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 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 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 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 #ifdef CONFIG_CFS_BANDWIDTH
101 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
102 * each time a cfs_rq requests quota.
104 * Note: in the case that the slice exceeds the runtime remaining (either due
105 * to consumption or the quota being specified to be smaller than the slice)
106 * we will always only issue the remaining available time.
108 * (default: 5 msec, units: microseconds)
110 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
114 * The margin used when comparing utilization with CPU capacity:
115 * util * margin < capacity * 1024
119 unsigned int capacity_margin = 1280;
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 thr' all leaf cfs_rq's on a runqueue */
356 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
357 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
360 /* Do the two (enqueued) entities belong to the same group ? */
361 static inline struct cfs_rq *
362 is_same_group(struct sched_entity *se, struct sched_entity *pse)
364 if (se->cfs_rq == pse->cfs_rq)
370 static inline struct sched_entity *parent_entity(struct sched_entity *se)
376 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
378 int se_depth, pse_depth;
381 * preemption test can be made between sibling entities who are in the
382 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
383 * both tasks until we find their ancestors who are siblings of common
387 /* First walk up until both entities are at same depth */
388 se_depth = (*se)->depth;
389 pse_depth = (*pse)->depth;
391 while (se_depth > pse_depth) {
393 *se = parent_entity(*se);
396 while (pse_depth > se_depth) {
398 *pse = parent_entity(*pse);
401 while (!is_same_group(*se, *pse)) {
402 *se = parent_entity(*se);
403 *pse = parent_entity(*pse);
407 #else /* !CONFIG_FAIR_GROUP_SCHED */
409 static inline struct task_struct *task_of(struct sched_entity *se)
411 return container_of(se, struct task_struct, se);
414 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
416 return container_of(cfs_rq, struct rq, cfs);
420 #define for_each_sched_entity(se) \
421 for (; se; se = NULL)
423 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
425 return &task_rq(p)->cfs;
428 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
430 struct task_struct *p = task_of(se);
431 struct rq *rq = task_rq(p);
436 /* runqueue "owned" by this group */
437 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
442 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
446 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
450 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
451 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
453 static inline struct sched_entity *parent_entity(struct sched_entity *se)
459 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
463 #endif /* CONFIG_FAIR_GROUP_SCHED */
465 static __always_inline
466 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
468 /**************************************************************
469 * Scheduling class tree data structure manipulation methods:
472 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
474 s64 delta = (s64)(vruntime - max_vruntime);
476 max_vruntime = vruntime;
481 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
483 s64 delta = (s64)(vruntime - min_vruntime);
485 min_vruntime = vruntime;
490 static inline int entity_before(struct sched_entity *a,
491 struct sched_entity *b)
493 return (s64)(a->vruntime - b->vruntime) < 0;
496 static void update_min_vruntime(struct cfs_rq *cfs_rq)
498 struct sched_entity *curr = cfs_rq->curr;
499 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
501 u64 vruntime = cfs_rq->min_vruntime;
505 vruntime = curr->vruntime;
510 if (leftmost) { /* non-empty tree */
511 struct sched_entity *se;
512 se = rb_entry(leftmost, struct sched_entity, run_node);
515 vruntime = se->vruntime;
517 vruntime = min_vruntime(vruntime, se->vruntime);
520 /* ensure we never gain time by being placed backwards. */
521 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
524 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
529 * Enqueue an entity into the rb-tree:
531 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
533 struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node;
534 struct rb_node *parent = NULL;
535 struct sched_entity *entry;
536 bool leftmost = true;
539 * Find the right place in the rbtree:
543 entry = rb_entry(parent, struct sched_entity, run_node);
545 * We dont care about collisions. Nodes with
546 * the same key stay together.
548 if (entity_before(se, entry)) {
549 link = &parent->rb_left;
551 link = &parent->rb_right;
556 rb_link_node(&se->run_node, parent, link);
557 rb_insert_color_cached(&se->run_node,
558 &cfs_rq->tasks_timeline, leftmost);
561 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
563 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
566 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
568 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
573 return rb_entry(left, struct sched_entity, run_node);
576 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
578 struct rb_node *next = rb_next(&se->run_node);
583 return rb_entry(next, struct sched_entity, run_node);
586 #ifdef CONFIG_SCHED_DEBUG
587 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
589 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
594 return rb_entry(last, struct sched_entity, run_node);
597 /**************************************************************
598 * Scheduling class statistics methods:
601 int sched_proc_update_handler(struct ctl_table *table, int write,
602 void __user *buffer, size_t *lenp,
605 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
606 unsigned int factor = get_update_sysctl_factor();
611 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
612 sysctl_sched_min_granularity);
614 #define WRT_SYSCTL(name) \
615 (normalized_sysctl_##name = sysctl_##name / (factor))
616 WRT_SYSCTL(sched_min_granularity);
617 WRT_SYSCTL(sched_latency);
618 WRT_SYSCTL(sched_wakeup_granularity);
628 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
630 if (unlikely(se->load.weight != NICE_0_LOAD))
631 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
637 * The idea is to set a period in which each task runs once.
639 * When there are too many tasks (sched_nr_latency) we have to stretch
640 * this period because otherwise the slices get too small.
642 * p = (nr <= nl) ? l : l*nr/nl
644 static u64 __sched_period(unsigned long nr_running)
646 if (unlikely(nr_running > sched_nr_latency))
647 return nr_running * sysctl_sched_min_granularity;
649 return sysctl_sched_latency;
653 * We calculate the wall-time slice from the period by taking a part
654 * proportional to the weight.
658 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
660 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
662 for_each_sched_entity(se) {
663 struct load_weight *load;
664 struct load_weight lw;
666 cfs_rq = cfs_rq_of(se);
667 load = &cfs_rq->load;
669 if (unlikely(!se->on_rq)) {
672 update_load_add(&lw, se->load.weight);
675 slice = __calc_delta(slice, se->load.weight, load);
681 * We calculate the vruntime slice of a to-be-inserted task.
685 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
687 return calc_delta_fair(sched_slice(cfs_rq, se), se);
692 #include "sched-pelt.h"
694 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
695 static unsigned long task_h_load(struct task_struct *p);
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 intialized 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 intialized 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);
1397 * Multi-stage node selection is used in conjunction with a periodic
1398 * migration fault to build a temporal task<->page relation. By using
1399 * a two-stage filter we remove short/unlikely relations.
1401 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1402 * a task's usage of a particular page (n_p) per total usage of this
1403 * page (n_t) (in a given time-span) to a probability.
1405 * Our periodic faults will sample this probability and getting the
1406 * same result twice in a row, given these samples are fully
1407 * independent, is then given by P(n)^2, provided our sample period
1408 * is sufficiently short compared to the usage pattern.
1410 * This quadric squishes small probabilities, making it less likely we
1411 * act on an unlikely task<->page relation.
1413 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1414 if (!cpupid_pid_unset(last_cpupid) &&
1415 cpupid_to_nid(last_cpupid) != dst_nid)
1418 /* Always allow migrate on private faults */
1419 if (cpupid_match_pid(p, last_cpupid))
1422 /* A shared fault, but p->numa_group has not been set up yet. */
1427 * Destination node is much more heavily used than the source
1428 * node? Allow migration.
1430 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1431 ACTIVE_NODE_FRACTION)
1435 * Distribute memory according to CPU & memory use on each node,
1436 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1438 * faults_cpu(dst) 3 faults_cpu(src)
1439 * --------------- * - > ---------------
1440 * faults_mem(dst) 4 faults_mem(src)
1442 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1443 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1446 static unsigned long weighted_cpuload(struct rq *rq);
1447 static unsigned long source_load(int cpu, int type);
1448 static unsigned long target_load(int cpu, int type);
1449 static unsigned long capacity_of(int cpu);
1451 /* Cached statistics for all CPUs within a node */
1455 /* Total compute capacity of CPUs on a node */
1456 unsigned long compute_capacity;
1458 unsigned int nr_running;
1462 * XXX borrowed from update_sg_lb_stats
1464 static void update_numa_stats(struct numa_stats *ns, int nid)
1466 int smt, cpu, cpus = 0;
1467 unsigned long capacity;
1469 memset(ns, 0, sizeof(*ns));
1470 for_each_cpu(cpu, cpumask_of_node(nid)) {
1471 struct rq *rq = cpu_rq(cpu);
1473 ns->nr_running += rq->nr_running;
1474 ns->load += weighted_cpuload(rq);
1475 ns->compute_capacity += capacity_of(cpu);
1481 * If we raced with hotplug and there are no CPUs left in our mask
1482 * the @ns structure is NULL'ed and task_numa_compare() will
1483 * not find this node attractive.
1485 * We'll detect a huge imbalance and bail there.
1490 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1491 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1492 capacity = cpus / smt; /* cores */
1494 capacity = min_t(unsigned, capacity,
1495 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1498 struct task_numa_env {
1499 struct task_struct *p;
1501 int src_cpu, src_nid;
1502 int dst_cpu, dst_nid;
1504 struct numa_stats src_stats, dst_stats;
1509 struct task_struct *best_task;
1514 static void task_numa_assign(struct task_numa_env *env,
1515 struct task_struct *p, long imp)
1517 struct rq *rq = cpu_rq(env->dst_cpu);
1519 /* Bail out if run-queue part of active NUMA balance. */
1520 if (xchg(&rq->numa_migrate_on, 1))
1524 * Clear previous best_cpu/rq numa-migrate flag, since task now
1525 * found a better CPU to move/swap.
1527 if (env->best_cpu != -1) {
1528 rq = cpu_rq(env->best_cpu);
1529 WRITE_ONCE(rq->numa_migrate_on, 0);
1533 put_task_struct(env->best_task);
1538 env->best_imp = imp;
1539 env->best_cpu = env->dst_cpu;
1542 static bool load_too_imbalanced(long src_load, long dst_load,
1543 struct task_numa_env *env)
1546 long orig_src_load, orig_dst_load;
1547 long src_capacity, dst_capacity;
1550 * The load is corrected for the CPU capacity available on each node.
1553 * ------------ vs ---------
1554 * src_capacity dst_capacity
1556 src_capacity = env->src_stats.compute_capacity;
1557 dst_capacity = env->dst_stats.compute_capacity;
1559 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1561 orig_src_load = env->src_stats.load;
1562 orig_dst_load = env->dst_stats.load;
1564 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1566 /* Would this change make things worse? */
1567 return (imb > old_imb);
1571 * This checks if the overall compute and NUMA accesses of the system would
1572 * be improved if the source tasks was migrated to the target dst_cpu taking
1573 * into account that it might be best if task running on the dst_cpu should
1574 * be exchanged with the source task
1576 static void task_numa_compare(struct task_numa_env *env,
1577 long taskimp, long groupimp, bool maymove)
1579 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1580 struct task_struct *cur;
1581 long src_load, dst_load;
1583 long imp = env->p->numa_group ? groupimp : taskimp;
1585 int dist = env->dist;
1587 if (READ_ONCE(dst_rq->numa_migrate_on))
1591 cur = task_rcu_dereference(&dst_rq->curr);
1592 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1596 * Because we have preemption enabled we can get migrated around and
1597 * end try selecting ourselves (current == env->p) as a swap candidate.
1603 if (maymove || imp > env->best_imp)
1610 * "imp" is the fault differential for the source task between the
1611 * source and destination node. Calculate the total differential for
1612 * the source task and potential destination task. The more negative
1613 * the value is, the more remote accesses that would be expected to
1614 * be incurred if the tasks were swapped.
1616 /* Skip this swap candidate if cannot move to the source cpu */
1617 if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed))
1621 * If dst and source tasks are in the same NUMA group, or not
1622 * in any group then look only at task weights.
1624 if (cur->numa_group == env->p->numa_group) {
1625 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1626 task_weight(cur, env->dst_nid, dist);
1628 * Add some hysteresis to prevent swapping the
1629 * tasks within a group over tiny differences.
1631 if (cur->numa_group)
1635 * Compare the group weights. If a task is all by itself
1636 * (not part of a group), use the task weight instead.
1638 if (cur->numa_group && env->p->numa_group)
1639 imp += group_weight(cur, env->src_nid, dist) -
1640 group_weight(cur, env->dst_nid, dist);
1642 imp += task_weight(cur, env->src_nid, dist) -
1643 task_weight(cur, env->dst_nid, dist);
1646 if (imp <= env->best_imp)
1649 if (maymove && moveimp > imp && moveimp > env->best_imp) {
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 task to preferred node migration periodically, in case it
2403 * case it 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 (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2623 if (src_nid != dst_nid)
2624 p->numa_scan_period = task_scan_start(p);
2628 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2632 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2636 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2640 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2644 #endif /* CONFIG_NUMA_BALANCING */
2647 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2649 update_load_add(&cfs_rq->load, se->load.weight);
2650 if (!parent_entity(se))
2651 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2653 if (entity_is_task(se)) {
2654 struct rq *rq = rq_of(cfs_rq);
2656 account_numa_enqueue(rq, task_of(se));
2657 list_add(&se->group_node, &rq->cfs_tasks);
2660 cfs_rq->nr_running++;
2664 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2666 update_load_sub(&cfs_rq->load, se->load.weight);
2667 if (!parent_entity(se))
2668 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2670 if (entity_is_task(se)) {
2671 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2672 list_del_init(&se->group_node);
2675 cfs_rq->nr_running--;
2679 * Signed add and clamp on underflow.
2681 * Explicitly do a load-store to ensure the intermediate value never hits
2682 * memory. This allows lockless observations without ever seeing the negative
2685 #define add_positive(_ptr, _val) do { \
2686 typeof(_ptr) ptr = (_ptr); \
2687 typeof(_val) val = (_val); \
2688 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2692 if (val < 0 && res > var) \
2695 WRITE_ONCE(*ptr, res); \
2699 * Unsigned subtract and clamp on underflow.
2701 * Explicitly do a load-store to ensure the intermediate value never hits
2702 * memory. This allows lockless observations without ever seeing the negative
2705 #define sub_positive(_ptr, _val) do { \
2706 typeof(_ptr) ptr = (_ptr); \
2707 typeof(*ptr) val = (_val); \
2708 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2712 WRITE_ONCE(*ptr, res); \
2717 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2719 cfs_rq->runnable_weight += se->runnable_weight;
2721 cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2722 cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2726 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2728 cfs_rq->runnable_weight -= se->runnable_weight;
2730 sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2731 sub_positive(&cfs_rq->avg.runnable_load_sum,
2732 se_runnable(se) * se->avg.runnable_load_sum);
2736 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2738 cfs_rq->avg.load_avg += se->avg.load_avg;
2739 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2743 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2745 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2746 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2750 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2752 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2754 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2756 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2759 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2760 unsigned long weight, unsigned long runnable)
2763 /* commit outstanding execution time */
2764 if (cfs_rq->curr == se)
2765 update_curr(cfs_rq);
2766 account_entity_dequeue(cfs_rq, se);
2767 dequeue_runnable_load_avg(cfs_rq, se);
2769 dequeue_load_avg(cfs_rq, se);
2771 se->runnable_weight = runnable;
2772 update_load_set(&se->load, weight);
2776 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2778 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2779 se->avg.runnable_load_avg =
2780 div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2784 enqueue_load_avg(cfs_rq, se);
2786 account_entity_enqueue(cfs_rq, se);
2787 enqueue_runnable_load_avg(cfs_rq, se);
2791 void reweight_task(struct task_struct *p, int prio)
2793 struct sched_entity *se = &p->se;
2794 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2795 struct load_weight *load = &se->load;
2796 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2798 reweight_entity(cfs_rq, se, weight, weight);
2799 load->inv_weight = sched_prio_to_wmult[prio];
2802 #ifdef CONFIG_FAIR_GROUP_SCHED
2805 * All this does is approximate the hierarchical proportion which includes that
2806 * global sum we all love to hate.
2808 * That is, the weight of a group entity, is the proportional share of the
2809 * group weight based on the group runqueue weights. That is:
2811 * tg->weight * grq->load.weight
2812 * ge->load.weight = ----------------------------- (1)
2813 * \Sum grq->load.weight
2815 * Now, because computing that sum is prohibitively expensive to compute (been
2816 * there, done that) we approximate it with this average stuff. The average
2817 * moves slower and therefore the approximation is cheaper and more stable.
2819 * So instead of the above, we substitute:
2821 * grq->load.weight -> grq->avg.load_avg (2)
2823 * which yields the following:
2825 * tg->weight * grq->avg.load_avg
2826 * ge->load.weight = ------------------------------ (3)
2829 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2831 * That is shares_avg, and it is right (given the approximation (2)).
2833 * The problem with it is that because the average is slow -- it was designed
2834 * to be exactly that of course -- this leads to transients in boundary
2835 * conditions. In specific, the case where the group was idle and we start the
2836 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2837 * yielding bad latency etc..
2839 * Now, in that special case (1) reduces to:
2841 * tg->weight * grq->load.weight
2842 * ge->load.weight = ----------------------------- = tg->weight (4)
2845 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2847 * So what we do is modify our approximation (3) to approach (4) in the (near)
2852 * tg->weight * grq->load.weight
2853 * --------------------------------------------------- (5)
2854 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2856 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2857 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2860 * tg->weight * grq->load.weight
2861 * ge->load.weight = ----------------------------- (6)
2866 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2867 * max(grq->load.weight, grq->avg.load_avg)
2869 * And that is shares_weight and is icky. In the (near) UP case it approaches
2870 * (4) while in the normal case it approaches (3). It consistently
2871 * overestimates the ge->load.weight and therefore:
2873 * \Sum ge->load.weight >= tg->weight
2877 static long calc_group_shares(struct cfs_rq *cfs_rq)
2879 long tg_weight, tg_shares, load, shares;
2880 struct task_group *tg = cfs_rq->tg;
2882 tg_shares = READ_ONCE(tg->shares);
2884 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
2886 tg_weight = atomic_long_read(&tg->load_avg);
2888 /* Ensure tg_weight >= load */
2889 tg_weight -= cfs_rq->tg_load_avg_contrib;
2892 shares = (tg_shares * load);
2894 shares /= tg_weight;
2897 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2898 * of a group with small tg->shares value. It is a floor value which is
2899 * assigned as a minimum load.weight to the sched_entity representing
2900 * the group on a CPU.
2902 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2903 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2904 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2905 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2908 return clamp_t(long, shares, MIN_SHARES, tg_shares);
2912 * This calculates the effective runnable weight for a group entity based on
2913 * the group entity weight calculated above.
2915 * Because of the above approximation (2), our group entity weight is
2916 * an load_avg based ratio (3). This means that it includes blocked load and
2917 * does not represent the runnable weight.
2919 * Approximate the group entity's runnable weight per ratio from the group
2922 * grq->avg.runnable_load_avg
2923 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
2926 * However, analogous to above, since the avg numbers are slow, this leads to
2927 * transients in the from-idle case. Instead we use:
2929 * ge->runnable_weight = ge->load.weight *
2931 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
2932 * ----------------------------------------------------- (8)
2933 * max(grq->avg.load_avg, grq->load.weight)
2935 * Where these max() serve both to use the 'instant' values to fix the slow
2936 * from-idle and avoid the /0 on to-idle, similar to (6).
2938 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
2940 long runnable, load_avg;
2942 load_avg = max(cfs_rq->avg.load_avg,
2943 scale_load_down(cfs_rq->load.weight));
2945 runnable = max(cfs_rq->avg.runnable_load_avg,
2946 scale_load_down(cfs_rq->runnable_weight));
2950 runnable /= load_avg;
2952 return clamp_t(long, runnable, MIN_SHARES, shares);
2954 #endif /* CONFIG_SMP */
2956 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2959 * Recomputes the group entity based on the current state of its group
2962 static void update_cfs_group(struct sched_entity *se)
2964 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
2965 long shares, runnable;
2970 if (throttled_hierarchy(gcfs_rq))
2974 runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
2976 if (likely(se->load.weight == shares))
2979 shares = calc_group_shares(gcfs_rq);
2980 runnable = calc_group_runnable(gcfs_rq, shares);
2983 reweight_entity(cfs_rq_of(se), se, shares, runnable);
2986 #else /* CONFIG_FAIR_GROUP_SCHED */
2987 static inline void update_cfs_group(struct sched_entity *se)
2990 #endif /* CONFIG_FAIR_GROUP_SCHED */
2992 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
2994 struct rq *rq = rq_of(cfs_rq);
2996 if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) {
2998 * There are a few boundary cases this might miss but it should
2999 * get called often enough that that should (hopefully) not be
3002 * It will not get called when we go idle, because the idle
3003 * thread is a different class (!fair), nor will the utilization
3004 * number include things like RT tasks.
3006 * As is, the util number is not freq-invariant (we'd have to
3007 * implement arch_scale_freq_capacity() for that).
3011 cpufreq_update_util(rq, flags);
3016 #ifdef CONFIG_FAIR_GROUP_SCHED
3018 * update_tg_load_avg - update the tg's load avg
3019 * @cfs_rq: the cfs_rq whose avg changed
3020 * @force: update regardless of how small the difference
3022 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3023 * However, because tg->load_avg is a global value there are performance
3026 * In order to avoid having to look at the other cfs_rq's, we use a
3027 * differential update where we store the last value we propagated. This in
3028 * turn allows skipping updates if the differential is 'small'.
3030 * Updating tg's load_avg is necessary before update_cfs_share().
3032 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3034 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3037 * No need to update load_avg for root_task_group as it is not used.
3039 if (cfs_rq->tg == &root_task_group)
3042 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3043 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3044 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3049 * Called within set_task_rq() right before setting a task's CPU. The
3050 * caller only guarantees p->pi_lock is held; no other assumptions,
3051 * including the state of rq->lock, should be made.
3053 void set_task_rq_fair(struct sched_entity *se,
3054 struct cfs_rq *prev, struct cfs_rq *next)
3056 u64 p_last_update_time;
3057 u64 n_last_update_time;
3059 if (!sched_feat(ATTACH_AGE_LOAD))
3063 * We are supposed to update the task to "current" time, then its up to
3064 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3065 * getting what current time is, so simply throw away the out-of-date
3066 * time. This will result in the wakee task is less decayed, but giving
3067 * the wakee more load sounds not bad.
3069 if (!(se->avg.last_update_time && prev))
3072 #ifndef CONFIG_64BIT
3074 u64 p_last_update_time_copy;
3075 u64 n_last_update_time_copy;
3078 p_last_update_time_copy = prev->load_last_update_time_copy;
3079 n_last_update_time_copy = next->load_last_update_time_copy;
3083 p_last_update_time = prev->avg.last_update_time;
3084 n_last_update_time = next->avg.last_update_time;
3086 } while (p_last_update_time != p_last_update_time_copy ||
3087 n_last_update_time != n_last_update_time_copy);
3090 p_last_update_time = prev->avg.last_update_time;
3091 n_last_update_time = next->avg.last_update_time;
3093 __update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se);
3094 se->avg.last_update_time = n_last_update_time;
3099 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3100 * propagate its contribution. The key to this propagation is the invariant
3101 * that for each group:
3103 * ge->avg == grq->avg (1)
3105 * _IFF_ we look at the pure running and runnable sums. Because they
3106 * represent the very same entity, just at different points in the hierarchy.
