1 // SPDX-License-Identifier: GPL-2.0
3 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
5 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
7 * Interactivity improvements by Mike Galbraith
8 * (C) 2007 Mike Galbraith <efault@gmx.de>
10 * Various enhancements by Dmitry Adamushko.
11 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
13 * Group scheduling enhancements by Srivatsa Vaddagiri
14 * Copyright IBM Corporation, 2007
15 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
17 * Scaled math optimizations by Thomas Gleixner
18 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
20 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
21 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
25 #include <trace/events/sched.h>
28 * Targeted preemption latency for CPU-bound tasks:
30 * NOTE: this latency value is not the same as the concept of
31 * 'timeslice length' - timeslices in CFS are of variable length
32 * and have no persistent notion like in traditional, time-slice
33 * based scheduling concepts.
35 * (to see the precise effective timeslice length of your workload,
36 * run vmstat and monitor the context-switches (cs) field)
38 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
40 unsigned int sysctl_sched_latency = 6000000ULL;
41 static unsigned int normalized_sysctl_sched_latency = 6000000ULL;
44 * The initial- and re-scaling of tunables is configurable
48 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
49 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
50 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
52 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
54 enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
57 * Minimal preemption granularity for CPU-bound tasks:
59 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
61 unsigned int sysctl_sched_min_granularity = 750000ULL;
62 static unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
65 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
67 static unsigned int sched_nr_latency = 8;
70 * After fork, child runs first. If set to 0 (default) then
71 * parent will (try to) run first.
73 unsigned int sysctl_sched_child_runs_first __read_mostly;
76 * SCHED_OTHER wake-up granularity.
78 * This option delays the preemption effects of decoupled workloads
79 * and reduces their over-scheduling. Synchronous workloads will still
80 * have immediate wakeup/sleep latencies.
82 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
84 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
85 static unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
87 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
91 * For asym packing, by default the lower numbered CPU has higher priority.
93 int __weak arch_asym_cpu_priority(int cpu)
99 * The margin used when comparing utilization with CPU capacity:
100 * util * margin < capacity * 1024
104 static unsigned int capacity_margin = 1280;
107 #ifdef CONFIG_CFS_BANDWIDTH
109 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
110 * each time a cfs_rq requests quota.
112 * Note: in the case that the slice exceeds the runtime remaining (either due
113 * to consumption or the quota being specified to be smaller than the slice)
114 * we will always only issue the remaining available time.
116 * (default: 5 msec, units: microseconds)
118 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
121 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
127 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
133 static inline void update_load_set(struct load_weight *lw, unsigned long w)
140 * Increase the granularity value when there are more CPUs,
141 * because with more CPUs the 'effective latency' as visible
142 * to users decreases. But the relationship is not linear,
143 * so pick a second-best guess by going with the log2 of the
146 * This idea comes from the SD scheduler of Con Kolivas:
148 static unsigned int get_update_sysctl_factor(void)
150 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
153 switch (sysctl_sched_tunable_scaling) {
154 case SCHED_TUNABLESCALING_NONE:
157 case SCHED_TUNABLESCALING_LINEAR:
160 case SCHED_TUNABLESCALING_LOG:
162 factor = 1 + ilog2(cpus);
169 static void update_sysctl(void)
171 unsigned int factor = get_update_sysctl_factor();
173 #define SET_SYSCTL(name) \
174 (sysctl_##name = (factor) * normalized_sysctl_##name)
175 SET_SYSCTL(sched_min_granularity);
176 SET_SYSCTL(sched_latency);
177 SET_SYSCTL(sched_wakeup_granularity);
181 void sched_init_granularity(void)
186 #define WMULT_CONST (~0U)
187 #define WMULT_SHIFT 32
189 static void __update_inv_weight(struct load_weight *lw)
193 if (likely(lw->inv_weight))
196 w = scale_load_down(lw->weight);
198 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
200 else if (unlikely(!w))
201 lw->inv_weight = WMULT_CONST;
203 lw->inv_weight = WMULT_CONST / w;
207 * delta_exec * weight / lw.weight
209 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
211 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
212 * we're guaranteed shift stays positive because inv_weight is guaranteed to
213 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
215 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
216 * weight/lw.weight <= 1, and therefore our shift will also be positive.
218 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
220 u64 fact = scale_load_down(weight);
221 int shift = WMULT_SHIFT;
223 __update_inv_weight(lw);
225 if (unlikely(fact >> 32)) {
232 /* hint to use a 32x32->64 mul */
233 fact = (u64)(u32)fact * lw->inv_weight;
240 return mul_u64_u32_shr(delta_exec, fact, shift);
244 const struct sched_class fair_sched_class;
246 /**************************************************************
247 * CFS operations on generic schedulable entities:
250 #ifdef CONFIG_FAIR_GROUP_SCHED
252 /* cpu runqueue to which this cfs_rq is attached */
253 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
258 static inline struct task_struct *task_of(struct sched_entity *se)
260 SCHED_WARN_ON(!entity_is_task(se));
261 return container_of(se, struct task_struct, se);
264 /* Walk up scheduling entities hierarchy */
265 #define for_each_sched_entity(se) \
266 for (; se; se = se->parent)
268 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
273 /* runqueue on which this entity is (to be) queued */
274 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
279 /* runqueue "owned" by this group */
280 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
285 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
287 if (!cfs_rq->on_list) {
288 struct rq *rq = rq_of(cfs_rq);
289 int cpu = cpu_of(rq);
291 * Ensure we either appear before our parent (if already
292 * enqueued) or force our parent to appear after us when it is
293 * enqueued. The fact that we always enqueue bottom-up
294 * reduces this to two cases and a special case for the root
295 * cfs_rq. Furthermore, it also means that we will always reset
296 * tmp_alone_branch either when the branch is connected
297 * to a tree or when we reach the beg of the tree
299 if (cfs_rq->tg->parent &&
300 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
302 * If parent is already on the list, we add the child
303 * just before. Thanks to circular linked property of
304 * the list, this means to put the child at the tail
305 * of the list that starts by parent.
307 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
308 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
310 * The branch is now connected to its tree so we can
311 * reset tmp_alone_branch to the beginning of the
314 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
315 } else if (!cfs_rq->tg->parent) {
317 * cfs rq without parent should be put
318 * at the tail of the list.
320 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
321 &rq->leaf_cfs_rq_list);
323 * We have reach the beg of a tree so we can reset
324 * tmp_alone_branch to the beginning of the list.
326 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
329 * The parent has not already been added so we want to
330 * make sure that it will be put after us.
331 * tmp_alone_branch points to the beg of the branch
332 * where we will add parent.
334 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
335 rq->tmp_alone_branch);
337 * update tmp_alone_branch to points to the new beg
340 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
347 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
349 if (cfs_rq->on_list) {
350 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
355 /* Iterate through all leaf cfs_rq's on a runqueue: */
356 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
357 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
359 /* Do the two (enqueued) entities belong to the same group ? */
360 static inline struct cfs_rq *
361 is_same_group(struct sched_entity *se, struct sched_entity *pse)
363 if (se->cfs_rq == pse->cfs_rq)
369 static inline struct sched_entity *parent_entity(struct sched_entity *se)
375 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
377 int se_depth, pse_depth;
380 * preemption test can be made between sibling entities who are in the
381 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
382 * both tasks until we find their ancestors who are siblings of common
386 /* First walk up until both entities are at same depth */
387 se_depth = (*se)->depth;
388 pse_depth = (*pse)->depth;
390 while (se_depth > pse_depth) {
392 *se = parent_entity(*se);
395 while (pse_depth > se_depth) {
397 *pse = parent_entity(*pse);
400 while (!is_same_group(*se, *pse)) {
401 *se = parent_entity(*se);
402 *pse = parent_entity(*pse);
406 #else /* !CONFIG_FAIR_GROUP_SCHED */
408 static inline struct task_struct *task_of(struct sched_entity *se)
410 return container_of(se, struct task_struct, se);
413 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
415 return container_of(cfs_rq, struct rq, cfs);
419 #define for_each_sched_entity(se) \
420 for (; se; se = NULL)
422 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
424 return &task_rq(p)->cfs;
427 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
429 struct task_struct *p = task_of(se);
430 struct rq *rq = task_rq(p);
435 /* runqueue "owned" by this group */
436 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
441 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
445 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
449 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
450 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
452 static inline struct sched_entity *parent_entity(struct sched_entity *se)
458 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
462 #endif /* CONFIG_FAIR_GROUP_SCHED */
464 static __always_inline
465 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
467 /**************************************************************
468 * Scheduling class tree data structure manipulation methods:
471 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
473 s64 delta = (s64)(vruntime - max_vruntime);
475 max_vruntime = vruntime;
480 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
482 s64 delta = (s64)(vruntime - min_vruntime);
484 min_vruntime = vruntime;
489 static inline int entity_before(struct sched_entity *a,
490 struct sched_entity *b)
492 return (s64)(a->vruntime - b->vruntime) < 0;
495 static void update_min_vruntime(struct cfs_rq *cfs_rq)
497 struct sched_entity *curr = cfs_rq->curr;
498 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
500 u64 vruntime = cfs_rq->min_vruntime;
504 vruntime = curr->vruntime;
509 if (leftmost) { /* non-empty tree */
510 struct sched_entity *se;
511 se = rb_entry(leftmost, struct sched_entity, run_node);
514 vruntime = se->vruntime;
516 vruntime = min_vruntime(vruntime, se->vruntime);
519 /* ensure we never gain time by being placed backwards. */
520 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
523 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
528 * Enqueue an entity into the rb-tree:
530 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
532 struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node;
533 struct rb_node *parent = NULL;
534 struct sched_entity *entry;
535 bool leftmost = true;
538 * Find the right place in the rbtree:
542 entry = rb_entry(parent, struct sched_entity, run_node);
544 * We dont care about collisions. Nodes with
545 * the same key stay together.
547 if (entity_before(se, entry)) {
548 link = &parent->rb_left;
550 link = &parent->rb_right;
555 rb_link_node(&se->run_node, parent, link);
556 rb_insert_color_cached(&se->run_node,
557 &cfs_rq->tasks_timeline, leftmost);
560 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
562 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
565 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
567 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
572 return rb_entry(left, struct sched_entity, run_node);
575 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
577 struct rb_node *next = rb_next(&se->run_node);
582 return rb_entry(next, struct sched_entity, run_node);
585 #ifdef CONFIG_SCHED_DEBUG
586 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
588 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
593 return rb_entry(last, struct sched_entity, run_node);
596 /**************************************************************
597 * Scheduling class statistics methods:
600 int sched_proc_update_handler(struct ctl_table *table, int write,
601 void __user *buffer, size_t *lenp,
604 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
605 unsigned int factor = get_update_sysctl_factor();
610 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
611 sysctl_sched_min_granularity);
613 #define WRT_SYSCTL(name) \
614 (normalized_sysctl_##name = sysctl_##name / (factor))
615 WRT_SYSCTL(sched_min_granularity);
616 WRT_SYSCTL(sched_latency);
617 WRT_SYSCTL(sched_wakeup_granularity);
627 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
629 if (unlikely(se->load.weight != NICE_0_LOAD))
630 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
636 * The idea is to set a period in which each task runs once.
638 * When there are too many tasks (sched_nr_latency) we have to stretch
639 * this period because otherwise the slices get too small.
641 * p = (nr <= nl) ? l : l*nr/nl
643 static u64 __sched_period(unsigned long nr_running)
645 if (unlikely(nr_running > sched_nr_latency))
646 return nr_running * sysctl_sched_min_granularity;
648 return sysctl_sched_latency;
652 * We calculate the wall-time slice from the period by taking a part
653 * proportional to the weight.
657 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
659 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
661 for_each_sched_entity(se) {
662 struct load_weight *load;
663 struct load_weight lw;
665 cfs_rq = cfs_rq_of(se);
666 load = &cfs_rq->load;
668 if (unlikely(!se->on_rq)) {
671 update_load_add(&lw, se->load.weight);
674 slice = __calc_delta(slice, se->load.weight, load);
680 * We calculate the vruntime slice of a to-be-inserted task.
684 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
686 return calc_delta_fair(sched_slice(cfs_rq, se), se);
691 #include "sched-pelt.h"
693 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
694 static unsigned long task_h_load(struct task_struct *p);
695 static unsigned long capacity_of(int cpu);
697 /* Give new sched_entity start runnable values to heavy its load in infant time */
698 void init_entity_runnable_average(struct sched_entity *se)
700 struct sched_avg *sa = &se->avg;
702 memset(sa, 0, sizeof(*sa));
705 * Tasks are initialized with full load to be seen as heavy tasks until
706 * they get a chance to stabilize to their real load level.
707 * Group entities are initialized with zero load to reflect the fact that
708 * nothing has been attached to the task group yet.
710 if (entity_is_task(se))
711 sa->runnable_load_avg = sa->load_avg = scale_load_down(se->load.weight);
713 se->runnable_weight = se->load.weight;
715 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
718 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
719 static void attach_entity_cfs_rq(struct sched_entity *se);
722 * With new tasks being created, their initial util_avgs are extrapolated
723 * based on the cfs_rq's current util_avg:
725 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
727 * However, in many cases, the above util_avg does not give a desired
728 * value. Moreover, the sum of the util_avgs may be divergent, such
729 * as when the series is a harmonic series.
731 * To solve this problem, we also cap the util_avg of successive tasks to
732 * only 1/2 of the left utilization budget:
734 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
736 * where n denotes the nth task and cpu_scale the CPU capacity.
738 * For example, for a CPU with 1024 of capacity, a simplest series from
739 * the beginning would be like:
741 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
742 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
744 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
745 * if util_avg > util_avg_cap.
747 void post_init_entity_util_avg(struct sched_entity *se)
749 struct cfs_rq *cfs_rq = cfs_rq_of(se);
750 struct sched_avg *sa = &se->avg;
751 long cpu_scale = arch_scale_cpu_capacity(NULL, cpu_of(rq_of(cfs_rq)));
752 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
755 if (cfs_rq->avg.util_avg != 0) {
756 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
757 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
759 if (sa->util_avg > cap)
766 if (entity_is_task(se)) {
767 struct task_struct *p = task_of(se);
768 if (p->sched_class != &fair_sched_class) {
770 * For !fair tasks do:
772 update_cfs_rq_load_avg(now, cfs_rq);
773 attach_entity_load_avg(cfs_rq, se, 0);
774 switched_from_fair(rq, p);
776 * such that the next switched_to_fair() has the
779 se->avg.last_update_time = cfs_rq_clock_task(cfs_rq);
784 attach_entity_cfs_rq(se);
787 #else /* !CONFIG_SMP */
788 void init_entity_runnable_average(struct sched_entity *se)
791 void post_init_entity_util_avg(struct sched_entity *se)
794 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
797 #endif /* CONFIG_SMP */
800 * Update the current task's runtime statistics.
802 static void update_curr(struct cfs_rq *cfs_rq)
804 struct sched_entity *curr = cfs_rq->curr;
805 u64 now = rq_clock_task(rq_of(cfs_rq));
811 delta_exec = now - curr->exec_start;
812 if (unlikely((s64)delta_exec <= 0))
815 curr->exec_start = now;
817 schedstat_set(curr->statistics.exec_max,
818 max(delta_exec, curr->statistics.exec_max));
820 curr->sum_exec_runtime += delta_exec;
821 schedstat_add(cfs_rq->exec_clock, delta_exec);
823 curr->vruntime += calc_delta_fair(delta_exec, curr);
824 update_min_vruntime(cfs_rq);
826 if (entity_is_task(curr)) {
827 struct task_struct *curtask = task_of(curr);
829 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
830 cgroup_account_cputime(curtask, delta_exec);
831 account_group_exec_runtime(curtask, delta_exec);
834 account_cfs_rq_runtime(cfs_rq, delta_exec);
837 static void update_curr_fair(struct rq *rq)
839 update_curr(cfs_rq_of(&rq->curr->se));
843 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
845 u64 wait_start, prev_wait_start;
847 if (!schedstat_enabled())
850 wait_start = rq_clock(rq_of(cfs_rq));
851 prev_wait_start = schedstat_val(se->statistics.wait_start);
853 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
854 likely(wait_start > prev_wait_start))
855 wait_start -= prev_wait_start;
857 __schedstat_set(se->statistics.wait_start, wait_start);
861 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
863 struct task_struct *p;
866 if (!schedstat_enabled())
869 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
871 if (entity_is_task(se)) {
873 if (task_on_rq_migrating(p)) {
875 * Preserve migrating task's wait time so wait_start
876 * time stamp can be adjusted to accumulate wait time
877 * prior to migration.
879 __schedstat_set(se->statistics.wait_start, delta);
882 trace_sched_stat_wait(p, delta);
885 __schedstat_set(se->statistics.wait_max,
886 max(schedstat_val(se->statistics.wait_max), delta));
887 __schedstat_inc(se->statistics.wait_count);
888 __schedstat_add(se->statistics.wait_sum, delta);
889 __schedstat_set(se->statistics.wait_start, 0);
893 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
895 struct task_struct *tsk = NULL;
896 u64 sleep_start, block_start;
898 if (!schedstat_enabled())
901 sleep_start = schedstat_val(se->statistics.sleep_start);
902 block_start = schedstat_val(se->statistics.block_start);
904 if (entity_is_task(se))
908 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
913 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
914 __schedstat_set(se->statistics.sleep_max, delta);
916 __schedstat_set(se->statistics.sleep_start, 0);
917 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
920 account_scheduler_latency(tsk, delta >> 10, 1);
921 trace_sched_stat_sleep(tsk, delta);
925 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
930 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
931 __schedstat_set(se->statistics.block_max, delta);
933 __schedstat_set(se->statistics.block_start, 0);
934 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
937 if (tsk->in_iowait) {
938 __schedstat_add(se->statistics.iowait_sum, delta);
939 __schedstat_inc(se->statistics.iowait_count);
940 trace_sched_stat_iowait(tsk, delta);
943 trace_sched_stat_blocked(tsk, delta);
946 * Blocking time is in units of nanosecs, so shift by
947 * 20 to get a milliseconds-range estimation of the
948 * amount of time that the task spent sleeping:
950 if (unlikely(prof_on == SLEEP_PROFILING)) {
951 profile_hits(SLEEP_PROFILING,
952 (void *)get_wchan(tsk),
955 account_scheduler_latency(tsk, delta >> 10, 0);
961 * Task is being enqueued - update stats:
964 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
966 if (!schedstat_enabled())
970 * Are we enqueueing a waiting task? (for current tasks
971 * a dequeue/enqueue event is a NOP)
973 if (se != cfs_rq->curr)
974 update_stats_wait_start(cfs_rq, se);
976 if (flags & ENQUEUE_WAKEUP)
977 update_stats_enqueue_sleeper(cfs_rq, se);
981 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
984 if (!schedstat_enabled())
988 * Mark the end of the wait period if dequeueing a
991 if (se != cfs_rq->curr)
992 update_stats_wait_end(cfs_rq, se);
994 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
995 struct task_struct *tsk = task_of(se);
997 if (tsk->state & TASK_INTERRUPTIBLE)
998 __schedstat_set(se->statistics.sleep_start,
999 rq_clock(rq_of(cfs_rq)));
1000 if (tsk->state & TASK_UNINTERRUPTIBLE)
1001 __schedstat_set(se->statistics.block_start,
1002 rq_clock(rq_of(cfs_rq)));
1007 * We are picking a new current task - update its stats:
1010 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1013 * We are starting a new run period:
1015 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1018 /**************************************************
1019 * Scheduling class queueing methods:
1022 #ifdef CONFIG_NUMA_BALANCING
1024 * Approximate time to scan a full NUMA task in ms. The task scan period is
1025 * calculated based on the tasks virtual memory size and
1026 * numa_balancing_scan_size.
1028 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1029 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1031 /* Portion of address space to scan in MB */
1032 unsigned int sysctl_numa_balancing_scan_size = 256;
1034 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1035 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1040 spinlock_t lock; /* nr_tasks, tasks */
1045 struct rcu_head rcu;
1046 unsigned long total_faults;
1047 unsigned long max_faults_cpu;
1049 * Faults_cpu is used to decide whether memory should move
1050 * towards the CPU. As a consequence, these stats are weighted
1051 * more by CPU use than by memory faults.
1053 unsigned long *faults_cpu;
1054 unsigned long faults[0];
1057 static inline unsigned long group_faults_priv(struct numa_group *ng);
1058 static inline unsigned long group_faults_shared(struct numa_group *ng);
1060 static unsigned int task_nr_scan_windows(struct task_struct *p)
1062 unsigned long rss = 0;
1063 unsigned long nr_scan_pages;
1066 * Calculations based on RSS as non-present and empty pages are skipped
1067 * by the PTE scanner and NUMA hinting faults should be trapped based
1070 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1071 rss = get_mm_rss(p->mm);
1073 rss = nr_scan_pages;
1075 rss = round_up(rss, nr_scan_pages);
1076 return rss / nr_scan_pages;
1079 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1080 #define MAX_SCAN_WINDOW 2560
1082 static unsigned int task_scan_min(struct task_struct *p)
1084 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1085 unsigned int scan, floor;
1086 unsigned int windows = 1;
1088 if (scan_size < MAX_SCAN_WINDOW)
1089 windows = MAX_SCAN_WINDOW / scan_size;
1090 floor = 1000 / windows;
1092 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1093 return max_t(unsigned int, floor, scan);
1096 static unsigned int task_scan_start(struct task_struct *p)
1098 unsigned long smin = task_scan_min(p);
1099 unsigned long period = smin;
1101 /* Scale the maximum scan period with the amount of shared memory. */
1102 if (p->numa_group) {
1103 struct numa_group *ng = p->numa_group;
1104 unsigned long shared = group_faults_shared(ng);
1105 unsigned long private = group_faults_priv(ng);
1107 period *= atomic_read(&ng->refcount);
1108 period *= shared + 1;
1109 period /= private + shared + 1;
1112 return max(smin, period);
1115 static unsigned int task_scan_max(struct task_struct *p)
1117 unsigned long smin = task_scan_min(p);
1120 /* Watch for min being lower than max due to floor calculations */
1121 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1123 /* Scale the maximum scan period with the amount of shared memory. */
1124 if (p->numa_group) {
1125 struct numa_group *ng = p->numa_group;
1126 unsigned long shared = group_faults_shared(ng);
1127 unsigned long private = group_faults_priv(ng);
1128 unsigned long period = smax;
1130 period *= atomic_read(&ng->refcount);
1131 period *= shared + 1;
1132 period /= private + shared + 1;
1134 smax = max(smax, period);
1137 return max(smin, smax);
1140 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
1143 struct mm_struct *mm = p->mm;
1146 mm_users = atomic_read(&mm->mm_users);
1147 if (mm_users == 1) {
1148 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1149 mm->numa_scan_seq = 0;
1153 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
1154 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1155 p->numa_work.next = &p->numa_work;
1156 p->numa_faults = NULL;
1157 p->numa_group = NULL;
1158 p->last_task_numa_placement = 0;
1159 p->last_sum_exec_runtime = 0;
1161 /* New address space, reset the preferred nid */
1162 if (!(clone_flags & CLONE_VM)) {
1163 p->numa_preferred_nid = -1;
1168 * New thread, keep existing numa_preferred_nid which should be copied
1169 * already by arch_dup_task_struct but stagger when scans start.
1174 delay = min_t(unsigned int, task_scan_max(current),
1175 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
1176 delay += 2 * TICK_NSEC;
1177 p->node_stamp = delay;
1181 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1183 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1184 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1187 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1189 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1190 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1193 /* Shared or private faults. */
1194 #define NR_NUMA_HINT_FAULT_TYPES 2
1196 /* Memory and CPU locality */
1197 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1199 /* Averaged statistics, and temporary buffers. */
1200 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1202 pid_t task_numa_group_id(struct task_struct *p)
1204 return p->numa_group ? p->numa_group->gid : 0;
1208 * The averaged statistics, shared & private, memory & CPU,
1209 * occupy the first half of the array. The second half of the
1210 * array is for current counters, which are averaged into the
1211 * first set by task_numa_placement.
1213 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1215 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1218 static inline unsigned long task_faults(struct task_struct *p, int nid)
1220 if (!p->numa_faults)
1223 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1224 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1227 static inline unsigned long group_faults(struct task_struct *p, int nid)
1232 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1233 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1236 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1238 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1239 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1242 static inline unsigned long group_faults_priv(struct numa_group *ng)
1244 unsigned long faults = 0;
1247 for_each_online_node(node) {
1248 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1254 static inline unsigned long group_faults_shared(struct numa_group *ng)
1256 unsigned long faults = 0;
1259 for_each_online_node(node) {
1260 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1267 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1268 * considered part of a numa group's pseudo-interleaving set. Migrations
1269 * between these nodes are slowed down, to allow things to settle down.
1271 #define ACTIVE_NODE_FRACTION 3
1273 static bool numa_is_active_node(int nid, struct numa_group *ng)
1275 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1278 /* Handle placement on systems where not all nodes are directly connected. */
1279 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1280 int maxdist, bool task)
1282 unsigned long score = 0;
1286 * All nodes are directly connected, and the same distance
1287 * from each other. No need for fancy placement algorithms.
1289 if (sched_numa_topology_type == NUMA_DIRECT)
1293 * This code is called for each node, introducing N^2 complexity,
1294 * which should be ok given the number of nodes rarely exceeds 8.
1296 for_each_online_node(node) {
1297 unsigned long faults;
1298 int dist = node_distance(nid, node);
1301 * The furthest away nodes in the system are not interesting
1302 * for placement; nid was already counted.
1304 if (dist == sched_max_numa_distance || node == nid)
1308 * On systems with a backplane NUMA topology, compare groups
1309 * of nodes, and move tasks towards the group with the most
1310 * memory accesses. When comparing two nodes at distance
1311 * "hoplimit", only nodes closer by than "hoplimit" are part
1312 * of each group. Skip other nodes.
1314 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1318 /* Add up the faults from nearby nodes. */
1320 faults = task_faults(p, node);
1322 faults = group_faults(p, node);
1325 * On systems with a glueless mesh NUMA topology, there are
1326 * no fixed "groups of nodes". Instead, nodes that are not
1327 * directly connected bounce traffic through intermediate
1328 * nodes; a numa_group can occupy any set of nodes.
1329 * The further away a node is, the less the faults count.
1330 * This seems to result in good task placement.
1332 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1333 faults *= (sched_max_numa_distance - dist);
1334 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1344 * These return the fraction of accesses done by a particular task, or
1345 * task group, on a particular numa node. The group weight is given a
1346 * larger multiplier, in order to group tasks together that are almost
1347 * evenly spread out between numa nodes.
1349 static inline unsigned long task_weight(struct task_struct *p, int nid,
1352 unsigned long faults, total_faults;
1354 if (!p->numa_faults)
1357 total_faults = p->total_numa_faults;
1362 faults = task_faults(p, nid);
1363 faults += score_nearby_nodes(p, nid, dist, true);
1365 return 1000 * faults / total_faults;
1368 static inline unsigned long group_weight(struct task_struct *p, int nid,
1371 unsigned long faults, total_faults;
1376 total_faults = p->numa_group->total_faults;
1381 faults = group_faults(p, nid);
1382 faults += score_nearby_nodes(p, nid, dist, false);
1384 return 1000 * faults / total_faults;
1387 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1388 int src_nid, int dst_cpu)
1390 struct numa_group *ng = p->numa_group;
1391 int dst_nid = cpu_to_node(dst_cpu);
1392 int last_cpupid, this_cpupid;
1394 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1395 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1398 * Allow first faults or private faults to migrate immediately early in
1399 * the lifetime of a task. The magic number 4 is based on waiting for
1400 * two full passes of the "multi-stage node selection" test that is
1403 if ((p->numa_preferred_nid == -1 || p->numa_scan_seq <= 4) &&
1404 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1408 * Multi-stage node selection is used in conjunction with a periodic
1409 * migration fault to build a temporal task<->page relation. By using
1410 * a two-stage filter we remove short/unlikely relations.
1412 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1413 * a task's usage of a particular page (n_p) per total usage of this
1414 * page (n_t) (in a given time-span) to a probability.
1416 * Our periodic faults will sample this probability and getting the
1417 * same result twice in a row, given these samples are fully
1418 * independent, is then given by P(n)^2, provided our sample period
1419 * is sufficiently short compared to the usage pattern.
1421 * This quadric squishes small probabilities, making it less likely we
1422 * act on an unlikely task<->page relation.
1424 if (!cpupid_pid_unset(last_cpupid) &&
1425 cpupid_to_nid(last_cpupid) != dst_nid)
1428 /* Always allow migrate on private faults */
1429 if (cpupid_match_pid(p, last_cpupid))
1432 /* A shared fault, but p->numa_group has not been set up yet. */
1437 * Destination node is much more heavily used than the source
1438 * node? Allow migration.
1440 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1441 ACTIVE_NODE_FRACTION)
1445 * Distribute memory according to CPU & memory use on each node,
1446 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1448 * faults_cpu(dst) 3 faults_cpu(src)
1449 * --------------- * - > ---------------
1450 * faults_mem(dst) 4 faults_mem(src)
1452 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1453 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1456 static unsigned long weighted_cpuload(struct rq *rq);
1457 static unsigned long source_load(int cpu, int type);
1458 static unsigned long target_load(int cpu, int type);
1460 /* Cached statistics for all CPUs within a node */
1464 /* Total compute capacity of CPUs on a node */
1465 unsigned long compute_capacity;
1469 * XXX borrowed from update_sg_lb_stats
1471 static void update_numa_stats(struct numa_stats *ns, int nid)
1475 memset(ns, 0, sizeof(*ns));
1476 for_each_cpu(cpu, cpumask_of_node(nid)) {
1477 struct rq *rq = cpu_rq(cpu);
1479 ns->load += weighted_cpuload(rq);
1480 ns->compute_capacity += capacity_of(cpu);
1485 struct task_numa_env {
1486 struct task_struct *p;
1488 int src_cpu, src_nid;
1489 int dst_cpu, dst_nid;
1491 struct numa_stats src_stats, dst_stats;
1496 struct task_struct *best_task;
1501 static void task_numa_assign(struct task_numa_env *env,
1502 struct task_struct *p, long imp)
1504 struct rq *rq = cpu_rq(env->dst_cpu);
1506 /* Bail out if run-queue part of active NUMA balance. */
1507 if (xchg(&rq->numa_migrate_on, 1))
1511 * Clear previous best_cpu/rq numa-migrate flag, since task now
1512 * found a better CPU to move/swap.
