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 */
1453 unsigned long nr_running;
1456 /* Total compute capacity of CPUs on a node */
1457 unsigned long compute_capacity;
1459 /* Approximate capacity in terms of runnable tasks on a node */
1460 unsigned long task_capacity;
1461 int has_free_capacity;
1465 * XXX borrowed from update_sg_lb_stats
1467 static void update_numa_stats(struct numa_stats *ns, int nid)
1469 int smt, cpu, cpus = 0;
1470 unsigned long capacity;
1472 memset(ns, 0, sizeof(*ns));
1473 for_each_cpu(cpu, cpumask_of_node(nid)) {
1474 struct rq *rq = cpu_rq(cpu);
1476 ns->nr_running += rq->nr_running;
1477 ns->load += weighted_cpuload(rq);
1478 ns->compute_capacity += capacity_of(cpu);
1484 * If we raced with hotplug and there are no CPUs left in our mask
1485 * the @ns structure is NULL'ed and task_numa_compare() will
1486 * not find this node attractive.
1488 * We'll either bail at !has_free_capacity, or we'll detect a huge
1489 * imbalance and bail there.
1494 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1495 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1496 capacity = cpus / smt; /* cores */
1498 ns->task_capacity = min_t(unsigned, capacity,
1499 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1500 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1503 struct task_numa_env {
1504 struct task_struct *p;
1506 int src_cpu, src_nid;
1507 int dst_cpu, dst_nid;
1509 struct numa_stats src_stats, dst_stats;
1514 struct task_struct *best_task;
1519 static void task_numa_assign(struct task_numa_env *env,
1520 struct task_struct *p, long imp)
1523 put_task_struct(env->best_task);
1528 env->best_imp = imp;
1529 env->best_cpu = env->dst_cpu;
1532 static bool load_too_imbalanced(long src_load, long dst_load,
1533 struct task_numa_env *env)
1536 long orig_src_load, orig_dst_load;
1537 long src_capacity, dst_capacity;
1540 * The load is corrected for the CPU capacity available on each node.
1543 * ------------ vs ---------
1544 * src_capacity dst_capacity
1546 src_capacity = env->src_stats.compute_capacity;
1547 dst_capacity = env->dst_stats.compute_capacity;
1549 /* We care about the slope of the imbalance, not the direction. */
1550 if (dst_load < src_load)
1551 swap(dst_load, src_load);
1553 /* Is the difference below the threshold? */
1554 imb = dst_load * src_capacity * 100 -
1555 src_load * dst_capacity * env->imbalance_pct;
1560 * The imbalance is above the allowed threshold.
1561 * Compare it with the old imbalance.
1563 orig_src_load = env->src_stats.load;
1564 orig_dst_load = env->dst_stats.load;
1566 if (orig_dst_load < orig_src_load)
1567 swap(orig_dst_load, orig_src_load);
1569 old_imb = orig_dst_load * src_capacity * 100 -
1570 orig_src_load * dst_capacity * env->imbalance_pct;
1572 /* Would this change make things worse? */
1573 return (imb > old_imb);
1577 * This checks if the overall compute and NUMA accesses of the system would
1578 * be improved if the source tasks was migrated to the target dst_cpu taking
1579 * into account that it might be best if task running on the dst_cpu should
1580 * be exchanged with the source task
1582 static void task_numa_compare(struct task_numa_env *env,
1583 long taskimp, long groupimp)
1585 struct rq *src_rq = cpu_rq(env->src_cpu);
1586 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1587 struct task_struct *cur;
1588 long src_load, dst_load;
1590 long imp = env->p->numa_group ? groupimp : taskimp;
1592 int dist = env->dist;
1595 cur = task_rcu_dereference(&dst_rq->curr);
1596 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1600 * Because we have preemption enabled we can get migrated around and
1601 * end try selecting ourselves (current == env->p) as a swap candidate.
1607 * "imp" is the fault differential for the source task between the
1608 * source and destination node. Calculate the total differential for
1609 * the source task and potential destination task. The more negative
1610 * the value is, the more rmeote accesses that would be expected to
1611 * be incurred if the tasks were swapped.
1614 /* Skip this swap candidate if cannot move to the source CPU: */
1615 if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed))
1619 * If dst and source tasks are in the same NUMA group, or not
1620 * in any group then look only at task weights.
1622 if (cur->numa_group == env->p->numa_group) {
1623 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1624 task_weight(cur, env->dst_nid, dist);
1626 * Add some hysteresis to prevent swapping the
1627 * tasks within a group over tiny differences.