3108 * Per the above update_tg_cfs_util() is trivial and simply copies the running
3109 * sum over (but still wrong, because the group entity and group rq do not have
3110 * their PELT windows aligned).
3112 * However, update_tg_cfs_runnable() is more complex. So we have:
3114 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3116 * And since, like util, the runnable part should be directly transferable,
3117 * the following would _appear_ to be the straight forward approach:
3119 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3121 * And per (1) we have:
3123 * ge->avg.runnable_avg == grq->avg.runnable_avg
3127 * ge->load.weight * grq->avg.load_avg
3128 * ge->avg.load_avg = ----------------------------------- (4)
3131 * Except that is wrong!
3133 * Because while for entities historical weight is not important and we
3134 * really only care about our future and therefore can consider a pure
3135 * runnable sum, runqueues can NOT do this.
3137 * We specifically want runqueues to have a load_avg that includes
3138 * historical weights. Those represent the blocked load, the load we expect
3139 * to (shortly) return to us. This only works by keeping the weights as
3140 * integral part of the sum. We therefore cannot decompose as per (3).
3142 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3143 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3144 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3145 * runnable section of these tasks overlap (or not). If they were to perfectly
3146 * align the rq as a whole would be runnable 2/3 of the time. If however we
3147 * always have at least 1 runnable task, the rq as a whole is always runnable.
3149 * So we'll have to approximate.. :/
3151 * Given the constraint:
3153 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3155 * We can construct a rule that adds runnable to a rq by assuming minimal
3158 * On removal, we'll assume each task is equally runnable; which yields:
3160 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3162 * XXX: only do this for the part of runnable > running ?
3167 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3169 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3171 /* Nothing to update */
3176 * The relation between sum and avg is:
3178 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3180 * however, the PELT windows are not aligned between grq and gse.
3183 /* Set new sched_entity's utilization */
3184 se->avg.util_avg = gcfs_rq->avg.util_avg;
3185 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3187 /* Update parent cfs_rq utilization */
3188 add_positive(&cfs_rq->avg.util_avg, delta);
3189 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3193 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3195 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3196 unsigned long runnable_load_avg, load_avg;
3197 u64 runnable_load_sum, load_sum = 0;
3203 gcfs_rq->prop_runnable_sum = 0;
3205 if (runnable_sum >= 0) {
3207 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3208 * the CPU is saturated running == runnable.
3210 runnable_sum += se->avg.load_sum;
3211 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3214 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3215 * assuming all tasks are equally runnable.
3217 if (scale_load_down(gcfs_rq->load.weight)) {
3218 load_sum = div_s64(gcfs_rq->avg.load_sum,
3219 scale_load_down(gcfs_rq->load.weight));
3222 /* But make sure to not inflate se's runnable */
3223 runnable_sum = min(se->avg.load_sum, load_sum);
3227 * runnable_sum can't be lower than running_sum
3228 * As running sum is scale with CPU capacity wehreas the runnable sum
3229 * is not we rescale running_sum 1st
3231 running_sum = se->avg.util_sum /
3232 arch_scale_cpu_capacity(NULL, cpu_of(rq_of(cfs_rq)));
3233 runnable_sum = max(runnable_sum, running_sum);
3235 load_sum = (s64)se_weight(se) * runnable_sum;
3236 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3238 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3239 delta_avg = load_avg - se->avg.load_avg;
3241 se->avg.load_sum = runnable_sum;
3242 se->avg.load_avg = load_avg;
3243 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3244 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3246 runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3247 runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3248 delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3249 delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3251 se->avg.runnable_load_sum = runnable_sum;
3252 se->avg.runnable_load_avg = runnable_load_avg;
3255 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3256 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3260 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3262 cfs_rq->propagate = 1;
3263 cfs_rq->prop_runnable_sum += runnable_sum;
3266 /* Update task and its cfs_rq load average */
3267 static inline int propagate_entity_load_avg(struct sched_entity *se)
3269 struct cfs_rq *cfs_rq, *gcfs_rq;
3271 if (entity_is_task(se))
3274 gcfs_rq = group_cfs_rq(se);
3275 if (!gcfs_rq->propagate)
3278 gcfs_rq->propagate = 0;
3280 cfs_rq = cfs_rq_of(se);
3282 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3284 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3285 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3291 * Check if we need to update the load and the utilization of a blocked
3294 static inline bool skip_blocked_update(struct sched_entity *se)
3296 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3299 * If sched_entity still have not zero load or utilization, we have to
3302 if (se->avg.load_avg || se->avg.util_avg)
3306 * If there is a pending propagation, we have to update the load and
3307 * the utilization of the sched_entity:
3309 if (gcfs_rq->propagate)
3313 * Otherwise, the load and the utilization of the sched_entity is
3314 * already zero and there is no pending propagation, so it will be a
3315 * waste of time to try to decay it:
3320 #else /* CONFIG_FAIR_GROUP_SCHED */
3322 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3324 static inline int propagate_entity_load_avg(struct sched_entity *se)
3329 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3331 #endif /* CONFIG_FAIR_GROUP_SCHED */
3334 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3335 * @now: current time, as per cfs_rq_clock_task()
3336 * @cfs_rq: cfs_rq to update
3338 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3339 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3340 * post_init_entity_util_avg().
3342 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3344 * Returns true if the load decayed or we removed load.
3346 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3347 * call update_tg_load_avg() when this function returns true.
3350 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3352 unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3353 struct sched_avg *sa = &cfs_rq->avg;
3356 if (cfs_rq->removed.nr) {
3358 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3360 raw_spin_lock(&cfs_rq->removed.lock);
3361 swap(cfs_rq->removed.util_avg, removed_util);
3362 swap(cfs_rq->removed.load_avg, removed_load);
3363 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3364 cfs_rq->removed.nr = 0;
3365 raw_spin_unlock(&cfs_rq->removed.lock);
3368 sub_positive(&sa->load_avg, r);
3369 sub_positive(&sa->load_sum, r * divider);
3372 sub_positive(&sa->util_avg, r);
3373 sub_positive(&sa->util_sum, r * divider);
3375 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3380 decayed |= __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq);
3382 #ifndef CONFIG_64BIT
3384 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3388 cfs_rq_util_change(cfs_rq, 0);
3394 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3395 * @cfs_rq: cfs_rq to attach to
3396 * @se: sched_entity to attach
3397 * @flags: migration hints
3399 * Must call update_cfs_rq_load_avg() before this, since we rely on
3400 * cfs_rq->avg.last_update_time being current.
3402 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3404 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3407 * When we attach the @se to the @cfs_rq, we must align the decay
3408 * window because without that, really weird and wonderful things can
3413 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3414 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3417 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3418 * period_contrib. This isn't strictly correct, but since we're
3419 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3422 se->avg.util_sum = se->avg.util_avg * divider;
3424 se->avg.load_sum = divider;
3425 if (se_weight(se)) {
3427 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3430 se->avg.runnable_load_sum = se->avg.load_sum;
3432 enqueue_load_avg(cfs_rq, se);
3433 cfs_rq->avg.util_avg += se->avg.util_avg;
3434 cfs_rq->avg.util_sum += se->avg.util_sum;
3436 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3438 cfs_rq_util_change(cfs_rq, flags);
3442 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3443 * @cfs_rq: cfs_rq to detach from
3444 * @se: sched_entity to detach
3446 * Must call update_cfs_rq_load_avg() before this, since we rely on
3447 * cfs_rq->avg.last_update_time being current.
3449 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3451 dequeue_load_avg(cfs_rq, se);
3452 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3453 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3455 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3457 cfs_rq_util_change(cfs_rq, 0);
3461 * Optional action to be done while updating the load average
3463 #define UPDATE_TG 0x1
3464 #define SKIP_AGE_LOAD 0x2
3465 #define DO_ATTACH 0x4
3467 /* Update task and its cfs_rq load average */
3468 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3470 u64 now = cfs_rq_clock_task(cfs_rq);
3471 struct rq *rq = rq_of(cfs_rq);
3472 int cpu = cpu_of(rq);
3476 * Track task load average for carrying it to new CPU after migrated, and
3477 * track group sched_entity load average for task_h_load calc in migration
3479 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3480 __update_load_avg_se(now, cpu, cfs_rq, se);
3482 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3483 decayed |= propagate_entity_load_avg(se);
3485 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3488 * DO_ATTACH means we're here from enqueue_entity().
3489 * !last_update_time means we've passed through
3490 * migrate_task_rq_fair() indicating we migrated.
3492 * IOW we're enqueueing a task on a new CPU.
3494 attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
3495 update_tg_load_avg(cfs_rq, 0);
3497 } else if (decayed && (flags & UPDATE_TG))
3498 update_tg_load_avg(cfs_rq, 0);
3501 #ifndef CONFIG_64BIT
3502 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3504 u64 last_update_time_copy;
3505 u64 last_update_time;
3508 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3510 last_update_time = cfs_rq->avg.last_update_time;
3511 } while (last_update_time != last_update_time_copy);
3513 return last_update_time;
3516 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3518 return cfs_rq->avg.last_update_time;
3523 * Synchronize entity load avg of dequeued entity without locking
3526 void sync_entity_load_avg(struct sched_entity *se)
3528 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3529 u64 last_update_time;
3531 last_update_time = cfs_rq_last_update_time(cfs_rq);
3532 __update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se);
3536 * Task first catches up with cfs_rq, and then subtract
3537 * itself from the cfs_rq (task must be off the queue now).
3539 void remove_entity_load_avg(struct sched_entity *se)
3541 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3542 unsigned long flags;
3545 * tasks cannot exit without having gone through wake_up_new_task() ->
3546 * post_init_entity_util_avg() which will have added things to the
3547 * cfs_rq, so we can remove unconditionally.
3549 * Similarly for groups, they will have passed through
3550 * post_init_entity_util_avg() before unregister_sched_fair_group()
3554 sync_entity_load_avg(se);
3556 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3557 ++cfs_rq->removed.nr;
3558 cfs_rq->removed.util_avg += se->avg.util_avg;
3559 cfs_rq->removed.load_avg += se->avg.load_avg;
3560 cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */
3561 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3564 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3566 return cfs_rq->avg.runnable_load_avg;
3569 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3571 return cfs_rq->avg.load_avg;
3574 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3576 static inline unsigned long task_util(struct task_struct *p)
3578 return READ_ONCE(p->se.avg.util_avg);
3581 static inline unsigned long _task_util_est(struct task_struct *p)
3583 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3585 return max(ue.ewma, ue.enqueued);
3588 static inline unsigned long task_util_est(struct task_struct *p)
3590 return max(task_util(p), _task_util_est(p));
3593 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3594 struct task_struct *p)
3596 unsigned int enqueued;
3598 if (!sched_feat(UTIL_EST))
3601 /* Update root cfs_rq's estimated utilization */
3602 enqueued = cfs_rq->avg.util_est.enqueued;
3603 enqueued += (_task_util_est(p) | UTIL_AVG_UNCHANGED);
3604 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3608 * Check if a (signed) value is within a specified (unsigned) margin,
3609 * based on the observation that:
3611 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3613 * NOTE: this only works when value + maring < INT_MAX.
3615 static inline bool within_margin(int value, int margin)
3617 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3621 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3623 long last_ewma_diff;
3626 if (!sched_feat(UTIL_EST))
3629 /* Update root cfs_rq's estimated utilization */
3630 ue.enqueued = cfs_rq->avg.util_est.enqueued;
3631 ue.enqueued -= min_t(unsigned int, ue.enqueued,
3632 (_task_util_est(p) | UTIL_AVG_UNCHANGED));
3633 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3636 * Skip update of task's estimated utilization when the task has not
3637 * yet completed an activation, e.g. being migrated.
3643 * If the PELT values haven't changed since enqueue time,
3644 * skip the util_est update.
3646 ue = p->se.avg.util_est;
3647 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3651 * Skip update of task's estimated utilization when its EWMA is
3652 * already ~1% close to its last activation value.
3654 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3655 last_ewma_diff = ue.enqueued - ue.ewma;
3656 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3660 * Update Task's estimated utilization
3662 * When *p completes an activation we can consolidate another sample
3663 * of the task size. This is done by storing the current PELT value
3664 * as ue.enqueued and by using this value to update the Exponential
3665 * Weighted Moving Average (EWMA):
3667 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
3668 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
3669 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
3670 * = w * ( last_ewma_diff ) + ewma(t-1)
3671 * = w * (last_ewma_diff + ewma(t-1) / w)
3673 * Where 'w' is the weight of new samples, which is configured to be
3674 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3676 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
3677 ue.ewma += last_ewma_diff;
3678 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
3679 WRITE_ONCE(p->se.avg.util_est, ue);
3682 #else /* CONFIG_SMP */
3684 #define UPDATE_TG 0x0
3685 #define SKIP_AGE_LOAD 0x0
3686 #define DO_ATTACH 0x0
3688 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3690 cfs_rq_util_change(cfs_rq, 0);
3693 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3696 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
3698 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3700 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3706 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
3709 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
3712 #endif /* CONFIG_SMP */
3714 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3716 #ifdef CONFIG_SCHED_DEBUG
3717 s64 d = se->vruntime - cfs_rq->min_vruntime;
3722 if (d > 3*sysctl_sched_latency)
3723 schedstat_inc(cfs_rq->nr_spread_over);
3728 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3730 u64 vruntime = cfs_rq->min_vruntime;
3733 * The 'current' period is already promised to the current tasks,
3734 * however the extra weight of the new task will slow them down a
3735 * little, place the new task so that it fits in the slot that
3736 * stays open at the end.
3738 if (initial && sched_feat(START_DEBIT))
3739 vruntime += sched_vslice(cfs_rq, se);
3741 /* sleeps up to a single latency don't count. */
3743 unsigned long thresh = sysctl_sched_latency;
3746 * Halve their sleep time's effect, to allow
3747 * for a gentler effect of sleepers:
3749 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3755 /* ensure we never gain time by being placed backwards. */
3756 se->vruntime = max_vruntime(se->vruntime, vruntime);
3759 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3761 static inline void check_schedstat_required(void)
3763 #ifdef CONFIG_SCHEDSTATS
3764 if (schedstat_enabled())
3767 /* Force schedstat enabled if a dependent tracepoint is active */
3768 if (trace_sched_stat_wait_enabled() ||
3769 trace_sched_stat_sleep_enabled() ||
3770 trace_sched_stat_iowait_enabled() ||
3771 trace_sched_stat_blocked_enabled() ||
3772 trace_sched_stat_runtime_enabled()) {
3773 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3774 "stat_blocked and stat_runtime require the "
3775 "kernel parameter schedstats=enable or "
3776 "kernel.sched_schedstats=1\n");
3787 * update_min_vruntime()
3788 * vruntime -= min_vruntime
3792 * update_min_vruntime()
3793 * vruntime += min_vruntime
3795 * this way the vruntime transition between RQs is done when both
3796 * min_vruntime are up-to-date.
3800 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3801 * vruntime -= min_vruntime
3805 * update_min_vruntime()
3806 * vruntime += min_vruntime
3808 * this way we don't have the most up-to-date min_vruntime on the originating
3809 * CPU and an up-to-date min_vruntime on the destination CPU.
3813 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3815 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3816 bool curr = cfs_rq->curr == se;
3819 * If we're the current task, we must renormalise before calling
3823 se->vruntime += cfs_rq->min_vruntime;
3825 update_curr(cfs_rq);
3828 * Otherwise, renormalise after, such that we're placed at the current
3829 * moment in time, instead of some random moment in the past. Being
3830 * placed in the past could significantly boost this task to the
3831 * fairness detriment of existing tasks.
3833 if (renorm && !curr)
3834 se->vruntime += cfs_rq->min_vruntime;
3837 * When enqueuing a sched_entity, we must:
3838 * - Update loads to have both entity and cfs_rq synced with now.
3839 * - Add its load to cfs_rq->runnable_avg
3840 * - For group_entity, update its weight to reflect the new share of
3842 * - Add its new weight to cfs_rq->load.weight
3844 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
3845 update_cfs_group(se);
3846 enqueue_runnable_load_avg(cfs_rq, se);
3847 account_entity_enqueue(cfs_rq, se);
3849 if (flags & ENQUEUE_WAKEUP)
3850 place_entity(cfs_rq, se, 0);
3852 check_schedstat_required();
3853 update_stats_enqueue(cfs_rq, se, flags);
3854 check_spread(cfs_rq, se);
3856 __enqueue_entity(cfs_rq, se);
3859 if (cfs_rq->nr_running == 1) {
3860 list_add_leaf_cfs_rq(cfs_rq);
3861 check_enqueue_throttle(cfs_rq);
3865 static void __clear_buddies_last(struct sched_entity *se)
3867 for_each_sched_entity(se) {
3868 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3869 if (cfs_rq->last != se)
3872 cfs_rq->last = NULL;
3876 static void __clear_buddies_next(struct sched_entity *se)
3878 for_each_sched_entity(se) {
3879 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3880 if (cfs_rq->next != se)
3883 cfs_rq->next = NULL;
3887 static void __clear_buddies_skip(struct sched_entity *se)
3889 for_each_sched_entity(se) {
3890 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3891 if (cfs_rq->skip != se)
3894 cfs_rq->skip = NULL;
3898 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3900 if (cfs_rq->last == se)
3901 __clear_buddies_last(se);
3903 if (cfs_rq->next == se)
3904 __clear_buddies_next(se);
3906 if (cfs_rq->skip == se)
3907 __clear_buddies_skip(se);
3910 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3913 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3916 * Update run-time statistics of the 'current'.
3918 update_curr(cfs_rq);
3921 * When dequeuing a sched_entity, we must:
3922 * - Update loads to have both entity and cfs_rq synced with now.
3923 * - Substract its load from the cfs_rq->runnable_avg.
3924 * - Substract its previous weight from cfs_rq->load.weight.
3925 * - For group entity, update its weight to reflect the new share
3926 * of its group cfs_rq.
3928 update_load_avg(cfs_rq, se, UPDATE_TG);
3929 dequeue_runnable_load_avg(cfs_rq, se);
3931 update_stats_dequeue(cfs_rq, se, flags);
3933 clear_buddies(cfs_rq, se);
3935 if (se != cfs_rq->curr)
3936 __dequeue_entity(cfs_rq, se);
3938 account_entity_dequeue(cfs_rq, se);
3941 * Normalize after update_curr(); which will also have moved
3942 * min_vruntime if @se is the one holding it back. But before doing
3943 * update_min_vruntime() again, which will discount @se's position and
3944 * can move min_vruntime forward still more.
3946 if (!(flags & DEQUEUE_SLEEP))
3947 se->vruntime -= cfs_rq->min_vruntime;
3949 /* return excess runtime on last dequeue */
3950 return_cfs_rq_runtime(cfs_rq);
3952 update_cfs_group(se);
3955 * Now advance min_vruntime if @se was the entity holding it back,
3956 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
3957 * put back on, and if we advance min_vruntime, we'll be placed back
3958 * further than we started -- ie. we'll be penalized.
3960 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
3961 update_min_vruntime(cfs_rq);
3965 * Preempt the current task with a newly woken task if needed:
3968 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3970 unsigned long ideal_runtime, delta_exec;
3971 struct sched_entity *se;
3974 ideal_runtime = sched_slice(cfs_rq, curr);
3975 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3976 if (delta_exec > ideal_runtime) {
3977 resched_curr(rq_of(cfs_rq));
3979 * The current task ran long enough, ensure it doesn't get
3980 * re-elected due to buddy favours.
3982 clear_buddies(cfs_rq, curr);
3987 * Ensure that a task that missed wakeup preemption by a
3988 * narrow margin doesn't have to wait for a full slice.
3989 * This also mitigates buddy induced latencies under load.
3991 if (delta_exec < sysctl_sched_min_granularity)
3994 se = __pick_first_entity(cfs_rq);
3995 delta = curr->vruntime - se->vruntime;
4000 if (delta > ideal_runtime)
4001 resched_curr(rq_of(cfs_rq));
4005 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4007 /* 'current' is not kept within the tree. */
4010 * Any task has to be enqueued before it get to execute on
4011 * a CPU. So account for the time it spent waiting on the
4014 update_stats_wait_end(cfs_rq, se);
4015 __dequeue_entity(cfs_rq, se);
4016 update_load_avg(cfs_rq, se, UPDATE_TG);
4019 update_stats_curr_start(cfs_rq, se);
4023 * Track our maximum slice length, if the CPU's load is at
4024 * least twice that of our own weight (i.e. dont track it
4025 * when there are only lesser-weight tasks around):
4027 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
4028 schedstat_set(se->statistics.slice_max,
4029 max((u64)schedstat_val(se->statistics.slice_max),
4030 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4033 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4037 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4040 * Pick the next process, keeping these things in mind, in this order:
4041 * 1) keep things fair between processes/task groups
4042 * 2) pick the "next" process, since someone really wants that to run
4043 * 3) pick the "last" process, for cache locality
4044 * 4) do not run the "skip" process, if something else is available
4046 static struct sched_entity *
4047 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4049 struct sched_entity *left = __pick_first_entity(cfs_rq);
4050 struct sched_entity *se;
4053 * If curr is set we have to see if its left of the leftmost entity
4054 * still in the tree, provided there was anything in the tree at all.
4056 if (!left || (curr && entity_before(curr, left)))
4059 se = left; /* ideally we run the leftmost entity */
4062 * Avoid running the skip buddy, if running something else can
4063 * be done without getting too unfair.
4065 if (cfs_rq->skip == se) {
4066 struct sched_entity *second;
4069 second = __pick_first_entity(cfs_rq);
4071 second = __pick_next_entity(se);
4072 if (!second || (curr && entity_before(curr, second)))
4076 if (second && wakeup_preempt_entity(second, left) < 1)
4081 * Prefer last buddy, try to return the CPU to a preempted task.
4083 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4087 * Someone really wants this to run. If it's not unfair, run it.
4089 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4092 clear_buddies(cfs_rq, se);
4097 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4099 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4102 * If still on the runqueue then deactivate_task()
4103 * was not called and update_curr() has to be done:
4106 update_curr(cfs_rq);
4108 /* throttle cfs_rqs exceeding runtime */
4109 check_cfs_rq_runtime(cfs_rq);
4111 check_spread(cfs_rq, prev);
4114 update_stats_wait_start(cfs_rq, prev);
4115 /* Put 'current' back into the tree. */
4116 __enqueue_entity(cfs_rq, prev);
4117 /* in !on_rq case, update occurred at dequeue */
4118 update_load_avg(cfs_rq, prev, 0);
4120 cfs_rq->curr = NULL;
4124 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4127 * Update run-time statistics of the 'current'.
4129 update_curr(cfs_rq);
4132 * Ensure that runnable average is periodically updated.