1514 if (env->best_cpu != -1) {
1515 rq = cpu_rq(env->best_cpu);
1516 WRITE_ONCE(rq->numa_migrate_on, 0);
1520 put_task_struct(env->best_task);
1525 env->best_imp = imp;
1526 env->best_cpu = env->dst_cpu;
1529 static bool load_too_imbalanced(long src_load, long dst_load,
1530 struct task_numa_env *env)
1533 long orig_src_load, orig_dst_load;
1534 long src_capacity, dst_capacity;
1537 * The load is corrected for the CPU capacity available on each node.
1540 * ------------ vs ---------
1541 * src_capacity dst_capacity
1543 src_capacity = env->src_stats.compute_capacity;
1544 dst_capacity = env->dst_stats.compute_capacity;
1546 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1548 orig_src_load = env->src_stats.load;
1549 orig_dst_load = env->dst_stats.load;
1551 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1553 /* Would this change make things worse? */
1554 return (imb > old_imb);
1558 * Maximum NUMA importance can be 1998 (2*999);
1559 * SMALLIMP @ 30 would be close to 1998/64.
1560 * Used to deter task migration.
1565 * This checks if the overall compute and NUMA accesses of the system would
1566 * be improved if the source tasks was migrated to the target dst_cpu taking
1567 * into account that it might be best if task running on the dst_cpu should
1568 * be exchanged with the source task
1570 static void task_numa_compare(struct task_numa_env *env,
1571 long taskimp, long groupimp, bool maymove)
1573 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1574 struct task_struct *cur;
1575 long src_load, dst_load;
1577 long imp = env->p->numa_group ? groupimp : taskimp;
1579 int dist = env->dist;
1581 if (READ_ONCE(dst_rq->numa_migrate_on))
1585 cur = task_rcu_dereference(&dst_rq->curr);
1586 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1590 * Because we have preemption enabled we can get migrated around and
1591 * end try selecting ourselves (current == env->p) as a swap candidate.
1597 if (maymove && moveimp >= env->best_imp)
1604 * "imp" is the fault differential for the source task between the
1605 * source and destination node. Calculate the total differential for
1606 * the source task and potential destination task. The more negative
1607 * the value is, the more remote accesses that would be expected to
1608 * be incurred if the tasks were swapped.
1610 /* Skip this swap candidate if cannot move to the source cpu */
1611 if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed))
1615 * If dst and source tasks are in the same NUMA group, or not
1616 * in any group then look only at task weights.
1618 if (cur->numa_group == env->p->numa_group) {
1619 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1620 task_weight(cur, env->dst_nid, dist);
1622 * Add some hysteresis to prevent swapping the
1623 * tasks within a group over tiny differences.
1625 if (cur->numa_group)
1629 * Compare the group weights. If a task is all by itself
1630 * (not part of a group), use the task weight instead.
1632 if (cur->numa_group && env->p->numa_group)
1633 imp += group_weight(cur, env->src_nid, dist) -
1634 group_weight(cur, env->dst_nid, dist);
1636 imp += task_weight(cur, env->src_nid, dist) -
1637 task_weight(cur, env->dst_nid, dist);
1640 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1647 * If the NUMA importance is less than SMALLIMP,
1648 * task migration might only result in ping pong
1649 * of tasks and also hurt performance due to cache
1652 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1656 * In the overloaded case, try and keep the load balanced.
1658 load = task_h_load(env->p) - task_h_load(cur);
1662 dst_load = env->dst_stats.load + load;
1663 src_load = env->src_stats.load - load;
1665 if (load_too_imbalanced(src_load, dst_load, env))
1670 * One idle CPU per node is evaluated for a task numa move.
1671 * Call select_idle_sibling to maybe find a better one.
1675 * select_idle_siblings() uses an per-CPU cpumask that
1676 * can be used from IRQ context.
1678 local_irq_disable();
1679 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1684 task_numa_assign(env, cur, imp);
1689 static void task_numa_find_cpu(struct task_numa_env *env,
1690 long taskimp, long groupimp)
1692 long src_load, dst_load, load;
1693 bool maymove = false;
1696 load = task_h_load(env->p);
1697 dst_load = env->dst_stats.load + load;
1698 src_load = env->src_stats.load - load;
1701 * If the improvement from just moving env->p direction is better
1702 * than swapping tasks around, check if a move is possible.
1704 maymove = !load_too_imbalanced(src_load, dst_load, env);
1706 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1707 /* Skip this CPU if the source task cannot migrate */
1708 if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
1712 task_numa_compare(env, taskimp, groupimp, maymove);
1716 static int task_numa_migrate(struct task_struct *p)
1718 struct task_numa_env env = {
1721 .src_cpu = task_cpu(p),
1722 .src_nid = task_node(p),
1724 .imbalance_pct = 112,
1730 struct sched_domain *sd;
1732 unsigned long taskweight, groupweight;
1734 long taskimp, groupimp;
1737 * Pick the lowest SD_NUMA domain, as that would have the smallest
1738 * imbalance and would be the first to start moving tasks about.
1740 * And we want to avoid any moving of tasks about, as that would create
1741 * random movement of tasks -- counter the numa conditions we're trying
1745 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1747 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1751 * Cpusets can break the scheduler domain tree into smaller
1752 * balance domains, some of which do not cross NUMA boundaries.
1753 * Tasks that are "trapped" in such domains cannot be migrated
1754 * elsewhere, so there is no point in (re)trying.
1756 if (unlikely(!sd)) {
1757 sched_setnuma(p, task_node(p));
1761 env.dst_nid = p->numa_preferred_nid;
1762 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1763 taskweight = task_weight(p, env.src_nid, dist);
1764 groupweight = group_weight(p, env.src_nid, dist);
1765 update_numa_stats(&env.src_stats, env.src_nid);
1766 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1767 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1768 update_numa_stats(&env.dst_stats, env.dst_nid);
1770 /* Try to find a spot on the preferred nid. */
1771 task_numa_find_cpu(&env, taskimp, groupimp);
1774 * Look at other nodes in these cases:
1775 * - there is no space available on the preferred_nid
1776 * - the task is part of a numa_group that is interleaved across
1777 * multiple NUMA nodes; in order to better consolidate the group,
1778 * we need to check other locations.
1780 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1781 for_each_online_node(nid) {
1782 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1785 dist = node_distance(env.src_nid, env.dst_nid);
1786 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1788 taskweight = task_weight(p, env.src_nid, dist);
1789 groupweight = group_weight(p, env.src_nid, dist);
1792 /* Only consider nodes where both task and groups benefit */
1793 taskimp = task_weight(p, nid, dist) - taskweight;
1794 groupimp = group_weight(p, nid, dist) - groupweight;
1795 if (taskimp < 0 && groupimp < 0)
1800 update_numa_stats(&env.dst_stats, env.dst_nid);
1801 task_numa_find_cpu(&env, taskimp, groupimp);
1806 * If the task is part of a workload that spans multiple NUMA nodes,
1807 * and is migrating into one of the workload's active nodes, remember
1808 * this node as the task's preferred numa node, so the workload can
1810 * A task that migrated to a second choice node will be better off
1811 * trying for a better one later. Do not set the preferred node here.
1813 if (p->numa_group) {
1814 if (env.best_cpu == -1)
1817 nid = cpu_to_node(env.best_cpu);
1819 if (nid != p->numa_preferred_nid)
1820 sched_setnuma(p, nid);
1823 /* No better CPU than the current one was found. */
1824 if (env.best_cpu == -1)
1827 best_rq = cpu_rq(env.best_cpu);
1828 if (env.best_task == NULL) {
1829 ret = migrate_task_to(p, env.best_cpu);
1830 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1832 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1836 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
1837 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1840 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1841 put_task_struct(env.best_task);
1845 /* Attempt to migrate a task to a CPU on the preferred node. */
1846 static void numa_migrate_preferred(struct task_struct *p)
1848 unsigned long interval = HZ;
1850 /* This task has no NUMA fault statistics yet */
1851 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1854 /* Periodically retry migrating the task to the preferred node */
1855 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1856 p->numa_migrate_retry = jiffies + interval;
1858 /* Success if task is already running on preferred CPU */
1859 if (task_node(p) == p->numa_preferred_nid)
1862 /* Otherwise, try migrate to a CPU on the preferred node */
1863 task_numa_migrate(p);
1867 * Find out how many nodes on the workload is actively running on. Do this by
1868 * tracking the nodes from which NUMA hinting faults are triggered. This can
1869 * be different from the set of nodes where the workload's memory is currently
1872 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1874 unsigned long faults, max_faults = 0;
1875 int nid, active_nodes = 0;
1877 for_each_online_node(nid) {
1878 faults = group_faults_cpu(numa_group, nid);
1879 if (faults > max_faults)
1880 max_faults = faults;
1883 for_each_online_node(nid) {
1884 faults = group_faults_cpu(numa_group, nid);
1885 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1889 numa_group->max_faults_cpu = max_faults;
1890 numa_group->active_nodes = active_nodes;
1894 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1895 * increments. The more local the fault statistics are, the higher the scan
1896 * period will be for the next scan window. If local/(local+remote) ratio is
1897 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1898 * the scan period will decrease. Aim for 70% local accesses.
1900 #define NUMA_PERIOD_SLOTS 10
1901 #define NUMA_PERIOD_THRESHOLD 7
1904 * Increase the scan period (slow down scanning) if the majority of
1905 * our memory is already on our local node, or if the majority of
1906 * the page accesses are shared with other processes.
1907 * Otherwise, decrease the scan period.
1909 static void update_task_scan_period(struct task_struct *p,
1910 unsigned long shared, unsigned long private)
1912 unsigned int period_slot;
1913 int lr_ratio, ps_ratio;
1916 unsigned long remote = p->numa_faults_locality[0];
1917 unsigned long local = p->numa_faults_locality[1];
1920 * If there were no record hinting faults then either the task is
1921 * completely idle or all activity is areas that are not of interest
1922 * to automatic numa balancing. Related to that, if there were failed
1923 * migration then it implies we are migrating too quickly or the local
1924 * node is overloaded. In either case, scan slower
1926 if (local + shared == 0 || p->numa_faults_locality[2]) {
1927 p->numa_scan_period = min(p->numa_scan_period_max,
1928 p->numa_scan_period << 1);
1930 p->mm->numa_next_scan = jiffies +
1931 msecs_to_jiffies(p->numa_scan_period);
1937 * Prepare to scale scan period relative to the current period.
1938 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1939 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1940 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1942 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1943 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1944 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1946 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1948 * Most memory accesses are local. There is no need to
1949 * do fast NUMA scanning, since memory is already local.
1951 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
1954 diff = slot * period_slot;
1955 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
1957 * Most memory accesses are shared with other tasks.
1958 * There is no point in continuing fast NUMA scanning,
1959 * since other tasks may just move the memory elsewhere.
1961 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
1964 diff = slot * period_slot;
1967 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
1968 * yet they are not on the local NUMA node. Speed up
1969 * NUMA scanning to get the memory moved over.
1971 int ratio = max(lr_ratio, ps_ratio);
1972 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1975 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1976 task_scan_min(p), task_scan_max(p));
1977 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1981 * Get the fraction of time the task has been running since the last
1982 * NUMA placement cycle. The scheduler keeps similar statistics, but
1983 * decays those on a 32ms period, which is orders of magnitude off
1984 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1985 * stats only if the task is so new there are no NUMA statistics yet.
1987 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1989 u64 runtime, delta, now;
1990 /* Use the start of this time slice to avoid calculations. */
1991 now = p->se.exec_start;
1992 runtime = p->se.sum_exec_runtime;
1994 if (p->last_task_numa_placement) {
1995 delta = runtime - p->last_sum_exec_runtime;
1996 *period = now - p->last_task_numa_placement;
1998 delta = p->se.avg.load_sum;
1999 *period = LOAD_AVG_MAX;
2002 p->last_sum_exec_runtime = runtime;
2003 p->last_task_numa_placement = now;
2009 * Determine the preferred nid for a task in a numa_group. This needs to
2010 * be done in a way that produces consistent results with group_weight,
2011 * otherwise workloads might not converge.
2013 static int preferred_group_nid(struct task_struct *p, int nid)
2018 /* Direct connections between all NUMA nodes. */
2019 if (sched_numa_topology_type == NUMA_DIRECT)
2023 * On a system with glueless mesh NUMA topology, group_weight
2024 * scores nodes according to the number of NUMA hinting faults on
2025 * both the node itself, and on nearby nodes.
2027 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2028 unsigned long score, max_score = 0;
2029 int node, max_node = nid;
2031 dist = sched_max_numa_distance;
2033 for_each_online_node(node) {
2034 score = group_weight(p, node, dist);
2035 if (score > max_score) {
2044 * Finding the preferred nid in a system with NUMA backplane
2045 * interconnect topology is more involved. The goal is to locate
2046 * tasks from numa_groups near each other in the system, and
2047 * untangle workloads from different sides of the system. This requires
2048 * searching down the hierarchy of node groups, recursively searching
2049 * inside the highest scoring group of nodes. The nodemask tricks
2050 * keep the complexity of the search down.
2052 nodes = node_online_map;
2053 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2054 unsigned long max_faults = 0;
2055 nodemask_t max_group = NODE_MASK_NONE;
2058 /* Are there nodes at this distance from each other? */
2059 if (!find_numa_distance(dist))
2062 for_each_node_mask(a, nodes) {
2063 unsigned long faults = 0;
2064 nodemask_t this_group;
2065 nodes_clear(this_group);
2067 /* Sum group's NUMA faults; includes a==b case. */
2068 for_each_node_mask(b, nodes) {
2069 if (node_distance(a, b) < dist) {
2070 faults += group_faults(p, b);
2071 node_set(b, this_group);
2072 node_clear(b, nodes);
2076 /* Remember the top group. */
2077 if (faults > max_faults) {
2078 max_faults = faults;
2079 max_group = this_group;
2081 * subtle: at the smallest distance there is
2082 * just one node left in each "group", the
2083 * winner is the preferred nid.
2088 /* Next round, evaluate the nodes within max_group. */
2096 static void task_numa_placement(struct task_struct *p)
2098 int seq, nid, max_nid = -1;
2099 unsigned long max_faults = 0;
2100 unsigned long fault_types[2] = { 0, 0 };
2101 unsigned long total_faults;
2102 u64 runtime, period;
2103 spinlock_t *group_lock = NULL;
2106 * The p->mm->numa_scan_seq field gets updated without
2107 * exclusive access. Use READ_ONCE() here to ensure
2108 * that the field is read in a single access:
2110 seq = READ_ONCE(p->mm->numa_scan_seq);
2111 if (p->numa_scan_seq == seq)
2113 p->numa_scan_seq = seq;
2114 p->numa_scan_period_max = task_scan_max(p);
2116 total_faults = p->numa_faults_locality[0] +
2117 p->numa_faults_locality[1];
2118 runtime = numa_get_avg_runtime(p, &period);
2120 /* If the task is part of a group prevent parallel updates to group stats */
2121 if (p->numa_group) {
2122 group_lock = &p->numa_group->lock;
2123 spin_lock_irq(group_lock);
2126 /* Find the node with the highest number of faults */
2127 for_each_online_node(nid) {
2128 /* Keep track of the offsets in numa_faults array */
2129 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2130 unsigned long faults = 0, group_faults = 0;
2133 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2134 long diff, f_diff, f_weight;
2136 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2137 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2138 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2139 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2141 /* Decay existing window, copy faults since last scan */
2142 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2143 fault_types[priv] += p->numa_faults[membuf_idx];
2144 p->numa_faults[membuf_idx] = 0;
2147 * Normalize the faults_from, so all tasks in a group
2148 * count according to CPU use, instead of by the raw
2149 * number of faults. Tasks with little runtime have
2150 * little over-all impact on throughput, and thus their
2151 * faults are less important.
2153 f_weight = div64_u64(runtime << 16, period + 1);
2154 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2156 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2157 p->numa_faults[cpubuf_idx] = 0;
2159 p->numa_faults[mem_idx] += diff;
2160 p->numa_faults[cpu_idx] += f_diff;
2161 faults += p->numa_faults[mem_idx];
2162 p->total_numa_faults += diff;
2163 if (p->numa_group) {
2165 * safe because we can only change our own group
2167 * mem_idx represents the offset for a given
2168 * nid and priv in a specific region because it
2169 * is at the beginning of the numa_faults array.
2171 p->numa_group->faults[mem_idx] += diff;
2172 p->numa_group->faults_cpu[mem_idx] += f_diff;
2173 p->numa_group->total_faults += diff;
2174 group_faults += p->numa_group->faults[mem_idx];
2178 if (!p->numa_group) {
2179 if (faults > max_faults) {
2180 max_faults = faults;
2183 } else if (group_faults > max_faults) {
2184 max_faults = group_faults;
2189 if (p->numa_group) {
2190 numa_group_count_active_nodes(p->numa_group);
2191 spin_unlock_irq(group_lock);
2192 max_nid = preferred_group_nid(p, max_nid);
2196 /* Set the new preferred node */
2197 if (max_nid != p->numa_preferred_nid)
2198 sched_setnuma(p, max_nid);
2201 update_task_scan_period(p, fault_types[0], fault_types[1]);
2204 static inline int get_numa_group(struct numa_group *grp)
2206 return atomic_inc_not_zero(&grp->refcount);
2209 static inline void put_numa_group(struct numa_group *grp)
2211 if (atomic_dec_and_test(&grp->refcount))
2212 kfree_rcu(grp, rcu);
2215 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2218 struct numa_group *grp, *my_grp;
2219 struct task_struct *tsk;
2221 int cpu = cpupid_to_cpu(cpupid);
2224 if (unlikely(!p->numa_group)) {
2225 unsigned int size = sizeof(struct numa_group) +
2226 4*nr_node_ids*sizeof(unsigned long);
2228 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2232 atomic_set(&grp->refcount, 1);
2233 grp->active_nodes = 1;
2234 grp->max_faults_cpu = 0;
2235 spin_lock_init(&grp->lock);
2237 /* Second half of the array tracks nids where faults happen */
2238 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2241 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2242 grp->faults[i] = p->numa_faults[i];
2244 grp->total_faults = p->total_numa_faults;
2247 rcu_assign_pointer(p->numa_group, grp);
2251 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2253 if (!cpupid_match_pid(tsk, cpupid))
2256 grp = rcu_dereference(tsk->numa_group);
2260 my_grp = p->numa_group;
2265 * Only join the other group if its bigger; if we're the bigger group,
2266 * the other task will join us.
2268 if (my_grp->nr_tasks > grp->nr_tasks)
2272 * Tie-break on the grp address.
2274 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2277 /* Always join threads in the same process. */
2278 if (tsk->mm == current->mm)
2281 /* Simple filter to avoid false positives due to PID collisions */
2282 if (flags & TNF_SHARED)
2285 /* Update priv based on whether false sharing was detected */
2288 if (join && !get_numa_group(grp))
2296 BUG_ON(irqs_disabled());
2297 double_lock_irq(&my_grp->lock, &grp->lock);
2299 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2300 my_grp->faults[i] -= p->numa_faults[i];
2301 grp->faults[i] += p->numa_faults[i];
2303 my_grp->total_faults -= p->total_numa_faults;
2304 grp->total_faults += p->total_numa_faults;
2309 spin_unlock(&my_grp->lock);
2310 spin_unlock_irq(&grp->lock);
2312 rcu_assign_pointer(p->numa_group, grp);
2314 put_numa_group(my_grp);
2322 void task_numa_free(struct task_struct *p)
2324 struct numa_group *grp = p->numa_group;
2325 void *numa_faults = p->numa_faults;
2326 unsigned long flags;
2330 spin_lock_irqsave(&grp->lock, flags);
2331 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2332 grp->faults[i] -= p->numa_faults[i];
2333 grp->total_faults -= p->total_numa_faults;
2336 spin_unlock_irqrestore(&grp->lock, flags);
2337 RCU_INIT_POINTER(p->numa_group, NULL);
2338 put_numa_group(grp);
2341 p->numa_faults = NULL;
2346 * Got a PROT_NONE fault for a page on @node.
2348 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2350 struct task_struct *p = current;
2351 bool migrated = flags & TNF_MIGRATED;
2352 int cpu_node = task_node(current);
2353 int local = !!(flags & TNF_FAULT_LOCAL);
2354 struct numa_group *ng;
2357 if (!static_branch_likely(&sched_numa_balancing))
2360 /* for example, ksmd faulting in a user's mm */
2364 /* Allocate buffer to track faults on a per-node basis */
2365 if (unlikely(!p->numa_faults)) {
2366 int size = sizeof(*p->numa_faults) *
2367 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2369 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2370 if (!p->numa_faults)
2373 p->total_numa_faults = 0;
2374 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2378 * First accesses are treated as private, otherwise consider accesses
2379 * to be private if the accessing pid has not changed
2381 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2384 priv = cpupid_match_pid(p, last_cpupid);
2385 if (!priv && !(flags & TNF_NO_GROUP))
2386 task_numa_group(p, last_cpupid, flags, &priv);
2390 * If a workload spans multiple NUMA nodes, a shared fault that
2391 * occurs wholly within the set of nodes that the workload is
2392 * actively using should be counted as local. This allows the
2393 * scan rate to slow down when a workload has settled down.
2396 if (!priv && !local && ng && ng->active_nodes > 1 &&
2397 numa_is_active_node(cpu_node, ng) &&
2398 numa_is_active_node(mem_node, ng))
2402 * Retry to migrate task to preferred node periodically, in case it
2403 * previously failed, or the scheduler moved us.
2405 if (time_after(jiffies, p->numa_migrate_retry)) {
2406 task_numa_placement(p);
2407 numa_migrate_preferred(p);
2411 p->numa_pages_migrated += pages;
2412 if (flags & TNF_MIGRATE_FAIL)
2413 p->numa_faults_locality[2] += pages;
2415 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2416 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2417 p->numa_faults_locality[local] += pages;
2420 static void reset_ptenuma_scan(struct task_struct *p)
2423 * We only did a read acquisition of the mmap sem, so
2424 * p->mm->numa_scan_seq is written to without exclusive access
2425 * and the update is not guaranteed to be atomic. That's not
2426 * much of an issue though, since this is just used for
2427 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2428 * expensive, to avoid any form of compiler optimizations:
2430 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2431 p->mm->numa_scan_offset = 0;
2435 * The expensive part of numa migration is done from task_work context.
2436 * Triggered from task_tick_numa().
2438 void task_numa_work(struct callback_head *work)
2440 unsigned long migrate, next_scan, now = jiffies;
2441 struct task_struct *p = current;
2442 struct mm_struct *mm = p->mm;
2443 u64 runtime = p->se.sum_exec_runtime;
2444 struct vm_area_struct *vma;
2445 unsigned long start, end;
2446 unsigned long nr_pte_updates = 0;
2447 long pages, virtpages;
2449 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2451 work->next = work; /* protect against double add */
2453 * Who cares about NUMA placement when they're dying.
2455 * NOTE: make sure not to dereference p->mm before this check,
2456 * exit_task_work() happens _after_ exit_mm() so we could be called
2457 * without p->mm even though we still had it when we enqueued this
2460 if (p->flags & PF_EXITING)
2463 if (!mm->numa_next_scan) {
2464 mm->numa_next_scan = now +
2465 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2469 * Enforce maximal scan/migration frequency..
2471 migrate = mm->numa_next_scan;
2472 if (time_before(now, migrate))
2475 if (p->numa_scan_period == 0) {
2476 p->numa_scan_period_max = task_scan_max(p);
2477 p->numa_scan_period = task_scan_start(p);
2480 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2481 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2485 * Delay this task enough that another task of this mm will likely win
2486 * the next time around.
2488 p->node_stamp += 2 * TICK_NSEC;
2490 start = mm->numa_scan_offset;
2491 pages = sysctl_numa_balancing_scan_size;
2492 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2493 virtpages = pages * 8; /* Scan up to this much virtual space */
2498 if (!down_read_trylock(&mm->mmap_sem))
2500 vma = find_vma(mm, start);
2502 reset_ptenuma_scan(p);
2506 for (; vma; vma = vma->vm_next) {
2507 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2508 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2513 * Shared library pages mapped by multiple processes are not
2514 * migrated as it is expected they are cache replicated. Avoid
2515 * hinting faults in read-only file-backed mappings or the vdso
2516 * as migrating the pages will be of marginal benefit.
2519 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2523 * Skip inaccessible VMAs to avoid any confusion between
2524 * PROT_NONE and NUMA hinting ptes
2526 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2530 start = max(start, vma->vm_start);
2531 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2532 end = min(end, vma->vm_end);
2533 nr_pte_updates = change_prot_numa(vma, start, end);
2536 * Try to scan sysctl_numa_balancing_size worth of
2537 * hpages that have at least one present PTE that
2538 * is not already pte-numa. If the VMA contains
2539 * areas that are unused or already full of prot_numa
2540 * PTEs, scan up to virtpages, to skip through those
2544 pages -= (end - start) >> PAGE_SHIFT;
2545 virtpages -= (end - start) >> PAGE_SHIFT;
2548 if (pages <= 0 || virtpages <= 0)
2552 } while (end != vma->vm_end);
2557 * It is possible to reach the end of the VMA list but the last few
2558 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2559 * would find the !migratable VMA on the next scan but not reset the
2560 * scanner to the start so check it now.
2563 mm->numa_scan_offset = start;
2565 reset_ptenuma_scan(p);
2566 up_read(&mm->mmap_sem);
2569 * Make sure tasks use at least 32x as much time to run other code
2570 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2571 * Usually update_task_scan_period slows down scanning enough; on an
2572 * overloaded system we need to limit overhead on a per task basis.
2574 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2575 u64 diff = p->se.sum_exec_runtime - runtime;
2576 p->node_stamp += 32 * diff;
2581 * Drive the periodic memory faults..
2583 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2585 struct callback_head *work = &curr->numa_work;
2589 * We don't care about NUMA placement if we don't have memory.
2591 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2595 * Using runtime rather than walltime has the dual advantage that
2596 * we (mostly) drive the selection from busy threads and that the
2597 * task needs to have done some actual work before we bother with
2600 now = curr->se.sum_exec_runtime;
2601 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2603 if (now > curr->node_stamp + period) {
2604 if (!curr->node_stamp)
2605 curr->numa_scan_period = task_scan_start(curr);
2606 curr->node_stamp += period;
2608 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2609 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2610 task_work_add(curr, work, true);
2615 static void update_scan_period(struct task_struct *p, int new_cpu)
2617 int src_nid = cpu_to_node(task_cpu(p));
2618 int dst_nid = cpu_to_node(new_cpu);
2620 if (!static_branch_likely(&sched_numa_balancing))
2623 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2626 if (src_nid == dst_nid)
2630 * Allow resets if faults have been trapped before one scan
2631 * has completed. This is most likely due to a new task that
2632 * is pulled cross-node due to wakeups or load balancing.
2634 if (p->numa_scan_seq) {
2636 * Avoid scan adjustments if moving to the preferred
2637 * node or if the task was not previously running on
2638 * the preferred node.
2640 if (dst_nid == p->numa_preferred_nid ||
2641 (p->numa_preferred_nid != -1 && src_nid != p->numa_preferred_nid))
2645 p->numa_scan_period = task_scan_start(p);
2649 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2653 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2657 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2661 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2665 #endif /* CONFIG_NUMA_BALANCING */
2668 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2670 update_load_add(&cfs_rq->load, se->load.weight);
2671 if (!parent_entity(se))
2672 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2674 if (entity_is_task(se)) {
2675 struct rq *rq = rq_of(cfs_rq);
2677 account_numa_enqueue(rq, task_of(se));
2678 list_add(&se->group_node, &rq->cfs_tasks);
2681 cfs_rq->nr_running++;
2685 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2687 update_load_sub(&cfs_rq->load, se->load.weight);
2688 if (!parent_entity(se))
2689 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2691 if (entity_is_task(se)) {
2692 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2693 list_del_init(&se->group_node);
2696 cfs_rq->nr_running--;
2700 * Signed add and clamp on underflow.
2702 * Explicitly do a load-store to ensure the intermediate value never hits
2703 * memory. This allows lockless observations without ever seeing the negative
2706 #define add_positive(_ptr, _val) do { \
2707 typeof(_ptr) ptr = (_ptr); \
2708 typeof(_val) val = (_val); \
2709 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2713 if (val < 0 && res > var) \
2716 WRITE_ONCE(*ptr, res); \
2720 * Unsigned subtract and clamp on underflow.
2722 * Explicitly do a load-store to ensure the intermediate value never hits
2723 * memory. This allows lockless observations without ever seeing the negative
2726 #define sub_positive(_ptr, _val) do { \
2727 typeof(_ptr) ptr = (_ptr); \
2728 typeof(*ptr) val = (_val); \
2729 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2733 WRITE_ONCE(*ptr, res); \
2737 * Remove and clamp on negative, from a local variable.
2739 * A variant of sub_positive(), which does not use explicit load-store
2740 * and is thus optimized for local variable updates.