1629 if (cur->numa_group)
1633 * Compare the group weights. If a task is all by
1634 * itself (not part of a group), use the task weight
1637 if (cur->numa_group)
1638 imp += group_weight(cur, env->src_nid, dist) -
1639 group_weight(cur, env->dst_nid, dist);
1641 imp += task_weight(cur, env->src_nid, dist) -
1642 task_weight(cur, env->dst_nid, dist);
1646 if (imp <= env->best_imp && moveimp <= env->best_imp)
1650 /* Is there capacity at our destination? */
1651 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1652 !env->dst_stats.has_free_capacity)
1658 /* Balance doesn't matter much if we're running a task per CPU: */
1659 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1660 dst_rq->nr_running == 1)
1664 * In the overloaded case, try and keep the load balanced.
1667 load = task_h_load(env->p);
1668 dst_load = env->dst_stats.load + load;
1669 src_load = env->src_stats.load - load;
1671 if (moveimp > imp && moveimp > env->best_imp) {
1673 * If the improvement from just moving env->p direction is
1674 * better than swapping tasks around, check if a move is
1675 * possible. Store a slightly smaller score than moveimp,
1676 * so an actually idle CPU will win.
1678 if (!load_too_imbalanced(src_load, dst_load, env)) {
1685 if (imp <= env->best_imp)
1689 load = task_h_load(cur);
1694 if (load_too_imbalanced(src_load, dst_load, env))
1698 * One idle CPU per node is evaluated for a task numa move.
1699 * Call select_idle_sibling to maybe find a better one.
1703 * select_idle_siblings() uses an per-CPU cpumask that
1704 * can be used from IRQ context.
1706 local_irq_disable();
1707 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1713 task_numa_assign(env, cur, imp);
1718 static void task_numa_find_cpu(struct task_numa_env *env,
1719 long taskimp, long groupimp)
1723 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1724 /* Skip this CPU if the source task cannot migrate */
1725 if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
1729 task_numa_compare(env, taskimp, groupimp);
1733 /* Only move tasks to a NUMA node less busy than the current node. */
1734 static bool numa_has_capacity(struct task_numa_env *env)
1736 struct numa_stats *src = &env->src_stats;
1737 struct numa_stats *dst = &env->dst_stats;
1739 if (src->has_free_capacity && !dst->has_free_capacity)
1743 * Only consider a task move if the source has a higher load
1744 * than the destination, corrected for CPU capacity on each node.
1746 * src->load dst->load
1747 * --------------------- vs ---------------------
1748 * src->compute_capacity dst->compute_capacity
1750 if (src->load * dst->compute_capacity * env->imbalance_pct >
1752 dst->load * src->compute_capacity * 100)
1758 static int task_numa_migrate(struct task_struct *p)
1760 struct task_numa_env env = {
1763 .src_cpu = task_cpu(p),
1764 .src_nid = task_node(p),
1766 .imbalance_pct = 112,
1772 struct sched_domain *sd;
1773 unsigned long taskweight, groupweight;
1775 long taskimp, groupimp;
1778 * Pick the lowest SD_NUMA domain, as that would have the smallest
1779 * imbalance and would be the first to start moving tasks about.
1781 * And we want to avoid any moving of tasks about, as that would create
1782 * random movement of tasks -- counter the numa conditions we're trying
1786 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1788 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1792 * Cpusets can break the scheduler domain tree into smaller
1793 * balance domains, some of which do not cross NUMA boundaries.
1794 * Tasks that are "trapped" in such domains cannot be migrated
1795 * elsewhere, so there is no point in (re)trying.
1797 if (unlikely(!sd)) {
1798 p->numa_preferred_nid = task_node(p);
1802 env.dst_nid = p->numa_preferred_nid;
1803 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1804 taskweight = task_weight(p, env.src_nid, dist);
1805 groupweight = group_weight(p, env.src_nid, dist);
1806 update_numa_stats(&env.src_stats, env.src_nid);
1807 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1808 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1809 update_numa_stats(&env.dst_stats, env.dst_nid);
1811 /* Try to find a spot on the preferred nid. */
1812 if (numa_has_capacity(&env))
1813 task_numa_find_cpu(&env, taskimp, groupimp);
1816 * Look at other nodes in these cases:
1817 * - there is no space available on the preferred_nid
1818 * - the task is part of a numa_group that is interleaved across
1819 * multiple NUMA nodes; in order to better consolidate the group,
1820 * we need to check other locations.