4134 update_load_avg(cfs_rq, curr, UPDATE_TG);
4135 update_cfs_group(curr);
4137 #ifdef CONFIG_SCHED_HRTICK
4139 * queued ticks are scheduled to match the slice, so don't bother
4140 * validating it and just reschedule.
4143 resched_curr(rq_of(cfs_rq));
4147 * don't let the period tick interfere with the hrtick preemption
4149 if (!sched_feat(DOUBLE_TICK) &&
4150 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4154 if (cfs_rq->nr_running > 1)
4155 check_preempt_tick(cfs_rq, curr);
4159 /**************************************************
4160 * CFS bandwidth control machinery
4163 #ifdef CONFIG_CFS_BANDWIDTH
4165 #ifdef HAVE_JUMP_LABEL
4166 static struct static_key __cfs_bandwidth_used;
4168 static inline bool cfs_bandwidth_used(void)
4170 return static_key_false(&__cfs_bandwidth_used);
4173 void cfs_bandwidth_usage_inc(void)
4175 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4178 void cfs_bandwidth_usage_dec(void)
4180 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4182 #else /* HAVE_JUMP_LABEL */
4183 static bool cfs_bandwidth_used(void)
4188 void cfs_bandwidth_usage_inc(void) {}
4189 void cfs_bandwidth_usage_dec(void) {}
4190 #endif /* HAVE_JUMP_LABEL */
4193 * default period for cfs group bandwidth.
4194 * default: 0.1s, units: nanoseconds
4196 static inline u64 default_cfs_period(void)
4198 return 100000000ULL;
4201 static inline u64 sched_cfs_bandwidth_slice(void)
4203 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4207 * Replenish runtime according to assigned quota and update expiration time.
4208 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4209 * additional synchronization around rq->lock.
4211 * requires cfs_b->lock
4213 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4217 if (cfs_b->quota == RUNTIME_INF)
4220 now = sched_clock_cpu(smp_processor_id());
4221 cfs_b->runtime = cfs_b->quota;
4222 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4223 cfs_b->expires_seq++;
4226 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4228 return &tg->cfs_bandwidth;
4231 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4232 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4234 if (unlikely(cfs_rq->throttle_count))
4235 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4237 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4240 /* returns 0 on failure to allocate runtime */
4241 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4243 struct task_group *tg = cfs_rq->tg;
4244 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4245 u64 amount = 0, min_amount, expires;
4248 /* note: this is a positive sum as runtime_remaining <= 0 */
4249 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4251 raw_spin_lock(&cfs_b->lock);
4252 if (cfs_b->quota == RUNTIME_INF)
4253 amount = min_amount;
4255 start_cfs_bandwidth(cfs_b);
4257 if (cfs_b->runtime > 0) {
4258 amount = min(cfs_b->runtime, min_amount);
4259 cfs_b->runtime -= amount;
4263 expires_seq = cfs_b->expires_seq;
4264 expires = cfs_b->runtime_expires;
4265 raw_spin_unlock(&cfs_b->lock);
4267 cfs_rq->runtime_remaining += amount;
4269 * we may have advanced our local expiration to account for allowed
4270 * spread between our sched_clock and the one on which runtime was
4273 if (cfs_rq->expires_seq != expires_seq) {
4274 cfs_rq->expires_seq = expires_seq;
4275 cfs_rq->runtime_expires = expires;
4278 return cfs_rq->runtime_remaining > 0;
4282 * Note: This depends on the synchronization provided by sched_clock and the
4283 * fact that rq->clock snapshots this value.
4285 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4287 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4289 /* if the deadline is ahead of our clock, nothing to do */
4290 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4293 if (cfs_rq->runtime_remaining < 0)
4297 * If the local deadline has passed we have to consider the
4298 * possibility that our sched_clock is 'fast' and the global deadline
4299 * has not truly expired.
4301 * Fortunately we can check determine whether this the case by checking
4302 * whether the global deadline(cfs_b->expires_seq) has advanced.
4304 if (cfs_rq->expires_seq == cfs_b->expires_seq) {
4305 /* extend local deadline, drift is bounded above by 2 ticks */
4306 cfs_rq->runtime_expires += TICK_NSEC;
4308 /* global deadline is ahead, expiration has passed */
4309 cfs_rq->runtime_remaining = 0;
4313 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4315 /* dock delta_exec before expiring quota (as it could span periods) */
4316 cfs_rq->runtime_remaining -= delta_exec;
4317 expire_cfs_rq_runtime(cfs_rq);
4319 if (likely(cfs_rq->runtime_remaining > 0))
4323 * if we're unable to extend our runtime we resched so that the active
4324 * hierarchy can be throttled
4326 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4327 resched_curr(rq_of(cfs_rq));
4330 static __always_inline
4331 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4333 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4336 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4339 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4341 return cfs_bandwidth_used() && cfs_rq->throttled;
4344 /* check whether cfs_rq, or any parent, is throttled */
4345 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4347 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4351 * Ensure that neither of the group entities corresponding to src_cpu or
4352 * dest_cpu are members of a throttled hierarchy when performing group
4353 * load-balance operations.
4355 static inline int throttled_lb_pair(struct task_group *tg,
4356 int src_cpu, int dest_cpu)
4358 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4360 src_cfs_rq = tg->cfs_rq[src_cpu];
4361 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4363 return throttled_hierarchy(src_cfs_rq) ||
4364 throttled_hierarchy(dest_cfs_rq);
4367 static int tg_unthrottle_up(struct task_group *tg, void *data)
4369 struct rq *rq = data;
4370 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4372 cfs_rq->throttle_count--;
4373 if (!cfs_rq->throttle_count) {
4374 /* adjust cfs_rq_clock_task() */
4375 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4376 cfs_rq->throttled_clock_task;
4382 static int tg_throttle_down(struct task_group *tg, void *data)
4384 struct rq *rq = data;
4385 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4387 /* group is entering throttled state, stop time */
4388 if (!cfs_rq->throttle_count)
4389 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4390 cfs_rq->throttle_count++;
4395 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4397 struct rq *rq = rq_of(cfs_rq);
4398 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4399 struct sched_entity *se;
4400 long task_delta, dequeue = 1;
4403 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4405 /* freeze hierarchy runnable averages while throttled */
4407 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4410 task_delta = cfs_rq->h_nr_running;
4411 for_each_sched_entity(se) {
4412 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4413 /* throttled entity or throttle-on-deactivate */
4418 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4419 qcfs_rq->h_nr_running -= task_delta;
4421 if (qcfs_rq->load.weight)
4426 sub_nr_running(rq, task_delta);
4428 cfs_rq->throttled = 1;
4429 cfs_rq->throttled_clock = rq_clock(rq);
4430 raw_spin_lock(&cfs_b->lock);
4431 empty = list_empty(&cfs_b->throttled_cfs_rq);
4434 * Add to the _head_ of the list, so that an already-started
4435 * distribute_cfs_runtime will not see us
4437 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4440 * If we're the first throttled task, make sure the bandwidth
4444 start_cfs_bandwidth(cfs_b);
4446 raw_spin_unlock(&cfs_b->lock);
4449 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4451 struct rq *rq = rq_of(cfs_rq);
4452 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4453 struct sched_entity *se;
4457 se = cfs_rq->tg->se[cpu_of(rq)];
4459 cfs_rq->throttled = 0;
4461 update_rq_clock(rq);
4463 raw_spin_lock(&cfs_b->lock);
4464 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4465 list_del_rcu(&cfs_rq->throttled_list);
4466 raw_spin_unlock(&cfs_b->lock);
4468 /* update hierarchical throttle state */
4469 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4471 if (!cfs_rq->load.weight)
4474 task_delta = cfs_rq->h_nr_running;
4475 for_each_sched_entity(se) {
4479 cfs_rq = cfs_rq_of(se);
4481 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4482 cfs_rq->h_nr_running += task_delta;
4484 if (cfs_rq_throttled(cfs_rq))
4489 add_nr_running(rq, task_delta);
4491 /* Determine whether we need to wake up potentially idle CPU: */
4492 if (rq->curr == rq->idle && rq->cfs.nr_running)
4496 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4497 u64 remaining, u64 expires)
4499 struct cfs_rq *cfs_rq;
4501 u64 starting_runtime = remaining;
4504 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4506 struct rq *rq = rq_of(cfs_rq);
4510 if (!cfs_rq_throttled(cfs_rq))
4513 runtime = -cfs_rq->runtime_remaining + 1;
4514 if (runtime > remaining)
4515 runtime = remaining;
4516 remaining -= runtime;
4518 cfs_rq->runtime_remaining += runtime;
4519 cfs_rq->runtime_expires = expires;
4521 /* we check whether we're throttled above */
4522 if (cfs_rq->runtime_remaining > 0)
4523 unthrottle_cfs_rq(cfs_rq);
4533 return starting_runtime - remaining;
4537 * Responsible for refilling a task_group's bandwidth and unthrottling its
4538 * cfs_rqs as appropriate. If there has been no activity within the last
4539 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4540 * used to track this state.
4542 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4544 u64 runtime, runtime_expires;
4547 /* no need to continue the timer with no bandwidth constraint */
4548 if (cfs_b->quota == RUNTIME_INF)
4549 goto out_deactivate;
4551 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4552 cfs_b->nr_periods += overrun;
4555 * idle depends on !throttled (for the case of a large deficit), and if
4556 * we're going inactive then everything else can be deferred
4558 if (cfs_b->idle && !throttled)
4559 goto out_deactivate;
4561 __refill_cfs_bandwidth_runtime(cfs_b);
4564 /* mark as potentially idle for the upcoming period */
4569 /* account preceding periods in which throttling occurred */
4570 cfs_b->nr_throttled += overrun;
4572 runtime_expires = cfs_b->runtime_expires;
4575 * This check is repeated as we are holding onto the new bandwidth while
4576 * we unthrottle. This can potentially race with an unthrottled group
4577 * trying to acquire new bandwidth from the global pool. This can result
4578 * in us over-using our runtime if it is all used during this loop, but
4579 * only by limited amounts in that extreme case.
4581 while (throttled && cfs_b->runtime > 0) {
4582 runtime = cfs_b->runtime;
4583 raw_spin_unlock(&cfs_b->lock);
4584 /* we can't nest cfs_b->lock while distributing bandwidth */
4585 runtime = distribute_cfs_runtime(cfs_b, runtime,
4587 raw_spin_lock(&cfs_b->lock);
4589 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4591 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4595 * While we are ensured activity in the period following an
4596 * unthrottle, this also covers the case in which the new bandwidth is
4597 * insufficient to cover the existing bandwidth deficit. (Forcing the
4598 * timer to remain active while there are any throttled entities.)
4608 /* a cfs_rq won't donate quota below this amount */
4609 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4610 /* minimum remaining period time to redistribute slack quota */
4611 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4612 /* how long we wait to gather additional slack before distributing */
4613 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4616 * Are we near the end of the current quota period?
4618 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4619 * hrtimer base being cleared by hrtimer_start. In the case of
4620 * migrate_hrtimers, base is never cleared, so we are fine.
4622 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4624 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4627 /* if the call-back is running a quota refresh is already occurring */
4628 if (hrtimer_callback_running(refresh_timer))
4631 /* is a quota refresh about to occur? */
4632 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4633 if (remaining < min_expire)
4639 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4641 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4643 /* if there's a quota refresh soon don't bother with slack */
4644 if (runtime_refresh_within(cfs_b, min_left))
4647 hrtimer_start(&cfs_b->slack_timer,
4648 ns_to_ktime(cfs_bandwidth_slack_period),
4652 /* we know any runtime found here is valid as update_curr() precedes return */
4653 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4655 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4656 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4658 if (slack_runtime <= 0)
4661 raw_spin_lock(&cfs_b->lock);
4662 if (cfs_b->quota != RUNTIME_INF &&
4663 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4664 cfs_b->runtime += slack_runtime;
4666 /* we are under rq->lock, defer unthrottling using a timer */
4667 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4668 !list_empty(&cfs_b->throttled_cfs_rq))
4669 start_cfs_slack_bandwidth(cfs_b);
4671 raw_spin_unlock(&cfs_b->lock);
4673 /* even if it's not valid for return we don't want to try again */
4674 cfs_rq->runtime_remaining -= slack_runtime;
4677 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4679 if (!cfs_bandwidth_used())
4682 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4685 __return_cfs_rq_runtime(cfs_rq);
4689 * This is done with a timer (instead of inline with bandwidth return) since
4690 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4692 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4694 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4697 /* confirm we're still not at a refresh boundary */
4698 raw_spin_lock(&cfs_b->lock);
4699 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4700 raw_spin_unlock(&cfs_b->lock);
4704 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4705 runtime = cfs_b->runtime;
4707 expires = cfs_b->runtime_expires;
4708 raw_spin_unlock(&cfs_b->lock);
4713 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4715 raw_spin_lock(&cfs_b->lock);
4716 if (expires == cfs_b->runtime_expires)
4717 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4718 raw_spin_unlock(&cfs_b->lock);
4722 * When a group wakes up we want to make sure that its quota is not already
4723 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4724 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4726 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4728 if (!cfs_bandwidth_used())
4731 /* an active group must be handled by the update_curr()->put() path */
4732 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4735 /* ensure the group is not already throttled */
4736 if (cfs_rq_throttled(cfs_rq))
4739 /* update runtime allocation */
4740 account_cfs_rq_runtime(cfs_rq, 0);
4741 if (cfs_rq->runtime_remaining <= 0)
4742 throttle_cfs_rq(cfs_rq);
4745 static void sync_throttle(struct task_group *tg, int cpu)
4747 struct cfs_rq *pcfs_rq, *cfs_rq;
4749 if (!cfs_bandwidth_used())
4755 cfs_rq = tg->cfs_rq[cpu];
4756 pcfs_rq = tg->parent->cfs_rq[cpu];
4758 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4759 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4762 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4763 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4765 if (!cfs_bandwidth_used())
4768 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4772 * it's possible for a throttled entity to be forced into a running
4773 * state (e.g. set_curr_task), in this case we're finished.
4775 if (cfs_rq_throttled(cfs_rq))
4778 throttle_cfs_rq(cfs_rq);
4782 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4784 struct cfs_bandwidth *cfs_b =
4785 container_of(timer, struct cfs_bandwidth, slack_timer);
4787 do_sched_cfs_slack_timer(cfs_b);
4789 return HRTIMER_NORESTART;
4792 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4794 struct cfs_bandwidth *cfs_b =
4795 container_of(timer, struct cfs_bandwidth, period_timer);
4799 raw_spin_lock(&cfs_b->lock);
4801 overrun = hrtimer_forward_now(timer, cfs_b->period);
4805 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4808 cfs_b->period_active = 0;
4809 raw_spin_unlock(&cfs_b->lock);
4811 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4814 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4816 raw_spin_lock_init(&cfs_b->lock);
4818 cfs_b->quota = RUNTIME_INF;
4819 cfs_b->period = ns_to_ktime(default_cfs_period());
4821 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4822 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4823 cfs_b->period_timer.function = sched_cfs_period_timer;
4824 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4825 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4828 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4830 cfs_rq->runtime_enabled = 0;
4831 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4834 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4838 lockdep_assert_held(&cfs_b->lock);
4840 if (cfs_b->period_active)
4843 cfs_b->period_active = 1;
4844 overrun = hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4845 cfs_b->runtime_expires += (overrun + 1) * ktime_to_ns(cfs_b->period);
4846 cfs_b->expires_seq++;
4847 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4850 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4852 /* init_cfs_bandwidth() was not called */
4853 if (!cfs_b->throttled_cfs_rq.next)
4856 hrtimer_cancel(&cfs_b->period_timer);
4857 hrtimer_cancel(&cfs_b->slack_timer);
4861 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
4863 * The race is harmless, since modifying bandwidth settings of unhooked group
4864 * bits doesn't do much.
4867 /* cpu online calback */
4868 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4870 struct task_group *tg;
4872 lockdep_assert_held(&rq->lock);
4875 list_for_each_entry_rcu(tg, &task_groups, list) {
4876 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
4877 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4879 raw_spin_lock(&cfs_b->lock);
4880 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4881 raw_spin_unlock(&cfs_b->lock);
4886 /* cpu offline callback */
4887 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4889 struct task_group *tg;
4891 lockdep_assert_held(&rq->lock);
4894 list_for_each_entry_rcu(tg, &task_groups, list) {
4895 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4897 if (!cfs_rq->runtime_enabled)
4901 * clock_task is not advancing so we just need to make sure
4902 * there's some valid quota amount
4904 cfs_rq->runtime_remaining = 1;
4906 * Offline rq is schedulable till CPU is completely disabled
4907 * in take_cpu_down(), so we prevent new cfs throttling here.
4909 cfs_rq->runtime_enabled = 0;
4911 if (cfs_rq_throttled(cfs_rq))
4912 unthrottle_cfs_rq(cfs_rq);
4917 #else /* CONFIG_CFS_BANDWIDTH */
4918 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4920 return rq_clock_task(rq_of(cfs_rq));
4923 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4924 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4925 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4926 static inline void sync_throttle(struct task_group *tg, int cpu) {}
4927 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4929 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4934 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4939 static inline int throttled_lb_pair(struct task_group *tg,
4940 int src_cpu, int dest_cpu)
4945 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4947 #ifdef CONFIG_FAIR_GROUP_SCHED
4948 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4951 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4955 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4956 static inline void update_runtime_enabled(struct rq *rq) {}
4957 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4959 #endif /* CONFIG_CFS_BANDWIDTH */
4961 /**************************************************
4962 * CFS operations on tasks:
4965 #ifdef CONFIG_SCHED_HRTICK
4966 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4968 struct sched_entity *se = &p->se;
4969 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4971 SCHED_WARN_ON(task_rq(p) != rq);
4973 if (rq->cfs.h_nr_running > 1) {
4974 u64 slice = sched_slice(cfs_rq, se);
4975 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4976 s64 delta = slice - ran;
4983 hrtick_start(rq, delta);
4988 * called from enqueue/dequeue and updates the hrtick when the
4989 * current task is from our class and nr_running is low enough
4992 static void hrtick_update(struct rq *rq)
4994 struct task_struct *curr = rq->curr;
4996 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4999 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5000 hrtick_start_fair(rq, curr);
5002 #else /* !CONFIG_SCHED_HRTICK */
5004 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5008 static inline void hrtick_update(struct rq *rq)
5014 * The enqueue_task method is called before nr_running is
5015 * increased. Here we update the fair scheduling stats and
5016 * then put the task into the rbtree:
5019 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5021 struct cfs_rq *cfs_rq;
5022 struct sched_entity *se = &p->se;
5025 * The code below (indirectly) updates schedutil which looks at
5026 * the cfs_rq utilization to select a frequency.
5027 * Let's add the task's estimated utilization to the cfs_rq's
5028 * estimated utilization, before we update schedutil.
5030 util_est_enqueue(&rq->cfs, p);
5033 * If in_iowait is set, the code below may not trigger any cpufreq
5034 * utilization updates, so do it here explicitly with the IOWAIT flag
5038 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5040 for_each_sched_entity(se) {
5043 cfs_rq = cfs_rq_of(se);
5044 enqueue_entity(cfs_rq, se, flags);
5047 * end evaluation on encountering a throttled cfs_rq
5049 * note: in the case of encountering a throttled cfs_rq we will
5050 * post the final h_nr_running increment below.
5052 if (cfs_rq_throttled(cfs_rq))
5054 cfs_rq->h_nr_running++;
5056 flags = ENQUEUE_WAKEUP;
5059 for_each_sched_entity(se) {
5060 cfs_rq = cfs_rq_of(se);
5061 cfs_rq->h_nr_running++;
5063 if (cfs_rq_throttled(cfs_rq))
5066 update_load_avg(cfs_rq, se, UPDATE_TG);
5067 update_cfs_group(se);
5071 add_nr_running(rq, 1);
5076 static void set_next_buddy(struct sched_entity *se);
5079 * The dequeue_task method is called before nr_running is
5080 * decreased. We remove the task from the rbtree and
5081 * update the fair scheduling stats:
5083 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5085 struct cfs_rq *cfs_rq;
5086 struct sched_entity *se = &p->se;
5087 int task_sleep = flags & DEQUEUE_SLEEP;
5089 for_each_sched_entity(se) {
5090 cfs_rq = cfs_rq_of(se);
5091 dequeue_entity(cfs_rq, se, flags);
5094 * end evaluation on encountering a throttled cfs_rq
5096 * note: in the case of encountering a throttled cfs_rq we will
5097 * post the final h_nr_running decrement below.
5099 if (cfs_rq_throttled(cfs_rq))
5101 cfs_rq->h_nr_running--;
5103 /* Don't dequeue parent if it has other entities besides us */
5104 if (cfs_rq->load.weight) {
5105 /* Avoid re-evaluating load for this entity: */
5106 se = parent_entity(se);
5108 * Bias pick_next to pick a task from this cfs_rq, as
5109 * p is sleeping when it is within its sched_slice.
5111 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5115 flags |= DEQUEUE_SLEEP;
5118 for_each_sched_entity(se) {
5119 cfs_rq = cfs_rq_of(se);
5120 cfs_rq->h_nr_running--;
5122 if (cfs_rq_throttled(cfs_rq))
5125 update_load_avg(cfs_rq, se, UPDATE_TG);
5126 update_cfs_group(se);
5130 sub_nr_running(rq, 1);
5132 util_est_dequeue(&rq->cfs, p, task_sleep);
5138 /* Working cpumask for: load_balance, load_balance_newidle. */
5139 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5140 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5142 #ifdef CONFIG_NO_HZ_COMMON
5144 * per rq 'load' arrray crap; XXX kill this.
5148 * The exact cpuload calculated at every tick would be:
5150 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5152 * If a CPU misses updates for n ticks (as it was idle) and update gets
5153 * called on the n+1-th tick when CPU may be busy, then we have:
5155 * load_n = (1 - 1/2^i)^n * load_0
5156 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
5158 * decay_load_missed() below does efficient calculation of
5160 * load' = (1 - 1/2^i)^n * load
5162 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5163 * This allows us to precompute the above in said factors, thereby allowing the
5164 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5165 * fixed_power_int())
5167 * The calculation is approximated on a 128 point scale.
5169 #define DEGRADE_SHIFT 7
5171 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5172 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5173 { 0, 0, 0, 0, 0, 0, 0, 0 },
5174 { 64, 32, 8, 0, 0, 0, 0, 0 },
5175 { 96, 72, 40, 12, 1, 0, 0, 0 },
5176 { 112, 98, 75, 43, 15, 1, 0, 0 },
5177 { 120, 112, 98, 76, 45, 16, 2, 0 }
5181 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5182 * would be when CPU is idle and so we just decay the old load without
5183 * adding any new load.