2742 #define lsub_positive(_ptr, _val) do { \
2743 typeof(_ptr) ptr = (_ptr); \
2744 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
2749 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2751 cfs_rq->runnable_weight += se->runnable_weight;
2753 cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2754 cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2758 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2760 cfs_rq->runnable_weight -= se->runnable_weight;
2762 sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2763 sub_positive(&cfs_rq->avg.runnable_load_sum,
2764 se_runnable(se) * se->avg.runnable_load_sum);
2768 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2770 cfs_rq->avg.load_avg += se->avg.load_avg;
2771 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2775 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2777 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2778 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2782 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2784 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2786 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2788 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2791 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2792 unsigned long weight, unsigned long runnable)
2795 /* commit outstanding execution time */
2796 if (cfs_rq->curr == se)
2797 update_curr(cfs_rq);
2798 account_entity_dequeue(cfs_rq, se);
2799 dequeue_runnable_load_avg(cfs_rq, se);
2801 dequeue_load_avg(cfs_rq, se);
2803 se->runnable_weight = runnable;
2804 update_load_set(&se->load, weight);
2808 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2810 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2811 se->avg.runnable_load_avg =
2812 div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2816 enqueue_load_avg(cfs_rq, se);
2818 account_entity_enqueue(cfs_rq, se);
2819 enqueue_runnable_load_avg(cfs_rq, se);
2823 void reweight_task(struct task_struct *p, int prio)
2825 struct sched_entity *se = &p->se;
2826 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2827 struct load_weight *load = &se->load;
2828 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2830 reweight_entity(cfs_rq, se, weight, weight);
2831 load->inv_weight = sched_prio_to_wmult[prio];
2834 #ifdef CONFIG_FAIR_GROUP_SCHED
2837 * All this does is approximate the hierarchical proportion which includes that
2838 * global sum we all love to hate.
2840 * That is, the weight of a group entity, is the proportional share of the
2841 * group weight based on the group runqueue weights. That is:
2843 * tg->weight * grq->load.weight
2844 * ge->load.weight = ----------------------------- (1)
2845 * \Sum grq->load.weight
2847 * Now, because computing that sum is prohibitively expensive to compute (been
2848 * there, done that) we approximate it with this average stuff. The average
2849 * moves slower and therefore the approximation is cheaper and more stable.
2851 * So instead of the above, we substitute:
2853 * grq->load.weight -> grq->avg.load_avg (2)
2855 * which yields the following:
2857 * tg->weight * grq->avg.load_avg
2858 * ge->load.weight = ------------------------------ (3)
2861 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2863 * That is shares_avg, and it is right (given the approximation (2)).
2865 * The problem with it is that because the average is slow -- it was designed
2866 * to be exactly that of course -- this leads to transients in boundary
2867 * conditions. In specific, the case where the group was idle and we start the
2868 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2869 * yielding bad latency etc..
2871 * Now, in that special case (1) reduces to:
2873 * tg->weight * grq->load.weight
2874 * ge->load.weight = ----------------------------- = tg->weight (4)
2877 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2879 * So what we do is modify our approximation (3) to approach (4) in the (near)
2884 * tg->weight * grq->load.weight
2885 * --------------------------------------------------- (5)
2886 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2888 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2889 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2892 * tg->weight * grq->load.weight
2893 * ge->load.weight = ----------------------------- (6)
2898 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2899 * max(grq->load.weight, grq->avg.load_avg)
2901 * And that is shares_weight and is icky. In the (near) UP case it approaches
2902 * (4) while in the normal case it approaches (3). It consistently
2903 * overestimates the ge->load.weight and therefore:
2905 * \Sum ge->load.weight >= tg->weight
2909 static long calc_group_shares(struct cfs_rq *cfs_rq)
2911 long tg_weight, tg_shares, load, shares;
2912 struct task_group *tg = cfs_rq->tg;
2914 tg_shares = READ_ONCE(tg->shares);
2916 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
2918 tg_weight = atomic_long_read(&tg->load_avg);
2920 /* Ensure tg_weight >= load */
2921 tg_weight -= cfs_rq->tg_load_avg_contrib;
2924 shares = (tg_shares * load);
2926 shares /= tg_weight;
2929 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2930 * of a group with small tg->shares value. It is a floor value which is
2931 * assigned as a minimum load.weight to the sched_entity representing
2932 * the group on a CPU.
2934 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2935 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2936 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2937 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2940 return clamp_t(long, shares, MIN_SHARES, tg_shares);
2944 * This calculates the effective runnable weight for a group entity based on
2945 * the group entity weight calculated above.
2947 * Because of the above approximation (2), our group entity weight is
2948 * an load_avg based ratio (3). This means that it includes blocked load and
2949 * does not represent the runnable weight.
2951 * Approximate the group entity's runnable weight per ratio from the group
2954 * grq->avg.runnable_load_avg
2955 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
2958 * However, analogous to above, since the avg numbers are slow, this leads to
2959 * transients in the from-idle case. Instead we use:
2961 * ge->runnable_weight = ge->load.weight *
2963 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
2964 * ----------------------------------------------------- (8)
2965 * max(grq->avg.load_avg, grq->load.weight)
2967 * Where these max() serve both to use the 'instant' values to fix the slow
2968 * from-idle and avoid the /0 on to-idle, similar to (6).
2970 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
2972 long runnable, load_avg;
2974 load_avg = max(cfs_rq->avg.load_avg,
2975 scale_load_down(cfs_rq->load.weight));
2977 runnable = max(cfs_rq->avg.runnable_load_avg,
2978 scale_load_down(cfs_rq->runnable_weight));
2982 runnable /= load_avg;
2984 return clamp_t(long, runnable, MIN_SHARES, shares);
2986 #endif /* CONFIG_SMP */
2988 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2991 * Recomputes the group entity based on the current state of its group
2994 static void update_cfs_group(struct sched_entity *se)
2996 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
2997 long shares, runnable;
3002 if (throttled_hierarchy(gcfs_rq))
3006 runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
3008 if (likely(se->load.weight == shares))
3011 shares = calc_group_shares(gcfs_rq);
3012 runnable = calc_group_runnable(gcfs_rq, shares);
3015 reweight_entity(cfs_rq_of(se), se, shares, runnable);
3018 #else /* CONFIG_FAIR_GROUP_SCHED */
3019 static inline void update_cfs_group(struct sched_entity *se)
3022 #endif /* CONFIG_FAIR_GROUP_SCHED */
3024 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3026 struct rq *rq = rq_of(cfs_rq);
3028 if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) {
3030 * There are a few boundary cases this might miss but it should
3031 * get called often enough that that should (hopefully) not be
3034 * It will not get called when we go idle, because the idle
3035 * thread is a different class (!fair), nor will the utilization
3036 * number include things like RT tasks.
3038 * As is, the util number is not freq-invariant (we'd have to
3039 * implement arch_scale_freq_capacity() for that).
3043 cpufreq_update_util(rq, flags);
3048 #ifdef CONFIG_FAIR_GROUP_SCHED
3050 * update_tg_load_avg - update the tg's load avg
3051 * @cfs_rq: the cfs_rq whose avg changed
3052 * @force: update regardless of how small the difference
3054 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3055 * However, because tg->load_avg is a global value there are performance
3058 * In order to avoid having to look at the other cfs_rq's, we use a
3059 * differential update where we store the last value we propagated. This in
3060 * turn allows skipping updates if the differential is 'small'.
3062 * Updating tg's load_avg is necessary before update_cfs_share().
3064 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3066 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3069 * No need to update load_avg for root_task_group as it is not used.
3071 if (cfs_rq->tg == &root_task_group)
3074 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3075 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3076 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3081 * Called within set_task_rq() right before setting a task's CPU. The
3082 * caller only guarantees p->pi_lock is held; no other assumptions,
3083 * including the state of rq->lock, should be made.
3085 void set_task_rq_fair(struct sched_entity *se,
3086 struct cfs_rq *prev, struct cfs_rq *next)
3088 u64 p_last_update_time;
3089 u64 n_last_update_time;
3091 if (!sched_feat(ATTACH_AGE_LOAD))
3095 * We are supposed to update the task to "current" time, then its up to
3096 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3097 * getting what current time is, so simply throw away the out-of-date
3098 * time. This will result in the wakee task is less decayed, but giving
3099 * the wakee more load sounds not bad.
3101 if (!(se->avg.last_update_time && prev))
3104 #ifndef CONFIG_64BIT
3106 u64 p_last_update_time_copy;
3107 u64 n_last_update_time_copy;
3110 p_last_update_time_copy = prev->load_last_update_time_copy;
3111 n_last_update_time_copy = next->load_last_update_time_copy;
3115 p_last_update_time = prev->avg.last_update_time;
3116 n_last_update_time = next->avg.last_update_time;
3118 } while (p_last_update_time != p_last_update_time_copy ||
3119 n_last_update_time != n_last_update_time_copy);
3122 p_last_update_time = prev->avg.last_update_time;
3123 n_last_update_time = next->avg.last_update_time;
3125 __update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se);
3126 se->avg.last_update_time = n_last_update_time;
3131 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3132 * propagate its contribution. The key to this propagation is the invariant
3133 * that for each group:
3135 * ge->avg == grq->avg (1)
3137 * _IFF_ we look at the pure running and runnable sums. Because they
3138 * represent the very same entity, just at different points in the hierarchy.
3140 * Per the above update_tg_cfs_util() is trivial and simply copies the running
3141 * sum over (but still wrong, because the group entity and group rq do not have
3142 * their PELT windows aligned).
3144 * However, update_tg_cfs_runnable() is more complex. So we have:
3146 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3148 * And since, like util, the runnable part should be directly transferable,
3149 * the following would _appear_ to be the straight forward approach:
3151 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3153 * And per (1) we have:
3155 * ge->avg.runnable_avg == grq->avg.runnable_avg
3159 * ge->load.weight * grq->avg.load_avg
3160 * ge->avg.load_avg = ----------------------------------- (4)
3163 * Except that is wrong!
3165 * Because while for entities historical weight is not important and we
3166 * really only care about our future and therefore can consider a pure
3167 * runnable sum, runqueues can NOT do this.
3169 * We specifically want runqueues to have a load_avg that includes
3170 * historical weights. Those represent the blocked load, the load we expect
3171 * to (shortly) return to us. This only works by keeping the weights as
3172 * integral part of the sum. We therefore cannot decompose as per (3).
3174 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3175 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3176 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3177 * runnable section of these tasks overlap (or not). If they were to perfectly
3178 * align the rq as a whole would be runnable 2/3 of the time. If however we
3179 * always have at least 1 runnable task, the rq as a whole is always runnable.
3181 * So we'll have to approximate.. :/
3183 * Given the constraint:
3185 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3187 * We can construct a rule that adds runnable to a rq by assuming minimal
3190 * On removal, we'll assume each task is equally runnable; which yields:
3192 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3194 * XXX: only do this for the part of runnable > running ?
3199 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3201 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3203 /* Nothing to update */
3208 * The relation between sum and avg is:
3210 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3212 * however, the PELT windows are not aligned between grq and gse.
3215 /* Set new sched_entity's utilization */
3216 se->avg.util_avg = gcfs_rq->avg.util_avg;
3217 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3219 /* Update parent cfs_rq utilization */
3220 add_positive(&cfs_rq->avg.util_avg, delta);
3221 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3225 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3227 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3228 unsigned long runnable_load_avg, load_avg;
3229 u64 runnable_load_sum, load_sum = 0;
3235 gcfs_rq->prop_runnable_sum = 0;
3237 if (runnable_sum >= 0) {
3239 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3240 * the CPU is saturated running == runnable.
3242 runnable_sum += se->avg.load_sum;
3243 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3246 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3247 * assuming all tasks are equally runnable.
3249 if (scale_load_down(gcfs_rq->load.weight)) {
3250 load_sum = div_s64(gcfs_rq->avg.load_sum,
3251 scale_load_down(gcfs_rq->load.weight));
3254 /* But make sure to not inflate se's runnable */
3255 runnable_sum = min(se->avg.load_sum, load_sum);
3259 * runnable_sum can't be lower than running_sum
3260 * As running sum is scale with CPU capacity wehreas the runnable sum
3261 * is not we rescale running_sum 1st
3263 running_sum = se->avg.util_sum /
3264 arch_scale_cpu_capacity(NULL, cpu_of(rq_of(cfs_rq)));
3265 runnable_sum = max(runnable_sum, running_sum);
3267 load_sum = (s64)se_weight(se) * runnable_sum;
3268 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3270 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3271 delta_avg = load_avg - se->avg.load_avg;
3273 se->avg.load_sum = runnable_sum;
3274 se->avg.load_avg = load_avg;
3275 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3276 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3278 runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3279 runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3280 delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3281 delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3283 se->avg.runnable_load_sum = runnable_sum;
3284 se->avg.runnable_load_avg = runnable_load_avg;
3287 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3288 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3292 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3294 cfs_rq->propagate = 1;
3295 cfs_rq->prop_runnable_sum += runnable_sum;
3298 /* Update task and its cfs_rq load average */
3299 static inline int propagate_entity_load_avg(struct sched_entity *se)
3301 struct cfs_rq *cfs_rq, *gcfs_rq;
3303 if (entity_is_task(se))
3306 gcfs_rq = group_cfs_rq(se);
3307 if (!gcfs_rq->propagate)
3310 gcfs_rq->propagate = 0;
3312 cfs_rq = cfs_rq_of(se);
3314 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3316 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3317 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3323 * Check if we need to update the load and the utilization of a blocked
3326 static inline bool skip_blocked_update(struct sched_entity *se)
3328 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3331 * If sched_entity still have not zero load or utilization, we have to
3334 if (se->avg.load_avg || se->avg.util_avg)
3338 * If there is a pending propagation, we have to update the load and
3339 * the utilization of the sched_entity:
3341 if (gcfs_rq->propagate)
3345 * Otherwise, the load and the utilization of the sched_entity is
3346 * already zero and there is no pending propagation, so it will be a
3347 * waste of time to try to decay it:
3352 #else /* CONFIG_FAIR_GROUP_SCHED */
3354 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3356 static inline int propagate_entity_load_avg(struct sched_entity *se)
3361 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3363 #endif /* CONFIG_FAIR_GROUP_SCHED */
3366 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3367 * @now: current time, as per cfs_rq_clock_task()
3368 * @cfs_rq: cfs_rq to update
3370 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3371 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3372 * post_init_entity_util_avg().
3374 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3376 * Returns true if the load decayed or we removed load.
3378 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3379 * call update_tg_load_avg() when this function returns true.
3382 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3384 unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3385 struct sched_avg *sa = &cfs_rq->avg;
3388 if (cfs_rq->removed.nr) {
3390 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3392 raw_spin_lock(&cfs_rq->removed.lock);
3393 swap(cfs_rq->removed.util_avg, removed_util);
3394 swap(cfs_rq->removed.load_avg, removed_load);
3395 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3396 cfs_rq->removed.nr = 0;
3397 raw_spin_unlock(&cfs_rq->removed.lock);
3400 sub_positive(&sa->load_avg, r);
3401 sub_positive(&sa->load_sum, r * divider);
3404 sub_positive(&sa->util_avg, r);
3405 sub_positive(&sa->util_sum, r * divider);
3407 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3412 decayed |= __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq);
3414 #ifndef CONFIG_64BIT
3416 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3420 cfs_rq_util_change(cfs_rq, 0);
3426 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3427 * @cfs_rq: cfs_rq to attach to
3428 * @se: sched_entity to attach
3429 * @flags: migration hints
3431 * Must call update_cfs_rq_load_avg() before this, since we rely on
3432 * cfs_rq->avg.last_update_time being current.
3434 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3436 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3439 * When we attach the @se to the @cfs_rq, we must align the decay
3440 * window because without that, really weird and wonderful things can
3445 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3446 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3449 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3450 * period_contrib. This isn't strictly correct, but since we're
3451 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3454 se->avg.util_sum = se->avg.util_avg * divider;
3456 se->avg.load_sum = divider;
3457 if (se_weight(se)) {
3459 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3462 se->avg.runnable_load_sum = se->avg.load_sum;
3464 enqueue_load_avg(cfs_rq, se);
3465 cfs_rq->avg.util_avg += se->avg.util_avg;
3466 cfs_rq->avg.util_sum += se->avg.util_sum;
3468 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3470 cfs_rq_util_change(cfs_rq, flags);
3474 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3475 * @cfs_rq: cfs_rq to detach from
3476 * @se: sched_entity to detach
3478 * Must call update_cfs_rq_load_avg() before this, since we rely on
3479 * cfs_rq->avg.last_update_time being current.
3481 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3483 dequeue_load_avg(cfs_rq, se);
3484 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3485 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3487 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3489 cfs_rq_util_change(cfs_rq, 0);
3493 * Optional action to be done while updating the load average
3495 #define UPDATE_TG 0x1
3496 #define SKIP_AGE_LOAD 0x2
3497 #define DO_ATTACH 0x4
3499 /* Update task and its cfs_rq load average */
3500 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3502 u64 now = cfs_rq_clock_task(cfs_rq);
3503 struct rq *rq = rq_of(cfs_rq);
3504 int cpu = cpu_of(rq);
3508 * Track task load average for carrying it to new CPU after migrated, and
3509 * track group sched_entity load average for task_h_load calc in migration
3511 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3512 __update_load_avg_se(now, cpu, cfs_rq, se);
3514 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3515 decayed |= propagate_entity_load_avg(se);
3517 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3520 * DO_ATTACH means we're here from enqueue_entity().
3521 * !last_update_time means we've passed through
3522 * migrate_task_rq_fair() indicating we migrated.
3524 * IOW we're enqueueing a task on a new CPU.
3526 attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
3527 update_tg_load_avg(cfs_rq, 0);
3529 } else if (decayed && (flags & UPDATE_TG))
3530 update_tg_load_avg(cfs_rq, 0);
3533 #ifndef CONFIG_64BIT
3534 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3536 u64 last_update_time_copy;
3537 u64 last_update_time;
3540 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3542 last_update_time = cfs_rq->avg.last_update_time;
3543 } while (last_update_time != last_update_time_copy);
3545 return last_update_time;
3548 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3550 return cfs_rq->avg.last_update_time;
3555 * Synchronize entity load avg of dequeued entity without locking
3558 void sync_entity_load_avg(struct sched_entity *se)
3560 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3561 u64 last_update_time;
3563 last_update_time = cfs_rq_last_update_time(cfs_rq);
3564 __update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se);
3568 * Task first catches up with cfs_rq, and then subtract
3569 * itself from the cfs_rq (task must be off the queue now).
3571 void remove_entity_load_avg(struct sched_entity *se)
3573 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3574 unsigned long flags;
3577 * tasks cannot exit without having gone through wake_up_new_task() ->
3578 * post_init_entity_util_avg() which will have added things to the
3579 * cfs_rq, so we can remove unconditionally.
3581 * Similarly for groups, they will have passed through
3582 * post_init_entity_util_avg() before unregister_sched_fair_group()
3586 sync_entity_load_avg(se);
3588 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3589 ++cfs_rq->removed.nr;
3590 cfs_rq->removed.util_avg += se->avg.util_avg;
3591 cfs_rq->removed.load_avg += se->avg.load_avg;
3592 cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */
3593 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3596 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3598 return cfs_rq->avg.runnable_load_avg;
3601 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3603 return cfs_rq->avg.load_avg;
3606 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3608 static inline unsigned long task_util(struct task_struct *p)
3610 return READ_ONCE(p->se.avg.util_avg);
3613 static inline unsigned long _task_util_est(struct task_struct *p)
3615 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3617 return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED);
3620 static inline unsigned long task_util_est(struct task_struct *p)
3622 return max(task_util(p), _task_util_est(p));
3625 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3626 struct task_struct *p)
3628 unsigned int enqueued;
3630 if (!sched_feat(UTIL_EST))
3633 /* Update root cfs_rq's estimated utilization */
3634 enqueued = cfs_rq->avg.util_est.enqueued;
3635 enqueued += _task_util_est(p);
3636 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3640 * Check if a (signed) value is within a specified (unsigned) margin,
3641 * based on the observation that:
3643 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3645 * NOTE: this only works when value + maring < INT_MAX.
3647 static inline bool within_margin(int value, int margin)
3649 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3653 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3655 long last_ewma_diff;
3658 if (!sched_feat(UTIL_EST))
3661 /* Update root cfs_rq's estimated utilization */
3662 ue.enqueued = cfs_rq->avg.util_est.enqueued;
3663 ue.enqueued -= min_t(unsigned int, ue.enqueued, _task_util_est(p));
3664 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3667 * Skip update of task's estimated utilization when the task has not
3668 * yet completed an activation, e.g. being migrated.
3674 * If the PELT values haven't changed since enqueue time,
3675 * skip the util_est update.
3677 ue = p->se.avg.util_est;
3678 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3682 * Skip update of task's estimated utilization when its EWMA is
3683 * already ~1% close to its last activation value.
3685 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3686 last_ewma_diff = ue.enqueued - ue.ewma;
3687 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3691 * Update Task's estimated utilization
3693 * When *p completes an activation we can consolidate another sample
3694 * of the task size. This is done by storing the current PELT value
3695 * as ue.enqueued and by using this value to update the Exponential
3696 * Weighted Moving Average (EWMA):
3698 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
3699 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
3700 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
3701 * = w * ( last_ewma_diff ) + ewma(t-1)
3702 * = w * (last_ewma_diff + ewma(t-1) / w)
3704 * Where 'w' is the weight of new samples, which is configured to be
3705 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3707 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
3708 ue.ewma += last_ewma_diff;
3709 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
3710 WRITE_ONCE(p->se.avg.util_est, ue);
3713 static inline int task_fits_capacity(struct task_struct *p, long capacity)
3715 return capacity * 1024 > task_util_est(p) * capacity_margin;
3718 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
3720 if (!static_branch_unlikely(&sched_asym_cpucapacity))
3724 rq->misfit_task_load = 0;
3728 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
3729 rq->misfit_task_load = 0;
3733 rq->misfit_task_load = task_h_load(p);
3736 #else /* CONFIG_SMP */
3738 #define UPDATE_TG 0x0
3739 #define SKIP_AGE_LOAD 0x0
3740 #define DO_ATTACH 0x0
3742 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3744 cfs_rq_util_change(cfs_rq, 0);
3747 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3750 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
3752 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3754 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3760 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
3763 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
3765 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
3767 #endif /* CONFIG_SMP */
3769 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3771 #ifdef CONFIG_SCHED_DEBUG
3772 s64 d = se->vruntime - cfs_rq->min_vruntime;
3777 if (d > 3*sysctl_sched_latency)
3778 schedstat_inc(cfs_rq->nr_spread_over);
3783 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3785 u64 vruntime = cfs_rq->min_vruntime;
3788 * The 'current' period is already promised to the current tasks,
3789 * however the extra weight of the new task will slow them down a
3790 * little, place the new task so that it fits in the slot that
3791 * stays open at the end.
3793 if (initial && sched_feat(START_DEBIT))
3794 vruntime += sched_vslice(cfs_rq, se);
3796 /* sleeps up to a single latency don't count. */
3798 unsigned long thresh = sysctl_sched_latency;
3801 * Halve their sleep time's effect, to allow
3802 * for a gentler effect of sleepers:
3804 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3810 /* ensure we never gain time by being placed backwards. */
3811 se->vruntime = max_vruntime(se->vruntime, vruntime);
3814 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3816 static inline void check_schedstat_required(void)
3818 #ifdef CONFIG_SCHEDSTATS
3819 if (schedstat_enabled())
3822 /* Force schedstat enabled if a dependent tracepoint is active */
3823 if (trace_sched_stat_wait_enabled() ||
3824 trace_sched_stat_sleep_enabled() ||
3825 trace_sched_stat_iowait_enabled() ||
3826 trace_sched_stat_blocked_enabled() ||
3827 trace_sched_stat_runtime_enabled()) {
3828 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3829 "stat_blocked and stat_runtime require the "
3830 "kernel parameter schedstats=enable or "
3831 "kernel.sched_schedstats=1\n");
3842 * update_min_vruntime()
3843 * vruntime -= min_vruntime
3847 * update_min_vruntime()
3848 * vruntime += min_vruntime
3850 * this way the vruntime transition between RQs is done when both
3851 * min_vruntime are up-to-date.
3855 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3856 * vruntime -= min_vruntime
3860 * update_min_vruntime()
3861 * vruntime += min_vruntime
3863 * this way we don't have the most up-to-date min_vruntime on the originating
3864 * CPU and an up-to-date min_vruntime on the destination CPU.
3868 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3870 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3871 bool curr = cfs_rq->curr == se;
3874 * If we're the current task, we must renormalise before calling
3878 se->vruntime += cfs_rq->min_vruntime;
3880 update_curr(cfs_rq);
3883 * Otherwise, renormalise after, such that we're placed at the current
3884 * moment in time, instead of some random moment in the past. Being
3885 * placed in the past could significantly boost this task to the
3886 * fairness detriment of existing tasks.
3888 if (renorm && !curr)
3889 se->vruntime += cfs_rq->min_vruntime;
3892 * When enqueuing a sched_entity, we must:
3893 * - Update loads to have both entity and cfs_rq synced with now.
3894 * - Add its load to cfs_rq->runnable_avg
3895 * - For group_entity, update its weight to reflect the new share of
3897 * - Add its new weight to cfs_rq->load.weight
3899 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
3900 update_cfs_group(se);
3901 enqueue_runnable_load_avg(cfs_rq, se);
3902 account_entity_enqueue(cfs_rq, se);
3904 if (flags & ENQUEUE_WAKEUP)
3905 place_entity(cfs_rq, se, 0);
3907 check_schedstat_required();
3908 update_stats_enqueue(cfs_rq, se, flags);
3909 check_spread(cfs_rq, se);
3911 __enqueue_entity(cfs_rq, se);
3914 if (cfs_rq->nr_running == 1) {
3915 list_add_leaf_cfs_rq(cfs_rq);
3916 check_enqueue_throttle(cfs_rq);
3920 static void __clear_buddies_last(struct sched_entity *se)
3922 for_each_sched_entity(se) {
3923 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3924 if (cfs_rq->last != se)
3927 cfs_rq->last = NULL;
3931 static void __clear_buddies_next(struct sched_entity *se)
3933 for_each_sched_entity(se) {
3934 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3935 if (cfs_rq->next != se)
3938 cfs_rq->next = NULL;
3942 static void __clear_buddies_skip(struct sched_entity *se)
3944 for_each_sched_entity(se) {
3945 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3946 if (cfs_rq->skip != se)
3949 cfs_rq->skip = NULL;
3953 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3955 if (cfs_rq->last == se)
3956 __clear_buddies_last(se);
3958 if (cfs_rq->next == se)
3959 __clear_buddies_next(se);
3961 if (cfs_rq->skip == se)
3962 __clear_buddies_skip(se);
3965 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3968 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3971 * Update run-time statistics of the 'current'.
3973 update_curr(cfs_rq);
3976 * When dequeuing a sched_entity, we must:
3977 * - Update loads to have both entity and cfs_rq synced with now.
3978 * - Subtract its load from the cfs_rq->runnable_avg.
3979 * - Subtract its previous weight from cfs_rq->load.weight.
3980 * - For group entity, update its weight to reflect the new share
3981 * of its group cfs_rq.
3983 update_load_avg(cfs_rq, se, UPDATE_TG);
3984 dequeue_runnable_load_avg(cfs_rq, se);
3986 update_stats_dequeue(cfs_rq, se, flags);
3988 clear_buddies(cfs_rq, se);
3990 if (se != cfs_rq->curr)
3991 __dequeue_entity(cfs_rq, se);
3993 account_entity_dequeue(cfs_rq, se);
3996 * Normalize after update_curr(); which will also have moved
3997 * min_vruntime if @se is the one holding it back. But before doing
3998 * update_min_vruntime() again, which will discount @se's position and
3999 * can move min_vruntime forward still more.
4001 if (!(flags & DEQUEUE_SLEEP))
4002 se->vruntime -= cfs_rq->min_vruntime;
4004 /* return excess runtime on last dequeue */
4005 return_cfs_rq_runtime(cfs_rq);
4007 update_cfs_group(se);
4010 * Now advance min_vruntime if @se was the entity holding it back,
4011 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4012 * put back on, and if we advance min_vruntime, we'll be placed back
4013 * further than we started -- ie. we'll be penalized.
4015 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4016 update_min_vruntime(cfs_rq);
4020 * Preempt the current task with a newly woken task if needed:
4023 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4025 unsigned long ideal_runtime, delta_exec;
4026 struct sched_entity *se;
4029 ideal_runtime = sched_slice(cfs_rq, curr);
4030 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4031 if (delta_exec > ideal_runtime) {
4032 resched_curr(rq_of(cfs_rq));
4034 * The current task ran long enough, ensure it doesn't get
4035 * re-elected due to buddy favours.
4037 clear_buddies(cfs_rq, curr);
4042 * Ensure that a task that missed wakeup preemption by a
4043 * narrow margin doesn't have to wait for a full slice.
4044 * This also mitigates buddy induced latencies under load.
4046 if (delta_exec < sysctl_sched_min_granularity)
4049 se = __pick_first_entity(cfs_rq);
4050 delta = curr->vruntime - se->vruntime;
4055 if (delta > ideal_runtime)
4056 resched_curr(rq_of(cfs_rq));
4060 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4062 /* 'current' is not kept within the tree. */
4065 * Any task has to be enqueued before it get to execute on
4066 * a CPU. So account for the time it spent waiting on the
4069 update_stats_wait_end(cfs_rq, se);
4070 __dequeue_entity(cfs_rq, se);
4071 update_load_avg(cfs_rq, se, UPDATE_TG);
4074 update_stats_curr_start(cfs_rq, se);
4078 * Track our maximum slice length, if the CPU's load is at
4079 * least twice that of our own weight (i.e. dont track it
4080 * when there are only lesser-weight tasks around):
4082 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
4083 schedstat_set(se->statistics.slice_max,
4084 max((u64)schedstat_val(se->statistics.slice_max),
4085 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4088 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4092 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4095 * Pick the next process, keeping these things in mind, in this order:
4096 * 1) keep things fair between processes/task groups
4097 * 2) pick the "next" process, since someone really wants that to run
4098 * 3) pick the "last" process, for cache locality
4099 * 4) do not run the "skip" process, if something else is available
4101 static struct sched_entity *
4102 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4104 struct sched_entity *left = __pick_first_entity(cfs_rq);
4105 struct sched_entity *se;
4108 * If curr is set we have to see if its left of the leftmost entity
4109 * still in the tree, provided there was anything in the tree at all.
4111 if (!left || (curr && entity_before(curr, left)))
4114 se = left; /* ideally we run the leftmost entity */
4117 * Avoid running the skip buddy, if running something else can
4118 * be done without getting too unfair.
4120 if (cfs_rq->skip == se) {
4121 struct sched_entity *second;
4124 second = __pick_first_entity(cfs_rq);
4126 second = __pick_next_entity(se);
4127 if (!second || (curr && entity_before(curr, second)))
4131 if (second && wakeup_preempt_entity(second, left) < 1)
4136 * Prefer last buddy, try to return the CPU to a preempted task.