1822 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1823 for_each_online_node(nid) {
1824 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1827 dist = node_distance(env.src_nid, env.dst_nid);
1828 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1830 taskweight = task_weight(p, env.src_nid, dist);
1831 groupweight = group_weight(p, env.src_nid, dist);
1834 /* Only consider nodes where both task and groups benefit */
1835 taskimp = task_weight(p, nid, dist) - taskweight;
1836 groupimp = group_weight(p, nid, dist) - groupweight;
1837 if (taskimp < 0 && groupimp < 0)
1842 update_numa_stats(&env.dst_stats, env.dst_nid);
1843 if (numa_has_capacity(&env))
1844 task_numa_find_cpu(&env, taskimp, groupimp);
1849 * If the task is part of a workload that spans multiple NUMA nodes,
1850 * and is migrating into one of the workload's active nodes, remember
1851 * this node as the task's preferred numa node, so the workload can
1853 * A task that migrated to a second choice node will be better off
1854 * trying for a better one later. Do not set the preferred node here.
1856 if (p->numa_group) {
1857 struct numa_group *ng = p->numa_group;
1859 if (env.best_cpu == -1)
1864 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1865 sched_setnuma(p, env.dst_nid);
1868 /* No better CPU than the current one was found. */
1869 if (env.best_cpu == -1)
1873 * Reset the scan period if the task is being rescheduled on an
1874 * alternative node to recheck if the tasks is now properly placed.
1876 p->numa_scan_period = task_scan_start(p);
1878 if (env.best_task == NULL) {
1879 ret = migrate_task_to(p, env.best_cpu);
1881 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1885 ret = migrate_swap(p, env.best_task);
1887 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1888 put_task_struct(env.best_task);
1892 /* Attempt to migrate a task to a CPU on the preferred node. */
1893 static void numa_migrate_preferred(struct task_struct *p)
1895 unsigned long interval = HZ;
1897 /* This task has no NUMA fault statistics yet */
1898 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1901 /* Periodically retry migrating the task to the preferred node */
1902 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1903 p->numa_migrate_retry = jiffies + interval;
1905 /* Success if task is already running on preferred CPU */
1906 if (task_node(p) == p->numa_preferred_nid)
1909 /* Otherwise, try migrate to a CPU on the preferred node */
1910 task_numa_migrate(p);
1914 * Find out how many nodes on the workload is actively running on. Do this by
1915 * tracking the nodes from which NUMA hinting faults are triggered. This can
1916 * be different from the set of nodes where the workload's memory is currently
1919 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1921 unsigned long faults, max_faults = 0;
1922 int nid, active_nodes = 0;
1924 for_each_online_node(nid) {
1925 faults = group_faults_cpu(numa_group, nid);
1926 if (faults > max_faults)
1927 max_faults = faults;
1930 for_each_online_node(nid) {
1931 faults = group_faults_cpu(numa_group, nid);
1932 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1936 numa_group->max_faults_cpu = max_faults;
1937 numa_group->active_nodes = active_nodes;
1941 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1942 * increments. The more local the fault statistics are, the higher the scan
1943 * period will be for the next scan window. If local/(local+remote) ratio is
1944 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1945 * the scan period will decrease. Aim for 70% local accesses.
1947 #define NUMA_PERIOD_SLOTS 10
1948 #define NUMA_PERIOD_THRESHOLD 7
1951 * Increase the scan period (slow down scanning) if the majority of
1952 * our memory is already on our local node, or if the majority of
1953 * the page accesses are shared with other processes.
1954 * Otherwise, decrease the scan period.
1956 static void update_task_scan_period(struct task_struct *p,
1957 unsigned long shared, unsigned long private)
1959 unsigned int period_slot;
1960 int lr_ratio, ps_ratio;
1963 unsigned long remote = p->numa_faults_locality[0];
1964 unsigned long local = p->numa_faults_locality[1];
1967 * If there were no record hinting faults then either the task is
1968 * completely idle or all activity is areas that are not of interest
1969 * to automatic numa balancing. Related to that, if there were failed
1970 * migration then it implies we are migrating too quickly or the local
1971 * node is overloaded. In either case, scan slower
1973 if (local + shared == 0 || p->numa_faults_locality[2]) {
1974 p->numa_scan_period = min(p->numa_scan_period_max,
1975 p->numa_scan_period << 1);
1977 p->mm->numa_next_scan = jiffies +
1978 msecs_to_jiffies(p->numa_scan_period);
1984 * Prepare to scale scan period relative to the current period.
1985 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1986 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1987 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1989 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1990 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1991 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1993 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1995 * Most memory accesses are local. There is no need to
1996 * do fast NUMA scanning, since memory is already local.