5185 static unsigned long
5186 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5190 if (!missed_updates)
5193 if (missed_updates >= degrade_zero_ticks[idx])
5197 return load >> missed_updates;
5199 while (missed_updates) {
5200 if (missed_updates % 2)
5201 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5203 missed_updates >>= 1;
5210 cpumask_var_t idle_cpus_mask;
5212 int has_blocked; /* Idle CPUS has blocked load */
5213 unsigned long next_balance; /* in jiffy units */
5214 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5215 } nohz ____cacheline_aligned;
5217 #endif /* CONFIG_NO_HZ_COMMON */
5220 * __cpu_load_update - update the rq->cpu_load[] statistics
5221 * @this_rq: The rq to update statistics for
5222 * @this_load: The current load
5223 * @pending_updates: The number of missed updates
5225 * Update rq->cpu_load[] statistics. This function is usually called every
5226 * scheduler tick (TICK_NSEC).
5228 * This function computes a decaying average:
5230 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5232 * Because of NOHZ it might not get called on every tick which gives need for
5233 * the @pending_updates argument.
5235 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5236 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5237 * = A * (A * load[i]_n-2 + B) + B
5238 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5239 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5240 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5241 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5242 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5244 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5245 * any change in load would have resulted in the tick being turned back on.
5247 * For regular NOHZ, this reduces to:
5249 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5251 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5254 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5255 unsigned long pending_updates)
5257 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5260 this_rq->nr_load_updates++;
5262 /* Update our load: */
5263 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5264 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5265 unsigned long old_load, new_load;
5267 /* scale is effectively 1 << i now, and >> i divides by scale */
5269 old_load = this_rq->cpu_load[i];
5270 #ifdef CONFIG_NO_HZ_COMMON
5271 old_load = decay_load_missed(old_load, pending_updates - 1, i);
5272 if (tickless_load) {
5273 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5275 * old_load can never be a negative value because a
5276 * decayed tickless_load cannot be greater than the
5277 * original tickless_load.
5279 old_load += tickless_load;
5282 new_load = this_load;
5284 * Round up the averaging division if load is increasing. This
5285 * prevents us from getting stuck on 9 if the load is 10, for
5288 if (new_load > old_load)
5289 new_load += scale - 1;
5291 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5295 /* Used instead of source_load when we know the type == 0 */
5296 static unsigned long weighted_cpuload(struct rq *rq)
5298 return cfs_rq_runnable_load_avg(&rq->cfs);
5301 #ifdef CONFIG_NO_HZ_COMMON
5303 * There is no sane way to deal with nohz on smp when using jiffies because the
5304 * CPU doing the jiffies update might drift wrt the CPU doing the jiffy reading
5305 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5307 * Therefore we need to avoid the delta approach from the regular tick when
5308 * possible since that would seriously skew the load calculation. This is why we
5309 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5310 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5311 * loop exit, nohz_idle_balance, nohz full exit...)
5313 * This means we might still be one tick off for nohz periods.
5316 static void cpu_load_update_nohz(struct rq *this_rq,
5317 unsigned long curr_jiffies,
5320 unsigned long pending_updates;
5322 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5323 if (pending_updates) {
5324 this_rq->last_load_update_tick = curr_jiffies;
5326 * In the regular NOHZ case, we were idle, this means load 0.
5327 * In the NOHZ_FULL case, we were non-idle, we should consider
5328 * its weighted load.
5330 cpu_load_update(this_rq, load, pending_updates);
5335 * Called from nohz_idle_balance() to update the load ratings before doing the
5338 static void cpu_load_update_idle(struct rq *this_rq)
5341 * bail if there's load or we're actually up-to-date.
5343 if (weighted_cpuload(this_rq))
5346 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5350 * Record CPU load on nohz entry so we know the tickless load to account
5351 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5352 * than other cpu_load[idx] but it should be fine as cpu_load readers
5353 * shouldn't rely into synchronized cpu_load[*] updates.
5355 void cpu_load_update_nohz_start(void)
5357 struct rq *this_rq = this_rq();
5360 * This is all lockless but should be fine. If weighted_cpuload changes
5361 * concurrently we'll exit nohz. And cpu_load write can race with
5362 * cpu_load_update_idle() but both updater would be writing the same.
5364 this_rq->cpu_load[0] = weighted_cpuload(this_rq);
5368 * Account the tickless load in the end of a nohz frame.
5370 void cpu_load_update_nohz_stop(void)
5372 unsigned long curr_jiffies = READ_ONCE(jiffies);
5373 struct rq *this_rq = this_rq();
5377 if (curr_jiffies == this_rq->last_load_update_tick)
5380 load = weighted_cpuload(this_rq);
5381 rq_lock(this_rq, &rf);
5382 update_rq_clock(this_rq);
5383 cpu_load_update_nohz(this_rq, curr_jiffies, load);
5384 rq_unlock(this_rq, &rf);
5386 #else /* !CONFIG_NO_HZ_COMMON */
5387 static inline void cpu_load_update_nohz(struct rq *this_rq,
5388 unsigned long curr_jiffies,
5389 unsigned long load) { }
5390 #endif /* CONFIG_NO_HZ_COMMON */
5392 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5394 #ifdef CONFIG_NO_HZ_COMMON
5395 /* See the mess around cpu_load_update_nohz(). */
5396 this_rq->last_load_update_tick = READ_ONCE(jiffies);
5398 cpu_load_update(this_rq, load, 1);
5402 * Called from scheduler_tick()
5404 void cpu_load_update_active(struct rq *this_rq)
5406 unsigned long load = weighted_cpuload(this_rq);
5408 if (tick_nohz_tick_stopped())
5409 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5411 cpu_load_update_periodic(this_rq, load);
5415 * Return a low guess at the load of a migration-source CPU weighted
5416 * according to the scheduling class and "nice" value.
5418 * We want to under-estimate the load of migration sources, to
5419 * balance conservatively.
5421 static unsigned long source_load(int cpu, int type)
5423 struct rq *rq = cpu_rq(cpu);
5424 unsigned long total = weighted_cpuload(rq);
5426 if (type == 0 || !sched_feat(LB_BIAS))
5429 return min(rq->cpu_load[type-1], total);
5433 * Return a high guess at the load of a migration-target CPU weighted
5434 * according to the scheduling class and "nice" value.
5436 static unsigned long target_load(int cpu, int type)
5438 struct rq *rq = cpu_rq(cpu);
5439 unsigned long total = weighted_cpuload(rq);
5441 if (type == 0 || !sched_feat(LB_BIAS))
5444 return max(rq->cpu_load[type-1], total);
5447 static unsigned long capacity_of(int cpu)
5449 return cpu_rq(cpu)->cpu_capacity;
5452 static unsigned long capacity_orig_of(int cpu)
5454 return cpu_rq(cpu)->cpu_capacity_orig;
5457 static unsigned long cpu_avg_load_per_task(int cpu)
5459 struct rq *rq = cpu_rq(cpu);
5460 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5461 unsigned long load_avg = weighted_cpuload(rq);
5464 return load_avg / nr_running;
5469 static void record_wakee(struct task_struct *p)
5472 * Only decay a single time; tasks that have less then 1 wakeup per
5473 * jiffy will not have built up many flips.
5475 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5476 current->wakee_flips >>= 1;
5477 current->wakee_flip_decay_ts = jiffies;
5480 if (current->last_wakee != p) {
5481 current->last_wakee = p;
5482 current->wakee_flips++;
5487 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5489 * A waker of many should wake a different task than the one last awakened
5490 * at a frequency roughly N times higher than one of its wakees.
5492 * In order to determine whether we should let the load spread vs consolidating
5493 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5494 * partner, and a factor of lls_size higher frequency in the other.
5496 * With both conditions met, we can be relatively sure that the relationship is
5497 * non-monogamous, with partner count exceeding socket size.
5499 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5500 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5503 static int wake_wide(struct task_struct *p)
5505 unsigned int master = current->wakee_flips;
5506 unsigned int slave = p->wakee_flips;
5507 int factor = this_cpu_read(sd_llc_size);
5510 swap(master, slave);
5511 if (slave < factor || master < slave * factor)
5517 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5518 * soonest. For the purpose of speed we only consider the waking and previous
5521 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5522 * cache-affine and is (or will be) idle.
5524 * wake_affine_weight() - considers the weight to reflect the average
5525 * scheduling latency of the CPUs. This seems to work
5526 * for the overloaded case.
5529 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5532 * If this_cpu is idle, it implies the wakeup is from interrupt
5533 * context. Only allow the move if cache is shared. Otherwise an
5534 * interrupt intensive workload could force all tasks onto one
5535 * node depending on the IO topology or IRQ affinity settings.
5537 * If the prev_cpu is idle and cache affine then avoid a migration.
5538 * There is no guarantee that the cache hot data from an interrupt
5539 * is more important than cache hot data on the prev_cpu and from
5540 * a cpufreq perspective, it's better to have higher utilisation
5543 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5544 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5546 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5549 return nr_cpumask_bits;
5553 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5554 int this_cpu, int prev_cpu, int sync)
5556 s64 this_eff_load, prev_eff_load;
5557 unsigned long task_load;
5559 this_eff_load = target_load(this_cpu, sd->wake_idx);
5562 unsigned long current_load = task_h_load(current);
5564 if (current_load > this_eff_load)
5567 this_eff_load -= current_load;
5570 task_load = task_h_load(p);
5572 this_eff_load += task_load;
5573 if (sched_feat(WA_BIAS))
5574 this_eff_load *= 100;
5575 this_eff_load *= capacity_of(prev_cpu);
5577 prev_eff_load = source_load(prev_cpu, sd->wake_idx);
5578 prev_eff_load -= task_load;
5579 if (sched_feat(WA_BIAS))
5580 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5581 prev_eff_load *= capacity_of(this_cpu);
5584 * If sync, adjust the weight of prev_eff_load such that if
5585 * prev_eff == this_eff that select_idle_sibling() will consider
5586 * stacking the wakee on top of the waker if no other CPU is
5592 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5595 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5596 int this_cpu, int prev_cpu, int sync)
5598 int target = nr_cpumask_bits;
5600 if (sched_feat(WA_IDLE))
5601 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5603 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5604 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5606 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5607 if (target == nr_cpumask_bits)
5610 schedstat_inc(sd->ttwu_move_affine);
5611 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5615 static unsigned long cpu_util_wake(int cpu, struct task_struct *p);
5617 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5619 return max_t(long, capacity_of(cpu) - cpu_util_wake(cpu, p), 0);
5623 * find_idlest_group finds and returns the least busy CPU group within the
5626 * Assumes p is allowed on at least one CPU in sd.
5628 static struct sched_group *
5629 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5630 int this_cpu, int sd_flag)
5632 struct sched_group *idlest = NULL, *group = sd->groups;
5633 struct sched_group *most_spare_sg = NULL;
5634 unsigned long min_runnable_load = ULONG_MAX;
5635 unsigned long this_runnable_load = ULONG_MAX;
5636 unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5637 unsigned long most_spare = 0, this_spare = 0;
5638 int load_idx = sd->forkexec_idx;
5639 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5640 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5641 (sd->imbalance_pct-100) / 100;
5643 if (sd_flag & SD_BALANCE_WAKE)
5644 load_idx = sd->wake_idx;
5647 unsigned long load, avg_load, runnable_load;
5648 unsigned long spare_cap, max_spare_cap;
5652 /* Skip over this group if it has no CPUs allowed */
5653 if (!cpumask_intersects(sched_group_span(group),
5657 local_group = cpumask_test_cpu(this_cpu,
5658 sched_group_span(group));
5661 * Tally up the load of all CPUs in the group and find
5662 * the group containing the CPU with most spare capacity.
5668 for_each_cpu(i, sched_group_span(group)) {
5669 /* Bias balancing toward CPUs of our domain */
5671 load = source_load(i, load_idx);
5673 load = target_load(i, load_idx);
5675 runnable_load += load;
5677 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5679 spare_cap = capacity_spare_wake(i, p);
5681 if (spare_cap > max_spare_cap)
5682 max_spare_cap = spare_cap;
5685 /* Adjust by relative CPU capacity of the group */
5686 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5687 group->sgc->capacity;
5688 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5689 group->sgc->capacity;
5692 this_runnable_load = runnable_load;
5693 this_avg_load = avg_load;
5694 this_spare = max_spare_cap;
5696 if (min_runnable_load > (runnable_load + imbalance)) {
5698 * The runnable load is significantly smaller
5699 * so we can pick this new CPU:
5701 min_runnable_load = runnable_load;
5702 min_avg_load = avg_load;
5704 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5705 (100*min_avg_load > imbalance_scale*avg_load)) {
5707 * The runnable loads are close so take the
5708 * blocked load into account through avg_load:
5710 min_avg_load = avg_load;
5714 if (most_spare < max_spare_cap) {
5715 most_spare = max_spare_cap;
5716 most_spare_sg = group;
5719 } while (group = group->next, group != sd->groups);
5722 * The cross-over point between using spare capacity or least load
5723 * is too conservative for high utilization tasks on partially
5724 * utilized systems if we require spare_capacity > task_util(p),
5725 * so we allow for some task stuffing by using
5726 * spare_capacity > task_util(p)/2.
5728 * Spare capacity can't be used for fork because the utilization has
5729 * not been set yet, we must first select a rq to compute the initial
5732 if (sd_flag & SD_BALANCE_FORK)
5735 if (this_spare > task_util(p) / 2 &&
5736 imbalance_scale*this_spare > 100*most_spare)
5739 if (most_spare > task_util(p) / 2)
5740 return most_spare_sg;
5747 * When comparing groups across NUMA domains, it's possible for the
5748 * local domain to be very lightly loaded relative to the remote
5749 * domains but "imbalance" skews the comparison making remote CPUs
5750 * look much more favourable. When considering cross-domain, add
5751 * imbalance to the runnable load on the remote node and consider
5754 if ((sd->flags & SD_NUMA) &&
5755 min_runnable_load + imbalance >= this_runnable_load)
5758 if (min_runnable_load > (this_runnable_load + imbalance))
5761 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5762 (100*this_avg_load < imbalance_scale*min_avg_load))
5769 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5772 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5774 unsigned long load, min_load = ULONG_MAX;
5775 unsigned int min_exit_latency = UINT_MAX;
5776 u64 latest_idle_timestamp = 0;
5777 int least_loaded_cpu = this_cpu;
5778 int shallowest_idle_cpu = -1;
5781 /* Check if we have any choice: */
5782 if (group->group_weight == 1)
5783 return cpumask_first(sched_group_span(group));
5785 /* Traverse only the allowed CPUs */
5786 for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
5787 if (available_idle_cpu(i)) {
5788 struct rq *rq = cpu_rq(i);
5789 struct cpuidle_state *idle = idle_get_state(rq);
5790 if (idle && idle->exit_latency < min_exit_latency) {
5792 * We give priority to a CPU whose idle state
5793 * has the smallest exit latency irrespective
5794 * of any idle timestamp.
5796 min_exit_latency = idle->exit_latency;
5797 latest_idle_timestamp = rq->idle_stamp;
5798 shallowest_idle_cpu = i;
5799 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5800 rq->idle_stamp > latest_idle_timestamp) {
5802 * If equal or no active idle state, then
5803 * the most recently idled CPU might have
5806 latest_idle_timestamp = rq->idle_stamp;
5807 shallowest_idle_cpu = i;
5809 } else if (shallowest_idle_cpu == -1) {
5810 load = weighted_cpuload(cpu_rq(i));
5811 if (load < min_load) {
5813 least_loaded_cpu = i;
5818 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5821 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5822 int cpu, int prev_cpu, int sd_flag)
5826 if (!cpumask_intersects(sched_domain_span(sd), &p->cpus_allowed))
5830 * We need task's util for capacity_spare_wake, sync it up to prev_cpu's
5833 if (!(sd_flag & SD_BALANCE_FORK))
5834 sync_entity_load_avg(&p->se);
5837 struct sched_group *group;
5838 struct sched_domain *tmp;
5841 if (!(sd->flags & sd_flag)) {
5846 group = find_idlest_group(sd, p, cpu, sd_flag);
5852 new_cpu = find_idlest_group_cpu(group, p, cpu);
5853 if (new_cpu == cpu) {
5854 /* Now try balancing at a lower domain level of 'cpu': */
5859 /* Now try balancing at a lower domain level of 'new_cpu': */
5861 weight = sd->span_weight;
5863 for_each_domain(cpu, tmp) {
5864 if (weight <= tmp->span_weight)
5866 if (tmp->flags & sd_flag)
5874 #ifdef CONFIG_SCHED_SMT
5875 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5877 static inline void set_idle_cores(int cpu, int val)
5879 struct sched_domain_shared *sds;
5881 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5883 WRITE_ONCE(sds->has_idle_cores, val);
5886 static inline bool test_idle_cores(int cpu, bool def)
5888 struct sched_domain_shared *sds;
5890 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5892 return READ_ONCE(sds->has_idle_cores);
5898 * Scans the local SMT mask to see if the entire core is idle, and records this
5899 * information in sd_llc_shared->has_idle_cores.
5901 * Since SMT siblings share all cache levels, inspecting this limited remote
5902 * state should be fairly cheap.
5904 void __update_idle_core(struct rq *rq)
5906 int core = cpu_of(rq);
5910 if (test_idle_cores(core, true))
5913 for_each_cpu(cpu, cpu_smt_mask(core)) {
5917 if (!available_idle_cpu(cpu))
5921 set_idle_cores(core, 1);
5927 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5928 * there are no idle cores left in the system; tracked through
5929 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5931 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5933 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5936 if (!static_branch_likely(&sched_smt_present))
5939 if (!test_idle_cores(target, false))
5942 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
5944 for_each_cpu_wrap(core, cpus, target) {
5947 for_each_cpu(cpu, cpu_smt_mask(core)) {
5948 cpumask_clear_cpu(cpu, cpus);
5949 if (!available_idle_cpu(cpu))
5958 * Failed to find an idle core; stop looking for one.
5960 set_idle_cores(target, 0);
5966 * Scan the local SMT mask for idle CPUs.
5968 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5972 if (!static_branch_likely(&sched_smt_present))
5975 for_each_cpu(cpu, cpu_smt_mask(target)) {
5976 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
5978 if (available_idle_cpu(cpu))
5985 #else /* CONFIG_SCHED_SMT */
5987 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5992 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5997 #endif /* CONFIG_SCHED_SMT */
6000 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6001 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6002 * average idle time for this rq (as found in rq->avg_idle).
6004 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
6006 struct sched_domain *this_sd;
6007 u64 avg_cost, avg_idle;
6010 int cpu, nr = INT_MAX;
6012 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6017 * Due to large variance we need a large fuzz factor; hackbench in
6018 * particularly is sensitive here.
6020 avg_idle = this_rq()->avg_idle / 512;
6021 avg_cost = this_sd->avg_scan_cost + 1;
6023 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
6026 if (sched_feat(SIS_PROP)) {
6027 u64 span_avg = sd->span_weight * avg_idle;
6028 if (span_avg > 4*avg_cost)
6029 nr = div_u64(span_avg, avg_cost);
6034 time = local_clock();
6036 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
6039 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6041 if (available_idle_cpu(cpu))
6045 time = local_clock() - time;
6046 cost = this_sd->avg_scan_cost;
6047 delta = (s64)(time - cost) / 8;
6048 this_sd->avg_scan_cost += delta;
6054 * Try and locate an idle core/thread in the LLC cache domain.
6056 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6058 struct sched_domain *sd;
6059 int i, recent_used_cpu;
6061 if (available_idle_cpu(target))
6065 * If the previous CPU is cache affine and idle, don't be stupid:
6067 if (prev != target && cpus_share_cache(prev, target) && available_idle_cpu(prev))
6070 /* Check a recently used CPU as a potential idle candidate: */
6071 recent_used_cpu = p->recent_used_cpu;
6072 if (recent_used_cpu != prev &&
6073 recent_used_cpu != target &&
6074 cpus_share_cache(recent_used_cpu, target) &&
6075 available_idle_cpu(recent_used_cpu) &&
6076 cpumask_test_cpu(p->recent_used_cpu, &p->cpus_allowed)) {
6078 * Replace recent_used_cpu with prev as it is a potential
6079 * candidate for the next wake:
6081 p->recent_used_cpu = prev;
6082 return recent_used_cpu;
6085 sd = rcu_dereference(per_cpu(sd_llc, target));
6089 i = select_idle_core(p, sd, target);
6090 if ((unsigned)i < nr_cpumask_bits)
6093 i = select_idle_cpu(p, sd, target);
6094 if ((unsigned)i < nr_cpumask_bits)
6097 i = select_idle_smt(p, sd, target);
6098 if ((unsigned)i < nr_cpumask_bits)
6105 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6106 * @cpu: the CPU to get the utilization of
6108 * The unit of the return value must be the one of capacity so we can compare
6109 * the utilization with the capacity of the CPU that is available for CFS task
6110 * (ie cpu_capacity).
6112 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6113 * recent utilization of currently non-runnable tasks on a CPU. It represents
6114 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6115 * capacity_orig is the cpu_capacity available at the highest frequency
6116 * (arch_scale_freq_capacity()).
6117 * The utilization of a CPU converges towards a sum equal to or less than the
6118 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6119 * the running time on this CPU scaled by capacity_curr.
6121 * The estimated utilization of a CPU is defined to be the maximum between its
6122 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6123 * currently RUNNABLE on that CPU.
6124 * This allows to properly represent the expected utilization of a CPU which
6125 * has just got a big task running since a long sleep period. At the same time
6126 * however it preserves the benefits of the "blocked utilization" in
6127 * describing the potential for other tasks waking up on the same CPU.
6129 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6130 * higher than capacity_orig because of unfortunate rounding in
6131 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6132 * the average stabilizes with the new running time. We need to check that the
6133 * utilization stays within the range of [0..capacity_orig] and cap it if
6134 * necessary. Without utilization capping, a group could be seen as overloaded
6135 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6136 * available capacity. We allow utilization to overshoot capacity_curr (but not
6137 * capacity_orig) as it useful for predicting the capacity required after task
6138 * migrations (scheduler-driven DVFS).
6140 * Return: the (estimated) utilization for the specified CPU
6142 static inline unsigned long cpu_util(int cpu)
6144 struct cfs_rq *cfs_rq;
6147 cfs_rq = &cpu_rq(cpu)->cfs;
6148 util = READ_ONCE(cfs_rq->avg.util_avg);
6150 if (sched_feat(UTIL_EST))
6151 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6153 return min_t(unsigned long, util, capacity_orig_of(cpu));
6157 * cpu_util_wake: Compute CPU utilization with any contributions from
6158 * the waking task p removed.
6160 static unsigned long cpu_util_wake(int cpu, struct task_struct *p)
6162 struct cfs_rq *cfs_rq;
6165 /* Task has no contribution or is new */
6166 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6167 return cpu_util(cpu);
6169 cfs_rq = &cpu_rq(cpu)->cfs;
6170 util = READ_ONCE(cfs_rq->avg.util_avg);
6172 /* Discount task's blocked util from CPU's util */
6173 util -= min_t(unsigned int, util, task_util(p));
6178 * a) if *p is the only task sleeping on this CPU, then:
6179 * cpu_util (== task_util) > util_est (== 0)
6180 * and thus we return:
6181 * cpu_util_wake = (cpu_util - task_util) = 0
6183 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6185 * cpu_util >= task_util
6186 * cpu_util > util_est (== 0)
6187 * and thus we discount *p's blocked utilization to return:
6188 * cpu_util_wake = (cpu_util - task_util) >= 0
6190 * c) if other tasks are RUNNABLE on that CPU and
6191 * util_est > cpu_util
6192 * then we use util_est since it returns a more restrictive
6193 * estimation of the spare capacity on that CPU, by just
6194 * considering the expected utilization of tasks already
6195 * runnable on that CPU.