4138 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4142 * Someone really wants this to run. If it's not unfair, run it.
4144 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4147 clear_buddies(cfs_rq, se);
4152 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4154 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4157 * If still on the runqueue then deactivate_task()
4158 * was not called and update_curr() has to be done:
4161 update_curr(cfs_rq);
4163 /* throttle cfs_rqs exceeding runtime */
4164 check_cfs_rq_runtime(cfs_rq);
4166 check_spread(cfs_rq, prev);
4169 update_stats_wait_start(cfs_rq, prev);
4170 /* Put 'current' back into the tree. */
4171 __enqueue_entity(cfs_rq, prev);
4172 /* in !on_rq case, update occurred at dequeue */
4173 update_load_avg(cfs_rq, prev, 0);
4175 cfs_rq->curr = NULL;
4179 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4182 * Update run-time statistics of the 'current'.
4184 update_curr(cfs_rq);
4187 * Ensure that runnable average is periodically updated.
4189 update_load_avg(cfs_rq, curr, UPDATE_TG);
4190 update_cfs_group(curr);
4192 #ifdef CONFIG_SCHED_HRTICK
4194 * queued ticks are scheduled to match the slice, so don't bother
4195 * validating it and just reschedule.
4198 resched_curr(rq_of(cfs_rq));
4202 * don't let the period tick interfere with the hrtick preemption
4204 if (!sched_feat(DOUBLE_TICK) &&
4205 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4209 if (cfs_rq->nr_running > 1)
4210 check_preempt_tick(cfs_rq, curr);
4214 /**************************************************
4215 * CFS bandwidth control machinery
4218 #ifdef CONFIG_CFS_BANDWIDTH
4220 #ifdef CONFIG_JUMP_LABEL
4221 static struct static_key __cfs_bandwidth_used;
4223 static inline bool cfs_bandwidth_used(void)
4225 return static_key_false(&__cfs_bandwidth_used);
4228 void cfs_bandwidth_usage_inc(void)
4230 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4233 void cfs_bandwidth_usage_dec(void)
4235 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4237 #else /* CONFIG_JUMP_LABEL */
4238 static bool cfs_bandwidth_used(void)
4243 void cfs_bandwidth_usage_inc(void) {}
4244 void cfs_bandwidth_usage_dec(void) {}
4245 #endif /* CONFIG_JUMP_LABEL */
4248 * default period for cfs group bandwidth.
4249 * default: 0.1s, units: nanoseconds
4251 static inline u64 default_cfs_period(void)
4253 return 100000000ULL;
4256 static inline u64 sched_cfs_bandwidth_slice(void)
4258 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4262 * Replenish runtime according to assigned quota and update expiration time.
4263 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4264 * additional synchronization around rq->lock.
4266 * requires cfs_b->lock
4268 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4272 if (cfs_b->quota == RUNTIME_INF)
4275 now = sched_clock_cpu(smp_processor_id());
4276 cfs_b->runtime = cfs_b->quota;
4277 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4278 cfs_b->expires_seq++;
4281 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4283 return &tg->cfs_bandwidth;
4286 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4287 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4289 if (unlikely(cfs_rq->throttle_count))
4290 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4292 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4295 /* returns 0 on failure to allocate runtime */
4296 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4298 struct task_group *tg = cfs_rq->tg;
4299 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4300 u64 amount = 0, min_amount, expires;
4303 /* note: this is a positive sum as runtime_remaining <= 0 */
4304 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4306 raw_spin_lock(&cfs_b->lock);
4307 if (cfs_b->quota == RUNTIME_INF)
4308 amount = min_amount;
4310 start_cfs_bandwidth(cfs_b);
4312 if (cfs_b->runtime > 0) {
4313 amount = min(cfs_b->runtime, min_amount);
4314 cfs_b->runtime -= amount;
4318 expires_seq = cfs_b->expires_seq;
4319 expires = cfs_b->runtime_expires;
4320 raw_spin_unlock(&cfs_b->lock);
4322 cfs_rq->runtime_remaining += amount;
4324 * we may have advanced our local expiration to account for allowed
4325 * spread between our sched_clock and the one on which runtime was
4328 if (cfs_rq->expires_seq != expires_seq) {
4329 cfs_rq->expires_seq = expires_seq;
4330 cfs_rq->runtime_expires = expires;
4333 return cfs_rq->runtime_remaining > 0;
4337 * Note: This depends on the synchronization provided by sched_clock and the
4338 * fact that rq->clock snapshots this value.
4340 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4342 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4344 /* if the deadline is ahead of our clock, nothing to do */
4345 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4348 if (cfs_rq->runtime_remaining < 0)
4352 * If the local deadline has passed we have to consider the
4353 * possibility that our sched_clock is 'fast' and the global deadline
4354 * has not truly expired.
4356 * Fortunately we can check determine whether this the case by checking
4357 * whether the global deadline(cfs_b->expires_seq) has advanced.
4359 if (cfs_rq->expires_seq == cfs_b->expires_seq) {
4360 /* extend local deadline, drift is bounded above by 2 ticks */
4361 cfs_rq->runtime_expires += TICK_NSEC;
4363 /* global deadline is ahead, expiration has passed */
4364 cfs_rq->runtime_remaining = 0;
4368 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4370 /* dock delta_exec before expiring quota (as it could span periods) */
4371 cfs_rq->runtime_remaining -= delta_exec;
4372 expire_cfs_rq_runtime(cfs_rq);
4374 if (likely(cfs_rq->runtime_remaining > 0))
4378 * if we're unable to extend our runtime we resched so that the active
4379 * hierarchy can be throttled
4381 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4382 resched_curr(rq_of(cfs_rq));
4385 static __always_inline
4386 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4388 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4391 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4394 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4396 return cfs_bandwidth_used() && cfs_rq->throttled;
4399 /* check whether cfs_rq, or any parent, is throttled */
4400 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4402 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4406 * Ensure that neither of the group entities corresponding to src_cpu or
4407 * dest_cpu are members of a throttled hierarchy when performing group
4408 * load-balance operations.
4410 static inline int throttled_lb_pair(struct task_group *tg,
4411 int src_cpu, int dest_cpu)
4413 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4415 src_cfs_rq = tg->cfs_rq[src_cpu];
4416 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4418 return throttled_hierarchy(src_cfs_rq) ||
4419 throttled_hierarchy(dest_cfs_rq);
4422 static int tg_unthrottle_up(struct task_group *tg, void *data)
4424 struct rq *rq = data;
4425 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4427 cfs_rq->throttle_count--;
4428 if (!cfs_rq->throttle_count) {
4429 /* adjust cfs_rq_clock_task() */
4430 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4431 cfs_rq->throttled_clock_task;
4437 static int tg_throttle_down(struct task_group *tg, void *data)
4439 struct rq *rq = data;
4440 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4442 /* group is entering throttled state, stop time */
4443 if (!cfs_rq->throttle_count)
4444 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4445 cfs_rq->throttle_count++;
4450 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4452 struct rq *rq = rq_of(cfs_rq);
4453 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4454 struct sched_entity *se;
4455 long task_delta, dequeue = 1;
4458 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4460 /* freeze hierarchy runnable averages while throttled */
4462 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4465 task_delta = cfs_rq->h_nr_running;
4466 for_each_sched_entity(se) {
4467 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4468 /* throttled entity or throttle-on-deactivate */
4473 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4474 qcfs_rq->h_nr_running -= task_delta;
4476 if (qcfs_rq->load.weight)
4481 sub_nr_running(rq, task_delta);
4483 cfs_rq->throttled = 1;
4484 cfs_rq->throttled_clock = rq_clock(rq);
4485 raw_spin_lock(&cfs_b->lock);
4486 empty = list_empty(&cfs_b->throttled_cfs_rq);
4489 * Add to the _head_ of the list, so that an already-started
4490 * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
4491 * not running add to the tail so that later runqueues don't get starved.
4493 if (cfs_b->distribute_running)
4494 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4496 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4499 * If we're the first throttled task, make sure the bandwidth
4503 start_cfs_bandwidth(cfs_b);
4505 raw_spin_unlock(&cfs_b->lock);
4508 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4510 struct rq *rq = rq_of(cfs_rq);
4511 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4512 struct sched_entity *se;
4516 se = cfs_rq->tg->se[cpu_of(rq)];
4518 cfs_rq->throttled = 0;
4520 update_rq_clock(rq);
4522 raw_spin_lock(&cfs_b->lock);
4523 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4524 list_del_rcu(&cfs_rq->throttled_list);
4525 raw_spin_unlock(&cfs_b->lock);
4527 /* update hierarchical throttle state */
4528 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4530 if (!cfs_rq->load.weight)
4533 task_delta = cfs_rq->h_nr_running;
4534 for_each_sched_entity(se) {
4538 cfs_rq = cfs_rq_of(se);
4540 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4541 cfs_rq->h_nr_running += task_delta;
4543 if (cfs_rq_throttled(cfs_rq))
4548 add_nr_running(rq, task_delta);
4550 /* Determine whether we need to wake up potentially idle CPU: */
4551 if (rq->curr == rq->idle && rq->cfs.nr_running)
4555 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4556 u64 remaining, u64 expires)
4558 struct cfs_rq *cfs_rq;
4560 u64 starting_runtime = remaining;
4563 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4565 struct rq *rq = rq_of(cfs_rq);
4569 if (!cfs_rq_throttled(cfs_rq))
4572 runtime = -cfs_rq->runtime_remaining + 1;
4573 if (runtime > remaining)
4574 runtime = remaining;
4575 remaining -= runtime;
4577 cfs_rq->runtime_remaining += runtime;
4578 cfs_rq->runtime_expires = expires;
4580 /* we check whether we're throttled above */
4581 if (cfs_rq->runtime_remaining > 0)
4582 unthrottle_cfs_rq(cfs_rq);
4592 return starting_runtime - remaining;
4596 * Responsible for refilling a task_group's bandwidth and unthrottling its
4597 * cfs_rqs as appropriate. If there has been no activity within the last
4598 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4599 * used to track this state.
4601 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4603 u64 runtime, runtime_expires;
4606 /* no need to continue the timer with no bandwidth constraint */
4607 if (cfs_b->quota == RUNTIME_INF)
4608 goto out_deactivate;
4610 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4611 cfs_b->nr_periods += overrun;
4614 * idle depends on !throttled (for the case of a large deficit), and if
4615 * we're going inactive then everything else can be deferred
4617 if (cfs_b->idle && !throttled)
4618 goto out_deactivate;
4620 __refill_cfs_bandwidth_runtime(cfs_b);
4623 /* mark as potentially idle for the upcoming period */
4628 /* account preceding periods in which throttling occurred */
4629 cfs_b->nr_throttled += overrun;
4631 runtime_expires = cfs_b->runtime_expires;
4634 * This check is repeated as we are holding onto the new bandwidth while
4635 * we unthrottle. This can potentially race with an unthrottled group
4636 * trying to acquire new bandwidth from the global pool. This can result
4637 * in us over-using our runtime if it is all used during this loop, but
4638 * only by limited amounts in that extreme case.
4640 while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
4641 runtime = cfs_b->runtime;
4642 cfs_b->distribute_running = 1;
4643 raw_spin_unlock(&cfs_b->lock);
4644 /* we can't nest cfs_b->lock while distributing bandwidth */
4645 runtime = distribute_cfs_runtime(cfs_b, runtime,
4647 raw_spin_lock(&cfs_b->lock);
4649 cfs_b->distribute_running = 0;
4650 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4652 lsub_positive(&cfs_b->runtime, runtime);
4656 * While we are ensured activity in the period following an
4657 * unthrottle, this also covers the case in which the new bandwidth is
4658 * insufficient to cover the existing bandwidth deficit. (Forcing the
4659 * timer to remain active while there are any throttled entities.)
4669 /* a cfs_rq won't donate quota below this amount */
4670 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4671 /* minimum remaining period time to redistribute slack quota */
4672 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4673 /* how long we wait to gather additional slack before distributing */
4674 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4677 * Are we near the end of the current quota period?
4679 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4680 * hrtimer base being cleared by hrtimer_start. In the case of
4681 * migrate_hrtimers, base is never cleared, so we are fine.
4683 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4685 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4688 /* if the call-back is running a quota refresh is already occurring */
4689 if (hrtimer_callback_running(refresh_timer))
4692 /* is a quota refresh about to occur? */
4693 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4694 if (remaining < min_expire)
4700 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4702 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4704 /* if there's a quota refresh soon don't bother with slack */
4705 if (runtime_refresh_within(cfs_b, min_left))
4708 hrtimer_start(&cfs_b->slack_timer,
4709 ns_to_ktime(cfs_bandwidth_slack_period),
4713 /* we know any runtime found here is valid as update_curr() precedes return */
4714 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4716 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4717 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4719 if (slack_runtime <= 0)
4722 raw_spin_lock(&cfs_b->lock);
4723 if (cfs_b->quota != RUNTIME_INF &&
4724 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4725 cfs_b->runtime += slack_runtime;
4727 /* we are under rq->lock, defer unthrottling using a timer */
4728 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4729 !list_empty(&cfs_b->throttled_cfs_rq))
4730 start_cfs_slack_bandwidth(cfs_b);
4732 raw_spin_unlock(&cfs_b->lock);
4734 /* even if it's not valid for return we don't want to try again */
4735 cfs_rq->runtime_remaining -= slack_runtime;
4738 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4740 if (!cfs_bandwidth_used())
4743 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4746 __return_cfs_rq_runtime(cfs_rq);
4750 * This is done with a timer (instead of inline with bandwidth return) since
4751 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4753 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4755 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4758 /* confirm we're still not at a refresh boundary */
4759 raw_spin_lock(&cfs_b->lock);
4760 if (cfs_b->distribute_running) {
4761 raw_spin_unlock(&cfs_b->lock);
4765 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4766 raw_spin_unlock(&cfs_b->lock);
4770 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4771 runtime = cfs_b->runtime;
4773 expires = cfs_b->runtime_expires;
4775 cfs_b->distribute_running = 1;
4777 raw_spin_unlock(&cfs_b->lock);
4782 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4784 raw_spin_lock(&cfs_b->lock);
4785 if (expires == cfs_b->runtime_expires)
4786 lsub_positive(&cfs_b->runtime, runtime);
4787 cfs_b->distribute_running = 0;
4788 raw_spin_unlock(&cfs_b->lock);
4792 * When a group wakes up we want to make sure that its quota is not already
4793 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4794 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4796 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4798 if (!cfs_bandwidth_used())
4801 /* an active group must be handled by the update_curr()->put() path */
4802 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4805 /* ensure the group is not already throttled */
4806 if (cfs_rq_throttled(cfs_rq))
4809 /* update runtime allocation */
4810 account_cfs_rq_runtime(cfs_rq, 0);
4811 if (cfs_rq->runtime_remaining <= 0)
4812 throttle_cfs_rq(cfs_rq);
4815 static void sync_throttle(struct task_group *tg, int cpu)
4817 struct cfs_rq *pcfs_rq, *cfs_rq;
4819 if (!cfs_bandwidth_used())
4825 cfs_rq = tg->cfs_rq[cpu];
4826 pcfs_rq = tg->parent->cfs_rq[cpu];
4828 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4829 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4832 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4833 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4835 if (!cfs_bandwidth_used())
4838 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4842 * it's possible for a throttled entity to be forced into a running
4843 * state (e.g. set_curr_task), in this case we're finished.
4845 if (cfs_rq_throttled(cfs_rq))
4848 throttle_cfs_rq(cfs_rq);
4852 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4854 struct cfs_bandwidth *cfs_b =
4855 container_of(timer, struct cfs_bandwidth, slack_timer);
4857 do_sched_cfs_slack_timer(cfs_b);
4859 return HRTIMER_NORESTART;
4862 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4864 struct cfs_bandwidth *cfs_b =
4865 container_of(timer, struct cfs_bandwidth, period_timer);
4869 raw_spin_lock(&cfs_b->lock);
4871 overrun = hrtimer_forward_now(timer, cfs_b->period);
4875 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4878 cfs_b->period_active = 0;
4879 raw_spin_unlock(&cfs_b->lock);
4881 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4884 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4886 raw_spin_lock_init(&cfs_b->lock);
4888 cfs_b->quota = RUNTIME_INF;
4889 cfs_b->period = ns_to_ktime(default_cfs_period());
4891 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4892 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4893 cfs_b->period_timer.function = sched_cfs_period_timer;
4894 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4895 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4896 cfs_b->distribute_running = 0;
4899 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4901 cfs_rq->runtime_enabled = 0;
4902 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4905 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4909 lockdep_assert_held(&cfs_b->lock);
4911 if (cfs_b->period_active)
4914 cfs_b->period_active = 1;
4915 overrun = hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4916 cfs_b->runtime_expires += (overrun + 1) * ktime_to_ns(cfs_b->period);
4917 cfs_b->expires_seq++;
4918 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4921 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4923 /* init_cfs_bandwidth() was not called */
4924 if (!cfs_b->throttled_cfs_rq.next)
4927 hrtimer_cancel(&cfs_b->period_timer);
4928 hrtimer_cancel(&cfs_b->slack_timer);
4932 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
4934 * The race is harmless, since modifying bandwidth settings of unhooked group
4935 * bits doesn't do much.
4938 /* cpu online calback */
4939 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4941 struct task_group *tg;
4943 lockdep_assert_held(&rq->lock);
4946 list_for_each_entry_rcu(tg, &task_groups, list) {
4947 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
4948 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4950 raw_spin_lock(&cfs_b->lock);
4951 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4952 raw_spin_unlock(&cfs_b->lock);
4957 /* cpu offline callback */
4958 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4960 struct task_group *tg;
4962 lockdep_assert_held(&rq->lock);
4965 list_for_each_entry_rcu(tg, &task_groups, list) {
4966 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4968 if (!cfs_rq->runtime_enabled)
4972 * clock_task is not advancing so we just need to make sure
4973 * there's some valid quota amount
4975 cfs_rq->runtime_remaining = 1;
4977 * Offline rq is schedulable till CPU is completely disabled
4978 * in take_cpu_down(), so we prevent new cfs throttling here.
4980 cfs_rq->runtime_enabled = 0;
4982 if (cfs_rq_throttled(cfs_rq))
4983 unthrottle_cfs_rq(cfs_rq);
4988 #else /* CONFIG_CFS_BANDWIDTH */
4989 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4991 return rq_clock_task(rq_of(cfs_rq));
4994 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4995 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4996 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4997 static inline void sync_throttle(struct task_group *tg, int cpu) {}
4998 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5000 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5005 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5010 static inline int throttled_lb_pair(struct task_group *tg,
5011 int src_cpu, int dest_cpu)
5016 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5018 #ifdef CONFIG_FAIR_GROUP_SCHED
5019 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5022 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5026 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5027 static inline void update_runtime_enabled(struct rq *rq) {}
5028 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5030 #endif /* CONFIG_CFS_BANDWIDTH */
5032 /**************************************************
5033 * CFS operations on tasks:
5036 #ifdef CONFIG_SCHED_HRTICK
5037 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5039 struct sched_entity *se = &p->se;
5040 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5042 SCHED_WARN_ON(task_rq(p) != rq);
5044 if (rq->cfs.h_nr_running > 1) {
5045 u64 slice = sched_slice(cfs_rq, se);
5046 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5047 s64 delta = slice - ran;
5054 hrtick_start(rq, delta);
5059 * called from enqueue/dequeue and updates the hrtick when the
5060 * current task is from our class and nr_running is low enough
5063 static void hrtick_update(struct rq *rq)
5065 struct task_struct *curr = rq->curr;
5067 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5070 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5071 hrtick_start_fair(rq, curr);
5073 #else /* !CONFIG_SCHED_HRTICK */
5075 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5079 static inline void hrtick_update(struct rq *rq)
5085 static inline unsigned long cpu_util(int cpu);
5086 static unsigned long capacity_of(int cpu);
5088 static inline bool cpu_overutilized(int cpu)
5090 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5093 static inline void update_overutilized_status(struct rq *rq)
5095 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu))
5096 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5099 static inline void update_overutilized_status(struct rq *rq) { }
5103 * The enqueue_task method is called before nr_running is
5104 * increased. Here we update the fair scheduling stats and
5105 * then put the task into the rbtree:
5108 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5110 struct cfs_rq *cfs_rq;
5111 struct sched_entity *se = &p->se;
5114 * The code below (indirectly) updates schedutil which looks at
5115 * the cfs_rq utilization to select a frequency.
5116 * Let's add the task's estimated utilization to the cfs_rq's
5117 * estimated utilization, before we update schedutil.
5119 util_est_enqueue(&rq->cfs, p);
5122 * If in_iowait is set, the code below may not trigger any cpufreq
5123 * utilization updates, so do it here explicitly with the IOWAIT flag
5127 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5129 for_each_sched_entity(se) {
5132 cfs_rq = cfs_rq_of(se);
5133 enqueue_entity(cfs_rq, se, flags);
5136 * end evaluation on encountering a throttled cfs_rq
5138 * note: in the case of encountering a throttled cfs_rq we will
5139 * post the final h_nr_running increment below.
5141 if (cfs_rq_throttled(cfs_rq))
5143 cfs_rq->h_nr_running++;
5145 flags = ENQUEUE_WAKEUP;
5148 for_each_sched_entity(se) {
5149 cfs_rq = cfs_rq_of(se);
5150 cfs_rq->h_nr_running++;
5152 if (cfs_rq_throttled(cfs_rq))
5155 update_load_avg(cfs_rq, se, UPDATE_TG);
5156 update_cfs_group(se);
5160 add_nr_running(rq, 1);
5162 * Since new tasks are assigned an initial util_avg equal to
5163 * half of the spare capacity of their CPU, tiny tasks have the
5164 * ability to cross the overutilized threshold, which will
5165 * result in the load balancer ruining all the task placement
5166 * done by EAS. As a way to mitigate that effect, do not account
5167 * for the first enqueue operation of new tasks during the
5168 * overutilized flag detection.
5170 * A better way of solving this problem would be to wait for
5171 * the PELT signals of tasks to converge before taking them
5172 * into account, but that is not straightforward to implement,
5173 * and the following generally works well enough in practice.
5175 if (flags & ENQUEUE_WAKEUP)
5176 update_overutilized_status(rq);
5183 static void set_next_buddy(struct sched_entity *se);
5186 * The dequeue_task method is called before nr_running is
5187 * decreased. We remove the task from the rbtree and
5188 * update the fair scheduling stats:
5190 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5192 struct cfs_rq *cfs_rq;
5193 struct sched_entity *se = &p->se;
5194 int task_sleep = flags & DEQUEUE_SLEEP;
5196 for_each_sched_entity(se) {
5197 cfs_rq = cfs_rq_of(se);
5198 dequeue_entity(cfs_rq, se, flags);
5201 * end evaluation on encountering a throttled cfs_rq
5203 * note: in the case of encountering a throttled cfs_rq we will
5204 * post the final h_nr_running decrement below.
5206 if (cfs_rq_throttled(cfs_rq))
5208 cfs_rq->h_nr_running--;
5210 /* Don't dequeue parent if it has other entities besides us */
5211 if (cfs_rq->load.weight) {
5212 /* Avoid re-evaluating load for this entity: */
5213 se = parent_entity(se);
5215 * Bias pick_next to pick a task from this cfs_rq, as
5216 * p is sleeping when it is within its sched_slice.
5218 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5222 flags |= DEQUEUE_SLEEP;
5225 for_each_sched_entity(se) {
5226 cfs_rq = cfs_rq_of(se);
5227 cfs_rq->h_nr_running--;
5229 if (cfs_rq_throttled(cfs_rq))
5232 update_load_avg(cfs_rq, se, UPDATE_TG);
5233 update_cfs_group(se);
5237 sub_nr_running(rq, 1);
5239 util_est_dequeue(&rq->cfs, p, task_sleep);
5245 /* Working cpumask for: load_balance, load_balance_newidle. */
5246 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5247 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5249 #ifdef CONFIG_NO_HZ_COMMON
5251 * per rq 'load' arrray crap; XXX kill this.
5255 * The exact cpuload calculated at every tick would be:
5257 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5259 * If a CPU misses updates for n ticks (as it was idle) and update gets
5260 * called on the n+1-th tick when CPU may be busy, then we have:
5262 * load_n = (1 - 1/2^i)^n * load_0
5263 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
5265 * decay_load_missed() below does efficient calculation of
5267 * load' = (1 - 1/2^i)^n * load
5269 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5270 * This allows us to precompute the above in said factors, thereby allowing the
5271 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5272 * fixed_power_int())
5274 * The calculation is approximated on a 128 point scale.
5276 #define DEGRADE_SHIFT 7
5278 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5279 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5280 { 0, 0, 0, 0, 0, 0, 0, 0 },
5281 { 64, 32, 8, 0, 0, 0, 0, 0 },
5282 { 96, 72, 40, 12, 1, 0, 0, 0 },
5283 { 112, 98, 75, 43, 15, 1, 0, 0 },
5284 { 120, 112, 98, 76, 45, 16, 2, 0 }
5288 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5289 * would be when CPU is idle and so we just decay the old load without
5290 * adding any new load.
5292 static unsigned long
5293 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5297 if (!missed_updates)
5300 if (missed_updates >= degrade_zero_ticks[idx])
5304 return load >> missed_updates;
5306 while (missed_updates) {
5307 if (missed_updates % 2)
5308 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5310 missed_updates >>= 1;
5317 cpumask_var_t idle_cpus_mask;
5319 int has_blocked; /* Idle CPUS has blocked load */
5320 unsigned long next_balance; /* in jiffy units */
5321 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5322 } nohz ____cacheline_aligned;
5324 #endif /* CONFIG_NO_HZ_COMMON */
5327 * __cpu_load_update - update the rq->cpu_load[] statistics
5328 * @this_rq: The rq to update statistics for
5329 * @this_load: The current load
5330 * @pending_updates: The number of missed updates
5332 * Update rq->cpu_load[] statistics. This function is usually called every
5333 * scheduler tick (TICK_NSEC).
5335 * This function computes a decaying average:
5337 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5339 * Because of NOHZ it might not get called on every tick which gives need for
5340 * the @pending_updates argument.
5342 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5343 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5344 * = A * (A * load[i]_n-2 + B) + B
5345 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5346 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5347 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5348 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5349 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5351 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5352 * any change in load would have resulted in the tick being turned back on.
5354 * For regular NOHZ, this reduces to:
5356 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5358 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5361 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5362 unsigned long pending_updates)
5364 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5367 this_rq->nr_load_updates++;
5369 /* Update our load: */
5370 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5371 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5372 unsigned long old_load, new_load;
5374 /* scale is effectively 1 << i now, and >> i divides by scale */
5376 old_load = this_rq->cpu_load[i];
5377 #ifdef CONFIG_NO_HZ_COMMON
5378 old_load = decay_load_missed(old_load, pending_updates - 1, i);
5379 if (tickless_load) {
5380 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5382 * old_load can never be a negative value because a
5383 * decayed tickless_load cannot be greater than the
5384 * original tickless_load.
5386 old_load += tickless_load;
5389 new_load = this_load;
5391 * Round up the averaging division if load is increasing. This
5392 * prevents us from getting stuck on 9 if the load is 10, for
5395 if (new_load > old_load)
5396 new_load += scale - 1;
5398 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5402 /* Used instead of source_load when we know the type == 0 */
5403 static unsigned long weighted_cpuload(struct rq *rq)
5405 return cfs_rq_runnable_load_avg(&rq->cfs);
5408 #ifdef CONFIG_NO_HZ_COMMON
5410 * There is no sane way to deal with nohz on smp when using jiffies because the
5411 * CPU doing the jiffies update might drift wrt the CPU doing the jiffy reading
5412 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5414 * Therefore we need to avoid the delta approach from the regular tick when
5415 * possible since that would seriously skew the load calculation. This is why we
5416 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5417 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5418 * loop exit, nohz_idle_balance, nohz full exit...)
5420 * This means we might still be one tick off for nohz periods.
5423 static void cpu_load_update_nohz(struct rq *this_rq,
5424 unsigned long curr_jiffies,
5427 unsigned long pending_updates;
5429 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5430 if (pending_updates) {
5431 this_rq->last_load_update_tick = curr_jiffies;
5433 * In the regular NOHZ case, we were idle, this means load 0.
5434 * In the NOHZ_FULL case, we were non-idle, we should consider
5435 * its weighted load.
5437 cpu_load_update(this_rq, load, pending_updates);
5442 * Called from nohz_idle_balance() to update the load ratings before doing the
5445 static void cpu_load_update_idle(struct rq *this_rq)
5448 * bail if there's load or we're actually up-to-date.
5450 if (weighted_cpuload(this_rq))
5453 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5457 * Record CPU load on nohz entry so we know the tickless load to account
5458 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5459 * than other cpu_load[idx] but it should be fine as cpu_load readers
5460 * shouldn't rely into synchronized cpu_load[*] updates.
5462 void cpu_load_update_nohz_start(void)
5464 struct rq *this_rq = this_rq();
5467 * This is all lockless but should be fine. If weighted_cpuload changes
5468 * concurrently we'll exit nohz. And cpu_load write can race with
5469 * cpu_load_update_idle() but both updater would be writing the same.
5471 this_rq->cpu_load[0] = weighted_cpuload(this_rq);
5475 * Account the tickless load in the end of a nohz frame.
5477 void cpu_load_update_nohz_stop(void)
5479 unsigned long curr_jiffies = READ_ONCE(jiffies);
5480 struct rq *this_rq = this_rq();
5484 if (curr_jiffies == this_rq->last_load_update_tick)
5487 load = weighted_cpuload(this_rq);
5488 rq_lock(this_rq, &rf);
5489 update_rq_clock(this_rq);
5490 cpu_load_update_nohz(this_rq, curr_jiffies, load);
5491 rq_unlock(this_rq, &rf);
5493 #else /* !CONFIG_NO_HZ_COMMON */
5494 static inline void cpu_load_update_nohz(struct rq *this_rq,
5495 unsigned long curr_jiffies,
5496 unsigned long load) { }
5497 #endif /* CONFIG_NO_HZ_COMMON */
5499 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5501 #ifdef CONFIG_NO_HZ_COMMON
5502 /* See the mess around cpu_load_update_nohz(). */
5503 this_rq->last_load_update_tick = READ_ONCE(jiffies);
5505 cpu_load_update(this_rq, load, 1);
5509 * Called from scheduler_tick()
5511 void cpu_load_update_active(struct rq *this_rq)
5513 unsigned long load = weighted_cpuload(this_rq);
5515 if (tick_nohz_tick_stopped())
5516 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5518 cpu_load_update_periodic(this_rq, load);
5522 * Return a low guess at the load of a migration-source CPU weighted
5523 * according to the scheduling class and "nice" value.
