
10334 check_preempt_curr(rq, p, 0);
10337 static inline bool vruntime_normalized(struct task_struct *p)
10339 struct sched_entity *se = &p->se;
10342 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
10343 * the dequeue_entity(.flags=0) will already have normalized the
10350 * When !on_rq, vruntime of the task has usually NOT been normalized.
10351 * But there are some cases where it has already been normalized:
10353 * - A forked child which is waiting for being woken up by
10354 * wake_up_new_task().
10355 * - A task which has been woken up by try_to_wake_up() and
10356 * waiting for actually being woken up by sched_ttwu_pending().
10358 if (!se->sum_exec_runtime ||
10359 (p->state == TASK_WAKING && p->sched_remote_wakeup))
10365 #ifdef CONFIG_FAIR_GROUP_SCHED
10367 * Propagate the changes of the sched_entity across the tg tree to make it
10368 * visible to the root
10370 static void propagate_entity_cfs_rq(struct sched_entity *se)
10372 struct cfs_rq *cfs_rq;
10374 /* Start to propagate at parent */
10377 for_each_sched_entity(se) {
10378 cfs_rq = cfs_rq_of(se);
10380 if (cfs_rq_throttled(cfs_rq))
10383 update_load_avg(cfs_rq, se, UPDATE_TG);
10387 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
10390 static void detach_entity_cfs_rq(struct sched_entity *se)
10392 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10394 /* Catch up with the cfs_rq and remove our load when we leave */
10395 update_load_avg(cfs_rq, se, 0);
10396 detach_entity_load_avg(cfs_rq, se);
10397 update_tg_load_avg(cfs_rq, false);
10398 propagate_entity_cfs_rq(se);
10401 static void attach_entity_cfs_rq(struct sched_entity *se)
10403 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10405 #ifdef CONFIG_FAIR_GROUP_SCHED
10407 * Since the real-depth could have been changed (only FAIR
10408 * class maintain depth value), reset depth properly.
10410 se->depth = se->parent ? se->parent->depth + 1 : 0;
10413 /* Synchronize entity with its cfs_rq */
10414 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
10415 attach_entity_load_avg(cfs_rq, se, 0);
10416 update_tg_load_avg(cfs_rq, false);
10417 propagate_entity_cfs_rq(se);
10420 static void detach_task_cfs_rq(struct task_struct *p)
10422 struct sched_entity *se = &p->se;
10423 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10425 if (!vruntime_normalized(p)) {
10427 * Fix up our vruntime so that the current sleep doesn't
10428 * cause 'unlimited' sleep bonus.
10430 place_entity(cfs_rq, se, 0);
10431 se->vruntime -= cfs_rq->min_vruntime;
10434 detach_entity_cfs_rq(se);
10437 static void attach_task_cfs_rq(struct task_struct *p)
10439 struct sched_entity *se = &p->se;
10440 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10442 attach_entity_cfs_rq(se);
10444 if (!vruntime_normalized(p))
10445 se->vruntime += cfs_rq->min_vruntime;
10448 static void switched_from_fair(struct rq *rq, struct task_struct *p)
10450 detach_task_cfs_rq(p);
10453 static void switched_to_fair(struct rq *rq, struct task_struct *p)
10455 attach_task_cfs_rq(p);
10457 if (task_on_rq_queued(p)) {
10459 * We were most likely switched from sched_rt, so
10460 * kick off the schedule if running, otherwise just see
10461 * if we can still preempt the current task.
10466 check_preempt_curr(rq, p, 0);
10470 /* Account for a task changing its policy or group.
10472 * This routine is mostly called to set cfs_rq->curr field when a task
10473 * migrates between groups/classes.
10475 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
10477 struct sched_entity *se = &p->se;
10480 if (task_on_rq_queued(p)) {
10482 * Move the next running task to the front of the list, so our
10483 * cfs_tasks list becomes MRU one.