6197 * Cases a) and b) are covered by the above code, while case c) is
6198 * covered by the following code when estimated utilization is
6201 if (sched_feat(UTIL_EST))
6202 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6205 * Utilization (estimated) can exceed the CPU capacity, thus let's
6206 * clamp to the maximum CPU capacity to ensure consistency with
6207 * the cpu_util call.
6209 return min_t(unsigned long, util, capacity_orig_of(cpu));
6213 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6214 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6216 * In that case WAKE_AFFINE doesn't make sense and we'll let
6217 * BALANCE_WAKE sort things out.
6219 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6221 long min_cap, max_cap;
6223 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6224 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6226 /* Minimum capacity is close to max, no need to abort wake_affine */
6227 if (max_cap - min_cap < max_cap >> 3)
6230 /* Bring task utilization in sync with prev_cpu */
6231 sync_entity_load_avg(&p->se);
6233 return min_cap * 1024 < task_util(p) * capacity_margin;
6237 * select_task_rq_fair: Select target runqueue for the waking task in domains
6238 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6239 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6241 * Balances load by selecting the idlest CPU in the idlest group, or under
6242 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6244 * Returns the target CPU number.
6246 * preempt must be disabled.
6249 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6251 struct sched_domain *tmp, *sd = NULL;
6252 int cpu = smp_processor_id();
6253 int new_cpu = prev_cpu;
6254 int want_affine = 0;
6255 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6257 if (sd_flag & SD_BALANCE_WAKE) {
6259 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
6260 && cpumask_test_cpu(cpu, &p->cpus_allowed);
6264 for_each_domain(cpu, tmp) {
6265 if (!(tmp->flags & SD_LOAD_BALANCE))
6269 * If both 'cpu' and 'prev_cpu' are part of this domain,
6270 * cpu is a valid SD_WAKE_AFFINE target.
6272 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6273 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6274 if (cpu != prev_cpu)
6275 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6277 sd = NULL; /* Prefer wake_affine over balance flags */
6281 if (tmp->flags & sd_flag)
6283 else if (!want_affine)
6289 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6290 } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6293 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6296 current->recent_used_cpu = cpu;
6303 static void detach_entity_cfs_rq(struct sched_entity *se);
6306 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6307 * cfs_rq_of(p) references at time of call are still valid and identify the
6308 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6310 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6313 * As blocked tasks retain absolute vruntime the migration needs to
6314 * deal with this by subtracting the old and adding the new
6315 * min_vruntime -- the latter is done by enqueue_entity() when placing
6316 * the task on the new runqueue.
6318 if (p->state == TASK_WAKING) {
6319 struct sched_entity *se = &p->se;
6320 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6323 #ifndef CONFIG_64BIT
6324 u64 min_vruntime_copy;
6327 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6329 min_vruntime = cfs_rq->min_vruntime;
6330 } while (min_vruntime != min_vruntime_copy);
6332 min_vruntime = cfs_rq->min_vruntime;
6335 se->vruntime -= min_vruntime;
6338 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6340 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6341 * rq->lock and can modify state directly.
6343 lockdep_assert_held(&task_rq(p)->lock);
6344 detach_entity_cfs_rq(&p->se);
6348 * We are supposed to update the task to "current" time, then
6349 * its up to date and ready to go to new CPU/cfs_rq. But we
6350 * have difficulty in getting what current time is, so simply
6351 * throw away the out-of-date time. This will result in the
6352 * wakee task is less decayed, but giving the wakee more load
6355 remove_entity_load_avg(&p->se);
6358 /* Tell new CPU we are migrated */
6359 p->se.avg.last_update_time = 0;
6361 /* We have migrated, no longer consider this task hot */
6362 p->se.exec_start = 0;
6364 update_scan_period(p, new_cpu);
6367 static void task_dead_fair(struct task_struct *p)
6369 remove_entity_load_avg(&p->se);
6371 #endif /* CONFIG_SMP */
6373 static unsigned long wakeup_gran(struct sched_entity *se)
6375 unsigned long gran = sysctl_sched_wakeup_granularity;
6378 * Since its curr running now, convert the gran from real-time
6379 * to virtual-time in his units.
6381 * By using 'se' instead of 'curr' we penalize light tasks, so
6382 * they get preempted easier. That is, if 'se' < 'curr' then
6383 * the resulting gran will be larger, therefore penalizing the
6384 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6385 * be smaller, again penalizing the lighter task.
6387 * This is especially important for buddies when the leftmost
6388 * task is higher priority than the buddy.
6390 return calc_delta_fair(gran, se);
6394 * Should 'se' preempt 'curr'.
6408 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6410 s64 gran, vdiff = curr->vruntime - se->vruntime;
6415 gran = wakeup_gran(se);
6422 static void set_last_buddy(struct sched_entity *se)
6424 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6427 for_each_sched_entity(se) {
6428 if (SCHED_WARN_ON(!se->on_rq))
6430 cfs_rq_of(se)->last = se;
6434 static void set_next_buddy(struct sched_entity *se)
6436 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6439 for_each_sched_entity(se) {
6440 if (SCHED_WARN_ON(!se->on_rq))
6442 cfs_rq_of(se)->next = se;
6446 static void set_skip_buddy(struct sched_entity *se)
6448 for_each_sched_entity(se)
6449 cfs_rq_of(se)->skip = se;
6453 * Preempt the current task with a newly woken task if needed:
6455 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6457 struct task_struct *curr = rq->curr;
6458 struct sched_entity *se = &curr->se, *pse = &p->se;
6459 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6460 int scale = cfs_rq->nr_running >= sched_nr_latency;
6461 int next_buddy_marked = 0;
6463 if (unlikely(se == pse))
6467 * This is possible from callers such as attach_tasks(), in which we
6468 * unconditionally check_prempt_curr() after an enqueue (which may have
6469 * lead to a throttle). This both saves work and prevents false
6470 * next-buddy nomination below.
6472 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6475 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6476 set_next_buddy(pse);
6477 next_buddy_marked = 1;
6481 * We can come here with TIF_NEED_RESCHED already set from new task
6484 * Note: this also catches the edge-case of curr being in a throttled
6485 * group (e.g. via set_curr_task), since update_curr() (in the
6486 * enqueue of curr) will have resulted in resched being set. This
6487 * prevents us from potentially nominating it as a false LAST_BUDDY
6490 if (test_tsk_need_resched(curr))
6493 /* Idle tasks are by definition preempted by non-idle tasks. */
6494 if (unlikely(curr->policy == SCHED_IDLE) &&
6495 likely(p->policy != SCHED_IDLE))
6499 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6500 * is driven by the tick):
6502 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6505 find_matching_se(&se, &pse);
6506 update_curr(cfs_rq_of(se));
6508 if (wakeup_preempt_entity(se, pse) == 1) {
6510 * Bias pick_next to pick the sched entity that is
6511 * triggering this preemption.
6513 if (!next_buddy_marked)
6514 set_next_buddy(pse);
6523 * Only set the backward buddy when the current task is still
6524 * on the rq. This can happen when a wakeup gets interleaved
6525 * with schedule on the ->pre_schedule() or idle_balance()
6526 * point, either of which can * drop the rq lock.
6528 * Also, during early boot the idle thread is in the fair class,
6529 * for obvious reasons its a bad idea to schedule back to it.
6531 if (unlikely(!se->on_rq || curr == rq->idle))
6534 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6538 static struct task_struct *
6539 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6541 struct cfs_rq *cfs_rq = &rq->cfs;
6542 struct sched_entity *se;
6543 struct task_struct *p;
6547 if (!cfs_rq->nr_running)
6550 #ifdef CONFIG_FAIR_GROUP_SCHED
6551 if (prev->sched_class != &fair_sched_class)
6555 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6556 * likely that a next task is from the same cgroup as the current.
6558 * Therefore attempt to avoid putting and setting the entire cgroup
6559 * hierarchy, only change the part that actually changes.
6563 struct sched_entity *curr = cfs_rq->curr;
6566 * Since we got here without doing put_prev_entity() we also
6567 * have to consider cfs_rq->curr. If it is still a runnable
6568 * entity, update_curr() will update its vruntime, otherwise
6569 * forget we've ever seen it.
6573 update_curr(cfs_rq);
6578 * This call to check_cfs_rq_runtime() will do the
6579 * throttle and dequeue its entity in the parent(s).
6580 * Therefore the nr_running test will indeed
6583 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6586 if (!cfs_rq->nr_running)
6593 se = pick_next_entity(cfs_rq, curr);
6594 cfs_rq = group_cfs_rq(se);
6600 * Since we haven't yet done put_prev_entity and if the selected task
6601 * is a different task than we started out with, try and touch the
6602 * least amount of cfs_rqs.
6605 struct sched_entity *pse = &prev->se;
6607 while (!(cfs_rq = is_same_group(se, pse))) {
6608 int se_depth = se->depth;
6609 int pse_depth = pse->depth;
6611 if (se_depth <= pse_depth) {
6612 put_prev_entity(cfs_rq_of(pse), pse);
6613 pse = parent_entity(pse);
6615 if (se_depth >= pse_depth) {
6616 set_next_entity(cfs_rq_of(se), se);
6617 se = parent_entity(se);
6621 put_prev_entity(cfs_rq, pse);
6622 set_next_entity(cfs_rq, se);
6629 put_prev_task(rq, prev);
6632 se = pick_next_entity(cfs_rq, NULL);
6633 set_next_entity(cfs_rq, se);
6634 cfs_rq = group_cfs_rq(se);
6639 done: __maybe_unused;
6642 * Move the next running task to the front of
6643 * the list, so our cfs_tasks list becomes MRU
6646 list_move(&p->se.group_node, &rq->cfs_tasks);
6649 if (hrtick_enabled(rq))
6650 hrtick_start_fair(rq, p);
6655 new_tasks = idle_balance(rq, rf);
6658 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6659 * possible for any higher priority task to appear. In that case we
6660 * must re-start the pick_next_entity() loop.
6672 * Account for a descheduled task:
6674 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6676 struct sched_entity *se = &prev->se;
6677 struct cfs_rq *cfs_rq;
6679 for_each_sched_entity(se) {
6680 cfs_rq = cfs_rq_of(se);
6681 put_prev_entity(cfs_rq, se);
6686 * sched_yield() is very simple
6688 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6690 static void yield_task_fair(struct rq *rq)
6692 struct task_struct *curr = rq->curr;
6693 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6694 struct sched_entity *se = &curr->se;
6697 * Are we the only task in the tree?
6699 if (unlikely(rq->nr_running == 1))
6702 clear_buddies(cfs_rq, se);
6704 if (curr->policy != SCHED_BATCH) {
6705 update_rq_clock(rq);
6707 * Update run-time statistics of the 'current'.
6709 update_curr(cfs_rq);
6711 * Tell update_rq_clock() that we've just updated,
6712 * so we don't do microscopic update in schedule()
6713 * and double the fastpath cost.
6715 rq_clock_skip_update(rq);
6721 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6723 struct sched_entity *se = &p->se;
6725 /* throttled hierarchies are not runnable */
6726 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6729 /* Tell the scheduler that we'd really like pse to run next. */
6732 yield_task_fair(rq);
6738 /**************************************************
6739 * Fair scheduling class load-balancing methods.
6743 * The purpose of load-balancing is to achieve the same basic fairness the
6744 * per-CPU scheduler provides, namely provide a proportional amount of compute
6745 * time to each task. This is expressed in the following equation:
6747 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6749 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
6750 * W_i,0 is defined as:
6752 * W_i,0 = \Sum_j w_i,j (2)
6754 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
6755 * is derived from the nice value as per sched_prio_to_weight[].
6757 * The weight average is an exponential decay average of the instantaneous
6760 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6762 * C_i is the compute capacity of CPU i, typically it is the
6763 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6764 * can also include other factors [XXX].
6766 * To achieve this balance we define a measure of imbalance which follows
6767 * directly from (1):
6769 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6771 * We them move tasks around to minimize the imbalance. In the continuous
6772 * function space it is obvious this converges, in the discrete case we get
6773 * a few fun cases generally called infeasible weight scenarios.
6776 * - infeasible weights;
6777 * - local vs global optima in the discrete case. ]
6782 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6783 * for all i,j solution, we create a tree of CPUs that follows the hardware
6784 * topology where each level pairs two lower groups (or better). This results
6785 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
6786 * tree to only the first of the previous level and we decrease the frequency
6787 * of load-balance at each level inv. proportional to the number of CPUs in
6793 * \Sum { --- * --- * 2^i } = O(n) (5)
6795 * `- size of each group
6796 * | | `- number of CPUs doing load-balance
6798 * `- sum over all levels
6800 * Coupled with a limit on how many tasks we can migrate every balance pass,
6801 * this makes (5) the runtime complexity of the balancer.
6803 * An important property here is that each CPU is still (indirectly) connected
6804 * to every other CPU in at most O(log n) steps:
6806 * The adjacency matrix of the resulting graph is given by:
6809 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6812 * And you'll find that:
6814 * A^(log_2 n)_i,j != 0 for all i,j (7)
6816 * Showing there's indeed a path between every CPU in at most O(log n) steps.
6817 * The task movement gives a factor of O(m), giving a convergence complexity
6820 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6825 * In order to avoid CPUs going idle while there's still work to do, new idle
6826 * balancing is more aggressive and has the newly idle CPU iterate up the domain
6827 * tree itself instead of relying on other CPUs to bring it work.
6829 * This adds some complexity to both (5) and (8) but it reduces the total idle
6837 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6840 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6845 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6847 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
6849 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6852 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6853 * rewrite all of this once again.]
6856 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6858 enum fbq_type { regular, remote, all };
6860 #define LBF_ALL_PINNED 0x01
6861 #define LBF_NEED_BREAK 0x02
6862 #define LBF_DST_PINNED 0x04
6863 #define LBF_SOME_PINNED 0x08
6864 #define LBF_NOHZ_STATS 0x10
6865 #define LBF_NOHZ_AGAIN 0x20
6868 struct sched_domain *sd;
6876 struct cpumask *dst_grpmask;
6878 enum cpu_idle_type idle;
6880 /* The set of CPUs under consideration for load-balancing */
6881 struct cpumask *cpus;
6886 unsigned int loop_break;
6887 unsigned int loop_max;
6889 enum fbq_type fbq_type;
6890 struct list_head tasks;
6894 * Is this task likely cache-hot:
6896 static int task_hot(struct task_struct *p, struct lb_env *env)
6900 lockdep_assert_held(&env->src_rq->lock);
6902 if (p->sched_class != &fair_sched_class)
6905 if (unlikely(p->policy == SCHED_IDLE))
6909 * Buddy candidates are cache hot:
6911 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6912 (&p->se == cfs_rq_of(&p->se)->next ||
6913 &p->se == cfs_rq_of(&p->se)->last))
6916 if (sysctl_sched_migration_cost == -1)
6918 if (sysctl_sched_migration_cost == 0)
6921 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6923 return delta < (s64)sysctl_sched_migration_cost;
6926 #ifdef CONFIG_NUMA_BALANCING
6928 * Returns 1, if task migration degrades locality
6929 * Returns 0, if task migration improves locality i.e migration preferred.
6930 * Returns -1, if task migration is not affected by locality.
6932 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6934 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6935 unsigned long src_weight, dst_weight;
6936 int src_nid, dst_nid, dist;
6938 if (!static_branch_likely(&sched_numa_balancing))
6941 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6944 src_nid = cpu_to_node(env->src_cpu);
6945 dst_nid = cpu_to_node(env->dst_cpu);
6947 if (src_nid == dst_nid)
6950 /* Migrating away from the preferred node is always bad. */
6951 if (src_nid == p->numa_preferred_nid) {
6952 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6958 /* Encourage migration to the preferred node. */
6959 if (dst_nid == p->numa_preferred_nid)
6962 /* Leaving a core idle is often worse than degrading locality. */
6963 if (env->idle == CPU_IDLE)
6966 dist = node_distance(src_nid, dst_nid);
6968 src_weight = group_weight(p, src_nid, dist);
6969 dst_weight = group_weight(p, dst_nid, dist);
6971 src_weight = task_weight(p, src_nid, dist);
6972 dst_weight = task_weight(p, dst_nid, dist);
6975 return dst_weight < src_weight;
6979 static inline int migrate_degrades_locality(struct task_struct *p,
6987 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6990 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6994 lockdep_assert_held(&env->src_rq->lock);
6997 * We do not migrate tasks that are:
6998 * 1) throttled_lb_pair, or
6999 * 2) cannot be migrated to this CPU due to cpus_allowed, or
7000 * 3) running (obviously), or
7001 * 4) are cache-hot on their current CPU.
7003 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7006 if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) {
7009 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7011 env->flags |= LBF_SOME_PINNED;
7014 * Remember if this task can be migrated to any other CPU in
7015 * our sched_group. We may want to revisit it if we couldn't
7016 * meet load balance goals by pulling other tasks on src_cpu.
7018 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7019 * already computed one in current iteration.
7021 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7024 /* Prevent to re-select dst_cpu via env's CPUs: */
7025 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7026 if (cpumask_test_cpu(cpu, &p->cpus_allowed)) {
7027 env->flags |= LBF_DST_PINNED;
7028 env->new_dst_cpu = cpu;
7036 /* Record that we found atleast one task that could run on dst_cpu */
7037 env->flags &= ~LBF_ALL_PINNED;
7039 if (task_running(env->src_rq, p)) {
7040 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7045 * Aggressive migration if:
7046 * 1) destination numa is preferred
7047 * 2) task is cache cold, or
7048 * 3) too many balance attempts have failed.
7050 tsk_cache_hot = migrate_degrades_locality(p, env);
7051 if (tsk_cache_hot == -1)
7052 tsk_cache_hot = task_hot(p, env);
7054 if (tsk_cache_hot <= 0 ||
7055 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7056 if (tsk_cache_hot == 1) {
7057 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7058 schedstat_inc(p->se.statistics.nr_forced_migrations);
7063 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7068 * detach_task() -- detach the task for the migration specified in env
7070 static void detach_task(struct task_struct *p, struct lb_env *env)
7072 lockdep_assert_held(&env->src_rq->lock);
7074 p->on_rq = TASK_ON_RQ_MIGRATING;
7075 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7076 set_task_cpu(p, env->dst_cpu);
7080 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7081 * part of active balancing operations within "domain".
7083 * Returns a task if successful and NULL otherwise.
7085 static struct task_struct *detach_one_task(struct lb_env *env)
7087 struct task_struct *p;
7089 lockdep_assert_held(&env->src_rq->lock);
7091 list_for_each_entry_reverse(p,
7092 &env->src_rq->cfs_tasks, se.group_node) {
7093 if (!can_migrate_task(p, env))
7096 detach_task(p, env);
7099 * Right now, this is only the second place where
7100 * lb_gained[env->idle] is updated (other is detach_tasks)
7101 * so we can safely collect stats here rather than
7102 * inside detach_tasks().
7104 schedstat_inc(env->sd->lb_gained[env->idle]);
7110 static const unsigned int sched_nr_migrate_break = 32;
7113 * detach_tasks() -- tries to detach up to imbalance weighted load from
7114 * busiest_rq, as part of a balancing operation within domain "sd".
7116 * Returns number of detached tasks if successful and 0 otherwise.
7118 static int detach_tasks(struct lb_env *env)
7120 struct list_head *tasks = &env->src_rq->cfs_tasks;
7121 struct task_struct *p;
7125 lockdep_assert_held(&env->src_rq->lock);
7127 if (env->imbalance <= 0)
7130 while (!list_empty(tasks)) {
7132 * We don't want to steal all, otherwise we may be treated likewise,
7133 * which could at worst lead to a livelock crash.
7135 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7138 p = list_last_entry(tasks, struct task_struct, se.group_node);
7141 /* We've more or less seen every task there is, call it quits */
7142 if (env->loop > env->loop_max)
7145 /* take a breather every nr_migrate tasks */
7146 if (env->loop > env->loop_break) {
7147 env->loop_break += sched_nr_migrate_break;
7148 env->flags |= LBF_NEED_BREAK;
7152 if (!can_migrate_task(p, env))
7155 load = task_h_load(p);
7157 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7160 if ((load / 2) > env->imbalance)
7163 detach_task(p, env);
7164 list_add(&p->se.group_node, &env->tasks);
7167 env->imbalance -= load;
7169 #ifdef CONFIG_PREEMPT
7171 * NEWIDLE balancing is a source of latency, so preemptible
7172 * kernels will stop after the first task is detached to minimize
7173 * the critical section.
7175 if (env->idle == CPU_NEWLY_IDLE)
7180 * We only want to steal up to the prescribed amount of
7183 if (env->imbalance <= 0)
7188 list_move(&p->se.group_node, tasks);
7192 * Right now, this is one of only two places we collect this stat
7193 * so we can safely collect detach_one_task() stats here rather
7194 * than inside detach_one_task().
7196 schedstat_add(env->sd->lb_gained[env->idle], detached);
7202 * attach_task() -- attach the task detached by detach_task() to its new rq.
7204 static void attach_task(struct rq *rq, struct task_struct *p)
7206 lockdep_assert_held(&rq->lock);
7208 BUG_ON(task_rq(p) != rq);
7209 activate_task(rq, p, ENQUEUE_NOCLOCK);
7210 p->on_rq = TASK_ON_RQ_QUEUED;
7211 check_preempt_curr(rq, p, 0);
7215 * attach_one_task() -- attaches the task returned from detach_one_task() to
7218 static void attach_one_task(struct rq *rq, struct task_struct *p)
7223 update_rq_clock(rq);
7229 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7232 static void attach_tasks(struct lb_env *env)
7234 struct list_head *tasks = &env->tasks;
7235 struct task_struct *p;
7238 rq_lock(env->dst_rq, &rf);
7239 update_rq_clock(env->dst_rq);
7241 while (!list_empty(tasks)) {
7242 p = list_first_entry(tasks, struct task_struct, se.group_node);
7243 list_del_init(&p->se.group_node);
7245 attach_task(env->dst_rq, p);
7248 rq_unlock(env->dst_rq, &rf);
7251 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7253 if (cfs_rq->avg.load_avg)
7256 if (cfs_rq->avg.util_avg)
7262 static inline bool others_have_blocked(struct rq *rq)
7264 if (READ_ONCE(rq->avg_rt.util_avg))
7267 if (READ_ONCE(rq->avg_dl.util_avg))
7270 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
7271 if (READ_ONCE(rq->avg_irq.util_avg))
7278 #ifdef CONFIG_FAIR_GROUP_SCHED
7280 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7282 if (cfs_rq->load.weight)
7285 if (cfs_rq->avg.load_sum)
7288 if (cfs_rq->avg.util_sum)
7291 if (cfs_rq->avg.runnable_load_sum)
7297 static void update_blocked_averages(int cpu)
7299 struct rq *rq = cpu_rq(cpu);
7300 struct cfs_rq *cfs_rq, *pos;
7301 const struct sched_class *curr_class;
7305 rq_lock_irqsave(rq, &rf);
7306 update_rq_clock(rq);
7309 * Iterates the task_group tree in a bottom up fashion, see
7310 * list_add_leaf_cfs_rq() for details.