5525 * We want to under-estimate the load of migration sources, to
5526 * balance conservatively.
5528 static unsigned long source_load(int cpu, int type)
5530 struct rq *rq = cpu_rq(cpu);
5531 unsigned long total = weighted_cpuload(rq);
5533 if (type == 0 || !sched_feat(LB_BIAS))
5536 return min(rq->cpu_load[type-1], total);
5540 * Return a high guess at the load of a migration-target CPU weighted
5541 * according to the scheduling class and "nice" value.
5543 static unsigned long target_load(int cpu, int type)
5545 struct rq *rq = cpu_rq(cpu);
5546 unsigned long total = weighted_cpuload(rq);
5548 if (type == 0 || !sched_feat(LB_BIAS))
5551 return max(rq->cpu_load[type-1], total);
5554 static unsigned long capacity_of(int cpu)
5556 return cpu_rq(cpu)->cpu_capacity;
5559 static unsigned long capacity_orig_of(int cpu)
5561 return cpu_rq(cpu)->cpu_capacity_orig;
5564 static unsigned long cpu_avg_load_per_task(int cpu)
5566 struct rq *rq = cpu_rq(cpu);
5567 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5568 unsigned long load_avg = weighted_cpuload(rq);
5571 return load_avg / nr_running;
5576 static void record_wakee(struct task_struct *p)
5579 * Only decay a single time; tasks that have less then 1 wakeup per
5580 * jiffy will not have built up many flips.
5582 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5583 current->wakee_flips >>= 1;
5584 current->wakee_flip_decay_ts = jiffies;
5587 if (current->last_wakee != p) {
5588 current->last_wakee = p;
5589 current->wakee_flips++;
5594 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5596 * A waker of many should wake a different task than the one last awakened
5597 * at a frequency roughly N times higher than one of its wakees.
5599 * In order to determine whether we should let the load spread vs consolidating
5600 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5601 * partner, and a factor of lls_size higher frequency in the other.
5603 * With both conditions met, we can be relatively sure that the relationship is
5604 * non-monogamous, with partner count exceeding socket size.
5606 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5607 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5610 static int wake_wide(struct task_struct *p)
5612 unsigned int master = current->wakee_flips;
5613 unsigned int slave = p->wakee_flips;
5614 int factor = this_cpu_read(sd_llc_size);
5617 swap(master, slave);
5618 if (slave < factor || master < slave * factor)
5624 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5625 * soonest. For the purpose of speed we only consider the waking and previous
5628 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5629 * cache-affine and is (or will be) idle.
5631 * wake_affine_weight() - considers the weight to reflect the average
5632 * scheduling latency of the CPUs. This seems to work
5633 * for the overloaded case.
5636 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5639 * If this_cpu is idle, it implies the wakeup is from interrupt
5640 * context. Only allow the move if cache is shared. Otherwise an
5641 * interrupt intensive workload could force all tasks onto one
5642 * node depending on the IO topology or IRQ affinity settings.
5644 * If the prev_cpu is idle and cache affine then avoid a migration.
5645 * There is no guarantee that the cache hot data from an interrupt
5646 * is more important than cache hot data on the prev_cpu and from
5647 * a cpufreq perspective, it's better to have higher utilisation
5650 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5651 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5653 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5656 return nr_cpumask_bits;
5660 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5661 int this_cpu, int prev_cpu, int sync)
5663 s64 this_eff_load, prev_eff_load;
5664 unsigned long task_load;
5666 this_eff_load = target_load(this_cpu, sd->wake_idx);
5669 unsigned long current_load = task_h_load(current);
5671 if (current_load > this_eff_load)
5674 this_eff_load -= current_load;
5677 task_load = task_h_load(p);
5679 this_eff_load += task_load;
5680 if (sched_feat(WA_BIAS))
5681 this_eff_load *= 100;
5682 this_eff_load *= capacity_of(prev_cpu);
5684 prev_eff_load = source_load(prev_cpu, sd->wake_idx);
5685 prev_eff_load -= task_load;
5686 if (sched_feat(WA_BIAS))
5687 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5688 prev_eff_load *= capacity_of(this_cpu);
5691 * If sync, adjust the weight of prev_eff_load such that if
5692 * prev_eff == this_eff that select_idle_sibling() will consider
5693 * stacking the wakee on top of the waker if no other CPU is
5699 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5702 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5703 int this_cpu, int prev_cpu, int sync)
5705 int target = nr_cpumask_bits;
5707 if (sched_feat(WA_IDLE))
5708 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5710 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5711 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5713 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5714 if (target == nr_cpumask_bits)
5717 schedstat_inc(sd->ttwu_move_affine);
5718 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5722 static unsigned long cpu_util_without(int cpu, struct task_struct *p);
5724 static unsigned long capacity_spare_without(int cpu, struct task_struct *p)
5726 return max_t(long, capacity_of(cpu) - cpu_util_without(cpu, p), 0);
5730 * find_idlest_group finds and returns the least busy CPU group within the
5733 * Assumes p is allowed on at least one CPU in sd.
5735 static struct sched_group *
5736 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5737 int this_cpu, int sd_flag)
5739 struct sched_group *idlest = NULL, *group = sd->groups;
5740 struct sched_group *most_spare_sg = NULL;
5741 unsigned long min_runnable_load = ULONG_MAX;
5742 unsigned long this_runnable_load = ULONG_MAX;
5743 unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5744 unsigned long most_spare = 0, this_spare = 0;
5745 int load_idx = sd->forkexec_idx;
5746 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5747 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5748 (sd->imbalance_pct-100) / 100;
5750 if (sd_flag & SD_BALANCE_WAKE)
5751 load_idx = sd->wake_idx;
5754 unsigned long load, avg_load, runnable_load;
5755 unsigned long spare_cap, max_spare_cap;
5759 /* Skip over this group if it has no CPUs allowed */
5760 if (!cpumask_intersects(sched_group_span(group),
5764 local_group = cpumask_test_cpu(this_cpu,
5765 sched_group_span(group));
5768 * Tally up the load of all CPUs in the group and find
5769 * the group containing the CPU with most spare capacity.
5775 for_each_cpu(i, sched_group_span(group)) {
5776 /* Bias balancing toward CPUs of our domain */
5778 load = source_load(i, load_idx);
5780 load = target_load(i, load_idx);
5782 runnable_load += load;
5784 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5786 spare_cap = capacity_spare_without(i, p);
5788 if (spare_cap > max_spare_cap)
5789 max_spare_cap = spare_cap;
5792 /* Adjust by relative CPU capacity of the group */
5793 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5794 group->sgc->capacity;
5795 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5796 group->sgc->capacity;
5799 this_runnable_load = runnable_load;
5800 this_avg_load = avg_load;
5801 this_spare = max_spare_cap;
5803 if (min_runnable_load > (runnable_load + imbalance)) {
5805 * The runnable load is significantly smaller
5806 * so we can pick this new CPU:
5808 min_runnable_load = runnable_load;
5809 min_avg_load = avg_load;
5811 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5812 (100*min_avg_load > imbalance_scale*avg_load)) {
5814 * The runnable loads are close so take the
5815 * blocked load into account through avg_load:
5817 min_avg_load = avg_load;
5821 if (most_spare < max_spare_cap) {
5822 most_spare = max_spare_cap;
5823 most_spare_sg = group;
5826 } while (group = group->next, group != sd->groups);
5829 * The cross-over point between using spare capacity or least load
5830 * is too conservative for high utilization tasks on partially
5831 * utilized systems if we require spare_capacity > task_util(p),
5832 * so we allow for some task stuffing by using
5833 * spare_capacity > task_util(p)/2.
5835 * Spare capacity can't be used for fork because the utilization has
5836 * not been set yet, we must first select a rq to compute the initial
5839 if (sd_flag & SD_BALANCE_FORK)
5842 if (this_spare > task_util(p) / 2 &&
5843 imbalance_scale*this_spare > 100*most_spare)
5846 if (most_spare > task_util(p) / 2)
5847 return most_spare_sg;
5854 * When comparing groups across NUMA domains, it's possible for the
5855 * local domain to be very lightly loaded relative to the remote
5856 * domains but "imbalance" skews the comparison making remote CPUs
5857 * look much more favourable. When considering cross-domain, add
5858 * imbalance to the runnable load on the remote node and consider
5861 if ((sd->flags & SD_NUMA) &&
5862 min_runnable_load + imbalance >= this_runnable_load)
5865 if (min_runnable_load > (this_runnable_load + imbalance))
5868 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5869 (100*this_avg_load < imbalance_scale*min_avg_load))
5876 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5879 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5881 unsigned long load, min_load = ULONG_MAX;
5882 unsigned int min_exit_latency = UINT_MAX;
5883 u64 latest_idle_timestamp = 0;
5884 int least_loaded_cpu = this_cpu;
5885 int shallowest_idle_cpu = -1;
5888 /* Check if we have any choice: */
5889 if (group->group_weight == 1)
5890 return cpumask_first(sched_group_span(group));
5892 /* Traverse only the allowed CPUs */
5893 for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
5894 if (available_idle_cpu(i)) {
5895 struct rq *rq = cpu_rq(i);
5896 struct cpuidle_state *idle = idle_get_state(rq);
5897 if (idle && idle->exit_latency < min_exit_latency) {
5899 * We give priority to a CPU whose idle state
5900 * has the smallest exit latency irrespective
5901 * of any idle timestamp.
5903 min_exit_latency = idle->exit_latency;
5904 latest_idle_timestamp = rq->idle_stamp;
5905 shallowest_idle_cpu = i;
5906 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5907 rq->idle_stamp > latest_idle_timestamp) {
5909 * If equal or no active idle state, then
5910 * the most recently idled CPU might have
5913 latest_idle_timestamp = rq->idle_stamp;
5914 shallowest_idle_cpu = i;
5916 } else if (shallowest_idle_cpu == -1) {
5917 load = weighted_cpuload(cpu_rq(i));
5918 if (load < min_load) {
5920 least_loaded_cpu = i;
5925 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5928 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5929 int cpu, int prev_cpu, int sd_flag)
5933 if (!cpumask_intersects(sched_domain_span(sd), &p->cpus_allowed))
5937 * We need task's util for capacity_spare_without, sync it up to
5938 * prev_cpu's last_update_time.
5940 if (!(sd_flag & SD_BALANCE_FORK))
5941 sync_entity_load_avg(&p->se);
5944 struct sched_group *group;
5945 struct sched_domain *tmp;
5948 if (!(sd->flags & sd_flag)) {
5953 group = find_idlest_group(sd, p, cpu, sd_flag);
5959 new_cpu = find_idlest_group_cpu(group, p, cpu);
5960 if (new_cpu == cpu) {
5961 /* Now try balancing at a lower domain level of 'cpu': */
5966 /* Now try balancing at a lower domain level of 'new_cpu': */
5968 weight = sd->span_weight;
5970 for_each_domain(cpu, tmp) {
5971 if (weight <= tmp->span_weight)
5973 if (tmp->flags & sd_flag)
5981 #ifdef CONFIG_SCHED_SMT
5982 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5984 static inline void set_idle_cores(int cpu, int val)
5986 struct sched_domain_shared *sds;
5988 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5990 WRITE_ONCE(sds->has_idle_cores, val);
5993 static inline bool test_idle_cores(int cpu, bool def)
5995 struct sched_domain_shared *sds;
5997 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5999 return READ_ONCE(sds->has_idle_cores);
6005 * Scans the local SMT mask to see if the entire core is idle, and records this
6006 * information in sd_llc_shared->has_idle_cores.
6008 * Since SMT siblings share all cache levels, inspecting this limited remote
6009 * state should be fairly cheap.
6011 void __update_idle_core(struct rq *rq)
6013 int core = cpu_of(rq);
6017 if (test_idle_cores(core, true))
6020 for_each_cpu(cpu, cpu_smt_mask(core)) {
6024 if (!available_idle_cpu(cpu))
6028 set_idle_cores(core, 1);
6034 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6035 * there are no idle cores left in the system; tracked through
6036 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6038 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6040 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6043 if (!static_branch_likely(&sched_smt_present))
6046 if (!test_idle_cores(target, false))
6049 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
6051 for_each_cpu_wrap(core, cpus, target) {
6054 for_each_cpu(cpu, cpu_smt_mask(core)) {
6055 cpumask_clear_cpu(cpu, cpus);
6056 if (!available_idle_cpu(cpu))
6065 * Failed to find an idle core; stop looking for one.
6067 set_idle_cores(target, 0);
6073 * Scan the local SMT mask for idle CPUs.
6075 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6079 if (!static_branch_likely(&sched_smt_present))
6082 for_each_cpu(cpu, cpu_smt_mask(target)) {
6083 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6085 if (available_idle_cpu(cpu))
6092 #else /* CONFIG_SCHED_SMT */
6094 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6099 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6104 #endif /* CONFIG_SCHED_SMT */
6107 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6108 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6109 * average idle time for this rq (as found in rq->avg_idle).
6111 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
6113 struct sched_domain *this_sd;
6114 u64 avg_cost, avg_idle;
6117 int cpu, nr = INT_MAX;
6119 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6124 * Due to large variance we need a large fuzz factor; hackbench in
6125 * particularly is sensitive here.
6127 avg_idle = this_rq()->avg_idle / 512;
6128 avg_cost = this_sd->avg_scan_cost + 1;
6130 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
6133 if (sched_feat(SIS_PROP)) {
6134 u64 span_avg = sd->span_weight * avg_idle;
6135 if (span_avg > 4*avg_cost)
6136 nr = div_u64(span_avg, avg_cost);
6141 time = local_clock();
6143 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
6146 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6148 if (available_idle_cpu(cpu))
6152 time = local_clock() - time;
6153 cost = this_sd->avg_scan_cost;
6154 delta = (s64)(time - cost) / 8;
6155 this_sd->avg_scan_cost += delta;
6161 * Try and locate an idle core/thread in the LLC cache domain.
6163 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6165 struct sched_domain *sd;
6166 int i, recent_used_cpu;
6168 if (available_idle_cpu(target))
6172 * If the previous CPU is cache affine and idle, don't be stupid:
6174 if (prev != target && cpus_share_cache(prev, target) && available_idle_cpu(prev))
6177 /* Check a recently used CPU as a potential idle candidate: */
6178 recent_used_cpu = p->recent_used_cpu;
6179 if (recent_used_cpu != prev &&
6180 recent_used_cpu != target &&
6181 cpus_share_cache(recent_used_cpu, target) &&
6182 available_idle_cpu(recent_used_cpu) &&
6183 cpumask_test_cpu(p->recent_used_cpu, &p->cpus_allowed)) {
6185 * Replace recent_used_cpu with prev as it is a potential
6186 * candidate for the next wake:
6188 p->recent_used_cpu = prev;
6189 return recent_used_cpu;
6192 sd = rcu_dereference(per_cpu(sd_llc, target));
6196 i = select_idle_core(p, sd, target);
6197 if ((unsigned)i < nr_cpumask_bits)
6200 i = select_idle_cpu(p, sd, target);
6201 if ((unsigned)i < nr_cpumask_bits)
6204 i = select_idle_smt(p, sd, target);
6205 if ((unsigned)i < nr_cpumask_bits)
6212 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6213 * @cpu: the CPU to get the utilization of
6215 * The unit of the return value must be the one of capacity so we can compare
6216 * the utilization with the capacity of the CPU that is available for CFS task
6217 * (ie cpu_capacity).
6219 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6220 * recent utilization of currently non-runnable tasks on a CPU. It represents
6221 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6222 * capacity_orig is the cpu_capacity available at the highest frequency
6223 * (arch_scale_freq_capacity()).
6224 * The utilization of a CPU converges towards a sum equal to or less than the
6225 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6226 * the running time on this CPU scaled by capacity_curr.
6228 * The estimated utilization of a CPU is defined to be the maximum between its
6229 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6230 * currently RUNNABLE on that CPU.
6231 * This allows to properly represent the expected utilization of a CPU which
6232 * has just got a big task running since a long sleep period. At the same time
6233 * however it preserves the benefits of the "blocked utilization" in
6234 * describing the potential for other tasks waking up on the same CPU.
6236 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6237 * higher than capacity_orig because of unfortunate rounding in
6238 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6239 * the average stabilizes with the new running time. We need to check that the
6240 * utilization stays within the range of [0..capacity_orig] and cap it if
6241 * necessary. Without utilization capping, a group could be seen as overloaded
6242 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6243 * available capacity. We allow utilization to overshoot capacity_curr (but not
6244 * capacity_orig) as it useful for predicting the capacity required after task
6245 * migrations (scheduler-driven DVFS).
6247 * Return: the (estimated) utilization for the specified CPU
6249 static inline unsigned long cpu_util(int cpu)
6251 struct cfs_rq *cfs_rq;
6254 cfs_rq = &cpu_rq(cpu)->cfs;
6255 util = READ_ONCE(cfs_rq->avg.util_avg);
6257 if (sched_feat(UTIL_EST))
6258 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6260 return min_t(unsigned long, util, capacity_orig_of(cpu));
6264 * cpu_util_without: compute cpu utilization without any contributions from *p
6265 * @cpu: the CPU which utilization is requested
6266 * @p: the task which utilization should be discounted
6268 * The utilization of a CPU is defined by the utilization of tasks currently
6269 * enqueued on that CPU as well as tasks which are currently sleeping after an
6270 * execution on that CPU.
6272 * This method returns the utilization of the specified CPU by discounting the
6273 * utilization of the specified task, whenever the task is currently
6274 * contributing to the CPU utilization.
6276 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6278 struct cfs_rq *cfs_rq;
6281 /* Task has no contribution or is new */
6282 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6283 return cpu_util(cpu);
6285 cfs_rq = &cpu_rq(cpu)->cfs;
6286 util = READ_ONCE(cfs_rq->avg.util_avg);
6288 /* Discount task's util from CPU's util */
6289 lsub_positive(&util, task_util(p));
6294 * a) if *p is the only task sleeping on this CPU, then:
6295 * cpu_util (== task_util) > util_est (== 0)
6296 * and thus we return:
6297 * cpu_util_without = (cpu_util - task_util) = 0
6299 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6301 * cpu_util >= task_util
6302 * cpu_util > util_est (== 0)
6303 * and thus we discount *p's blocked utilization to return:
6304 * cpu_util_without = (cpu_util - task_util) >= 0
6306 * c) if other tasks are RUNNABLE on that CPU and
6307 * util_est > cpu_util
6308 * then we use util_est since it returns a more restrictive
6309 * estimation of the spare capacity on that CPU, by just
6310 * considering the expected utilization of tasks already
6311 * runnable on that CPU.
6313 * Cases a) and b) are covered by the above code, while case c) is
6314 * covered by the following code when estimated utilization is
6317 if (sched_feat(UTIL_EST)) {
6318 unsigned int estimated =
6319 READ_ONCE(cfs_rq->avg.util_est.enqueued);
6322 * Despite the following checks we still have a small window
6323 * for a possible race, when an execl's select_task_rq_fair()
6324 * races with LB's detach_task():
6327 * p->on_rq = TASK_ON_RQ_MIGRATING;
6328 * ---------------------------------- A
6329 * deactivate_task() \
6330 * dequeue_task() + RaceTime
6331 * util_est_dequeue() /
6332 * ---------------------------------- B
6334 * The additional check on "current == p" it's required to
6335 * properly fix the execl regression and it helps in further
6336 * reducing the chances for the above race.
6338 if (unlikely(task_on_rq_queued(p) || current == p))
6339 lsub_positive(&estimated, _task_util_est(p));
6341 util = max(util, estimated);
6345 * Utilization (estimated) can exceed the CPU capacity, thus let's
6346 * clamp to the maximum CPU capacity to ensure consistency with
6347 * the cpu_util call.
6349 return min_t(unsigned long, util, capacity_orig_of(cpu));
6353 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6354 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6356 * In that case WAKE_AFFINE doesn't make sense and we'll let
6357 * BALANCE_WAKE sort things out.
6359 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6361 long min_cap, max_cap;
6363 if (!static_branch_unlikely(&sched_asym_cpucapacity))
6366 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6367 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6369 /* Minimum capacity is close to max, no need to abort wake_affine */
6370 if (max_cap - min_cap < max_cap >> 3)
6373 /* Bring task utilization in sync with prev_cpu */
6374 sync_entity_load_avg(&p->se);
6376 return !task_fits_capacity(p, min_cap);
6380 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6383 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6385 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6386 unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
6389 * If @p migrates from @cpu to another, remove its contribution. Or,
6390 * if @p migrates from another CPU to @cpu, add its contribution. In
6391 * the other cases, @cpu is not impacted by the migration, so the
6392 * util_avg should already be correct.
6394 if (task_cpu(p) == cpu && dst_cpu != cpu)
6395 sub_positive(&util, task_util(p));
6396 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6397 util += task_util(p);
6399 if (sched_feat(UTIL_EST)) {
6400 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6403 * During wake-up, the task isn't enqueued yet and doesn't
6404 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6405 * so just add it (if needed) to "simulate" what will be
6406 * cpu_util() after the task has been enqueued.
6409 util_est += _task_util_est(p);
6411 util = max(util, util_est);
6414 return min(util, capacity_orig_of(cpu));
6418 * compute_energy(): Estimates the energy that would be consumed if @p was
6419 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6420 * landscape of the * CPUs after the task migration, and uses the Energy Model
6421 * to compute what would be the energy if we decided to actually migrate that
6425 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
6427 long util, max_util, sum_util, energy = 0;
6430 for (; pd; pd = pd->next) {
6431 max_util = sum_util = 0;
6433 * The capacity state of CPUs of the current rd can be driven by
6434 * CPUs of another rd if they belong to the same performance
6435 * domain. So, account for the utilization of these CPUs too
6436 * by masking pd with cpu_online_mask instead of the rd span.
6438 * If an entire performance domain is outside of the current rd,
6439 * it will not appear in its pd list and will not be accounted
6440 * by compute_energy().
6442 for_each_cpu_and(cpu, perf_domain_span(pd), cpu_online_mask) {
6443 util = cpu_util_next(cpu, p, dst_cpu);
6444 util = schedutil_energy_util(cpu, util);
6445 max_util = max(util, max_util);
6449 energy += em_pd_energy(pd->em_pd, max_util, sum_util);
6456 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6457 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6458 * spare capacity in each performance domain and uses it as a potential
6459 * candidate to execute the task. Then, it uses the Energy Model to figure
6460 * out which of the CPU candidates is the most energy-efficient.
6462 * The rationale for this heuristic is as follows. In a performance domain,
6463 * all the most energy efficient CPU candidates (according to the Energy
6464 * Model) are those for which we'll request a low frequency. When there are
6465 * several CPUs for which the frequency request will be the same, we don't
6466 * have enough data to break the tie between them, because the Energy Model
6467 * only includes active power costs. With this model, if we assume that
6468 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6469 * the maximum spare capacity in a performance domain is guaranteed to be among
6470 * the best candidates of the performance domain.
6472 * In practice, it could be preferable from an energy standpoint to pack
6473 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6474 * but that could also hurt our chances to go cluster idle, and we have no
6475 * ways to tell with the current Energy Model if this is actually a good
6476 * idea or not. So, find_energy_efficient_cpu() basically favors
6477 * cluster-packing, and spreading inside a cluster. That should at least be
6478 * a good thing for latency, and this is consistent with the idea that most
6479 * of the energy savings of EAS come from the asymmetry of the system, and
6480 * not so much from breaking the tie between identical CPUs. That's also the
6481 * reason why EAS is enabled in the topology code only for systems where
6482 * SD_ASYM_CPUCAPACITY is set.
6484 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6485 * they don't have any useful utilization data yet and it's not possible to
6486 * forecast their impact on energy consumption. Consequently, they will be
6487 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6488 * to be energy-inefficient in some use-cases. The alternative would be to
6489 * bias new tasks towards specific types of CPUs first, or to try to infer
6490 * their util_avg from the parent task, but those heuristics could hurt
6491 * other use-cases too. So, until someone finds a better way to solve this,
6492 * let's keep things simple by re-using the existing slow path.
6495 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6497 unsigned long prev_energy = ULONG_MAX, best_energy = ULONG_MAX;
6498 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6499 int cpu, best_energy_cpu = prev_cpu;
6500 struct perf_domain *head, *pd;
6501 unsigned long cpu_cap, util;
6502 struct sched_domain *sd;
6505 pd = rcu_dereference(rd->pd);
6506 if (!pd || READ_ONCE(rd->overutilized))
6511 * Energy-aware wake-up happens on the lowest sched_domain starting
6512 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6514 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6515 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6520 sync_entity_load_avg(&p->se);
6521 if (!task_util_est(p))
6524 for (; pd; pd = pd->next) {
6525 unsigned long cur_energy, spare_cap, max_spare_cap = 0;
6526 int max_spare_cap_cpu = -1;
6528 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
6529 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6532 /* Skip CPUs that will be overutilized. */
6533 util = cpu_util_next(cpu, p, cpu);
6534 cpu_cap = capacity_of(cpu);
6535 if (cpu_cap * 1024 < util * capacity_margin)
6538 /* Always use prev_cpu as a candidate. */
6539 if (cpu == prev_cpu) {
6540 prev_energy = compute_energy(p, prev_cpu, head);
6541 best_energy = min(best_energy, prev_energy);
6546 * Find the CPU with the maximum spare capacity in
6547 * the performance domain
6549 spare_cap = cpu_cap - util;
6550 if (spare_cap > max_spare_cap) {
6551 max_spare_cap = spare_cap;
6552 max_spare_cap_cpu = cpu;
6556 /* Evaluate the energy impact of using this CPU. */
6557 if (max_spare_cap_cpu >= 0) {
6558 cur_energy = compute_energy(p, max_spare_cap_cpu, head);
6559 if (cur_energy < best_energy) {
6560 best_energy = cur_energy;
6561 best_energy_cpu = max_spare_cap_cpu;
6569 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6570 * least 6% of the energy used by prev_cpu.
6572 if (prev_energy == ULONG_MAX)
6573 return best_energy_cpu;
6575 if ((prev_energy - best_energy) > (prev_energy >> 4))
6576 return best_energy_cpu;
6587 * select_task_rq_fair: Select target runqueue for the waking task in domains
6588 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6589 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6591 * Balances load by selecting the idlest CPU in the idlest group, or under
6592 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6594 * Returns the target CPU number.
6596 * preempt must be disabled.
6599 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6601 struct sched_domain *tmp, *sd = NULL;
6602 int cpu = smp_processor_id();
6603 int new_cpu = prev_cpu;
6604 int want_affine = 0;
6605 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6607 if (sd_flag & SD_BALANCE_WAKE) {
6610 if (static_branch_unlikely(&sched_energy_present)) {
6611 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
6617 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu) &&
6618 cpumask_test_cpu(cpu, &p->cpus_allowed);
6622 for_each_domain(cpu, tmp) {
6623 if (!(tmp->flags & SD_LOAD_BALANCE))
6627 * If both 'cpu' and 'prev_cpu' are part of this domain,
6628 * cpu is a valid SD_WAKE_AFFINE target.
6630 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6631 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6632 if (cpu != prev_cpu)
6633 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6635 sd = NULL; /* Prefer wake_affine over balance flags */
6639 if (tmp->flags & sd_flag)
6641 else if (!want_affine)
6647 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6648 } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6651 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6654 current->recent_used_cpu = cpu;
6661 static void detach_entity_cfs_rq(struct sched_entity *se);
6664 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6665 * cfs_rq_of(p) references at time of call are still valid and identify the
6666 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6668 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6671 * As blocked tasks retain absolute vruntime the migration needs to
6672 * deal with this by subtracting the old and adding the new
6673 * min_vruntime -- the latter is done by enqueue_entity() when placing
6674 * the task on the new runqueue.
6676 if (p->state == TASK_WAKING) {
6677 struct sched_entity *se = &p->se;
6678 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6681 #ifndef CONFIG_64BIT
6682 u64 min_vruntime_copy;
6685 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6687 min_vruntime = cfs_rq->min_vruntime;
6688 } while (min_vruntime != min_vruntime_copy);
6690 min_vruntime = cfs_rq->min_vruntime;
6693 se->vruntime -= min_vruntime;
6696 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6698 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6699 * rq->lock and can modify state directly.
6701 lockdep_assert_held(&task_rq(p)->lock);
6702 detach_entity_cfs_rq(&p->se);
6706 * We are supposed to update the task to "current" time, then
6707 * its up to date and ready to go to new CPU/cfs_rq. But we
6708 * have difficulty in getting what current time is, so simply
6709 * throw away the out-of-date time. This will result in the
6710 * wakee task is less decayed, but giving the wakee more load
6713 remove_entity_load_avg(&p->se);
6716 /* Tell new CPU we are migrated */
6717 p->se.avg.last_update_time = 0;
6719 /* We have migrated, no longer consider this task hot */
6720 p->se.exec_start = 0;
6722 update_scan_period(p, new_cpu);
6725 static void task_dead_fair(struct task_struct *p)
6727 remove_entity_load_avg(&p->se);
6729 #endif /* CONFIG_SMP */
6731 static unsigned long wakeup_gran(struct sched_entity *se)
6733 unsigned long gran = sysctl_sched_wakeup_granularity;
6736 * Since its curr running now, convert the gran from real-time
6737 * to virtual-time in his units.
6739 * By using 'se' instead of 'curr' we penalize light tasks, so
6740 * they get preempted easier. That is, if 'se' < 'curr' then
6741 * the resulting gran will be larger, therefore penalizing the
6742 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6743 * be smaller, again penalizing the lighter task.
6745 * This is especially important for buddies when the leftmost
6746 * task is higher priority than the buddy.