10485 list_move(&se->group_node, &rq->cfs_tasks);
10489 for_each_sched_entity(se) {
10490 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10492 set_next_entity(cfs_rq, se);
10493 /* ensure bandwidth has been allocated on our new cfs_rq */
10494 account_cfs_rq_runtime(cfs_rq, 0);
10498 void init_cfs_rq(struct cfs_rq *cfs_rq)
10500 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
10501 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
10502 #ifndef CONFIG_64BIT
10503 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
10506 raw_spin_lock_init(&cfs_rq->removed.lock);
10510 #ifdef CONFIG_FAIR_GROUP_SCHED
10511 static void task_set_group_fair(struct task_struct *p)
10513 struct sched_entity *se = &p->se;
10515 set_task_rq(p, task_cpu(p));
10516 se->depth = se->parent ? se->parent->depth + 1 : 0;
10519 static void task_move_group_fair(struct task_struct *p)
10521 detach_task_cfs_rq(p);
10522 set_task_rq(p, task_cpu(p));
10525 /* Tell se's cfs_rq has been changed -- migrated */
10526 p->se.avg.last_update_time = 0;
10528 attach_task_cfs_rq(p);
10531 static void task_change_group_fair(struct task_struct *p, int type)
10534 case TASK_SET_GROUP:
10535 task_set_group_fair(p);
10538 case TASK_MOVE_GROUP:
10539 task_move_group_fair(p);
10544 void free_fair_sched_group(struct task_group *tg)
10548 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
10550 for_each_possible_cpu(i) {
10552 kfree(tg->cfs_rq[i]);
10561 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10563 struct sched_entity *se;
10564 struct cfs_rq *cfs_rq;
10567 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
10570 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
10574 tg->shares = NICE_0_LOAD;
10576 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
10578 for_each_possible_cpu(i) {
10579 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
10580 GFP_KERNEL, cpu_to_node(i));
10584 se = kzalloc_node(sizeof(struct sched_entity),
10585 GFP_KERNEL, cpu_to_node(i));
10589 init_cfs_rq(cfs_rq);
10590 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
10591 init_entity_runnable_average(se);
10602 void online_fair_sched_group(struct task_group *tg)
10604 struct sched_entity *se;
10605 struct rq_flags rf;
10609 for_each_possible_cpu(i) {
10612 rq_lock_irq(rq, &rf);
10613 update_rq_clock(rq);
10614 attach_entity_cfs_rq(se);
10615 sync_throttle(tg, i);
10616 rq_unlock_irq(rq, &rf);
10620 void unregister_fair_sched_group(struct task_group *tg)
10622 unsigned long flags;
10626 for_each_possible_cpu(cpu) {
10628 remove_entity_load_avg(tg->se[cpu]);
10631 * Only empty task groups can be destroyed; so we can speculatively
10632 * check on_list without danger of it being re-added.
10634 if (!tg->cfs_rq[cpu]->on_list)
10639 raw_spin_lock_irqsave(&rq->lock, flags);
10640 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
10641 raw_spin_unlock_irqrestore(&rq->lock, flags);
10645 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
10646 struct sched_entity *se, int cpu,
10647 struct sched_entity *parent)
10649 struct rq *rq = cpu_rq(cpu);
10653 init_cfs_rq_runtime(cfs_rq);
10655 tg->cfs_rq[cpu] = cfs_rq;
10658 /* se could be NULL for root_task_group */
10663 se->cfs_rq = &rq->cfs;
10666 se->cfs_rq = parent->my_q;
10667 se->depth = parent->depth + 1;
10671 /* guarantee group entities always have weight */
10672 update_load_set(&se->load, NICE_0_LOAD);
10673 se->parent = parent;
10676 static DEFINE_MUTEX(shares_mutex);
10678 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10683 * We can't change the weight of the root cgroup.