7312 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7313 struct sched_entity *se;
7315 /* throttled entities do not contribute to load */
7316 if (throttled_hierarchy(cfs_rq))
7319 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
7320 update_tg_load_avg(cfs_rq, 0);
7322 /* Propagate pending load changes to the parent, if any: */
7323 se = cfs_rq->tg->se[cpu];
7324 if (se && !skip_blocked_update(se))
7325 update_load_avg(cfs_rq_of(se), se, 0);
7328 * There can be a lot of idle CPU cgroups. Don't let fully
7329 * decayed cfs_rqs linger on the list.
7331 if (cfs_rq_is_decayed(cfs_rq))
7332 list_del_leaf_cfs_rq(cfs_rq);
7334 /* Don't need periodic decay once load/util_avg are null */
7335 if (cfs_rq_has_blocked(cfs_rq))
7339 curr_class = rq->curr->sched_class;
7340 update_rt_rq_load_avg(rq_clock_task(rq), rq, curr_class == &rt_sched_class);
7341 update_dl_rq_load_avg(rq_clock_task(rq), rq, curr_class == &dl_sched_class);
7342 update_irq_load_avg(rq, 0);
7343 /* Don't need periodic decay once load/util_avg are null */
7344 if (others_have_blocked(rq))
7347 #ifdef CONFIG_NO_HZ_COMMON
7348 rq->last_blocked_load_update_tick = jiffies;
7350 rq->has_blocked_load = 0;
7352 rq_unlock_irqrestore(rq, &rf);
7356 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7357 * This needs to be done in a top-down fashion because the load of a child
7358 * group is a fraction of its parents load.
7360 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7362 struct rq *rq = rq_of(cfs_rq);
7363 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7364 unsigned long now = jiffies;
7367 if (cfs_rq->last_h_load_update == now)
7370 cfs_rq->h_load_next = NULL;
7371 for_each_sched_entity(se) {
7372 cfs_rq = cfs_rq_of(se);
7373 cfs_rq->h_load_next = se;
7374 if (cfs_rq->last_h_load_update == now)
7379 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7380 cfs_rq->last_h_load_update = now;
7383 while ((se = cfs_rq->h_load_next) != NULL) {
7384 load = cfs_rq->h_load;
7385 load = div64_ul(load * se->avg.load_avg,
7386 cfs_rq_load_avg(cfs_rq) + 1);
7387 cfs_rq = group_cfs_rq(se);
7388 cfs_rq->h_load = load;
7389 cfs_rq->last_h_load_update = now;
7393 static unsigned long task_h_load(struct task_struct *p)
7395 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7397 update_cfs_rq_h_load(cfs_rq);
7398 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7399 cfs_rq_load_avg(cfs_rq) + 1);
7402 static inline void update_blocked_averages(int cpu)
7404 struct rq *rq = cpu_rq(cpu);
7405 struct cfs_rq *cfs_rq = &rq->cfs;
7406 const struct sched_class *curr_class;
7409 rq_lock_irqsave(rq, &rf);
7410 update_rq_clock(rq);
7411 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
7413 curr_class = rq->curr->sched_class;
7414 update_rt_rq_load_avg(rq_clock_task(rq), rq, curr_class == &rt_sched_class);
7415 update_dl_rq_load_avg(rq_clock_task(rq), rq, curr_class == &dl_sched_class);
7416 update_irq_load_avg(rq, 0);
7417 #ifdef CONFIG_NO_HZ_COMMON
7418 rq->last_blocked_load_update_tick = jiffies;
7419 if (!cfs_rq_has_blocked(cfs_rq) && !others_have_blocked(rq))
7420 rq->has_blocked_load = 0;
7422 rq_unlock_irqrestore(rq, &rf);
7425 static unsigned long task_h_load(struct task_struct *p)
7427 return p->se.avg.load_avg;
7431 /********** Helpers for find_busiest_group ************************/
7440 * sg_lb_stats - stats of a sched_group required for load_balancing
7442 struct sg_lb_stats {
7443 unsigned long avg_load; /*Avg load across the CPUs of the group */
7444 unsigned long group_load; /* Total load over the CPUs of the group */
7445 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7446 unsigned long load_per_task;
7447 unsigned long group_capacity;
7448 unsigned long group_util; /* Total utilization of the group */
7449 unsigned int sum_nr_running; /* Nr tasks running in the group */
7450 unsigned int idle_cpus;
7451 unsigned int group_weight;
7452 enum group_type group_type;
7453 int group_no_capacity;
7454 #ifdef CONFIG_NUMA_BALANCING
7455 unsigned int nr_numa_running;
7456 unsigned int nr_preferred_running;
7461 * sd_lb_stats - Structure to store the statistics of a sched_domain
7462 * during load balancing.
7464 struct sd_lb_stats {
7465 struct sched_group *busiest; /* Busiest group in this sd */
7466 struct sched_group *local; /* Local group in this sd */
7467 unsigned long total_running;
7468 unsigned long total_load; /* Total load of all groups in sd */
7469 unsigned long total_capacity; /* Total capacity of all groups in sd */
7470 unsigned long avg_load; /* Average load across all groups in sd */
7472 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7473 struct sg_lb_stats local_stat; /* Statistics of the local group */
7476 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7479 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7480 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7481 * We must however clear busiest_stat::avg_load because
7482 * update_sd_pick_busiest() reads this before assignment.
7484 *sds = (struct sd_lb_stats){
7487 .total_running = 0UL,
7489 .total_capacity = 0UL,
7492 .sum_nr_running = 0,
7493 .group_type = group_other,
7499 * get_sd_load_idx - Obtain the load index for a given sched domain.
7500 * @sd: The sched_domain whose load_idx is to be obtained.
7501 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7503 * Return: The load index.
7505 static inline int get_sd_load_idx(struct sched_domain *sd,
7506 enum cpu_idle_type idle)
7512 load_idx = sd->busy_idx;
7515 case CPU_NEWLY_IDLE:
7516 load_idx = sd->newidle_idx;
7519 load_idx = sd->idle_idx;
7526 static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu)
7528 struct rq *rq = cpu_rq(cpu);
7529 unsigned long max = arch_scale_cpu_capacity(sd, cpu);
7530 unsigned long used, free;
7533 irq = cpu_util_irq(rq);
7535 if (unlikely(irq >= max))
7538 used = READ_ONCE(rq->avg_rt.util_avg);
7539 used += READ_ONCE(rq->avg_dl.util_avg);
7541 if (unlikely(used >= max))
7546 return scale_irq_capacity(free, irq, max);
7549 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7551 unsigned long capacity = scale_rt_capacity(sd, cpu);
7552 struct sched_group *sdg = sd->groups;
7554 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(sd, cpu);
7559 cpu_rq(cpu)->cpu_capacity = capacity;
7560 sdg->sgc->capacity = capacity;
7561 sdg->sgc->min_capacity = capacity;
7564 void update_group_capacity(struct sched_domain *sd, int cpu)
7566 struct sched_domain *child = sd->child;
7567 struct sched_group *group, *sdg = sd->groups;
7568 unsigned long capacity, min_capacity;
7569 unsigned long interval;
7571 interval = msecs_to_jiffies(sd->balance_interval);
7572 interval = clamp(interval, 1UL, max_load_balance_interval);
7573 sdg->sgc->next_update = jiffies + interval;
7576 update_cpu_capacity(sd, cpu);
7581 min_capacity = ULONG_MAX;
7583 if (child->flags & SD_OVERLAP) {
7585 * SD_OVERLAP domains cannot assume that child groups
7586 * span the current group.
7589 for_each_cpu(cpu, sched_group_span(sdg)) {
7590 struct sched_group_capacity *sgc;
7591 struct rq *rq = cpu_rq(cpu);
7594 * build_sched_domains() -> init_sched_groups_capacity()
7595 * gets here before we've attached the domains to the
7598 * Use capacity_of(), which is set irrespective of domains
7599 * in update_cpu_capacity().
7601 * This avoids capacity from being 0 and
7602 * causing divide-by-zero issues on boot.
7604 if (unlikely(!rq->sd)) {
7605 capacity += capacity_of(cpu);
7607 sgc = rq->sd->groups->sgc;
7608 capacity += sgc->capacity;
7611 min_capacity = min(capacity, min_capacity);
7615 * !SD_OVERLAP domains can assume that child groups
7616 * span the current group.
7619 group = child->groups;
7621 struct sched_group_capacity *sgc = group->sgc;
7623 capacity += sgc->capacity;
7624 min_capacity = min(sgc->min_capacity, min_capacity);
7625 group = group->next;
7626 } while (group != child->groups);
7629 sdg->sgc->capacity = capacity;
7630 sdg->sgc->min_capacity = min_capacity;
7634 * Check whether the capacity of the rq has been noticeably reduced by side
7635 * activity. The imbalance_pct is used for the threshold.
7636 * Return true is the capacity is reduced
7639 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7641 return ((rq->cpu_capacity * sd->imbalance_pct) <
7642 (rq->cpu_capacity_orig * 100));
7646 * Group imbalance indicates (and tries to solve) the problem where balancing
7647 * groups is inadequate due to ->cpus_allowed constraints.
7649 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
7650 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
7653 * { 0 1 2 3 } { 4 5 6 7 }
7656 * If we were to balance group-wise we'd place two tasks in the first group and
7657 * two tasks in the second group. Clearly this is undesired as it will overload
7658 * cpu 3 and leave one of the CPUs in the second group unused.
7660 * The current solution to this issue is detecting the skew in the first group
7661 * by noticing the lower domain failed to reach balance and had difficulty
7662 * moving tasks due to affinity constraints.
7664 * When this is so detected; this group becomes a candidate for busiest; see
7665 * update_sd_pick_busiest(). And calculate_imbalance() and
7666 * find_busiest_group() avoid some of the usual balance conditions to allow it
7667 * to create an effective group imbalance.
7669 * This is a somewhat tricky proposition since the next run might not find the
7670 * group imbalance and decide the groups need to be balanced again. A most
7671 * subtle and fragile situation.
7674 static inline int sg_imbalanced(struct sched_group *group)
7676 return group->sgc->imbalance;
7680 * group_has_capacity returns true if the group has spare capacity that could
7681 * be used by some tasks.
7682 * We consider that a group has spare capacity if the * number of task is
7683 * smaller than the number of CPUs or if the utilization is lower than the
7684 * available capacity for CFS tasks.
7685 * For the latter, we use a threshold to stabilize the state, to take into
7686 * account the variance of the tasks' load and to return true if the available
7687 * capacity in meaningful for the load balancer.
7688 * As an example, an available capacity of 1% can appear but it doesn't make
7689 * any benefit for the load balance.
7692 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7694 if (sgs->sum_nr_running < sgs->group_weight)
7697 if ((sgs->group_capacity * 100) >
7698 (sgs->group_util * env->sd->imbalance_pct))
7705 * group_is_overloaded returns true if the group has more tasks than it can
7707 * group_is_overloaded is not equals to !group_has_capacity because a group
7708 * with the exact right number of tasks, has no more spare capacity but is not
7709 * overloaded so both group_has_capacity and group_is_overloaded return
7713 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7715 if (sgs->sum_nr_running <= sgs->group_weight)
7718 if ((sgs->group_capacity * 100) <
7719 (sgs->group_util * env->sd->imbalance_pct))
7726 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7727 * per-CPU capacity than sched_group ref.
7730 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7732 return sg->sgc->min_capacity * capacity_margin <
7733 ref->sgc->min_capacity * 1024;
7737 group_type group_classify(struct sched_group *group,
7738 struct sg_lb_stats *sgs)
7740 if (sgs->group_no_capacity)
7741 return group_overloaded;
7743 if (sg_imbalanced(group))
7744 return group_imbalanced;
7749 static bool update_nohz_stats(struct rq *rq, bool force)
7751 #ifdef CONFIG_NO_HZ_COMMON
7752 unsigned int cpu = rq->cpu;
7754 if (!rq->has_blocked_load)
7757 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
7760 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
7763 update_blocked_averages(cpu);
7765 return rq->has_blocked_load;
7772 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7773 * @env: The load balancing environment.
7774 * @group: sched_group whose statistics are to be updated.
7775 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7776 * @local_group: Does group contain this_cpu.
7777 * @sgs: variable to hold the statistics for this group.
7778 * @overload: Indicate more than one runnable task for any CPU.
7780 static inline void update_sg_lb_stats(struct lb_env *env,
7781 struct sched_group *group, int load_idx,
7782 int local_group, struct sg_lb_stats *sgs,
7788 memset(sgs, 0, sizeof(*sgs));
7790 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
7791 struct rq *rq = cpu_rq(i);
7793 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
7794 env->flags |= LBF_NOHZ_AGAIN;
7796 /* Bias balancing toward CPUs of our domain: */
7798 load = target_load(i, load_idx);
7800 load = source_load(i, load_idx);
7802 sgs->group_load += load;
7803 sgs->group_util += cpu_util(i);
7804 sgs->sum_nr_running += rq->cfs.h_nr_running;
7806 nr_running = rq->nr_running;
7810 #ifdef CONFIG_NUMA_BALANCING
7811 sgs->nr_numa_running += rq->nr_numa_running;
7812 sgs->nr_preferred_running += rq->nr_preferred_running;
7814 sgs->sum_weighted_load += weighted_cpuload(rq);
7816 * No need to call idle_cpu() if nr_running is not 0
7818 if (!nr_running && idle_cpu(i))
7822 /* Adjust by relative CPU capacity of the group */
7823 sgs->group_capacity = group->sgc->capacity;
7824 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7826 if (sgs->sum_nr_running)
7827 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7829 sgs->group_weight = group->group_weight;
7831 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7832 sgs->group_type = group_classify(group, sgs);
7836 * update_sd_pick_busiest - return 1 on busiest group
7837 * @env: The load balancing environment.
7838 * @sds: sched_domain statistics
7839 * @sg: sched_group candidate to be checked for being the busiest
7840 * @sgs: sched_group statistics
7842 * Determine if @sg is a busier group than the previously selected
7845 * Return: %true if @sg is a busier group than the previously selected
7846 * busiest group. %false otherwise.
7848 static bool update_sd_pick_busiest(struct lb_env *env,
7849 struct sd_lb_stats *sds,
7850 struct sched_group *sg,
7851 struct sg_lb_stats *sgs)
7853 struct sg_lb_stats *busiest = &sds->busiest_stat;
7855 if (sgs->group_type > busiest->group_type)
7858 if (sgs->group_type < busiest->group_type)
7861 if (sgs->avg_load <= busiest->avg_load)
7864 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7868 * Candidate sg has no more than one task per CPU and
7869 * has higher per-CPU capacity. Migrating tasks to less
7870 * capable CPUs may harm throughput. Maximize throughput,
7871 * power/energy consequences are not considered.
7873 if (sgs->sum_nr_running <= sgs->group_weight &&
7874 group_smaller_cpu_capacity(sds->local, sg))
7878 /* This is the busiest node in its class. */
7879 if (!(env->sd->flags & SD_ASYM_PACKING))
7882 /* No ASYM_PACKING if target CPU is already busy */
7883 if (env->idle == CPU_NOT_IDLE)
7886 * ASYM_PACKING needs to move all the work to the highest
7887 * prority CPUs in the group, therefore mark all groups
7888 * of lower priority than ourself as busy.
7890 if (sgs->sum_nr_running &&
7891 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
7895 /* Prefer to move from lowest priority CPU's work */
7896 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
7897 sg->asym_prefer_cpu))
7904 #ifdef CONFIG_NUMA_BALANCING
7905 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7907 if (sgs->sum_nr_running > sgs->nr_numa_running)
7909 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7914 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7916 if (rq->nr_running > rq->nr_numa_running)
7918 if (rq->nr_running > rq->nr_preferred_running)
7923 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7928 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7932 #endif /* CONFIG_NUMA_BALANCING */
7935 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7936 * @env: The load balancing environment.
7937 * @sds: variable to hold the statistics for this sched_domain.
7939 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7941 struct sched_domain *child = env->sd->child;
7942 struct sched_group *sg = env->sd->groups;
7943 struct sg_lb_stats *local = &sds->local_stat;
7944 struct sg_lb_stats tmp_sgs;
7945 int load_idx, prefer_sibling = 0;
7946 bool overload = false;
7948 if (child && child->flags & SD_PREFER_SIBLING)
7951 #ifdef CONFIG_NO_HZ_COMMON
7952 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
7953 env->flags |= LBF_NOHZ_STATS;
7956 load_idx = get_sd_load_idx(env->sd, env->idle);
7959 struct sg_lb_stats *sgs = &tmp_sgs;
7962 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
7967 if (env->idle != CPU_NEWLY_IDLE ||
7968 time_after_eq(jiffies, sg->sgc->next_update))
7969 update_group_capacity(env->sd, env->dst_cpu);
7972 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7979 * In case the child domain prefers tasks go to siblings
7980 * first, lower the sg capacity so that we'll try
7981 * and move all the excess tasks away. We lower the capacity
7982 * of a group only if the local group has the capacity to fit
7983 * these excess tasks. The extra check prevents the case where
7984 * you always pull from the heaviest group when it is already
7985 * under-utilized (possible with a large weight task outweighs
7986 * the tasks on the system).
7988 if (prefer_sibling && sds->local &&
7989 group_has_capacity(env, local) &&
7990 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
7991 sgs->group_no_capacity = 1;
7992 sgs->group_type = group_classify(sg, sgs);
7995 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7997 sds->busiest_stat = *sgs;
8001 /* Now, start updating sd_lb_stats */
8002 sds->total_running += sgs->sum_nr_running;
8003 sds->total_load += sgs->group_load;
8004 sds->total_capacity += sgs->group_capacity;
8007 } while (sg != env->sd->groups);
8009 #ifdef CONFIG_NO_HZ_COMMON
8010 if ((env->flags & LBF_NOHZ_AGAIN) &&
8011 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8013 WRITE_ONCE(nohz.next_blocked,
8014 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8018 if (env->sd->flags & SD_NUMA)
8019 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8021 if (!env->sd->parent) {
8022 /* update overload indicator if we are at root domain */
8023 if (env->dst_rq->rd->overload != overload)
8024 env->dst_rq->rd->overload = overload;
8029 * check_asym_packing - Check to see if the group is packed into the
8032 * This is primarily intended to used at the sibling level. Some
8033 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8034 * case of POWER7, it can move to lower SMT modes only when higher
8035 * threads are idle. When in lower SMT modes, the threads will
8036 * perform better since they share less core resources. Hence when we
8037 * have idle threads, we want them to be the higher ones.
8039 * This packing function is run on idle threads. It checks to see if
8040 * the busiest CPU in this domain (core in the P7 case) has a higher
8041 * CPU number than the packing function is being run on. Here we are
8042 * assuming lower CPU number will be equivalent to lower a SMT thread
8045 * Return: 1 when packing is required and a task should be moved to
8046 * this CPU. The amount of the imbalance is returned in env->imbalance.
8048 * @env: The load balancing environment.
8049 * @sds: Statistics of the sched_domain which is to be packed
8051 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8055 if (!(env->sd->flags & SD_ASYM_PACKING))
8058 if (env->idle == CPU_NOT_IDLE)
8064 busiest_cpu = sds->busiest->asym_prefer_cpu;
8065 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
8068 env->imbalance = DIV_ROUND_CLOSEST(
8069 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
8070 SCHED_CAPACITY_SCALE);
8076 * fix_small_imbalance - Calculate the minor imbalance that exists
8077 * amongst the groups of a sched_domain, during
8079 * @env: The load balancing environment.
8080 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8083 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8085 unsigned long tmp, capa_now = 0, capa_move = 0;
8086 unsigned int imbn = 2;
8087 unsigned long scaled_busy_load_per_task;
8088 struct sg_lb_stats *local, *busiest;
8090 local = &sds->local_stat;
8091 busiest = &sds->busiest_stat;
8093 if (!local->sum_nr_running)
8094 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8095 else if (busiest->load_per_task > local->load_per_task)
8098 scaled_busy_load_per_task =
8099 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8100 busiest->group_capacity;
8102 if (busiest->avg_load + scaled_busy_load_per_task >=
8103 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8104 env->imbalance = busiest->load_per_task;
8109 * OK, we don't have enough imbalance to justify moving tasks,
8110 * however we may be able to increase total CPU capacity used by
8114 capa_now += busiest->group_capacity *
8115 min(busiest->load_per_task, busiest->avg_load);
8116 capa_now += local->group_capacity *
8117 min(local->load_per_task, local->avg_load);
8118 capa_now /= SCHED_CAPACITY_SCALE;
8120 /* Amount of load we'd subtract */
8121 if (busiest->avg_load > scaled_busy_load_per_task) {
8122 capa_move += busiest->group_capacity *
8123 min(busiest->load_per_task,
8124 busiest->avg_load - scaled_busy_load_per_task);
8127 /* Amount of load we'd add */
8128 if (busiest->avg_load * busiest->group_capacity <
8129 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8130 tmp = (busiest->avg_load * busiest->group_capacity) /
8131 local->group_capacity;
8133 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8134 local->group_capacity;
8136 capa_move += local->group_capacity *
8137 min(local->load_per_task, local->avg_load + tmp);
8138 capa_move /= SCHED_CAPACITY_SCALE;
8140 /* Move if we gain throughput */
8141 if (capa_move > capa_now)
8142 env->imbalance = busiest->load_per_task;
8146 * calculate_imbalance - Calculate the amount of imbalance present within the
8147 * groups of a given sched_domain during load balance.
8148 * @env: load balance environment
8149 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8151 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8153 unsigned long max_pull, load_above_capacity = ~0UL;
8154 struct sg_lb_stats *local, *busiest;
8156 local = &sds->local_stat;
8157 busiest = &sds->busiest_stat;
8159 if (busiest->group_type == group_imbalanced) {
8161 * In the group_imb case we cannot rely on group-wide averages
8162 * to ensure CPU-load equilibrium, look at wider averages. XXX
8164 busiest->load_per_task =
8165 min(busiest->load_per_task, sds->avg_load);
8169 * Avg load of busiest sg can be less and avg load of local sg can
8170 * be greater than avg load across all sgs of sd because avg load
8171 * factors in sg capacity and sgs with smaller group_type are
8172 * skipped when updating the busiest sg:
8174 if (busiest->avg_load <= sds->avg_load ||
8175 local->avg_load >= sds->avg_load) {
8177 return fix_small_imbalance(env, sds);
8181 * If there aren't any idle CPUs, avoid creating some.