6748 return calc_delta_fair(gran, se);
6752 * Should 'se' preempt 'curr'.
6766 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6768 s64 gran, vdiff = curr->vruntime - se->vruntime;
6773 gran = wakeup_gran(se);
6780 static void set_last_buddy(struct sched_entity *se)
6782 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6785 for_each_sched_entity(se) {
6786 if (SCHED_WARN_ON(!se->on_rq))
6788 cfs_rq_of(se)->last = se;
6792 static void set_next_buddy(struct sched_entity *se)
6794 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6797 for_each_sched_entity(se) {
6798 if (SCHED_WARN_ON(!se->on_rq))
6800 cfs_rq_of(se)->next = se;
6804 static void set_skip_buddy(struct sched_entity *se)
6806 for_each_sched_entity(se)
6807 cfs_rq_of(se)->skip = se;
6811 * Preempt the current task with a newly woken task if needed:
6813 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6815 struct task_struct *curr = rq->curr;
6816 struct sched_entity *se = &curr->se, *pse = &p->se;
6817 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6818 int scale = cfs_rq->nr_running >= sched_nr_latency;
6819 int next_buddy_marked = 0;
6821 if (unlikely(se == pse))
6825 * This is possible from callers such as attach_tasks(), in which we
6826 * unconditionally check_prempt_curr() after an enqueue (which may have
6827 * lead to a throttle). This both saves work and prevents false
6828 * next-buddy nomination below.
6830 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6833 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6834 set_next_buddy(pse);
6835 next_buddy_marked = 1;
6839 * We can come here with TIF_NEED_RESCHED already set from new task
6842 * Note: this also catches the edge-case of curr being in a throttled
6843 * group (e.g. via set_curr_task), since update_curr() (in the
6844 * enqueue of curr) will have resulted in resched being set. This
6845 * prevents us from potentially nominating it as a false LAST_BUDDY
6848 if (test_tsk_need_resched(curr))
6851 /* Idle tasks are by definition preempted by non-idle tasks. */
6852 if (unlikely(task_has_idle_policy(curr)) &&
6853 likely(!task_has_idle_policy(p)))
6857 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6858 * is driven by the tick):
6860 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6863 find_matching_se(&se, &pse);
6864 update_curr(cfs_rq_of(se));
6866 if (wakeup_preempt_entity(se, pse) == 1) {
6868 * Bias pick_next to pick the sched entity that is
6869 * triggering this preemption.
6871 if (!next_buddy_marked)
6872 set_next_buddy(pse);
6881 * Only set the backward buddy when the current task is still
6882 * on the rq. This can happen when a wakeup gets interleaved
6883 * with schedule on the ->pre_schedule() or idle_balance()
6884 * point, either of which can * drop the rq lock.
6886 * Also, during early boot the idle thread is in the fair class,
6887 * for obvious reasons its a bad idea to schedule back to it.
6889 if (unlikely(!se->on_rq || curr == rq->idle))
6892 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6896 static struct task_struct *
6897 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6899 struct cfs_rq *cfs_rq = &rq->cfs;
6900 struct sched_entity *se;
6901 struct task_struct *p;
6905 if (!cfs_rq->nr_running)
6908 #ifdef CONFIG_FAIR_GROUP_SCHED
6909 if (prev->sched_class != &fair_sched_class)
6913 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6914 * likely that a next task is from the same cgroup as the current.
6916 * Therefore attempt to avoid putting and setting the entire cgroup
6917 * hierarchy, only change the part that actually changes.
6921 struct sched_entity *curr = cfs_rq->curr;
6924 * Since we got here without doing put_prev_entity() we also
6925 * have to consider cfs_rq->curr. If it is still a runnable
6926 * entity, update_curr() will update its vruntime, otherwise
6927 * forget we've ever seen it.
6931 update_curr(cfs_rq);
6936 * This call to check_cfs_rq_runtime() will do the
6937 * throttle and dequeue its entity in the parent(s).
6938 * Therefore the nr_running test will indeed
6941 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6944 if (!cfs_rq->nr_running)
6951 se = pick_next_entity(cfs_rq, curr);
6952 cfs_rq = group_cfs_rq(se);
6958 * Since we haven't yet done put_prev_entity and if the selected task
6959 * is a different task than we started out with, try and touch the
6960 * least amount of cfs_rqs.
6963 struct sched_entity *pse = &prev->se;
6965 while (!(cfs_rq = is_same_group(se, pse))) {
6966 int se_depth = se->depth;
6967 int pse_depth = pse->depth;
6969 if (se_depth <= pse_depth) {
6970 put_prev_entity(cfs_rq_of(pse), pse);
6971 pse = parent_entity(pse);
6973 if (se_depth >= pse_depth) {
6974 set_next_entity(cfs_rq_of(se), se);
6975 se = parent_entity(se);
6979 put_prev_entity(cfs_rq, pse);
6980 set_next_entity(cfs_rq, se);
6987 put_prev_task(rq, prev);
6990 se = pick_next_entity(cfs_rq, NULL);
6991 set_next_entity(cfs_rq, se);
6992 cfs_rq = group_cfs_rq(se);
6997 done: __maybe_unused;
7000 * Move the next running task to the front of
7001 * the list, so our cfs_tasks list becomes MRU
7004 list_move(&p->se.group_node, &rq->cfs_tasks);
7007 if (hrtick_enabled(rq))
7008 hrtick_start_fair(rq, p);
7010 update_misfit_status(p, rq);
7015 update_misfit_status(NULL, rq);
7016 new_tasks = idle_balance(rq, rf);
7019 * Because idle_balance() releases (and re-acquires) rq->lock, it is
7020 * possible for any higher priority task to appear. In that case we
7021 * must re-start the pick_next_entity() loop.
7033 * Account for a descheduled task:
7035 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
7037 struct sched_entity *se = &prev->se;
7038 struct cfs_rq *cfs_rq;
7040 for_each_sched_entity(se) {
7041 cfs_rq = cfs_rq_of(se);
7042 put_prev_entity(cfs_rq, se);
7047 * sched_yield() is very simple
7049 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7051 static void yield_task_fair(struct rq *rq)
7053 struct task_struct *curr = rq->curr;
7054 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7055 struct sched_entity *se = &curr->se;
7058 * Are we the only task in the tree?
7060 if (unlikely(rq->nr_running == 1))
7063 clear_buddies(cfs_rq, se);
7065 if (curr->policy != SCHED_BATCH) {
7066 update_rq_clock(rq);
7068 * Update run-time statistics of the 'current'.
7070 update_curr(cfs_rq);
7072 * Tell update_rq_clock() that we've just updated,
7073 * so we don't do microscopic update in schedule()
7074 * and double the fastpath cost.
7076 rq_clock_skip_update(rq);
7082 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
7084 struct sched_entity *se = &p->se;
7086 /* throttled hierarchies are not runnable */
7087 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
7090 /* Tell the scheduler that we'd really like pse to run next. */
7093 yield_task_fair(rq);
7099 /**************************************************
7100 * Fair scheduling class load-balancing methods.
7104 * The purpose of load-balancing is to achieve the same basic fairness the
7105 * per-CPU scheduler provides, namely provide a proportional amount of compute
7106 * time to each task. This is expressed in the following equation:
7108 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7110 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7111 * W_i,0 is defined as:
7113 * W_i,0 = \Sum_j w_i,j (2)
7115 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7116 * is derived from the nice value as per sched_prio_to_weight[].
7118 * The weight average is an exponential decay average of the instantaneous
7121 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7123 * C_i is the compute capacity of CPU i, typically it is the
7124 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7125 * can also include other factors [XXX].
7127 * To achieve this balance we define a measure of imbalance which follows
7128 * directly from (1):
7130 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7132 * We them move tasks around to minimize the imbalance. In the continuous
7133 * function space it is obvious this converges, in the discrete case we get
7134 * a few fun cases generally called infeasible weight scenarios.
7137 * - infeasible weights;
7138 * - local vs global optima in the discrete case. ]
7143 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7144 * for all i,j solution, we create a tree of CPUs that follows the hardware
7145 * topology where each level pairs two lower groups (or better). This results
7146 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7147 * tree to only the first of the previous level and we decrease the frequency
7148 * of load-balance at each level inv. proportional to the number of CPUs in
7154 * \Sum { --- * --- * 2^i } = O(n) (5)
7156 * `- size of each group
7157 * | | `- number of CPUs doing load-balance
7159 * `- sum over all levels
7161 * Coupled with a limit on how many tasks we can migrate every balance pass,
7162 * this makes (5) the runtime complexity of the balancer.
7164 * An important property here is that each CPU is still (indirectly) connected
7165 * to every other CPU in at most O(log n) steps:
7167 * The adjacency matrix of the resulting graph is given by:
7170 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7173 * And you'll find that:
7175 * A^(log_2 n)_i,j != 0 for all i,j (7)
7177 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7178 * The task movement gives a factor of O(m), giving a convergence complexity
7181 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7186 * In order to avoid CPUs going idle while there's still work to do, new idle
7187 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7188 * tree itself instead of relying on other CPUs to bring it work.
7190 * This adds some complexity to both (5) and (8) but it reduces the total idle
7198 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7201 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7206 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7208 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7210 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7213 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7214 * rewrite all of this once again.]
7217 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7219 enum fbq_type { regular, remote, all };
7228 #define LBF_ALL_PINNED 0x01
7229 #define LBF_NEED_BREAK 0x02
7230 #define LBF_DST_PINNED 0x04
7231 #define LBF_SOME_PINNED 0x08
7232 #define LBF_NOHZ_STATS 0x10
7233 #define LBF_NOHZ_AGAIN 0x20
7236 struct sched_domain *sd;
7244 struct cpumask *dst_grpmask;
7246 enum cpu_idle_type idle;
7248 /* The set of CPUs under consideration for load-balancing */
7249 struct cpumask *cpus;
7254 unsigned int loop_break;
7255 unsigned int loop_max;
7257 enum fbq_type fbq_type;
7258 enum group_type src_grp_type;
7259 struct list_head tasks;
7263 * Is this task likely cache-hot:
7265 static int task_hot(struct task_struct *p, struct lb_env *env)
7269 lockdep_assert_held(&env->src_rq->lock);
7271 if (p->sched_class != &fair_sched_class)
7274 if (unlikely(task_has_idle_policy(p)))
7278 * Buddy candidates are cache hot:
7280 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7281 (&p->se == cfs_rq_of(&p->se)->next ||
7282 &p->se == cfs_rq_of(&p->se)->last))
7285 if (sysctl_sched_migration_cost == -1)
7287 if (sysctl_sched_migration_cost == 0)
7290 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7292 return delta < (s64)sysctl_sched_migration_cost;
7295 #ifdef CONFIG_NUMA_BALANCING
7297 * Returns 1, if task migration degrades locality
7298 * Returns 0, if task migration improves locality i.e migration preferred.
7299 * Returns -1, if task migration is not affected by locality.
7301 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7303 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7304 unsigned long src_weight, dst_weight;
7305 int src_nid, dst_nid, dist;
7307 if (!static_branch_likely(&sched_numa_balancing))
7310 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7313 src_nid = cpu_to_node(env->src_cpu);
7314 dst_nid = cpu_to_node(env->dst_cpu);
7316 if (src_nid == dst_nid)
7319 /* Migrating away from the preferred node is always bad. */
7320 if (src_nid == p->numa_preferred_nid) {
7321 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7327 /* Encourage migration to the preferred node. */
7328 if (dst_nid == p->numa_preferred_nid)
7331 /* Leaving a core idle is often worse than degrading locality. */
7332 if (env->idle == CPU_IDLE)
7335 dist = node_distance(src_nid, dst_nid);
7337 src_weight = group_weight(p, src_nid, dist);
7338 dst_weight = group_weight(p, dst_nid, dist);
7340 src_weight = task_weight(p, src_nid, dist);
7341 dst_weight = task_weight(p, dst_nid, dist);
7344 return dst_weight < src_weight;
7348 static inline int migrate_degrades_locality(struct task_struct *p,
7356 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7359 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7363 lockdep_assert_held(&env->src_rq->lock);
7366 * We do not migrate tasks that are:
7367 * 1) throttled_lb_pair, or
7368 * 2) cannot be migrated to this CPU due to cpus_allowed, or
7369 * 3) running (obviously), or
7370 * 4) are cache-hot on their current CPU.
7372 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7375 if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) {
7378 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7380 env->flags |= LBF_SOME_PINNED;
7383 * Remember if this task can be migrated to any other CPU in
7384 * our sched_group. We may want to revisit it if we couldn't
7385 * meet load balance goals by pulling other tasks on src_cpu.
7387 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7388 * already computed one in current iteration.
7390 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7393 /* Prevent to re-select dst_cpu via env's CPUs: */
7394 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7395 if (cpumask_test_cpu(cpu, &p->cpus_allowed)) {
7396 env->flags |= LBF_DST_PINNED;
7397 env->new_dst_cpu = cpu;
7405 /* Record that we found atleast one task that could run on dst_cpu */
7406 env->flags &= ~LBF_ALL_PINNED;
7408 if (task_running(env->src_rq, p)) {
7409 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7414 * Aggressive migration if:
7415 * 1) destination numa is preferred
7416 * 2) task is cache cold, or
7417 * 3) too many balance attempts have failed.
7419 tsk_cache_hot = migrate_degrades_locality(p, env);
7420 if (tsk_cache_hot == -1)
7421 tsk_cache_hot = task_hot(p, env);
7423 if (tsk_cache_hot <= 0 ||
7424 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7425 if (tsk_cache_hot == 1) {
7426 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7427 schedstat_inc(p->se.statistics.nr_forced_migrations);
7432 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7437 * detach_task() -- detach the task for the migration specified in env
7439 static void detach_task(struct task_struct *p, struct lb_env *env)
7441 lockdep_assert_held(&env->src_rq->lock);
7443 p->on_rq = TASK_ON_RQ_MIGRATING;
7444 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7445 set_task_cpu(p, env->dst_cpu);
7449 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7450 * part of active balancing operations within "domain".
7452 * Returns a task if successful and NULL otherwise.
7454 static struct task_struct *detach_one_task(struct lb_env *env)
7456 struct task_struct *p;
7458 lockdep_assert_held(&env->src_rq->lock);
7460 list_for_each_entry_reverse(p,
7461 &env->src_rq->cfs_tasks, se.group_node) {
7462 if (!can_migrate_task(p, env))
7465 detach_task(p, env);
7468 * Right now, this is only the second place where
7469 * lb_gained[env->idle] is updated (other is detach_tasks)
7470 * so we can safely collect stats here rather than
7471 * inside detach_tasks().
7473 schedstat_inc(env->sd->lb_gained[env->idle]);
7479 static const unsigned int sched_nr_migrate_break = 32;
7482 * detach_tasks() -- tries to detach up to imbalance weighted load from
7483 * busiest_rq, as part of a balancing operation within domain "sd".
7485 * Returns number of detached tasks if successful and 0 otherwise.
7487 static int detach_tasks(struct lb_env *env)
7489 struct list_head *tasks = &env->src_rq->cfs_tasks;
7490 struct task_struct *p;
7494 lockdep_assert_held(&env->src_rq->lock);
7496 if (env->imbalance <= 0)
7499 while (!list_empty(tasks)) {
7501 * We don't want to steal all, otherwise we may be treated likewise,
7502 * which could at worst lead to a livelock crash.
7504 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7507 p = list_last_entry(tasks, struct task_struct, se.group_node);
7510 /* We've more or less seen every task there is, call it quits */
7511 if (env->loop > env->loop_max)
7514 /* take a breather every nr_migrate tasks */
7515 if (env->loop > env->loop_break) {
7516 env->loop_break += sched_nr_migrate_break;
7517 env->flags |= LBF_NEED_BREAK;
7521 if (!can_migrate_task(p, env))
7524 load = task_h_load(p);
7526 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7529 if ((load / 2) > env->imbalance)
7532 detach_task(p, env);
7533 list_add(&p->se.group_node, &env->tasks);
7536 env->imbalance -= load;
7538 #ifdef CONFIG_PREEMPT
7540 * NEWIDLE balancing is a source of latency, so preemptible
7541 * kernels will stop after the first task is detached to minimize
7542 * the critical section.
7544 if (env->idle == CPU_NEWLY_IDLE)
7549 * We only want to steal up to the prescribed amount of
7552 if (env->imbalance <= 0)
7557 list_move(&p->se.group_node, tasks);
7561 * Right now, this is one of only two places we collect this stat
7562 * so we can safely collect detach_one_task() stats here rather
7563 * than inside detach_one_task().
7565 schedstat_add(env->sd->lb_gained[env->idle], detached);
7571 * attach_task() -- attach the task detached by detach_task() to its new rq.
7573 static void attach_task(struct rq *rq, struct task_struct *p)
7575 lockdep_assert_held(&rq->lock);
7577 BUG_ON(task_rq(p) != rq);
7578 activate_task(rq, p, ENQUEUE_NOCLOCK);
7579 p->on_rq = TASK_ON_RQ_QUEUED;
7580 check_preempt_curr(rq, p, 0);
7584 * attach_one_task() -- attaches the task returned from detach_one_task() to
7587 static void attach_one_task(struct rq *rq, struct task_struct *p)
7592 update_rq_clock(rq);
7598 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7601 static void attach_tasks(struct lb_env *env)
7603 struct list_head *tasks = &env->tasks;
7604 struct task_struct *p;
7607 rq_lock(env->dst_rq, &rf);
7608 update_rq_clock(env->dst_rq);
7610 while (!list_empty(tasks)) {
7611 p = list_first_entry(tasks, struct task_struct, se.group_node);
7612 list_del_init(&p->se.group_node);
7614 attach_task(env->dst_rq, p);
7617 rq_unlock(env->dst_rq, &rf);
7620 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7622 if (cfs_rq->avg.load_avg)
7625 if (cfs_rq->avg.util_avg)
7631 static inline bool others_have_blocked(struct rq *rq)
7633 if (READ_ONCE(rq->avg_rt.util_avg))
7636 if (READ_ONCE(rq->avg_dl.util_avg))
7639 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7640 if (READ_ONCE(rq->avg_irq.util_avg))
7647 #ifdef CONFIG_FAIR_GROUP_SCHED
7649 static void update_blocked_averages(int cpu)
7651 struct rq *rq = cpu_rq(cpu);
7652 struct cfs_rq *cfs_rq;
7653 const struct sched_class *curr_class;
7657 rq_lock_irqsave(rq, &rf);
7658 update_rq_clock(rq);
7661 * Iterates the task_group tree in a bottom up fashion, see
7662 * list_add_leaf_cfs_rq() for details.
7664 for_each_leaf_cfs_rq(rq, cfs_rq) {
7665 struct sched_entity *se;
7667 /* throttled entities do not contribute to load */
7668 if (throttled_hierarchy(cfs_rq))
7671 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
7672 update_tg_load_avg(cfs_rq, 0);
7674 /* Propagate pending load changes to the parent, if any: */
7675 se = cfs_rq->tg->se[cpu];
7676 if (se && !skip_blocked_update(se))
7677 update_load_avg(cfs_rq_of(se), se, 0);
7679 /* Don't need periodic decay once load/util_avg are null */
7680 if (cfs_rq_has_blocked(cfs_rq))
7684 curr_class = rq->curr->sched_class;
7685 update_rt_rq_load_avg(rq_clock_task(rq), rq, curr_class == &rt_sched_class);
7686 update_dl_rq_load_avg(rq_clock_task(rq), rq, curr_class == &dl_sched_class);
7687 update_irq_load_avg(rq, 0);
7688 /* Don't need periodic decay once load/util_avg are null */
7689 if (others_have_blocked(rq))
7692 #ifdef CONFIG_NO_HZ_COMMON
7693 rq->last_blocked_load_update_tick = jiffies;
7695 rq->has_blocked_load = 0;
7697 rq_unlock_irqrestore(rq, &rf);
7701 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7702 * This needs to be done in a top-down fashion because the load of a child
7703 * group is a fraction of its parents load.
7705 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7707 struct rq *rq = rq_of(cfs_rq);
7708 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7709 unsigned long now = jiffies;
7712 if (cfs_rq->last_h_load_update == now)
7715 cfs_rq->h_load_next = NULL;
7716 for_each_sched_entity(se) {
7717 cfs_rq = cfs_rq_of(se);
7718 cfs_rq->h_load_next = se;
7719 if (cfs_rq->last_h_load_update == now)
7724 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7725 cfs_rq->last_h_load_update = now;
7728 while ((se = cfs_rq->h_load_next) != NULL) {
7729 load = cfs_rq->h_load;
7730 load = div64_ul(load * se->avg.load_avg,
7731 cfs_rq_load_avg(cfs_rq) + 1);
7732 cfs_rq = group_cfs_rq(se);
7733 cfs_rq->h_load = load;
7734 cfs_rq->last_h_load_update = now;
7738 static unsigned long task_h_load(struct task_struct *p)
7740 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7742 update_cfs_rq_h_load(cfs_rq);
7743 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7744 cfs_rq_load_avg(cfs_rq) + 1);
7747 static inline void update_blocked_averages(int cpu)
7749 struct rq *rq = cpu_rq(cpu);
7750 struct cfs_rq *cfs_rq = &rq->cfs;
7751 const struct sched_class *curr_class;
7754 rq_lock_irqsave(rq, &rf);
7755 update_rq_clock(rq);
7756 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
7758 curr_class = rq->curr->sched_class;
7759 update_rt_rq_load_avg(rq_clock_task(rq), rq, curr_class == &rt_sched_class);
7760 update_dl_rq_load_avg(rq_clock_task(rq), rq, curr_class == &dl_sched_class);
7761 update_irq_load_avg(rq, 0);
7762 #ifdef CONFIG_NO_HZ_COMMON
7763 rq->last_blocked_load_update_tick = jiffies;
7764 if (!cfs_rq_has_blocked(cfs_rq) && !others_have_blocked(rq))
7765 rq->has_blocked_load = 0;
7767 rq_unlock_irqrestore(rq, &rf);
7770 static unsigned long task_h_load(struct task_struct *p)
7772 return p->se.avg.load_avg;
7776 /********** Helpers for find_busiest_group ************************/
7779 * sg_lb_stats - stats of a sched_group required for load_balancing
7781 struct sg_lb_stats {
7782 unsigned long avg_load; /*Avg load across the CPUs of the group */
7783 unsigned long group_load; /* Total load over the CPUs of the group */
7784 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7785 unsigned long load_per_task;
7786 unsigned long group_capacity;
7787 unsigned long group_util; /* Total utilization of the group */
7788 unsigned int sum_nr_running; /* Nr tasks running in the group */
7789 unsigned int idle_cpus;
7790 unsigned int group_weight;
7791 enum group_type group_type;
7792 int group_no_capacity;
7793 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
7794 #ifdef CONFIG_NUMA_BALANCING
7795 unsigned int nr_numa_running;
7796 unsigned int nr_preferred_running;
7801 * sd_lb_stats - Structure to store the statistics of a sched_domain
7802 * during load balancing.
7804 struct sd_lb_stats {
7805 struct sched_group *busiest; /* Busiest group in this sd */
7806 struct sched_group *local; /* Local group in this sd */
7807 unsigned long total_running;
7808 unsigned long total_load; /* Total load of all groups in sd */
7809 unsigned long total_capacity; /* Total capacity of all groups in sd */
7810 unsigned long avg_load; /* Average load across all groups in sd */
7812 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7813 struct sg_lb_stats local_stat; /* Statistics of the local group */
7816 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7819 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7820 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7821 * We must however clear busiest_stat::avg_load because
7822 * update_sd_pick_busiest() reads this before assignment.
7824 *sds = (struct sd_lb_stats){
7827 .total_running = 0UL,
7829 .total_capacity = 0UL,
7832 .sum_nr_running = 0,
7833 .group_type = group_other,
7839 * get_sd_load_idx - Obtain the load index for a given sched domain.
7840 * @sd: The sched_domain whose load_idx is to be obtained.
7841 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7843 * Return: The load index.
7845 static inline int get_sd_load_idx(struct sched_domain *sd,
7846 enum cpu_idle_type idle)
7852 load_idx = sd->busy_idx;
7855 case CPU_NEWLY_IDLE:
7856 load_idx = sd->newidle_idx;
7859 load_idx = sd->idle_idx;
7866 static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu)
7868 struct rq *rq = cpu_rq(cpu);
7869 unsigned long max = arch_scale_cpu_capacity(sd, cpu);
7870 unsigned long used, free;
7873 irq = cpu_util_irq(rq);
7875 if (unlikely(irq >= max))
7878 used = READ_ONCE(rq->avg_rt.util_avg);
7879 used += READ_ONCE(rq->avg_dl.util_avg);
7881 if (unlikely(used >= max))
7886 return scale_irq_capacity(free, irq, max);
7889 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7891 unsigned long capacity = scale_rt_capacity(sd, cpu);
7892 struct sched_group *sdg = sd->groups;
7894 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(sd, cpu);
7899 cpu_rq(cpu)->cpu_capacity = capacity;
7900 sdg->sgc->capacity = capacity;
7901 sdg->sgc->min_capacity = capacity;
7902 sdg->sgc->max_capacity = capacity;
7905 void update_group_capacity(struct sched_domain *sd, int cpu)
7907 struct sched_domain *child = sd->child;
7908 struct sched_group *group, *sdg = sd->groups;
7909 unsigned long capacity, min_capacity, max_capacity;
7910 unsigned long interval;
7912 interval = msecs_to_jiffies(sd->balance_interval);
7913 interval = clamp(interval, 1UL, max_load_balance_interval);
7914 sdg->sgc->next_update = jiffies + interval;
7917 update_cpu_capacity(sd, cpu);
7922 min_capacity = ULONG_MAX;
7925 if (child->flags & SD_OVERLAP) {
7927 * SD_OVERLAP domains cannot assume that child groups
7928 * span the current group.
7931 for_each_cpu(cpu, sched_group_span(sdg)) {
7932 struct sched_group_capacity *sgc;
7933 struct rq *rq = cpu_rq(cpu);
7936 * build_sched_domains() -> init_sched_groups_capacity()
7937 * gets here before we've attached the domains to the
7940 * Use capacity_of(), which is set irrespective of domains
7941 * in update_cpu_capacity().
7943 * This avoids capacity from being 0 and
7944 * causing divide-by-zero issues on boot.
7946 if (unlikely(!rq->sd)) {
7947 capacity += capacity_of(cpu);
7949 sgc = rq->sd->groups->sgc;
7950 capacity += sgc->capacity;
7953 min_capacity = min(capacity, min_capacity);
7954 max_capacity = max(capacity, max_capacity);
7958 * !SD_OVERLAP domains can assume that child groups
7959 * span the current group.
7962 group = child->groups;
7964 struct sched_group_capacity *sgc = group->sgc;
7966 capacity += sgc->capacity;
7967 min_capacity = min(sgc->min_capacity, min_capacity);
7968 max_capacity = max(sgc->max_capacity, max_capacity);
7969 group = group->next;
7970 } while (group != child->groups);
7973 sdg->sgc->capacity = capacity;
7974 sdg->sgc->min_capacity = min_capacity;
7975 sdg->sgc->max_capacity = max_capacity;
7979 * Check whether the capacity of the rq has been noticeably reduced by side
7980 * activity. The imbalance_pct is used for the threshold.
7981 * Return true is the capacity is reduced
7984 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7986 return ((rq->cpu_capacity * sd->imbalance_pct) <
7987 (rq->cpu_capacity_orig * 100));
7991 * Group imbalance indicates (and tries to solve) the problem where balancing
7992 * groups is inadequate due to ->cpus_allowed constraints.
7994 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
7995 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
7998 * { 0 1 2 3 } { 4 5 6 7 }
8001 * If we were to balance group-wise we'd place two tasks in the first group and
8002 * two tasks in the second group. Clearly this is undesired as it will overload
8003 * cpu 3 and leave one of the CPUs in the second group unused.
8005 * The current solution to this issue is detecting the skew in the first group
8006 * by noticing the lower domain failed to reach balance and had difficulty
8007 * moving tasks due to affinity constraints.
8009 * When this is so detected; this group becomes a candidate for busiest; see
8010 * update_sd_pick_busiest(). And calculate_imbalance() and
8011 * find_busiest_group() avoid some of the usual balance conditions to allow it
8012 * to create an effective group imbalance.
8014 * This is a somewhat tricky proposition since the next run might not find the
8015 * group imbalance and decide the groups need to be balanced again. A most
8016 * subtle and fragile situation.
8019 static inline int sg_imbalanced(struct sched_group *group)
8021 return group->sgc->imbalance;
8025 * group_has_capacity returns true if the group has spare capacity that could
8026 * be used by some tasks.
8027 * We consider that a group has spare capacity if the * number of task is
8028 * smaller than the number of CPUs or if the utilization is lower than the
8029 * available capacity for CFS tasks.
8030 * For the latter, we use a threshold to stabilize the state, to take into
8031 * account the variance of the tasks' load and to return true if the available
8032 * capacity in meaningful for the load balancer.
8033 * As an example, an available capacity of 1% can appear but it doesn't make
8034 * any benefit for the load balance.
8037 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
8039 if (sgs->sum_nr_running < sgs->group_weight)
8042 if ((sgs->group_capacity * 100) >
8043 (sgs->group_util * env->sd->imbalance_pct))
8050 * group_is_overloaded returns true if the group has more tasks than it can
8052 * group_is_overloaded is not equals to !group_has_capacity because a group
8053 * with the exact right number of tasks, has no more spare capacity but is not
8054 * overloaded so both group_has_capacity and group_is_overloaded return
8058 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
8060 if (sgs->sum_nr_running <= sgs->group_weight)
8063 if ((sgs->group_capacity * 100) <
8064 (sgs->group_util * env->sd->imbalance_pct))
8071 * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller
8072 * per-CPU capacity than sched_group ref.
8075 group_smaller_min_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
8077 return sg->sgc->min_capacity * capacity_margin <
8078 ref->sgc->min_capacity * 1024;
8082 * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller
8083 * per-CPU capacity_orig than sched_group ref.