10688 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10690 mutex_lock(&shares_mutex);
10691 if (tg->shares == shares)
10694 tg->shares = shares;
10695 for_each_possible_cpu(i) {
10696 struct rq *rq = cpu_rq(i);
10697 struct sched_entity *se = tg->se[i];
10698 struct rq_flags rf;
10700 /* Propagate contribution to hierarchy */
10701 rq_lock_irqsave(rq, &rf);
10702 update_rq_clock(rq);
10703 for_each_sched_entity(se) {
10704 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
10705 update_cfs_group(se);
10707 rq_unlock_irqrestore(rq, &rf);
10711 mutex_unlock(&shares_mutex);
10714 #else /* CONFIG_FAIR_GROUP_SCHED */
10716 void free_fair_sched_group(struct task_group *tg) { }
10718 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10723 void online_fair_sched_group(struct task_group *tg) { }
10725 void unregister_fair_sched_group(struct task_group *tg) { }
10727 #endif /* CONFIG_FAIR_GROUP_SCHED */
10730 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10732 struct sched_entity *se = &task->se;
10733 unsigned int rr_interval = 0;
10736 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10739 if (rq->cfs.load.weight)
10740 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10742 return rr_interval;
10746 * All the scheduling class methods:
10748 const struct sched_class fair_sched_class = {
10749 .next = &idle_sched_class,
10750 .enqueue_task = enqueue_task_fair,
10751 .dequeue_task = dequeue_task_fair,
10752 .yield_task = yield_task_fair,
10753 .yield_to_task = yield_to_task_fair,
10755 .check_preempt_curr = check_preempt_wakeup,
10757 .pick_next_task = __pick_next_task_fair,
10758 .put_prev_task = put_prev_task_fair,
10759 .set_next_task = set_next_task_fair,
10762 .balance = balance_fair,
10763 .select_task_rq = select_task_rq_fair,
10764 .migrate_task_rq = migrate_task_rq_fair,
10766 .rq_online = rq_online_fair,
10767 .rq_offline = rq_offline_fair,
10769 .task_dead = task_dead_fair,
10770 .set_cpus_allowed = set_cpus_allowed_common,
10773 .task_tick = task_tick_fair,
10774 .task_fork = task_fork_fair,
10776 .prio_changed = prio_changed_fair,
10777 .switched_from = switched_from_fair,
10778 .switched_to = switched_to_fair,
10780 .get_rr_interval = get_rr_interval_fair,
10782 .update_curr = update_curr_fair,
10784 #ifdef CONFIG_FAIR_GROUP_SCHED
10785 .task_change_group = task_change_group_fair,
10788 #ifdef CONFIG_UCLAMP_TASK
10789 .uclamp_enabled = 1,
10793 #ifdef CONFIG_SCHED_DEBUG
10794 void print_cfs_stats(struct seq_file *m, int cpu)
10796 struct cfs_rq *cfs_rq, *pos;
10799 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
10800 print_cfs_rq(m, cpu, cfs_rq);
10804 #ifdef CONFIG_NUMA_BALANCING
10805 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10808 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10809 struct numa_group *ng;
10812 ng = rcu_dereference(p->numa_group);
10813 for_each_online_node(node) {
10814 if (p->numa_faults) {
10815 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10816 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10819 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
10820 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
10822 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10826 #endif /* CONFIG_NUMA_BALANCING */
10827 #endif /* CONFIG_SCHED_DEBUG */
10829 __init void init_sched_fair_class(void)
10832 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10834 #ifdef CONFIG_NO_HZ_COMMON
10835 nohz.next_balance = jiffies;
10836 nohz.next_blocked = jiffies;
10837 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
10844 * Helper functions to facilitate extracting info from tracepoints.
10847 const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq)
10850 return cfs_rq ? &cfs_rq->avg : NULL;
10855 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg);
10857 char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len)
10861 strlcpy(str, "(null)", len);
10866 cfs_rq_tg_path(cfs_rq, str, len);
10869 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path);
10871 int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq)
10873 return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1;
10875 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu);
10877 const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq)
10880 return rq ? &rq->avg_rt : NULL;
10885 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt);
10887 const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq)
10890 return rq ? &rq->avg_dl : NULL;
10895 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl);
10897 const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq)
10899 #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ)
10900 return rq ? &rq->avg_irq : NULL;
10905 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq);
10907 int sched_trace_rq_cpu(struct rq *rq)
10909 return rq ? cpu_of(rq) : -1;
10911 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu);
10913 const struct cpumask *sched_trace_rd_span(struct root_domain *rd)
10916 return rd ? rd->span : NULL;
10921 EXPORT_SYMBOL_GPL(sched_trace_rd_span);