8183 if (busiest->group_type == group_overloaded &&
8184 local->group_type == group_overloaded) {
8185 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
8186 if (load_above_capacity > busiest->group_capacity) {
8187 load_above_capacity -= busiest->group_capacity;
8188 load_above_capacity *= scale_load_down(NICE_0_LOAD);
8189 load_above_capacity /= busiest->group_capacity;
8191 load_above_capacity = ~0UL;
8195 * We're trying to get all the CPUs to the average_load, so we don't
8196 * want to push ourselves above the average load, nor do we wish to
8197 * reduce the max loaded CPU below the average load. At the same time,
8198 * we also don't want to reduce the group load below the group
8199 * capacity. Thus we look for the minimum possible imbalance.
8201 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8203 /* How much load to actually move to equalise the imbalance */
8204 env->imbalance = min(
8205 max_pull * busiest->group_capacity,
8206 (sds->avg_load - local->avg_load) * local->group_capacity
8207 ) / SCHED_CAPACITY_SCALE;
8210 * if *imbalance is less than the average load per runnable task
8211 * there is no guarantee that any tasks will be moved so we'll have
8212 * a think about bumping its value to force at least one task to be
8215 if (env->imbalance < busiest->load_per_task)
8216 return fix_small_imbalance(env, sds);
8219 /******* find_busiest_group() helpers end here *********************/
8222 * find_busiest_group - Returns the busiest group within the sched_domain
8223 * if there is an imbalance.
8225 * Also calculates the amount of weighted load which should be moved
8226 * to restore balance.
8228 * @env: The load balancing environment.
8230 * Return: - The busiest group if imbalance exists.
8232 static struct sched_group *find_busiest_group(struct lb_env *env)
8234 struct sg_lb_stats *local, *busiest;
8235 struct sd_lb_stats sds;
8237 init_sd_lb_stats(&sds);
8240 * Compute the various statistics relavent for load balancing at
8243 update_sd_lb_stats(env, &sds);
8244 local = &sds.local_stat;
8245 busiest = &sds.busiest_stat;
8247 /* ASYM feature bypasses nice load balance check */
8248 if (check_asym_packing(env, &sds))
8251 /* There is no busy sibling group to pull tasks from */
8252 if (!sds.busiest || busiest->sum_nr_running == 0)
8255 /* XXX broken for overlapping NUMA groups */
8256 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8257 / sds.total_capacity;
8260 * If the busiest group is imbalanced the below checks don't
8261 * work because they assume all things are equal, which typically
8262 * isn't true due to cpus_allowed constraints and the like.
8264 if (busiest->group_type == group_imbalanced)
8268 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8269 * capacities from resulting in underutilization due to avg_load.
8271 if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8272 busiest->group_no_capacity)
8276 * If the local group is busier than the selected busiest group
8277 * don't try and pull any tasks.
8279 if (local->avg_load >= busiest->avg_load)
8283 * Don't pull any tasks if this group is already above the domain
8286 if (local->avg_load >= sds.avg_load)
8289 if (env->idle == CPU_IDLE) {
8291 * This CPU is idle. If the busiest group is not overloaded
8292 * and there is no imbalance between this and busiest group
8293 * wrt idle CPUs, it is balanced. The imbalance becomes
8294 * significant if the diff is greater than 1 otherwise we
8295 * might end up to just move the imbalance on another group
8297 if ((busiest->group_type != group_overloaded) &&
8298 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8302 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8303 * imbalance_pct to be conservative.
8305 if (100 * busiest->avg_load <=
8306 env->sd->imbalance_pct * local->avg_load)
8311 /* Looks like there is an imbalance. Compute it */
8312 calculate_imbalance(env, &sds);
8313 return env->imbalance ? sds.busiest : NULL;
8321 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
8323 static struct rq *find_busiest_queue(struct lb_env *env,
8324 struct sched_group *group)
8326 struct rq *busiest = NULL, *rq;
8327 unsigned long busiest_load = 0, busiest_capacity = 1;
8330 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8331 unsigned long capacity, wl;
8335 rt = fbq_classify_rq(rq);
8338 * We classify groups/runqueues into three groups:
8339 * - regular: there are !numa tasks
8340 * - remote: there are numa tasks that run on the 'wrong' node
8341 * - all: there is no distinction
8343 * In order to avoid migrating ideally placed numa tasks,
8344 * ignore those when there's better options.
8346 * If we ignore the actual busiest queue to migrate another
8347 * task, the next balance pass can still reduce the busiest
8348 * queue by moving tasks around inside the node.
8350 * If we cannot move enough load due to this classification
8351 * the next pass will adjust the group classification and
8352 * allow migration of more tasks.
8354 * Both cases only affect the total convergence complexity.
8356 if (rt > env->fbq_type)
8359 capacity = capacity_of(i);
8361 wl = weighted_cpuload(rq);
8364 * When comparing with imbalance, use weighted_cpuload()
8365 * which is not scaled with the CPU capacity.
8368 if (rq->nr_running == 1 && wl > env->imbalance &&
8369 !check_cpu_capacity(rq, env->sd))
8373 * For the load comparisons with the other CPU's, consider
8374 * the weighted_cpuload() scaled with the CPU capacity, so
8375 * that the load can be moved away from the CPU that is
8376 * potentially running at a lower capacity.
8378 * Thus we're looking for max(wl_i / capacity_i), crosswise
8379 * multiplication to rid ourselves of the division works out
8380 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8381 * our previous maximum.
8383 if (wl * busiest_capacity > busiest_load * capacity) {
8385 busiest_capacity = capacity;
8394 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8395 * so long as it is large enough.
8397 #define MAX_PINNED_INTERVAL 512
8399 static int need_active_balance(struct lb_env *env)
8401 struct sched_domain *sd = env->sd;
8403 if (env->idle == CPU_NEWLY_IDLE) {
8406 * ASYM_PACKING needs to force migrate tasks from busy but
8407 * lower priority CPUs in order to pack all tasks in the
8408 * highest priority CPUs.
8410 if ((sd->flags & SD_ASYM_PACKING) &&
8411 sched_asym_prefer(env->dst_cpu, env->src_cpu))
8416 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8417 * It's worth migrating the task if the src_cpu's capacity is reduced
8418 * because of other sched_class or IRQs if more capacity stays
8419 * available on dst_cpu.
8421 if ((env->idle != CPU_NOT_IDLE) &&
8422 (env->src_rq->cfs.h_nr_running == 1)) {
8423 if ((check_cpu_capacity(env->src_rq, sd)) &&
8424 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8428 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8431 static int active_load_balance_cpu_stop(void *data);
8433 static int should_we_balance(struct lb_env *env)
8435 struct sched_group *sg = env->sd->groups;
8436 int cpu, balance_cpu = -1;
8439 * Ensure the balancing environment is consistent; can happen
8440 * when the softirq triggers 'during' hotplug.
8442 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8446 * In the newly idle case, we will allow all the CPUs
8447 * to do the newly idle load balance.
8449 if (env->idle == CPU_NEWLY_IDLE)
8452 /* Try to find first idle CPU */
8453 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8461 if (balance_cpu == -1)
8462 balance_cpu = group_balance_cpu(sg);
8465 * First idle CPU or the first CPU(busiest) in this sched group
8466 * is eligible for doing load balancing at this and above domains.
8468 return balance_cpu == env->dst_cpu;
8472 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8473 * tasks if there is an imbalance.
8475 static int load_balance(int this_cpu, struct rq *this_rq,
8476 struct sched_domain *sd, enum cpu_idle_type idle,
8477 int *continue_balancing)
8479 int ld_moved, cur_ld_moved, active_balance = 0;
8480 struct sched_domain *sd_parent = sd->parent;
8481 struct sched_group *group;
8484 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8486 struct lb_env env = {
8488 .dst_cpu = this_cpu,
8490 .dst_grpmask = sched_group_span(sd->groups),
8492 .loop_break = sched_nr_migrate_break,
8495 .tasks = LIST_HEAD_INIT(env.tasks),
8498 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8500 schedstat_inc(sd->lb_count[idle]);
8503 if (!should_we_balance(&env)) {
8504 *continue_balancing = 0;
8508 group = find_busiest_group(&env);
8510 schedstat_inc(sd->lb_nobusyg[idle]);
8514 busiest = find_busiest_queue(&env, group);
8516 schedstat_inc(sd->lb_nobusyq[idle]);
8520 BUG_ON(busiest == env.dst_rq);
8522 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8524 env.src_cpu = busiest->cpu;
8525 env.src_rq = busiest;
8528 if (busiest->nr_running > 1) {
8530 * Attempt to move tasks. If find_busiest_group has found
8531 * an imbalance but busiest->nr_running <= 1, the group is
8532 * still unbalanced. ld_moved simply stays zero, so it is
8533 * correctly treated as an imbalance.
8535 env.flags |= LBF_ALL_PINNED;
8536 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8539 rq_lock_irqsave(busiest, &rf);
8540 update_rq_clock(busiest);
8543 * cur_ld_moved - load moved in current iteration
8544 * ld_moved - cumulative load moved across iterations
8546 cur_ld_moved = detach_tasks(&env);
8549 * We've detached some tasks from busiest_rq. Every
8550 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8551 * unlock busiest->lock, and we are able to be sure
8552 * that nobody can manipulate the tasks in parallel.
8553 * See task_rq_lock() family for the details.
8556 rq_unlock(busiest, &rf);
8560 ld_moved += cur_ld_moved;
8563 local_irq_restore(rf.flags);
8565 if (env.flags & LBF_NEED_BREAK) {
8566 env.flags &= ~LBF_NEED_BREAK;
8571 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8572 * us and move them to an alternate dst_cpu in our sched_group
8573 * where they can run. The upper limit on how many times we
8574 * iterate on same src_cpu is dependent on number of CPUs in our
8577 * This changes load balance semantics a bit on who can move
8578 * load to a given_cpu. In addition to the given_cpu itself
8579 * (or a ilb_cpu acting on its behalf where given_cpu is
8580 * nohz-idle), we now have balance_cpu in a position to move
8581 * load to given_cpu. In rare situations, this may cause
8582 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8583 * _independently_ and at _same_ time to move some load to
8584 * given_cpu) causing exceess load to be moved to given_cpu.
8585 * This however should not happen so much in practice and
8586 * moreover subsequent load balance cycles should correct the
8587 * excess load moved.
8589 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8591 /* Prevent to re-select dst_cpu via env's CPUs */
8592 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8594 env.dst_rq = cpu_rq(env.new_dst_cpu);
8595 env.dst_cpu = env.new_dst_cpu;
8596 env.flags &= ~LBF_DST_PINNED;
8598 env.loop_break = sched_nr_migrate_break;
8601 * Go back to "more_balance" rather than "redo" since we
8602 * need to continue with same src_cpu.
8608 * We failed to reach balance because of affinity.
8611 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8613 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8614 *group_imbalance = 1;
8617 /* All tasks on this runqueue were pinned by CPU affinity */
8618 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8619 cpumask_clear_cpu(cpu_of(busiest), cpus);
8621 * Attempting to continue load balancing at the current
8622 * sched_domain level only makes sense if there are
8623 * active CPUs remaining as possible busiest CPUs to
8624 * pull load from which are not contained within the
8625 * destination group that is receiving any migrated
8628 if (!cpumask_subset(cpus, env.dst_grpmask)) {
8630 env.loop_break = sched_nr_migrate_break;
8633 goto out_all_pinned;
8638 schedstat_inc(sd->lb_failed[idle]);
8640 * Increment the failure counter only on periodic balance.
8641 * We do not want newidle balance, which can be very
8642 * frequent, pollute the failure counter causing
8643 * excessive cache_hot migrations and active balances.
8645 if (idle != CPU_NEWLY_IDLE)
8646 sd->nr_balance_failed++;
8648 if (need_active_balance(&env)) {
8649 unsigned long flags;
8651 raw_spin_lock_irqsave(&busiest->lock, flags);
8654 * Don't kick the active_load_balance_cpu_stop,
8655 * if the curr task on busiest CPU can't be
8656 * moved to this_cpu:
8658 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
8659 raw_spin_unlock_irqrestore(&busiest->lock,
8661 env.flags |= LBF_ALL_PINNED;
8662 goto out_one_pinned;
8666 * ->active_balance synchronizes accesses to
8667 * ->active_balance_work. Once set, it's cleared
8668 * only after active load balance is finished.
8670 if (!busiest->active_balance) {
8671 busiest->active_balance = 1;
8672 busiest->push_cpu = this_cpu;
8675 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8677 if (active_balance) {
8678 stop_one_cpu_nowait(cpu_of(busiest),
8679 active_load_balance_cpu_stop, busiest,
8680 &busiest->active_balance_work);
8683 /* We've kicked active balancing, force task migration. */
8684 sd->nr_balance_failed = sd->cache_nice_tries+1;
8687 sd->nr_balance_failed = 0;
8689 if (likely(!active_balance)) {
8690 /* We were unbalanced, so reset the balancing interval */
8691 sd->balance_interval = sd->min_interval;
8694 * If we've begun active balancing, start to back off. This
8695 * case may not be covered by the all_pinned logic if there
8696 * is only 1 task on the busy runqueue (because we don't call
8699 if (sd->balance_interval < sd->max_interval)
8700 sd->balance_interval *= 2;
8707 * We reach balance although we may have faced some affinity
8708 * constraints. Clear the imbalance flag if it was set.
8711 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8713 if (*group_imbalance)
8714 *group_imbalance = 0;
8719 * We reach balance because all tasks are pinned at this level so
8720 * we can't migrate them. Let the imbalance flag set so parent level
8721 * can try to migrate them.
8723 schedstat_inc(sd->lb_balanced[idle]);
8725 sd->nr_balance_failed = 0;
8728 /* tune up the balancing interval */
8729 if (((env.flags & LBF_ALL_PINNED) &&
8730 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8731 (sd->balance_interval < sd->max_interval))
8732 sd->balance_interval *= 2;
8739 static inline unsigned long
8740 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8742 unsigned long interval = sd->balance_interval;
8745 interval *= sd->busy_factor;
8747 /* scale ms to jiffies */
8748 interval = msecs_to_jiffies(interval);
8749 interval = clamp(interval, 1UL, max_load_balance_interval);
8755 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
8757 unsigned long interval, next;
8759 /* used by idle balance, so cpu_busy = 0 */
8760 interval = get_sd_balance_interval(sd, 0);
8761 next = sd->last_balance + interval;
8763 if (time_after(*next_balance, next))
8764 *next_balance = next;
8768 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
8769 * running tasks off the busiest CPU onto idle CPUs. It requires at
8770 * least 1 task to be running on each physical CPU where possible, and
8771 * avoids physical / logical imbalances.
8773 static int active_load_balance_cpu_stop(void *data)
8775 struct rq *busiest_rq = data;
8776 int busiest_cpu = cpu_of(busiest_rq);
8777 int target_cpu = busiest_rq->push_cpu;
8778 struct rq *target_rq = cpu_rq(target_cpu);
8779 struct sched_domain *sd;
8780 struct task_struct *p = NULL;
8783 rq_lock_irq(busiest_rq, &rf);
8785 * Between queueing the stop-work and running it is a hole in which
8786 * CPUs can become inactive. We should not move tasks from or to
8789 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
8792 /* Make sure the requested CPU hasn't gone down in the meantime: */
8793 if (unlikely(busiest_cpu != smp_processor_id() ||
8794 !busiest_rq->active_balance))
8797 /* Is there any task to move? */
8798 if (busiest_rq->nr_running <= 1)
8802 * This condition is "impossible", if it occurs
8803 * we need to fix it. Originally reported by
8804 * Bjorn Helgaas on a 128-CPU setup.
8806 BUG_ON(busiest_rq == target_rq);
8808 /* Search for an sd spanning us and the target CPU. */
8810 for_each_domain(target_cpu, sd) {
8811 if ((sd->flags & SD_LOAD_BALANCE) &&
8812 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8817 struct lb_env env = {
8819 .dst_cpu = target_cpu,
8820 .dst_rq = target_rq,
8821 .src_cpu = busiest_rq->cpu,
8822 .src_rq = busiest_rq,
8825 * can_migrate_task() doesn't need to compute new_dst_cpu
8826 * for active balancing. Since we have CPU_IDLE, but no
8827 * @dst_grpmask we need to make that test go away with lying
8830 .flags = LBF_DST_PINNED,
8833 schedstat_inc(sd->alb_count);
8834 update_rq_clock(busiest_rq);
8836 p = detach_one_task(&env);
8838 schedstat_inc(sd->alb_pushed);
8839 /* Active balancing done, reset the failure counter. */
8840 sd->nr_balance_failed = 0;
8842 schedstat_inc(sd->alb_failed);
8847 busiest_rq->active_balance = 0;
8848 rq_unlock(busiest_rq, &rf);
8851 attach_one_task(target_rq, p);
8858 static DEFINE_SPINLOCK(balancing);
8861 * Scale the max load_balance interval with the number of CPUs in the system.
8862 * This trades load-balance latency on larger machines for less cross talk.
8864 void update_max_interval(void)
8866 max_load_balance_interval = HZ*num_online_cpus()/10;
8870 * It checks each scheduling domain to see if it is due to be balanced,
8871 * and initiates a balancing operation if so.
8873 * Balancing parameters are set up in init_sched_domains.
8875 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8877 int continue_balancing = 1;
8879 unsigned long interval;
8880 struct sched_domain *sd;
8881 /* Earliest time when we have to do rebalance again */
8882 unsigned long next_balance = jiffies + 60*HZ;
8883 int update_next_balance = 0;
8884 int need_serialize, need_decay = 0;
8888 for_each_domain(cpu, sd) {
8890 * Decay the newidle max times here because this is a regular
8891 * visit to all the domains. Decay ~1% per second.
8893 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8894 sd->max_newidle_lb_cost =
8895 (sd->max_newidle_lb_cost * 253) / 256;
8896 sd->next_decay_max_lb_cost = jiffies + HZ;
8899 max_cost += sd->max_newidle_lb_cost;
8901 if (!(sd->flags & SD_LOAD_BALANCE))
8905 * Stop the load balance at this level. There is another
8906 * CPU in our sched group which is doing load balancing more
8909 if (!continue_balancing) {
8915 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8917 need_serialize = sd->flags & SD_SERIALIZE;
8918 if (need_serialize) {
8919 if (!spin_trylock(&balancing))
8923 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8924 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8926 * The LBF_DST_PINNED logic could have changed
8927 * env->dst_cpu, so we can't know our idle
8928 * state even if we migrated tasks. Update it.
8930 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8932 sd->last_balance = jiffies;
8933 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8936 spin_unlock(&balancing);
8938 if (time_after(next_balance, sd->last_balance + interval)) {
8939 next_balance = sd->last_balance + interval;
8940 update_next_balance = 1;
8945 * Ensure the rq-wide value also decays but keep it at a
8946 * reasonable floor to avoid funnies with rq->avg_idle.
8948 rq->max_idle_balance_cost =
8949 max((u64)sysctl_sched_migration_cost, max_cost);
8954 * next_balance will be updated only when there is a need.
8955 * When the cpu is attached to null domain for ex, it will not be
8958 if (likely(update_next_balance)) {
8959 rq->next_balance = next_balance;
8961 #ifdef CONFIG_NO_HZ_COMMON
8963 * If this CPU has been elected to perform the nohz idle
8964 * balance. Other idle CPUs have already rebalanced with
8965 * nohz_idle_balance() and nohz.next_balance has been
8966 * updated accordingly. This CPU is now running the idle load
8967 * balance for itself and we need to update the
8968 * nohz.next_balance accordingly.
8970 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8971 nohz.next_balance = rq->next_balance;
8976 static inline int on_null_domain(struct rq *rq)
8978 return unlikely(!rcu_dereference_sched(rq->sd));
8981 #ifdef CONFIG_NO_HZ_COMMON
8983 * idle load balancing details
8984 * - When one of the busy CPUs notice that there may be an idle rebalancing
8985 * needed, they will kick the idle load balancer, which then does idle
8986 * load balancing for all the idle CPUs.
8989 static inline int find_new_ilb(void)
8991 int ilb = cpumask_first(nohz.idle_cpus_mask);
8993 if (ilb < nr_cpu_ids && idle_cpu(ilb))
9000 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
9001 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
9002 * CPU (if there is one).
9004 static void kick_ilb(unsigned int flags)
9008 nohz.next_balance++;
9010 ilb_cpu = find_new_ilb();
9012 if (ilb_cpu >= nr_cpu_ids)
9015 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
9016 if (flags & NOHZ_KICK_MASK)
9020 * Use smp_send_reschedule() instead of resched_cpu().
9021 * This way we generate a sched IPI on the target CPU which
9022 * is idle. And the softirq performing nohz idle load balance
9023 * will be run before returning from the IPI.
9025 smp_send_reschedule(ilb_cpu);
9029 * Current heuristic for kicking the idle load balancer in the presence
9030 * of an idle cpu in the system.
9031 * - This rq has more than one task.
9032 * - This rq has at least one CFS task and the capacity of the CPU is
9033 * significantly reduced because of RT tasks or IRQs.
9034 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9035 * multiple busy cpu.
9036 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9037 * domain span are idle.
9039 static void nohz_balancer_kick(struct rq *rq)
9041 unsigned long now = jiffies;
9042 struct sched_domain_shared *sds;
9043 struct sched_domain *sd;
9044 int nr_busy, i, cpu = rq->cpu;
9045 unsigned int flags = 0;
9047 if (unlikely(rq->idle_balance))
9051 * We may be recently in ticked or tickless idle mode. At the first
9052 * busy tick after returning from idle, we will update the busy stats.
9054 nohz_balance_exit_idle(rq);
9057 * None are in tickless mode and hence no need for NOHZ idle load
9060 if (likely(!atomic_read(&nohz.nr_cpus)))
9063 if (READ_ONCE(nohz.has_blocked) &&
9064 time_after(now, READ_ONCE(nohz.next_blocked)))
9065 flags = NOHZ_STATS_KICK;
9067 if (time_before(now, nohz.next_balance))
9070 if (rq->nr_running >= 2) {
9071 flags = NOHZ_KICK_MASK;
9076 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9079 * XXX: write a coherent comment on why we do this.
9080 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
9082 nr_busy = atomic_read(&sds->nr_busy_cpus);
9084 flags = NOHZ_KICK_MASK;
9090 sd = rcu_dereference(rq->sd);
9092 if ((rq->cfs.h_nr_running >= 1) &&
9093 check_cpu_capacity(rq, sd)) {
9094 flags = NOHZ_KICK_MASK;
9099 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9101 for_each_cpu(i, sched_domain_span(sd)) {
9103 !cpumask_test_cpu(i, nohz.idle_cpus_mask))
9106 if (sched_asym_prefer(i, cpu)) {
9107 flags = NOHZ_KICK_MASK;
9119 static void set_cpu_sd_state_busy(int cpu)
9121 struct sched_domain *sd;
9124 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9126 if (!sd || !sd->nohz_idle)
9130 atomic_inc(&sd->shared->nr_busy_cpus);
9135 void nohz_balance_exit_idle(struct rq *rq)
9137 SCHED_WARN_ON(rq != this_rq());
9139 if (likely(!rq->nohz_tick_stopped))
9142 rq->nohz_tick_stopped = 0;
9143 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
9144 atomic_dec(&nohz.nr_cpus);
9146 set_cpu_sd_state_busy(rq->cpu);
9149 static void set_cpu_sd_state_idle(int cpu)
9151 struct sched_domain *sd;
9154 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9156 if (!sd || sd->nohz_idle)
9160 atomic_dec(&sd->shared->nr_busy_cpus);
9166 * This routine will record that the CPU is going idle with tick stopped.