8086 group_smaller_max_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
8088 return sg->sgc->max_capacity * capacity_margin <
8089 ref->sgc->max_capacity * 1024;
8093 group_type group_classify(struct sched_group *group,
8094 struct sg_lb_stats *sgs)
8096 if (sgs->group_no_capacity)
8097 return group_overloaded;
8099 if (sg_imbalanced(group))
8100 return group_imbalanced;
8102 if (sgs->group_misfit_task_load)
8103 return group_misfit_task;
8108 static bool update_nohz_stats(struct rq *rq, bool force)
8110 #ifdef CONFIG_NO_HZ_COMMON
8111 unsigned int cpu = rq->cpu;
8113 if (!rq->has_blocked_load)
8116 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
8119 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
8122 update_blocked_averages(cpu);
8124 return rq->has_blocked_load;
8131 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8132 * @env: The load balancing environment.
8133 * @group: sched_group whose statistics are to be updated.
8134 * @sgs: variable to hold the statistics for this group.
8135 * @sg_status: Holds flag indicating the status of the sched_group
8137 static inline void update_sg_lb_stats(struct lb_env *env,
8138 struct sched_group *group,
8139 struct sg_lb_stats *sgs,
8142 int local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(group));
8143 int load_idx = get_sd_load_idx(env->sd, env->idle);
8147 memset(sgs, 0, sizeof(*sgs));
8149 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8150 struct rq *rq = cpu_rq(i);
8152 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
8153 env->flags |= LBF_NOHZ_AGAIN;
8155 /* Bias balancing toward CPUs of our domain: */
8157 load = target_load(i, load_idx);
8159 load = source_load(i, load_idx);
8161 sgs->group_load += load;
8162 sgs->group_util += cpu_util(i);
8163 sgs->sum_nr_running += rq->cfs.h_nr_running;
8165 nr_running = rq->nr_running;
8167 *sg_status |= SG_OVERLOAD;
8169 if (cpu_overutilized(i))
8170 *sg_status |= SG_OVERUTILIZED;
8172 #ifdef CONFIG_NUMA_BALANCING
8173 sgs->nr_numa_running += rq->nr_numa_running;
8174 sgs->nr_preferred_running += rq->nr_preferred_running;
8176 sgs->sum_weighted_load += weighted_cpuload(rq);
8178 * No need to call idle_cpu() if nr_running is not 0
8180 if (!nr_running && idle_cpu(i))
8183 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8184 sgs->group_misfit_task_load < rq->misfit_task_load) {
8185 sgs->group_misfit_task_load = rq->misfit_task_load;
8186 *sg_status |= SG_OVERLOAD;
8190 /* Adjust by relative CPU capacity of the group */
8191 sgs->group_capacity = group->sgc->capacity;
8192 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
8194 if (sgs->sum_nr_running)
8195 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
8197 sgs->group_weight = group->group_weight;
8199 sgs->group_no_capacity = group_is_overloaded(env, sgs);
8200 sgs->group_type = group_classify(group, sgs);
8204 * update_sd_pick_busiest - return 1 on busiest group
8205 * @env: The load balancing environment.
8206 * @sds: sched_domain statistics
8207 * @sg: sched_group candidate to be checked for being the busiest
8208 * @sgs: sched_group statistics
8210 * Determine if @sg is a busier group than the previously selected
8213 * Return: %true if @sg is a busier group than the previously selected
8214 * busiest group. %false otherwise.
8216 static bool update_sd_pick_busiest(struct lb_env *env,
8217 struct sd_lb_stats *sds,
8218 struct sched_group *sg,
8219 struct sg_lb_stats *sgs)
8221 struct sg_lb_stats *busiest = &sds->busiest_stat;
8224 * Don't try to pull misfit tasks we can't help.
8225 * We can use max_capacity here as reduction in capacity on some
8226 * CPUs in the group should either be possible to resolve
8227 * internally or be covered by avg_load imbalance (eventually).
8229 if (sgs->group_type == group_misfit_task &&
8230 (!group_smaller_max_cpu_capacity(sg, sds->local) ||
8231 !group_has_capacity(env, &sds->local_stat)))
8234 if (sgs->group_type > busiest->group_type)
8237 if (sgs->group_type < busiest->group_type)
8240 if (sgs->avg_load <= busiest->avg_load)
8243 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
8247 * Candidate sg has no more than one task per CPU and
8248 * has higher per-CPU capacity. Migrating tasks to less
8249 * capable CPUs may harm throughput. Maximize throughput,
8250 * power/energy consequences are not considered.
8252 if (sgs->sum_nr_running <= sgs->group_weight &&
8253 group_smaller_min_cpu_capacity(sds->local, sg))
8257 * If we have more than one misfit sg go with the biggest misfit.
8259 if (sgs->group_type == group_misfit_task &&
8260 sgs->group_misfit_task_load < busiest->group_misfit_task_load)
8264 /* This is the busiest node in its class. */
8265 if (!(env->sd->flags & SD_ASYM_PACKING))
8268 /* No ASYM_PACKING if target CPU is already busy */
8269 if (env->idle == CPU_NOT_IDLE)
8272 * ASYM_PACKING needs to move all the work to the highest
8273 * prority CPUs in the group, therefore mark all groups
8274 * of lower priority than ourself as busy.
8276 if (sgs->sum_nr_running &&
8277 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
8281 /* Prefer to move from lowest priority CPU's work */
8282 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
8283 sg->asym_prefer_cpu))
8290 #ifdef CONFIG_NUMA_BALANCING
8291 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8293 if (sgs->sum_nr_running > sgs->nr_numa_running)
8295 if (sgs->sum_nr_running > sgs->nr_preferred_running)
8300 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8302 if (rq->nr_running > rq->nr_numa_running)
8304 if (rq->nr_running > rq->nr_preferred_running)
8309 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8314 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8318 #endif /* CONFIG_NUMA_BALANCING */
8321 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8322 * @env: The load balancing environment.
8323 * @sds: variable to hold the statistics for this sched_domain.
8325 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8327 struct sched_domain *child = env->sd->child;
8328 struct sched_group *sg = env->sd->groups;
8329 struct sg_lb_stats *local = &sds->local_stat;
8330 struct sg_lb_stats tmp_sgs;
8331 bool prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
8334 #ifdef CONFIG_NO_HZ_COMMON
8335 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
8336 env->flags |= LBF_NOHZ_STATS;
8340 struct sg_lb_stats *sgs = &tmp_sgs;
8343 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
8348 if (env->idle != CPU_NEWLY_IDLE ||
8349 time_after_eq(jiffies, sg->sgc->next_update))
8350 update_group_capacity(env->sd, env->dst_cpu);
8353 update_sg_lb_stats(env, sg, sgs, &sg_status);
8359 * In case the child domain prefers tasks go to siblings
8360 * first, lower the sg capacity so that we'll try
8361 * and move all the excess tasks away. We lower the capacity
8362 * of a group only if the local group has the capacity to fit
8363 * these excess tasks. The extra check prevents the case where
8364 * you always pull from the heaviest group when it is already
8365 * under-utilized (possible with a large weight task outweighs
8366 * the tasks on the system).
8368 if (prefer_sibling && sds->local &&
8369 group_has_capacity(env, local) &&
8370 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
8371 sgs->group_no_capacity = 1;
8372 sgs->group_type = group_classify(sg, sgs);
8375 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8377 sds->busiest_stat = *sgs;
8381 /* Now, start updating sd_lb_stats */
8382 sds->total_running += sgs->sum_nr_running;
8383 sds->total_load += sgs->group_load;
8384 sds->total_capacity += sgs->group_capacity;
8387 } while (sg != env->sd->groups);
8389 #ifdef CONFIG_NO_HZ_COMMON
8390 if ((env->flags & LBF_NOHZ_AGAIN) &&
8391 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8393 WRITE_ONCE(nohz.next_blocked,
8394 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8398 if (env->sd->flags & SD_NUMA)
8399 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8401 if (!env->sd->parent) {
8402 struct root_domain *rd = env->dst_rq->rd;
8404 /* update overload indicator if we are at root domain */
8405 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
8407 /* Update over-utilization (tipping point, U >= 0) indicator */
8408 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
8409 } else if (sg_status & SG_OVERUTILIZED) {
8410 WRITE_ONCE(env->dst_rq->rd->overutilized, SG_OVERUTILIZED);
8415 * check_asym_packing - Check to see if the group is packed into the
8418 * This is primarily intended to used at the sibling level. Some
8419 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8420 * case of POWER7, it can move to lower SMT modes only when higher
8421 * threads are idle. When in lower SMT modes, the threads will
8422 * perform better since they share less core resources. Hence when we
8423 * have idle threads, we want them to be the higher ones.
8425 * This packing function is run on idle threads. It checks to see if
8426 * the busiest CPU in this domain (core in the P7 case) has a higher
8427 * CPU number than the packing function is being run on. Here we are
8428 * assuming lower CPU number will be equivalent to lower a SMT thread
8431 * Return: 1 when packing is required and a task should be moved to
8432 * this CPU. The amount of the imbalance is returned in env->imbalance.
8434 * @env: The load balancing environment.
8435 * @sds: Statistics of the sched_domain which is to be packed
8437 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8441 if (!(env->sd->flags & SD_ASYM_PACKING))
8444 if (env->idle == CPU_NOT_IDLE)
8450 busiest_cpu = sds->busiest->asym_prefer_cpu;
8451 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
8454 env->imbalance = DIV_ROUND_CLOSEST(
8455 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
8456 SCHED_CAPACITY_SCALE);
8462 * fix_small_imbalance - Calculate the minor imbalance that exists
8463 * amongst the groups of a sched_domain, during
8465 * @env: The load balancing environment.
8466 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8469 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8471 unsigned long tmp, capa_now = 0, capa_move = 0;
8472 unsigned int imbn = 2;
8473 unsigned long scaled_busy_load_per_task;
8474 struct sg_lb_stats *local, *busiest;
8476 local = &sds->local_stat;
8477 busiest = &sds->busiest_stat;
8479 if (!local->sum_nr_running)
8480 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8481 else if (busiest->load_per_task > local->load_per_task)
8484 scaled_busy_load_per_task =
8485 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8486 busiest->group_capacity;
8488 if (busiest->avg_load + scaled_busy_load_per_task >=
8489 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8490 env->imbalance = busiest->load_per_task;
8495 * OK, we don't have enough imbalance to justify moving tasks,
8496 * however we may be able to increase total CPU capacity used by
8500 capa_now += busiest->group_capacity *
8501 min(busiest->load_per_task, busiest->avg_load);
8502 capa_now += local->group_capacity *
8503 min(local->load_per_task, local->avg_load);
8504 capa_now /= SCHED_CAPACITY_SCALE;
8506 /* Amount of load we'd subtract */
8507 if (busiest->avg_load > scaled_busy_load_per_task) {
8508 capa_move += busiest->group_capacity *
8509 min(busiest->load_per_task,
8510 busiest->avg_load - scaled_busy_load_per_task);
8513 /* Amount of load we'd add */
8514 if (busiest->avg_load * busiest->group_capacity <
8515 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8516 tmp = (busiest->avg_load * busiest->group_capacity) /
8517 local->group_capacity;
8519 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8520 local->group_capacity;
8522 capa_move += local->group_capacity *
8523 min(local->load_per_task, local->avg_load + tmp);
8524 capa_move /= SCHED_CAPACITY_SCALE;
8526 /* Move if we gain throughput */
8527 if (capa_move > capa_now)
8528 env->imbalance = busiest->load_per_task;
8532 * calculate_imbalance - Calculate the amount of imbalance present within the
8533 * groups of a given sched_domain during load balance.
8534 * @env: load balance environment
8535 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8537 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8539 unsigned long max_pull, load_above_capacity = ~0UL;
8540 struct sg_lb_stats *local, *busiest;
8542 local = &sds->local_stat;
8543 busiest = &sds->busiest_stat;
8545 if (busiest->group_type == group_imbalanced) {
8547 * In the group_imb case we cannot rely on group-wide averages
8548 * to ensure CPU-load equilibrium, look at wider averages. XXX
8550 busiest->load_per_task =
8551 min(busiest->load_per_task, sds->avg_load);
8555 * Avg load of busiest sg can be less and avg load of local sg can
8556 * be greater than avg load across all sgs of sd because avg load
8557 * factors in sg capacity and sgs with smaller group_type are
8558 * skipped when updating the busiest sg:
8560 if (busiest->group_type != group_misfit_task &&
8561 (busiest->avg_load <= sds->avg_load ||
8562 local->avg_load >= sds->avg_load)) {
8564 return fix_small_imbalance(env, sds);
8568 * If there aren't any idle CPUs, avoid creating some.
8570 if (busiest->group_type == group_overloaded &&
8571 local->group_type == group_overloaded) {
8572 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
8573 if (load_above_capacity > busiest->group_capacity) {
8574 load_above_capacity -= busiest->group_capacity;
8575 load_above_capacity *= scale_load_down(NICE_0_LOAD);
8576 load_above_capacity /= busiest->group_capacity;
8578 load_above_capacity = ~0UL;
8582 * We're trying to get all the CPUs to the average_load, so we don't
8583 * want to push ourselves above the average load, nor do we wish to
8584 * reduce the max loaded CPU below the average load. At the same time,
8585 * we also don't want to reduce the group load below the group
8586 * capacity. Thus we look for the minimum possible imbalance.
8588 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8590 /* How much load to actually move to equalise the imbalance */
8591 env->imbalance = min(
8592 max_pull * busiest->group_capacity,
8593 (sds->avg_load - local->avg_load) * local->group_capacity
8594 ) / SCHED_CAPACITY_SCALE;
8596 /* Boost imbalance to allow misfit task to be balanced. */
8597 if (busiest->group_type == group_misfit_task) {
8598 env->imbalance = max_t(long, env->imbalance,
8599 busiest->group_misfit_task_load);
8603 * if *imbalance is less than the average load per runnable task
8604 * there is no guarantee that any tasks will be moved so we'll have
8605 * a think about bumping its value to force at least one task to be
8608 if (env->imbalance < busiest->load_per_task)
8609 return fix_small_imbalance(env, sds);
8612 /******* find_busiest_group() helpers end here *********************/
8615 * find_busiest_group - Returns the busiest group within the sched_domain
8616 * if there is an imbalance.
8618 * Also calculates the amount of weighted load which should be moved
8619 * to restore balance.
8621 * @env: The load balancing environment.
8623 * Return: - The busiest group if imbalance exists.
8625 static struct sched_group *find_busiest_group(struct lb_env *env)
8627 struct sg_lb_stats *local, *busiest;
8628 struct sd_lb_stats sds;
8630 init_sd_lb_stats(&sds);
8633 * Compute the various statistics relavent for load balancing at
8636 update_sd_lb_stats(env, &sds);
8638 if (static_branch_unlikely(&sched_energy_present)) {
8639 struct root_domain *rd = env->dst_rq->rd;
8641 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
8645 local = &sds.local_stat;
8646 busiest = &sds.busiest_stat;
8648 /* ASYM feature bypasses nice load balance check */
8649 if (check_asym_packing(env, &sds))
8652 /* There is no busy sibling group to pull tasks from */
8653 if (!sds.busiest || busiest->sum_nr_running == 0)
8656 /* XXX broken for overlapping NUMA groups */
8657 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8658 / sds.total_capacity;
8661 * If the busiest group is imbalanced the below checks don't
8662 * work because they assume all things are equal, which typically
8663 * isn't true due to cpus_allowed constraints and the like.
8665 if (busiest->group_type == group_imbalanced)
8669 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8670 * capacities from resulting in underutilization due to avg_load.
8672 if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8673 busiest->group_no_capacity)
8676 /* Misfit tasks should be dealt with regardless of the avg load */
8677 if (busiest->group_type == group_misfit_task)
8681 * If the local group is busier than the selected busiest group
8682 * don't try and pull any tasks.
8684 if (local->avg_load >= busiest->avg_load)
8688 * Don't pull any tasks if this group is already above the domain
8691 if (local->avg_load >= sds.avg_load)
8694 if (env->idle == CPU_IDLE) {
8696 * This CPU is idle. If the busiest group is not overloaded
8697 * and there is no imbalance between this and busiest group
8698 * wrt idle CPUs, it is balanced. The imbalance becomes
8699 * significant if the diff is greater than 1 otherwise we
8700 * might end up to just move the imbalance on another group
8702 if ((busiest->group_type != group_overloaded) &&
8703 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8707 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8708 * imbalance_pct to be conservative.
8710 if (100 * busiest->avg_load <=
8711 env->sd->imbalance_pct * local->avg_load)
8716 /* Looks like there is an imbalance. Compute it */
8717 env->src_grp_type = busiest->group_type;
8718 calculate_imbalance(env, &sds);
8719 return env->imbalance ? sds.busiest : NULL;
8727 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
8729 static struct rq *find_busiest_queue(struct lb_env *env,
8730 struct sched_group *group)
8732 struct rq *busiest = NULL, *rq;
8733 unsigned long busiest_load = 0, busiest_capacity = 1;
8736 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8737 unsigned long capacity, wl;
8741 rt = fbq_classify_rq(rq);
8744 * We classify groups/runqueues into three groups:
8745 * - regular: there are !numa tasks
8746 * - remote: there are numa tasks that run on the 'wrong' node
8747 * - all: there is no distinction
8749 * In order to avoid migrating ideally placed numa tasks,
8750 * ignore those when there's better options.
8752 * If we ignore the actual busiest queue to migrate another
8753 * task, the next balance pass can still reduce the busiest
8754 * queue by moving tasks around inside the node.
8756 * If we cannot move enough load due to this classification
8757 * the next pass will adjust the group classification and
8758 * allow migration of more tasks.
8760 * Both cases only affect the total convergence complexity.
8762 if (rt > env->fbq_type)
8766 * For ASYM_CPUCAPACITY domains with misfit tasks we simply
8767 * seek the "biggest" misfit task.
8769 if (env->src_grp_type == group_misfit_task) {
8770 if (rq->misfit_task_load > busiest_load) {
8771 busiest_load = rq->misfit_task_load;
8778 capacity = capacity_of(i);
8781 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
8782 * eventually lead to active_balancing high->low capacity.
8783 * Higher per-CPU capacity is considered better than balancing
8786 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8787 capacity_of(env->dst_cpu) < capacity &&
8788 rq->nr_running == 1)
8791 wl = weighted_cpuload(rq);
8794 * When comparing with imbalance, use weighted_cpuload()
8795 * which is not scaled with the CPU capacity.
8798 if (rq->nr_running == 1 && wl > env->imbalance &&
8799 !check_cpu_capacity(rq, env->sd))
8803 * For the load comparisons with the other CPU's, consider
8804 * the weighted_cpuload() scaled with the CPU capacity, so
8805 * that the load can be moved away from the CPU that is
8806 * potentially running at a lower capacity.
8808 * Thus we're looking for max(wl_i / capacity_i), crosswise
8809 * multiplication to rid ourselves of the division works out
8810 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8811 * our previous maximum.
8813 if (wl * busiest_capacity > busiest_load * capacity) {
8815 busiest_capacity = capacity;
8824 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8825 * so long as it is large enough.
8827 #define MAX_PINNED_INTERVAL 512
8829 static int need_active_balance(struct lb_env *env)
8831 struct sched_domain *sd = env->sd;
8833 if (env->idle == CPU_NEWLY_IDLE) {
8836 * ASYM_PACKING needs to force migrate tasks from busy but
8837 * lower priority CPUs in order to pack all tasks in the
8838 * highest priority CPUs.
8840 if ((sd->flags & SD_ASYM_PACKING) &&
8841 sched_asym_prefer(env->dst_cpu, env->src_cpu))
8846 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8847 * It's worth migrating the task if the src_cpu's capacity is reduced
8848 * because of other sched_class or IRQs if more capacity stays
8849 * available on dst_cpu.
8851 if ((env->idle != CPU_NOT_IDLE) &&
8852 (env->src_rq->cfs.h_nr_running == 1)) {
8853 if ((check_cpu_capacity(env->src_rq, sd)) &&
8854 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8858 if (env->src_grp_type == group_misfit_task)
8861 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8864 static int active_load_balance_cpu_stop(void *data);
8866 static int should_we_balance(struct lb_env *env)
8868 struct sched_group *sg = env->sd->groups;
8869 int cpu, balance_cpu = -1;
8872 * Ensure the balancing environment is consistent; can happen
8873 * when the softirq triggers 'during' hotplug.
8875 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8879 * In the newly idle case, we will allow all the CPUs
8880 * to do the newly idle load balance.
8882 if (env->idle == CPU_NEWLY_IDLE)
8885 /* Try to find first idle CPU */
8886 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8894 if (balance_cpu == -1)
8895 balance_cpu = group_balance_cpu(sg);
8898 * First idle CPU or the first CPU(busiest) in this sched group
8899 * is eligible for doing load balancing at this and above domains.
8901 return balance_cpu == env->dst_cpu;
8905 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8906 * tasks if there is an imbalance.
8908 static int load_balance(int this_cpu, struct rq *this_rq,
8909 struct sched_domain *sd, enum cpu_idle_type idle,
8910 int *continue_balancing)
8912 int ld_moved, cur_ld_moved, active_balance = 0;
8913 struct sched_domain *sd_parent = sd->parent;
8914 struct sched_group *group;
8917 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8919 struct lb_env env = {
8921 .dst_cpu = this_cpu,
8923 .dst_grpmask = sched_group_span(sd->groups),
8925 .loop_break = sched_nr_migrate_break,
8928 .tasks = LIST_HEAD_INIT(env.tasks),
8931 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8933 schedstat_inc(sd->lb_count[idle]);
8936 if (!should_we_balance(&env)) {
8937 *continue_balancing = 0;
8941 group = find_busiest_group(&env);
8943 schedstat_inc(sd->lb_nobusyg[idle]);
8947 busiest = find_busiest_queue(&env, group);
8949 schedstat_inc(sd->lb_nobusyq[idle]);
8953 BUG_ON(busiest == env.dst_rq);
8955 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8957 env.src_cpu = busiest->cpu;
8958 env.src_rq = busiest;
8961 if (busiest->nr_running > 1) {
8963 * Attempt to move tasks. If find_busiest_group has found
8964 * an imbalance but busiest->nr_running <= 1, the group is
8965 * still unbalanced. ld_moved simply stays zero, so it is
8966 * correctly treated as an imbalance.
8968 env.flags |= LBF_ALL_PINNED;
8969 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8972 rq_lock_irqsave(busiest, &rf);
8973 update_rq_clock(busiest);
8976 * cur_ld_moved - load moved in current iteration
8977 * ld_moved - cumulative load moved across iterations
8979 cur_ld_moved = detach_tasks(&env);
8982 * We've detached some tasks from busiest_rq. Every
8983 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8984 * unlock busiest->lock, and we are able to be sure
8985 * that nobody can manipulate the tasks in parallel.
8986 * See task_rq_lock() family for the details.
8989 rq_unlock(busiest, &rf);
8993 ld_moved += cur_ld_moved;
8996 local_irq_restore(rf.flags);
8998 if (env.flags & LBF_NEED_BREAK) {
8999 env.flags &= ~LBF_NEED_BREAK;
9004 * Revisit (affine) tasks on src_cpu that couldn't be moved to
9005 * us and move them to an alternate dst_cpu in our sched_group
9006 * where they can run. The upper limit on how many times we
9007 * iterate on same src_cpu is dependent on number of CPUs in our
9010 * This changes load balance semantics a bit on who can move
9011 * load to a given_cpu. In addition to the given_cpu itself
9012 * (or a ilb_cpu acting on its behalf where given_cpu is
9013 * nohz-idle), we now have balance_cpu in a position to move
9014 * load to given_cpu. In rare situations, this may cause
9015 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
9016 * _independently_ and at _same_ time to move some load to
9017 * given_cpu) causing exceess load to be moved to given_cpu.
9018 * This however should not happen so much in practice and
9019 * moreover subsequent load balance cycles should correct the
9020 * excess load moved.
9022 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
9024 /* Prevent to re-select dst_cpu via env's CPUs */
9025 cpumask_clear_cpu(env.dst_cpu, env.cpus);
9027 env.dst_rq = cpu_rq(env.new_dst_cpu);
9028 env.dst_cpu = env.new_dst_cpu;
9029 env.flags &= ~LBF_DST_PINNED;
9031 env.loop_break = sched_nr_migrate_break;
9034 * Go back to "more_balance" rather than "redo" since we
9035 * need to continue with same src_cpu.
9041 * We failed to reach balance because of affinity.
9044 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9046 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
9047 *group_imbalance = 1;
9050 /* All tasks on this runqueue were pinned by CPU affinity */
9051 if (unlikely(env.flags & LBF_ALL_PINNED)) {
9052 cpumask_clear_cpu(cpu_of(busiest), cpus);
9054 * Attempting to continue load balancing at the current
9055 * sched_domain level only makes sense if there are
9056 * active CPUs remaining as possible busiest CPUs to
9057 * pull load from which are not contained within the
9058 * destination group that is receiving any migrated
9061 if (!cpumask_subset(cpus, env.dst_grpmask)) {
9063 env.loop_break = sched_nr_migrate_break;
9066 goto out_all_pinned;
9071 schedstat_inc(sd->lb_failed[idle]);
9073 * Increment the failure counter only on periodic balance.
9074 * We do not want newidle balance, which can be very
9075 * frequent, pollute the failure counter causing
9076 * excessive cache_hot migrations and active balances.
9078 if (idle != CPU_NEWLY_IDLE)
9079 sd->nr_balance_failed++;
9081 if (need_active_balance(&env)) {
9082 unsigned long flags;
9084 raw_spin_lock_irqsave(&busiest->lock, flags);
9087 * Don't kick the active_load_balance_cpu_stop,
9088 * if the curr task on busiest CPU can't be
9089 * moved to this_cpu:
9091 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
9092 raw_spin_unlock_irqrestore(&busiest->lock,
9094 env.flags |= LBF_ALL_PINNED;
9095 goto out_one_pinned;
9099 * ->active_balance synchronizes accesses to
9100 * ->active_balance_work. Once set, it's cleared
9101 * only after active load balance is finished.
9103 if (!busiest->active_balance) {
9104 busiest->active_balance = 1;
9105 busiest->push_cpu = this_cpu;
9108 raw_spin_unlock_irqrestore(&busiest->lock, flags);
9110 if (active_balance) {
9111 stop_one_cpu_nowait(cpu_of(busiest),
9112 active_load_balance_cpu_stop, busiest,
9113 &busiest->active_balance_work);
9116 /* We've kicked active balancing, force task migration. */
9117 sd->nr_balance_failed = sd->cache_nice_tries+1;
9120 sd->nr_balance_failed = 0;
9122 if (likely(!active_balance)) {
9123 /* We were unbalanced, so reset the balancing interval */
9124 sd->balance_interval = sd->min_interval;
9127 * If we've begun active balancing, start to back off. This
9128 * case may not be covered by the all_pinned logic if there
9129 * is only 1 task on the busy runqueue (because we don't call
9132 if (sd->balance_interval < sd->max_interval)
9133 sd->balance_interval *= 2;
9140 * We reach balance although we may have faced some affinity
9141 * constraints. Clear the imbalance flag if it was set.
9144 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9146 if (*group_imbalance)
9147 *group_imbalance = 0;
9152 * We reach balance because all tasks are pinned at this level so
9153 * we can't migrate them. Let the imbalance flag set so parent level
9154 * can try to migrate them.
9156 schedstat_inc(sd->lb_balanced[idle]);
9158 sd->nr_balance_failed = 0;
9164 * idle_balance() disregards balance intervals, so we could repeatedly
9165 * reach this code, which would lead to balance_interval skyrocketting
9166 * in a short amount of time. Skip the balance_interval increase logic
9169 if (env.idle == CPU_NEWLY_IDLE)
9172 /* tune up the balancing interval */
9173 if ((env.flags & LBF_ALL_PINNED &&
9174 sd->balance_interval < MAX_PINNED_INTERVAL) ||
9175 sd->balance_interval < sd->max_interval)
9176 sd->balance_interval *= 2;
9181 static inline unsigned long
9182 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
9184 unsigned long interval = sd->balance_interval;
9187 interval *= sd->busy_factor;
9189 /* scale ms to jiffies */
9190 interval = msecs_to_jiffies(interval);
9191 interval = clamp(interval, 1UL, max_load_balance_interval);
9197 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
9199 unsigned long interval, next;
9201 /* used by idle balance, so cpu_busy = 0 */
9202 interval = get_sd_balance_interval(sd, 0);
9203 next = sd->last_balance + interval;
9205 if (time_after(*next_balance, next))
9206 *next_balance = next;
9210 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9211 * running tasks off the busiest CPU onto idle CPUs. It requires at
9212 * least 1 task to be running on each physical CPU where possible, and
9213 * avoids physical / logical imbalances.
9215 static int active_load_balance_cpu_stop(void *data)
9217 struct rq *busiest_rq = data;
9218 int busiest_cpu = cpu_of(busiest_rq);
9219 int target_cpu = busiest_rq->push_cpu;
9220 struct rq *target_rq = cpu_rq(target_cpu);
9221 struct sched_domain *sd;
9222 struct task_struct *p = NULL;
9225 rq_lock_irq(busiest_rq, &rf);
9227 * Between queueing the stop-work and running it is a hole in which
9228 * CPUs can become inactive. We should not move tasks from or to
9231 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
9234 /* Make sure the requested CPU hasn't gone down in the meantime: */
9235 if (unlikely(busiest_cpu != smp_processor_id() ||
9236 !busiest_rq->active_balance))
9239 /* Is there any task to move? */
9240 if (busiest_rq->nr_running <= 1)
9244 * This condition is "impossible", if it occurs
9245 * we need to fix it. Originally reported by
9246 * Bjorn Helgaas on a 128-CPU setup.
9248 BUG_ON(busiest_rq == target_rq);
9250 /* Search for an sd spanning us and the target CPU. */
9252 for_each_domain(target_cpu, sd) {
9253 if ((sd->flags & SD_LOAD_BALANCE) &&
9254 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9259 struct lb_env env = {
9261 .dst_cpu = target_cpu,
9262 .dst_rq = target_rq,
9263 .src_cpu = busiest_rq->cpu,
9264 .src_rq = busiest_rq,
9267 * can_migrate_task() doesn't need to compute new_dst_cpu
9268 * for active balancing. Since we have CPU_IDLE, but no
9269 * @dst_grpmask we need to make that test go away with lying
9272 .flags = LBF_DST_PINNED,
9275 schedstat_inc(sd->alb_count);
9276 update_rq_clock(busiest_rq);
9278 p = detach_one_task(&env);
9280 schedstat_inc(sd->alb_pushed);
9281 /* Active balancing done, reset the failure counter. */
9282 sd->nr_balance_failed = 0;
9284 schedstat_inc(sd->alb_failed);
9289 busiest_rq->active_balance = 0;
9290 rq_unlock(busiest_rq, &rf);
9293 attach_one_task(target_rq, p);
9300 static DEFINE_SPINLOCK(balancing);
9303 * Scale the max load_balance interval with the number of CPUs in the system.
9304 * This trades load-balance latency on larger machines for less cross talk.
9306 void update_max_interval(void)
9308 max_load_balance_interval = HZ*num_online_cpus()/10;
9312 * It checks each scheduling domain to see if it is due to be balanced,
9313 * and initiates a balancing operation if so.
9315 * Balancing parameters are set up in init_sched_domains.
9317 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9319 int continue_balancing = 1;
9321 unsigned long interval;
9322 struct sched_domain *sd;
9323 /* Earliest time when we have to do rebalance again */
9324 unsigned long next_balance = jiffies + 60*HZ;
9325 int update_next_balance = 0;
9326 int need_serialize, need_decay = 0;
9330 for_each_domain(cpu, sd) {
9332 * Decay the newidle max times here because this is a regular
9333 * visit to all the domains. Decay ~1% per second.
9335 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9336 sd->max_newidle_lb_cost =
9337 (sd->max_newidle_lb_cost * 253) / 256;
9338 sd->next_decay_max_lb_cost = jiffies + HZ;
9341 max_cost += sd->max_newidle_lb_cost;
9343 if (!(sd->flags & SD_LOAD_BALANCE))
9347 * Stop the load balance at this level. There is another
9348 * CPU in our sched group which is doing load balancing more
9351 if (!continue_balancing) {
9357 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9359 need_serialize = sd->flags & SD_SERIALIZE;
9360 if (need_serialize) {
9361 if (!spin_trylock(&balancing))
9365 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9366 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9368 * The LBF_DST_PINNED logic could have changed
9369 * env->dst_cpu, so we can't know our idle
9370 * state even if we migrated tasks. Update it.
9372 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9374 sd->last_balance = jiffies;
9375 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9378 spin_unlock(&balancing);
9380 if (time_after(next_balance, sd->last_balance + interval)) {
9381 next_balance = sd->last_balance + interval;
9382 update_next_balance = 1;
9387 * Ensure the rq-wide value also decays but keep it at a
9388 * reasonable floor to avoid funnies with rq->avg_idle.
9390 rq->max_idle_balance_cost =
9391 max((u64)sysctl_sched_migration_cost, max_cost);
9396 * next_balance will be updated only when there is a need.
9397 * When the cpu is attached to null domain for ex, it will not be
9400 if (likely(update_next_balance)) {
9401 rq->next_balance = next_balance;
9403 #ifdef CONFIG_NO_HZ_COMMON
9405 * If this CPU has been elected to perform the nohz idle
9406 * balance. Other idle CPUs have already rebalanced with
9407 * nohz_idle_balance() and nohz.next_balance has been
9408 * updated accordingly. This CPU is now running the idle load
9409 * balance for itself and we need to update the
9410 * nohz.next_balance accordingly.
9412 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9413 nohz.next_balance = rq->next_balance;
9418 static inline int on_null_domain(struct rq *rq)
9420 return unlikely(!rcu_dereference_sched(rq->sd));
9423 #ifdef CONFIG_NO_HZ_COMMON
9425 * idle load balancing details
9426 * - When one of the busy CPUs notice that there may be an idle rebalancing
9427 * needed, they will kick the idle load balancer, which then does idle
9428 * load balancing for all the idle CPUs.
9431 static inline int find_new_ilb(void)
9433 int ilb = cpumask_first(nohz.idle_cpus_mask);
9435 if (ilb < nr_cpu_ids && idle_cpu(ilb))
9442 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
9443 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
9444 * CPU (if there is one).
9446 static void kick_ilb(unsigned int flags)
9450 nohz.next_balance++;
9452 ilb_cpu = find_new_ilb();
9454 if (ilb_cpu >= nr_cpu_ids)
9457 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
9458 if (flags & NOHZ_KICK_MASK)
9462 * Use smp_send_reschedule() instead of resched_cpu().
9463 * This way we generate a sched IPI on the target CPU which
9464 * is idle. And the softirq performing nohz idle load balance
9465 * will be run before returning from the IPI.
9467 smp_send_reschedule(ilb_cpu);
9471 * Current heuristic for kicking the idle load balancer in the presence
9472 * of an idle cpu in the system.
9473 * - This rq has more than one task.
9474 * - This rq has at least one CFS task and the capacity of the CPU is
9475 * significantly reduced because of RT tasks or IRQs.
9476 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9477 * multiple busy cpu.
9478 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9479 * domain span are idle.
9481 static void nohz_balancer_kick(struct rq *rq)
9483 unsigned long now = jiffies;
9484 struct sched_domain_shared *sds;
9485 struct sched_domain *sd;
9486 int nr_busy, i, cpu = rq->cpu;
9487 unsigned int flags = 0;
9489 if (unlikely(rq->idle_balance))
9493 * We may be recently in ticked or tickless idle mode. At the first
9494 * busy tick after returning from idle, we will update the busy stats.
9496 nohz_balance_exit_idle(rq);
9499 * None are in tickless mode and hence no need for NOHZ idle load
9502 if (likely(!atomic_read(&nohz.nr_cpus)))
9505 if (READ_ONCE(nohz.has_blocked) &&
9506 time_after(now, READ_ONCE(nohz.next_blocked)))
9507 flags = NOHZ_STATS_KICK;
9509 if (time_before(now, nohz.next_balance))
9512 if (rq->nr_running >= 2 || rq->misfit_task_load) {
9513 flags = NOHZ_KICK_MASK;
9518 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9521 * XXX: write a coherent comment on why we do this.
9522 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
9524 nr_busy = atomic_read(&sds->nr_busy_cpus);
9526 flags = NOHZ_KICK_MASK;
9532 sd = rcu_dereference(rq->sd);
9534 if ((rq->cfs.h_nr_running >= 1) &&
9535 check_cpu_capacity(rq, sd)) {
9536 flags = NOHZ_KICK_MASK;
9541 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
9543 for_each_cpu(i, sched_domain_span(sd)) {
9545 !cpumask_test_cpu(i, nohz.idle_cpus_mask))
9548 if (sched_asym_prefer(i, cpu)) {
9549 flags = NOHZ_KICK_MASK;
9561 static void set_cpu_sd_state_busy(int cpu)
9563 struct sched_domain *sd;
9566 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9568 if (!sd || !sd->nohz_idle)
9572 atomic_inc(&sd->shared->nr_busy_cpus);
9577 void nohz_balance_exit_idle(struct rq *rq)
9579 SCHED_WARN_ON(rq != this_rq());
9581 if (likely(!rq->nohz_tick_stopped))
9584 rq->nohz_tick_stopped = 0;
9585 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
9586 atomic_dec(&nohz.nr_cpus);
9588 set_cpu_sd_state_busy(rq->cpu);
9591 static void set_cpu_sd_state_idle(int cpu)
9593 struct sched_domain *sd;
9596 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9598 if (!sd || sd->nohz_idle)
9602 atomic_dec(&sd->shared->nr_busy_cpus);
9608 * This routine will record that the CPU is going idle with tick stopped.
9609 * This info will be used in performing idle load balancing in the future.
9611 void nohz_balance_enter_idle(int cpu)
9613 struct rq *rq = cpu_rq(cpu);
9615 SCHED_WARN_ON(cpu != smp_processor_id());
9617 /* If this CPU is going down, then nothing needs to be done: */
9618 if (!cpu_active(cpu))
9621 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9622 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9626 * Can be set safely without rq->lock held
9627 * If a clear happens, it will have evaluated last additions because
9628 * rq->lock is held during the check and the clear
9630 rq->has_blocked_load = 1;
9633 * The tick is still stopped but load could have been added in the
9634 * meantime. We set the nohz.has_blocked flag to trig a check of the
9635 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
9636 * of nohz.has_blocked can only happen after checking the new load
9638 if (rq->nohz_tick_stopped)
9641 /* If we're a completely isolated CPU, we don't play: */
9642 if (on_null_domain(rq))
9645 rq->nohz_tick_stopped = 1;
9647 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9648 atomic_inc(&nohz.nr_cpus);
9651 * Ensures that if nohz_idle_balance() fails to observe our
9652 * @idle_cpus_mask store, it must observe the @has_blocked
9655 smp_mb__after_atomic();
9657 set_cpu_sd_state_idle(cpu);
9661 * Each time a cpu enter idle, we assume that it has blocked load and
9662 * enable the periodic update of the load of idle cpus
9664 WRITE_ONCE(nohz.has_blocked, 1);
9668 * Internal function that runs load balance for all idle cpus. The load balance
9669 * can be a simple update of blocked load or a complete load balance with
9670 * tasks movement depending of flags.
9671 * The function returns false if the loop has stopped before running
9672 * through all idle CPUs.
9674 static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
9675 enum cpu_idle_type idle)
9677 /* Earliest time when we have to do rebalance again */
9678 unsigned long now = jiffies;
9679 unsigned long next_balance = now + 60*HZ;
9680 bool has_blocked_load = false;
9681 int update_next_balance = 0;
9682 int this_cpu = this_rq->cpu;
9687 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
9690 * We assume there will be no idle load after this update and clear
9691 * the has_blocked flag. If a cpu enters idle in the mean time, it will
9692 * set the has_blocked flag and trig another update of idle load.
9693 * Because a cpu that becomes idle, is added to idle_cpus_mask before
9694 * setting the flag, we are sure to not clear the state and not
9695 * check the load of an idle cpu.
9697 WRITE_ONCE(nohz.has_blocked, 0);
9700 * Ensures that if we miss the CPU, we must see the has_blocked
9701 * store from nohz_balance_enter_idle().
9705 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9706 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9710 * If this CPU gets work to do, stop the load balancing
9711 * work being done for other CPUs. Next load
9712 * balancing owner will pick it up.
9714 if (need_resched()) {
9715 has_blocked_load = true;
9719 rq = cpu_rq(balance_cpu);
9721 has_blocked_load |= update_nohz_stats(rq, true);
9724 * If time for next balance is due,
9727 if (time_after_eq(jiffies, rq->next_balance)) {
9730 rq_lock_irqsave(rq, &rf);
9731 update_rq_clock(rq);
9732 cpu_load_update_idle(rq);
9733 rq_unlock_irqrestore(rq, &rf);
9735 if (flags & NOHZ_BALANCE_KICK)
9736 rebalance_domains(rq, CPU_IDLE);
9739 if (time_after(next_balance, rq->next_balance)) {
9740 next_balance = rq->next_balance;
9741 update_next_balance = 1;
9745 /* Newly idle CPU doesn't need an update */
9746 if (idle != CPU_NEWLY_IDLE) {
9747 update_blocked_averages(this_cpu);
9748 has_blocked_load |= this_rq->has_blocked_load;
9751 if (flags & NOHZ_BALANCE_KICK)
9752 rebalance_domains(this_rq, CPU_IDLE);
9754 WRITE_ONCE(nohz.next_blocked,
9755 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
9757 /* The full idle balance loop has been done */
9761 /* There is still blocked load, enable periodic update */
9762 if (has_blocked_load)
9763 WRITE_ONCE(nohz.has_blocked, 1);
9766 * next_balance will be updated only when there is a need.
9767 * When the CPU is attached to null domain for ex, it will not be
9770 if (likely(update_next_balance))
9771 nohz.next_balance = next_balance;
9777 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9778 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9780 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9782 int this_cpu = this_rq->cpu;
9785 if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
9788 if (idle != CPU_IDLE) {
9789 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9793 /* could be _relaxed() */
9794 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9795 if (!(flags & NOHZ_KICK_MASK))
9798 _nohz_idle_balance(this_rq, flags, idle);
9803 static void nohz_newidle_balance(struct rq *this_rq)
9805 int this_cpu = this_rq->cpu;
9808 * This CPU doesn't want to be disturbed by scheduler
9811 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
9814 /* Will wake up very soon. No time for doing anything else*/
9815 if (this_rq->avg_idle < sysctl_sched_migration_cost)
9818 /* Don't need to update blocked load of idle CPUs*/
9819 if (!READ_ONCE(nohz.has_blocked) ||
9820 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
9823 raw_spin_unlock(&this_rq->lock);
9825 * This CPU is going to be idle and blocked load of idle CPUs
9826 * need to be updated. Run the ilb locally as it is a good
9827 * candidate for ilb instead of waking up another idle CPU.
9828 * Kick an normal ilb if we failed to do the update.
9830 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
9831 kick_ilb(NOHZ_STATS_KICK);
9832 raw_spin_lock(&this_rq->lock);
9835 #else /* !CONFIG_NO_HZ_COMMON */
9836 static inline void nohz_balancer_kick(struct rq *rq) { }
9838 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9843 static inline void nohz_newidle_balance(struct rq *this_rq) { }
9844 #endif /* CONFIG_NO_HZ_COMMON */
9847 * idle_balance is called by schedule() if this_cpu is about to become
9848 * idle. Attempts to pull tasks from other CPUs.
9850 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
9852 unsigned long next_balance = jiffies + HZ;
9853 int this_cpu = this_rq->cpu;
9854 struct sched_domain *sd;
9855 int pulled_task = 0;
9859 * We must set idle_stamp _before_ calling idle_balance(), such that we
9860 * measure the duration of idle_balance() as idle time.
9862 this_rq->idle_stamp = rq_clock(this_rq);
9865 * Do not pull tasks towards !active CPUs...
9867 if (!cpu_active(this_cpu))
9871 * This is OK, because current is on_cpu, which avoids it being picked
9872 * for load-balance and preemption/IRQs are still disabled avoiding
9873 * further scheduler activity on it and we're being very careful to
9874 * re-start the picking loop.
9876 rq_unpin_lock(this_rq, rf);
9878 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
9879 !READ_ONCE(this_rq->rd->overload)) {
9882 sd = rcu_dereference_check_sched_domain(this_rq->sd);
9884 update_next_balance(sd, &next_balance);
9887 nohz_newidle_balance(this_rq);
9892 raw_spin_unlock(&this_rq->lock);
9894 update_blocked_averages(this_cpu);
9896 for_each_domain(this_cpu, sd) {
9897 int continue_balancing = 1;
9898 u64 t0, domain_cost;
9900 if (!(sd->flags & SD_LOAD_BALANCE))
9903 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
9904 update_next_balance(sd, &next_balance);
9908 if (sd->flags & SD_BALANCE_NEWIDLE) {
9909 t0 = sched_clock_cpu(this_cpu);
9911 pulled_task = load_balance(this_cpu, this_rq,
9913 &continue_balancing);
9915 domain_cost = sched_clock_cpu(this_cpu) - t0;
9916 if (domain_cost > sd->max_newidle_lb_cost)
9917 sd->max_newidle_lb_cost = domain_cost;
9919 curr_cost += domain_cost;
9922 update_next_balance(sd, &next_balance);
9925 * Stop searching for tasks to pull if there are
9926 * now runnable tasks on this rq.
9928 if (pulled_task || this_rq->nr_running > 0)
9933 raw_spin_lock(&this_rq->lock);
9935 if (curr_cost > this_rq->max_idle_balance_cost)
9936 this_rq->max_idle_balance_cost = curr_cost;
9940 * While browsing the domains, we released the rq lock, a task could
9941 * have been enqueued in the meantime. Since we're not going idle,
9942 * pretend we pulled a task.
9944 if (this_rq->cfs.h_nr_running && !pulled_task)
9947 /* Move the next balance forward */
9948 if (time_after(this_rq->next_balance, next_balance))
9949 this_rq->next_balance = next_balance;
9951 /* Is there a task of a high priority class? */
9952 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
9956 this_rq->idle_stamp = 0;
9958 rq_repin_lock(this_rq, rf);
9964 * run_rebalance_domains is triggered when needed from the scheduler tick.
9965 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9967 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9969 struct rq *this_rq = this_rq();
9970 enum cpu_idle_type idle = this_rq->idle_balance ?
9971 CPU_IDLE : CPU_NOT_IDLE;
9974 * If this CPU has a pending nohz_balance_kick, then do the
9975 * balancing on behalf of the other idle CPUs whose ticks are
9976 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9977 * give the idle CPUs a chance to load balance. Else we may
9978 * load balance only within the local sched_domain hierarchy
9979 * and abort nohz_idle_balance altogether if we pull some load.
9981 if (nohz_idle_balance(this_rq, idle))
9984 /* normal load balance */
9985 update_blocked_averages(this_rq->cpu);
9986 rebalance_domains(this_rq, idle);
9990 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9992 void trigger_load_balance(struct rq *rq)
9994 /* Don't need to rebalance while attached to NULL domain */
9995 if (unlikely(on_null_domain(rq)))
9998 if (time_after_eq(jiffies, rq->next_balance))
9999 raise_softirq(SCHED_SOFTIRQ);
10001 nohz_balancer_kick(rq);
10004 static void rq_online_fair(struct rq *rq)
10008 update_runtime_enabled(rq);
10011 static void rq_offline_fair(struct rq *rq)
10015 /* Ensure any throttled groups are reachable by pick_next_task */
10016 unthrottle_offline_cfs_rqs(rq);
10019 #endif /* CONFIG_SMP */
10022 * scheduler tick hitting a task of our scheduling class.
10024 * NOTE: This function can be called remotely by the tick offload that
10025 * goes along full dynticks. Therefore no local assumption can be made
10026 * and everything must be accessed through the @rq and @curr passed in
10029 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
10031 struct cfs_rq *cfs_rq;
10032 struct sched_entity *se = &curr->se;
10034 for_each_sched_entity(se) {
10035 cfs_rq = cfs_rq_of(se);
10036 entity_tick(cfs_rq, se, queued);
10039 if (static_branch_unlikely(&sched_numa_balancing))
10040 task_tick_numa(rq, curr);
10042 update_misfit_status(curr, rq);
10043 update_overutilized_status(task_rq(curr));
10047 * called on fork with the child task as argument from the parent's context
10048 * - child not yet on the tasklist
10049 * - preemption disabled
10051 static void task_fork_fair(struct task_struct *p)
10053 struct cfs_rq *cfs_rq;
10054 struct sched_entity *se = &p->se, *curr;
10055 struct rq *rq = this_rq();
10056 struct rq_flags rf;
10059 update_rq_clock(rq);
10061 cfs_rq = task_cfs_rq(current);
10062 curr = cfs_rq->curr;
10064 update_curr(cfs_rq);
10065 se->vruntime = curr->vruntime;
10067 place_entity(cfs_rq, se, 1);
10069 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
10071 * Upon rescheduling, sched_class::put_prev_task() will place
10072 * 'current' within the tree based on its new key value.
10074 swap(curr->vruntime, se->vruntime);
10078 se->vruntime -= cfs_rq->min_vruntime;
10079 rq_unlock(rq, &rf);
10083 * Priority of the task has changed. Check to see if we preempt
10084 * the current task.
10087 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
10089 if (!task_on_rq_queued(p))
10093 * Reschedule if we are currently running on this runqueue and
10094 * our priority decreased, or if we are not currently running on
10095 * this runqueue and our priority is higher than the current's
10097 if (rq->curr == p) {
10098 if (p->prio > oldprio)
10101 check_preempt_curr(rq, p, 0);
10104 static inline bool vruntime_normalized(struct task_struct *p)
10106 struct sched_entity *se = &p->se;
10109 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
10110 * the dequeue_entity(.flags=0) will already have normalized the
10117 * When !on_rq, vruntime of the task has usually NOT been normalized.
10118 * But there are some cases where it has already been normalized:
10120 * - A forked child which is waiting for being woken up by
10121 * wake_up_new_task().
10122 * - A task which has been woken up by try_to_wake_up() and
10123 * waiting for actually being woken up by sched_ttwu_pending().
10125 if (!se->sum_exec_runtime ||
10126 (p->state == TASK_WAKING && p->sched_remote_wakeup))
10132 #ifdef CONFIG_FAIR_GROUP_SCHED
10134 * Propagate the changes of the sched_entity across the tg tree to make it
10135 * visible to the root
10137 static void propagate_entity_cfs_rq(struct sched_entity *se)
10139 struct cfs_rq *cfs_rq;
10141 /* Start to propagate at parent */
10144 for_each_sched_entity(se) {
10145 cfs_rq = cfs_rq_of(se);
10147 if (cfs_rq_throttled(cfs_rq))
10150 update_load_avg(cfs_rq, se, UPDATE_TG);
10154 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
10157 static void detach_entity_cfs_rq(struct sched_entity *se)
10159 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10161 /* Catch up with the cfs_rq and remove our load when we leave */
10162 update_load_avg(cfs_rq, se, 0);
10163 detach_entity_load_avg(cfs_rq, se);
10164 update_tg_load_avg(cfs_rq, false);
10165 propagate_entity_cfs_rq(se);
10168 static void attach_entity_cfs_rq(struct sched_entity *se)
10170 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10172 #ifdef CONFIG_FAIR_GROUP_SCHED
10174 * Since the real-depth could have been changed (only FAIR
10175 * class maintain depth value), reset depth properly.
10177 se->depth = se->parent ? se->parent->depth + 1 : 0;
10180 /* Synchronize entity with its cfs_rq */
10181 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
10182 attach_entity_load_avg(cfs_rq, se, 0);
10183 update_tg_load_avg(cfs_rq, false);
10184 propagate_entity_cfs_rq(se);
10187 static void detach_task_cfs_rq(struct task_struct *p)
10189 struct sched_entity *se = &p->se;
10190 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10192 if (!vruntime_normalized(p)) {
10194 * Fix up our vruntime so that the current sleep doesn't
10195 * cause 'unlimited' sleep bonus.
10197 place_entity(cfs_rq, se, 0);
10198 se->vruntime -= cfs_rq->min_vruntime;
10201 detach_entity_cfs_rq(se);
10204 static void attach_task_cfs_rq(struct task_struct *p)
10206 struct sched_entity *se = &p->se;
10207 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10209 attach_entity_cfs_rq(se);
10211 if (!vruntime_normalized(p))
10212 se->vruntime += cfs_rq->min_vruntime;
10215 static void switched_from_fair(struct rq *rq, struct task_struct *p)
10217 detach_task_cfs_rq(p);
10220 static void switched_to_fair(struct rq *rq, struct task_struct *p)
10222 attach_task_cfs_rq(p);
10224 if (task_on_rq_queued(p)) {
10226 * We were most likely switched from sched_rt, so
10227 * kick off the schedule if running, otherwise just see
10228 * if we can still preempt the current task.
10233 check_preempt_curr(rq, p, 0);
10237 /* Account for a task changing its policy or group.
10239 * This routine is mostly called to set cfs_rq->curr field when a task
10240 * migrates between groups/classes.
10242 static void set_curr_task_fair(struct rq *rq)
10244 struct sched_entity *se = &rq->curr->se;
10246 for_each_sched_entity(se) {
10247 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10249 set_next_entity(cfs_rq, se);
10250 /* ensure bandwidth has been allocated on our new cfs_rq */
10251 account_cfs_rq_runtime(cfs_rq, 0);
10255 void init_cfs_rq(struct cfs_rq *cfs_rq)
10257 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
10258 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
10259 #ifndef CONFIG_64BIT
10260 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
10263 raw_spin_lock_init(&cfs_rq->removed.lock);
10267 #ifdef CONFIG_FAIR_GROUP_SCHED
10268 static void task_set_group_fair(struct task_struct *p)
10270 struct sched_entity *se = &p->se;
10272 set_task_rq(p, task_cpu(p));
10273 se->depth = se->parent ? se->parent->depth + 1 : 0;
10276 static void task_move_group_fair(struct task_struct *p)
10278 detach_task_cfs_rq(p);
10279 set_task_rq(p, task_cpu(p));
10282 /* Tell se's cfs_rq has been changed -- migrated */
10283 p->se.avg.last_update_time = 0;
10285 attach_task_cfs_rq(p);
10288 static void task_change_group_fair(struct task_struct *p, int type)
10291 case TASK_SET_GROUP:
10292 task_set_group_fair(p);
10295 case TASK_MOVE_GROUP:
10296 task_move_group_fair(p);
10301 void free_fair_sched_group(struct task_group *tg)
10305 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
10307 for_each_possible_cpu(i) {
10309 kfree(tg->cfs_rq[i]);
10318 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10320 struct sched_entity *se;
10321 struct cfs_rq *cfs_rq;
10324 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
10327 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
10331 tg->shares = NICE_0_LOAD;
10333 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
10335 for_each_possible_cpu(i) {
10336 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
10337 GFP_KERNEL, cpu_to_node(i));
10341 se = kzalloc_node(sizeof(struct sched_entity),
10342 GFP_KERNEL, cpu_to_node(i));
10346 init_cfs_rq(cfs_rq);
10347 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
10348 init_entity_runnable_average(se);
10359 void online_fair_sched_group(struct task_group *tg)
10361 struct sched_entity *se;
10365 for_each_possible_cpu(i) {
10369 raw_spin_lock_irq(&rq->lock);
10370 update_rq_clock(rq);
10371 attach_entity_cfs_rq(se);
10372 sync_throttle(tg, i);
10373 raw_spin_unlock_irq(&rq->lock);
10377 void unregister_fair_sched_group(struct task_group *tg)
10379 unsigned long flags;
10383 for_each_possible_cpu(cpu) {
10385 remove_entity_load_avg(tg->se[cpu]);
10388 * Only empty task groups can be destroyed; so we can speculatively
10389 * check on_list without danger of it being re-added.
10391 if (!tg->cfs_rq[cpu]->on_list)
10396 raw_spin_lock_irqsave(&rq->lock, flags);
10397 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
10398 raw_spin_unlock_irqrestore(&rq->lock, flags);
10402 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
10403 struct sched_entity *se, int cpu,
10404 struct sched_entity *parent)
10406 struct rq *rq = cpu_rq(cpu);
10410 init_cfs_rq_runtime(cfs_rq);
10412 tg->cfs_rq[cpu] = cfs_rq;
10415 /* se could be NULL for root_task_group */
10420 se->cfs_rq = &rq->cfs;
10423 se->cfs_rq = parent->my_q;
10424 se->depth = parent->depth + 1;
10428 /* guarantee group entities always have weight */
10429 update_load_set(&se->load, NICE_0_LOAD);
10430 se->parent = parent;
10433 static DEFINE_MUTEX(shares_mutex);
10435 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10440 * We can't change the weight of the root cgroup.
10445 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10447 mutex_lock(&shares_mutex);
10448 if (tg->shares == shares)
10451 tg->shares = shares;
10452 for_each_possible_cpu(i) {
10453 struct rq *rq = cpu_rq(i);
10454 struct sched_entity *se = tg->se[i];
10455 struct rq_flags rf;
10457 /* Propagate contribution to hierarchy */
10458 rq_lock_irqsave(rq, &rf);
10459 update_rq_clock(rq);
10460 for_each_sched_entity(se) {
10461 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
10462 update_cfs_group(se);
10464 rq_unlock_irqrestore(rq, &rf);
10468 mutex_unlock(&shares_mutex);
10471 #else /* CONFIG_FAIR_GROUP_SCHED */
10473 void free_fair_sched_group(struct task_group *tg) { }
10475 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10480 void online_fair_sched_group(struct task_group *tg) { }
10482 void unregister_fair_sched_group(struct task_group *tg) { }
10484 #endif /* CONFIG_FAIR_GROUP_SCHED */
10487 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10489 struct sched_entity *se = &task->se;
10490 unsigned int rr_interval = 0;
10493 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10496 if (rq->cfs.load.weight)
10497 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10499 return rr_interval;
10503 * All the scheduling class methods:
10505 const struct sched_class fair_sched_class = {
10506 .next = &idle_sched_class,
10507 .enqueue_task = enqueue_task_fair,
10508 .dequeue_task = dequeue_task_fair,
10509 .yield_task = yield_task_fair,
10510 .yield_to_task = yield_to_task_fair,
10512 .check_preempt_curr = check_preempt_wakeup,
10514 .pick_next_task = pick_next_task_fair,
10515 .put_prev_task = put_prev_task_fair,
10518 .select_task_rq = select_task_rq_fair,
10519 .migrate_task_rq = migrate_task_rq_fair,
10521 .rq_online = rq_online_fair,
10522 .rq_offline = rq_offline_fair,
10524 .task_dead = task_dead_fair,
10525 .set_cpus_allowed = set_cpus_allowed_common,
10528 .set_curr_task = set_curr_task_fair,
10529 .task_tick = task_tick_fair,
10530 .task_fork = task_fork_fair,
10532 .prio_changed = prio_changed_fair,
10533 .switched_from = switched_from_fair,
10534 .switched_to = switched_to_fair,
10536 .get_rr_interval = get_rr_interval_fair,
10538 .update_curr = update_curr_fair,
10540 #ifdef CONFIG_FAIR_GROUP_SCHED
10541 .task_change_group = task_change_group_fair,
10545 #ifdef CONFIG_SCHED_DEBUG
10546 void print_cfs_stats(struct seq_file *m, int cpu)
10548 struct cfs_rq *cfs_rq;
10551 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
10552 print_cfs_rq(m, cpu, cfs_rq);
10556 #ifdef CONFIG_NUMA_BALANCING
10557 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10560 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10562 for_each_online_node(node) {
10563 if (p->numa_faults) {
10564 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10565 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10567 if (p->numa_group) {
10568 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
10569 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
10571 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10574 #endif /* CONFIG_NUMA_BALANCING */
10575 #endif /* CONFIG_SCHED_DEBUG */
10577 __init void init_sched_fair_class(void)
10580 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10582 #ifdef CONFIG_NO_HZ_COMMON
10583 nohz.next_balance = jiffies;
10584 nohz.next_blocked = jiffies;
10585 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);