9167 * This info will be used in performing idle load balancing in the future.
9169 void nohz_balance_enter_idle(int cpu)
9171 struct rq *rq = cpu_rq(cpu);
9173 SCHED_WARN_ON(cpu != smp_processor_id());
9175 /* If this CPU is going down, then nothing needs to be done: */
9176 if (!cpu_active(cpu))
9179 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9180 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9184 * Can be set safely without rq->lock held
9185 * If a clear happens, it will have evaluated last additions because
9186 * rq->lock is held during the check and the clear
9188 rq->has_blocked_load = 1;
9191 * The tick is still stopped but load could have been added in the
9192 * meantime. We set the nohz.has_blocked flag to trig a check of the
9193 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
9194 * of nohz.has_blocked can only happen after checking the new load
9196 if (rq->nohz_tick_stopped)
9199 /* If we're a completely isolated CPU, we don't play: */
9200 if (on_null_domain(rq))
9203 rq->nohz_tick_stopped = 1;
9205 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9206 atomic_inc(&nohz.nr_cpus);
9209 * Ensures that if nohz_idle_balance() fails to observe our
9210 * @idle_cpus_mask store, it must observe the @has_blocked
9213 smp_mb__after_atomic();
9215 set_cpu_sd_state_idle(cpu);
9219 * Each time a cpu enter idle, we assume that it has blocked load and
9220 * enable the periodic update of the load of idle cpus
9222 WRITE_ONCE(nohz.has_blocked, 1);
9226 * Internal function that runs load balance for all idle cpus. The load balance
9227 * can be a simple update of blocked load or a complete load balance with
9228 * tasks movement depending of flags.
9229 * The function returns false if the loop has stopped before running
9230 * through all idle CPUs.
9232 static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
9233 enum cpu_idle_type idle)
9235 /* Earliest time when we have to do rebalance again */
9236 unsigned long now = jiffies;
9237 unsigned long next_balance = now + 60*HZ;
9238 bool has_blocked_load = false;
9239 int update_next_balance = 0;
9240 int this_cpu = this_rq->cpu;
9245 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
9248 * We assume there will be no idle load after this update and clear
9249 * the has_blocked flag. If a cpu enters idle in the mean time, it will
9250 * set the has_blocked flag and trig another update of idle load.
9251 * Because a cpu that becomes idle, is added to idle_cpus_mask before
9252 * setting the flag, we are sure to not clear the state and not
9253 * check the load of an idle cpu.
9255 WRITE_ONCE(nohz.has_blocked, 0);
9258 * Ensures that if we miss the CPU, we must see the has_blocked
9259 * store from nohz_balance_enter_idle().
9263 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9264 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9268 * If this CPU gets work to do, stop the load balancing
9269 * work being done for other CPUs. Next load
9270 * balancing owner will pick it up.
9272 if (need_resched()) {
9273 has_blocked_load = true;
9277 rq = cpu_rq(balance_cpu);
9279 has_blocked_load |= update_nohz_stats(rq, true);
9282 * If time for next balance is due,
9285 if (time_after_eq(jiffies, rq->next_balance)) {
9288 rq_lock_irqsave(rq, &rf);
9289 update_rq_clock(rq);
9290 cpu_load_update_idle(rq);
9291 rq_unlock_irqrestore(rq, &rf);
9293 if (flags & NOHZ_BALANCE_KICK)
9294 rebalance_domains(rq, CPU_IDLE);
9297 if (time_after(next_balance, rq->next_balance)) {
9298 next_balance = rq->next_balance;
9299 update_next_balance = 1;
9303 /* Newly idle CPU doesn't need an update */
9304 if (idle != CPU_NEWLY_IDLE) {
9305 update_blocked_averages(this_cpu);
9306 has_blocked_load |= this_rq->has_blocked_load;
9309 if (flags & NOHZ_BALANCE_KICK)
9310 rebalance_domains(this_rq, CPU_IDLE);
9312 WRITE_ONCE(nohz.next_blocked,
9313 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
9315 /* The full idle balance loop has been done */
9319 /* There is still blocked load, enable periodic update */
9320 if (has_blocked_load)
9321 WRITE_ONCE(nohz.has_blocked, 1);
9324 * next_balance will be updated only when there is a need.
9325 * When the CPU is attached to null domain for ex, it will not be
9328 if (likely(update_next_balance))
9329 nohz.next_balance = next_balance;
9335 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9336 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9338 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9340 int this_cpu = this_rq->cpu;
9343 if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
9346 if (idle != CPU_IDLE) {
9347 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9352 * barrier, pairs with nohz_balance_enter_idle(), ensures ...
9354 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9355 if (!(flags & NOHZ_KICK_MASK))
9358 _nohz_idle_balance(this_rq, flags, idle);
9363 static void nohz_newidle_balance(struct rq *this_rq)
9365 int this_cpu = this_rq->cpu;
9368 * This CPU doesn't want to be disturbed by scheduler
9371 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
9374 /* Will wake up very soon. No time for doing anything else*/
9375 if (this_rq->avg_idle < sysctl_sched_migration_cost)
9378 /* Don't need to update blocked load of idle CPUs*/
9379 if (!READ_ONCE(nohz.has_blocked) ||
9380 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
9383 raw_spin_unlock(&this_rq->lock);
9385 * This CPU is going to be idle and blocked load of idle CPUs
9386 * need to be updated. Run the ilb locally as it is a good
9387 * candidate for ilb instead of waking up another idle CPU.
9388 * Kick an normal ilb if we failed to do the update.
9390 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
9391 kick_ilb(NOHZ_STATS_KICK);
9392 raw_spin_lock(&this_rq->lock);
9395 #else /* !CONFIG_NO_HZ_COMMON */
9396 static inline void nohz_balancer_kick(struct rq *rq) { }
9398 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9403 static inline void nohz_newidle_balance(struct rq *this_rq) { }
9404 #endif /* CONFIG_NO_HZ_COMMON */
9407 * idle_balance is called by schedule() if this_cpu is about to become
9408 * idle. Attempts to pull tasks from other CPUs.
9410 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
9412 unsigned long next_balance = jiffies + HZ;
9413 int this_cpu = this_rq->cpu;
9414 struct sched_domain *sd;
9415 int pulled_task = 0;
9419 * We must set idle_stamp _before_ calling idle_balance(), such that we
9420 * measure the duration of idle_balance() as idle time.
9422 this_rq->idle_stamp = rq_clock(this_rq);
9425 * Do not pull tasks towards !active CPUs...
9427 if (!cpu_active(this_cpu))
9431 * This is OK, because current is on_cpu, which avoids it being picked
9432 * for load-balance and preemption/IRQs are still disabled avoiding
9433 * further scheduler activity on it and we're being very careful to
9434 * re-start the picking loop.
9436 rq_unpin_lock(this_rq, rf);
9438 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
9439 !this_rq->rd->overload) {
9442 sd = rcu_dereference_check_sched_domain(this_rq->sd);
9444 update_next_balance(sd, &next_balance);
9447 nohz_newidle_balance(this_rq);
9452 raw_spin_unlock(&this_rq->lock);
9454 update_blocked_averages(this_cpu);
9456 for_each_domain(this_cpu, sd) {
9457 int continue_balancing = 1;
9458 u64 t0, domain_cost;
9460 if (!(sd->flags & SD_LOAD_BALANCE))
9463 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
9464 update_next_balance(sd, &next_balance);
9468 if (sd->flags & SD_BALANCE_NEWIDLE) {
9469 t0 = sched_clock_cpu(this_cpu);
9471 pulled_task = load_balance(this_cpu, this_rq,
9473 &continue_balancing);
9475 domain_cost = sched_clock_cpu(this_cpu) - t0;
9476 if (domain_cost > sd->max_newidle_lb_cost)
9477 sd->max_newidle_lb_cost = domain_cost;
9479 curr_cost += domain_cost;
9482 update_next_balance(sd, &next_balance);
9485 * Stop searching for tasks to pull if there are
9486 * now runnable tasks on this rq.
9488 if (pulled_task || this_rq->nr_running > 0)
9493 raw_spin_lock(&this_rq->lock);
9495 if (curr_cost > this_rq->max_idle_balance_cost)
9496 this_rq->max_idle_balance_cost = curr_cost;
9500 * While browsing the domains, we released the rq lock, a task could
9501 * have been enqueued in the meantime. Since we're not going idle,
9502 * pretend we pulled a task.
9504 if (this_rq->cfs.h_nr_running && !pulled_task)
9507 /* Move the next balance forward */
9508 if (time_after(this_rq->next_balance, next_balance))
9509 this_rq->next_balance = next_balance;
9511 /* Is there a task of a high priority class? */
9512 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
9516 this_rq->idle_stamp = 0;
9518 rq_repin_lock(this_rq, rf);
9524 * run_rebalance_domains is triggered when needed from the scheduler tick.
9525 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9527 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9529 struct rq *this_rq = this_rq();
9530 enum cpu_idle_type idle = this_rq->idle_balance ?
9531 CPU_IDLE : CPU_NOT_IDLE;
9534 * If this CPU has a pending nohz_balance_kick, then do the
9535 * balancing on behalf of the other idle CPUs whose ticks are
9536 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9537 * give the idle CPUs a chance to load balance. Else we may
9538 * load balance only within the local sched_domain hierarchy
9539 * and abort nohz_idle_balance altogether if we pull some load.
9541 if (nohz_idle_balance(this_rq, idle))
9544 /* normal load balance */
9545 update_blocked_averages(this_rq->cpu);
9546 rebalance_domains(this_rq, idle);
9550 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9552 void trigger_load_balance(struct rq *rq)
9554 /* Don't need to rebalance while attached to NULL domain */
9555 if (unlikely(on_null_domain(rq)))
9558 if (time_after_eq(jiffies, rq->next_balance))
9559 raise_softirq(SCHED_SOFTIRQ);
9561 nohz_balancer_kick(rq);
9564 static void rq_online_fair(struct rq *rq)
9568 update_runtime_enabled(rq);
9571 static void rq_offline_fair(struct rq *rq)
9575 /* Ensure any throttled groups are reachable by pick_next_task */
9576 unthrottle_offline_cfs_rqs(rq);
9579 #endif /* CONFIG_SMP */
9582 * scheduler tick hitting a task of our scheduling class.
9584 * NOTE: This function can be called remotely by the tick offload that
9585 * goes along full dynticks. Therefore no local assumption can be made
9586 * and everything must be accessed through the @rq and @curr passed in
9589 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9591 struct cfs_rq *cfs_rq;
9592 struct sched_entity *se = &curr->se;
9594 for_each_sched_entity(se) {
9595 cfs_rq = cfs_rq_of(se);
9596 entity_tick(cfs_rq, se, queued);
9599 if (static_branch_unlikely(&sched_numa_balancing))
9600 task_tick_numa(rq, curr);
9604 * called on fork with the child task as argument from the parent's context
9605 * - child not yet on the tasklist
9606 * - preemption disabled
9608 static void task_fork_fair(struct task_struct *p)
9610 struct cfs_rq *cfs_rq;
9611 struct sched_entity *se = &p->se, *curr;
9612 struct rq *rq = this_rq();
9616 update_rq_clock(rq);
9618 cfs_rq = task_cfs_rq(current);
9619 curr = cfs_rq->curr;
9621 update_curr(cfs_rq);
9622 se->vruntime = curr->vruntime;
9624 place_entity(cfs_rq, se, 1);
9626 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9628 * Upon rescheduling, sched_class::put_prev_task() will place
9629 * 'current' within the tree based on its new key value.
9631 swap(curr->vruntime, se->vruntime);
9635 se->vruntime -= cfs_rq->min_vruntime;
9640 * Priority of the task has changed. Check to see if we preempt
9644 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9646 if (!task_on_rq_queued(p))
9650 * Reschedule if we are currently running on this runqueue and
9651 * our priority decreased, or if we are not currently running on
9652 * this runqueue and our priority is higher than the current's
9654 if (rq->curr == p) {
9655 if (p->prio > oldprio)
9658 check_preempt_curr(rq, p, 0);
9661 static inline bool vruntime_normalized(struct task_struct *p)
9663 struct sched_entity *se = &p->se;
9666 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9667 * the dequeue_entity(.flags=0) will already have normalized the
9674 * When !on_rq, vruntime of the task has usually NOT been normalized.
9675 * But there are some cases where it has already been normalized:
9677 * - A forked child which is waiting for being woken up by
9678 * wake_up_new_task().
9679 * - A task which has been woken up by try_to_wake_up() and
9680 * waiting for actually being woken up by sched_ttwu_pending().
9682 if (!se->sum_exec_runtime ||
9683 (p->state == TASK_WAKING && p->sched_remote_wakeup))
9689 #ifdef CONFIG_FAIR_GROUP_SCHED
9691 * Propagate the changes of the sched_entity across the tg tree to make it
9692 * visible to the root
9694 static void propagate_entity_cfs_rq(struct sched_entity *se)
9696 struct cfs_rq *cfs_rq;
9698 /* Start to propagate at parent */
9701 for_each_sched_entity(se) {
9702 cfs_rq = cfs_rq_of(se);
9704 if (cfs_rq_throttled(cfs_rq))
9707 update_load_avg(cfs_rq, se, UPDATE_TG);
9711 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
9714 static void detach_entity_cfs_rq(struct sched_entity *se)
9716 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9718 /* Catch up with the cfs_rq and remove our load when we leave */
9719 update_load_avg(cfs_rq, se, 0);
9720 detach_entity_load_avg(cfs_rq, se);
9721 update_tg_load_avg(cfs_rq, false);
9722 propagate_entity_cfs_rq(se);
9725 static void attach_entity_cfs_rq(struct sched_entity *se)
9727 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9729 #ifdef CONFIG_FAIR_GROUP_SCHED
9731 * Since the real-depth could have been changed (only FAIR
9732 * class maintain depth value), reset depth properly.
9734 se->depth = se->parent ? se->parent->depth + 1 : 0;
9737 /* Synchronize entity with its cfs_rq */
9738 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
9739 attach_entity_load_avg(cfs_rq, se, 0);
9740 update_tg_load_avg(cfs_rq, false);
9741 propagate_entity_cfs_rq(se);
9744 static void detach_task_cfs_rq(struct task_struct *p)
9746 struct sched_entity *se = &p->se;
9747 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9749 if (!vruntime_normalized(p)) {
9751 * Fix up our vruntime so that the current sleep doesn't
9752 * cause 'unlimited' sleep bonus.
9754 place_entity(cfs_rq, se, 0);
9755 se->vruntime -= cfs_rq->min_vruntime;
9758 detach_entity_cfs_rq(se);
9761 static void attach_task_cfs_rq(struct task_struct *p)
9763 struct sched_entity *se = &p->se;
9764 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9766 attach_entity_cfs_rq(se);
9768 if (!vruntime_normalized(p))
9769 se->vruntime += cfs_rq->min_vruntime;
9772 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9774 detach_task_cfs_rq(p);
9777 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9779 attach_task_cfs_rq(p);
9781 if (task_on_rq_queued(p)) {
9783 * We were most likely switched from sched_rt, so
9784 * kick off the schedule if running, otherwise just see
9785 * if we can still preempt the current task.
9790 check_preempt_curr(rq, p, 0);
9794 /* Account for a task changing its policy or group.
9796 * This routine is mostly called to set cfs_rq->curr field when a task
9797 * migrates between groups/classes.
9799 static void set_curr_task_fair(struct rq *rq)
9801 struct sched_entity *se = &rq->curr->se;
9803 for_each_sched_entity(se) {
9804 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9806 set_next_entity(cfs_rq, se);
9807 /* ensure bandwidth has been allocated on our new cfs_rq */
9808 account_cfs_rq_runtime(cfs_rq, 0);
9812 void init_cfs_rq(struct cfs_rq *cfs_rq)
9814 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
9815 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9816 #ifndef CONFIG_64BIT
9817 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9820 raw_spin_lock_init(&cfs_rq->removed.lock);
9824 #ifdef CONFIG_FAIR_GROUP_SCHED
9825 static void task_set_group_fair(struct task_struct *p)
9827 struct sched_entity *se = &p->se;
9829 set_task_rq(p, task_cpu(p));
9830 se->depth = se->parent ? se->parent->depth + 1 : 0;
9833 static void task_move_group_fair(struct task_struct *p)
9835 detach_task_cfs_rq(p);
9836 set_task_rq(p, task_cpu(p));
9839 /* Tell se's cfs_rq has been changed -- migrated */
9840 p->se.avg.last_update_time = 0;
9842 attach_task_cfs_rq(p);
9845 static void task_change_group_fair(struct task_struct *p, int type)
9848 case TASK_SET_GROUP:
9849 task_set_group_fair(p);
9852 case TASK_MOVE_GROUP:
9853 task_move_group_fair(p);
9858 void free_fair_sched_group(struct task_group *tg)
9862 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9864 for_each_possible_cpu(i) {
9866 kfree(tg->cfs_rq[i]);
9875 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9877 struct sched_entity *se;
9878 struct cfs_rq *cfs_rq;
9881 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
9884 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
9888 tg->shares = NICE_0_LOAD;
9890 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9892 for_each_possible_cpu(i) {
9893 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9894 GFP_KERNEL, cpu_to_node(i));
9898 se = kzalloc_node(sizeof(struct sched_entity),
9899 GFP_KERNEL, cpu_to_node(i));
9903 init_cfs_rq(cfs_rq);
9904 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9905 init_entity_runnable_average(se);
9916 void online_fair_sched_group(struct task_group *tg)
9918 struct sched_entity *se;
9922 for_each_possible_cpu(i) {
9926 raw_spin_lock_irq(&rq->lock);
9927 update_rq_clock(rq);
9928 attach_entity_cfs_rq(se);
9929 sync_throttle(tg, i);
9930 raw_spin_unlock_irq(&rq->lock);
9934 void unregister_fair_sched_group(struct task_group *tg)
9936 unsigned long flags;
9940 for_each_possible_cpu(cpu) {
9942 remove_entity_load_avg(tg->se[cpu]);
9945 * Only empty task groups can be destroyed; so we can speculatively
9946 * check on_list without danger of it being re-added.
9948 if (!tg->cfs_rq[cpu]->on_list)
9953 raw_spin_lock_irqsave(&rq->lock, flags);
9954 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9955 raw_spin_unlock_irqrestore(&rq->lock, flags);
9959 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9960 struct sched_entity *se, int cpu,
9961 struct sched_entity *parent)
9963 struct rq *rq = cpu_rq(cpu);
9967 init_cfs_rq_runtime(cfs_rq);
9969 tg->cfs_rq[cpu] = cfs_rq;
9972 /* se could be NULL for root_task_group */
9977 se->cfs_rq = &rq->cfs;
9980 se->cfs_rq = parent->my_q;
9981 se->depth = parent->depth + 1;
9985 /* guarantee group entities always have weight */
9986 update_load_set(&se->load, NICE_0_LOAD);
9987 se->parent = parent;
9990 static DEFINE_MUTEX(shares_mutex);
9992 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9997 * We can't change the weight of the root cgroup.
10002 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10004 mutex_lock(&shares_mutex);
10005 if (tg->shares == shares)
10008 tg->shares = shares;
10009 for_each_possible_cpu(i) {
10010 struct rq *rq = cpu_rq(i);
10011 struct sched_entity *se = tg->se[i];
10012 struct rq_flags rf;
10014 /* Propagate contribution to hierarchy */
10015 rq_lock_irqsave(rq, &rf);
10016 update_rq_clock(rq);
10017 for_each_sched_entity(se) {
10018 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
10019 update_cfs_group(se);
10021 rq_unlock_irqrestore(rq, &rf);
10025 mutex_unlock(&shares_mutex);
10028 #else /* CONFIG_FAIR_GROUP_SCHED */
10030 void free_fair_sched_group(struct task_group *tg) { }
10032 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10037 void online_fair_sched_group(struct task_group *tg) { }
10039 void unregister_fair_sched_group(struct task_group *tg) { }
10041 #endif /* CONFIG_FAIR_GROUP_SCHED */
10044 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10046 struct sched_entity *se = &task->se;
10047 unsigned int rr_interval = 0;
10050 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10053 if (rq->cfs.load.weight)
10054 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10056 return rr_interval;
10060 * All the scheduling class methods:
10062 const struct sched_class fair_sched_class = {
10063 .next = &idle_sched_class,
10064 .enqueue_task = enqueue_task_fair,
10065 .dequeue_task = dequeue_task_fair,
10066 .yield_task = yield_task_fair,
10067 .yield_to_task = yield_to_task_fair,
10069 .check_preempt_curr = check_preempt_wakeup,
10071 .pick_next_task = pick_next_task_fair,
10072 .put_prev_task = put_prev_task_fair,
10075 .select_task_rq = select_task_rq_fair,
10076 .migrate_task_rq = migrate_task_rq_fair,
10078 .rq_online = rq_online_fair,
10079 .rq_offline = rq_offline_fair,
10081 .task_dead = task_dead_fair,
10082 .set_cpus_allowed = set_cpus_allowed_common,
10085 .set_curr_task = set_curr_task_fair,
10086 .task_tick = task_tick_fair,
10087 .task_fork = task_fork_fair,
10089 .prio_changed = prio_changed_fair,
10090 .switched_from = switched_from_fair,
10091 .switched_to = switched_to_fair,
10093 .get_rr_interval = get_rr_interval_fair,
10095 .update_curr = update_curr_fair,
10097 #ifdef CONFIG_FAIR_GROUP_SCHED
10098 .task_change_group = task_change_group_fair,
10102 #ifdef CONFIG_SCHED_DEBUG
10103 void print_cfs_stats(struct seq_file *m, int cpu)
10105 struct cfs_rq *cfs_rq, *pos;
10108 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
10109 print_cfs_rq(m, cpu, cfs_rq);
10113 #ifdef CONFIG_NUMA_BALANCING
10114 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10117 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10119 for_each_online_node(node) {
10120 if (p->numa_faults) {
10121 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10122 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10124 if (p->numa_group) {
10125 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
10126 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
10128 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10131 #endif /* CONFIG_NUMA_BALANCING */
10132 #endif /* CONFIG_SCHED_DEBUG */
10134 __init void init_sched_fair_class(void)
10137 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10139 #ifdef CONFIG_NO_HZ_COMMON
10140 nohz.next_balance = jiffies;
10141 nohz.next_blocked = jiffies;
10142 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);