1998 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2001 diff = slot * period_slot;
2002 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2004 * Most memory accesses are shared with other tasks.
2005 * There is no point in continuing fast NUMA scanning,
2006 * since other tasks may just move the memory elsewhere.
2008 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2011 diff = slot * period_slot;
2014 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2015 * yet they are not on the local NUMA node. Speed up
2016 * NUMA scanning to get the memory moved over.
2018 int ratio = max(lr_ratio, ps_ratio);
2019 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2022 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2023 task_scan_min(p), task_scan_max(p));
2024 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2028 * Get the fraction of time the task has been running since the last
2029 * NUMA placement cycle. The scheduler keeps similar statistics, but
2030 * decays those on a 32ms period, which is orders of magnitude off
2031 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2032 * stats only if the task is so new there are no NUMA statistics yet.
2034 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2036 u64 runtime, delta, now;
2037 /* Use the start of this time slice to avoid calculations. */
2038 now = p->se.exec_start;
2039 runtime = p->se.sum_exec_runtime;
2041 if (p->last_task_numa_placement) {
2042 delta = runtime - p->last_sum_exec_runtime;
2043 *period = now - p->last_task_numa_placement;
2045 delta = p->se.avg.load_sum;
2046 *period = LOAD_AVG_MAX;
2049 p->last_sum_exec_runtime = runtime;
2050 p->last_task_numa_placement = now;
2056 * Determine the preferred nid for a task in a numa_group. This needs to
2057 * be done in a way that produces consistent results with group_weight,
2058 * otherwise workloads might not converge.
2060 static int preferred_group_nid(struct task_struct *p, int nid)
2065 /* Direct connections between all NUMA nodes. */
2066 if (sched_numa_topology_type == NUMA_DIRECT)
2070 * On a system with glueless mesh NUMA topology, group_weight
2071 * scores nodes according to the number of NUMA hinting faults on
2072 * both the node itself, and on nearby nodes.
2074 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2075 unsigned long score, max_score = 0;
2076 int node, max_node = nid;
2078 dist = sched_max_numa_distance;
2080 for_each_online_node(node) {
2081 score = group_weight(p, node, dist);
2082 if (score > max_score) {
2091 * Finding the preferred nid in a system with NUMA backplane
2092 * interconnect topology is more involved. The goal is to locate
2093 * tasks from numa_groups near each other in the system, and
2094 * untangle workloads from different sides of the system. This requires
2095 * searching down the hierarchy of node groups, recursively searching
2096 * inside the highest scoring group of nodes. The nodemask tricks
2097 * keep the complexity of the search down.
2099 nodes = node_online_map;
2100 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2101 unsigned long max_faults = 0;
2102 nodemask_t max_group = NODE_MASK_NONE;
2105 /* Are there nodes at this distance from each other? */
2106 if (!find_numa_distance(dist))
2109 for_each_node_mask(a, nodes) {
2110 unsigned long faults = 0;
2111 nodemask_t this_group;
2112 nodes_clear(this_group);
2114 /* Sum group's NUMA faults; includes a==b case. */
2115 for_each_node_mask(b, nodes) {
2116 if (node_distance(a, b) < dist) {
2117 faults += group_faults(p, b);
2118 node_set(b, this_group);
2119 node_clear(b, nodes);
2123 /* Remember the top group. */
2124 if (faults > max_faults) {
2125 max_faults = faults;
2126 max_group = this_group;
2128 * subtle: at the smallest distance there is
2129 * just one node left in each "group", the
2130 * winner is the preferred nid.
2135 /* Next round, evaluate the nodes within max_group. */
2143 static void task_numa_placement(struct task_struct *p)
2145 int seq, nid, max_nid = -1, max_group_nid = -1;
2146 unsigned long max_faults = 0, max_group_faults = 0;
2147 unsigned long fault_types[2] = { 0, 0 };
2148 unsigned long total_faults;
2149 u64 runtime, period;
2150 spinlock_t *group_lock = NULL;
2153 * The p->mm->numa_scan_seq field gets updated without
2154 * exclusive access. Use READ_ONCE() here to ensure
2155 * that the field is read in a single access:
2157 seq = READ_ONCE(p->mm->numa_scan_seq);
2158 if (p->numa_scan_seq == seq)
2160 p->numa_scan_seq = seq;
2161 p->numa_scan_period_max = task_scan_max(p);
2163 total_faults = p->numa_faults_locality[0] +
2164 p->numa_faults_locality[1];
2165 runtime = numa_get_avg_runtime(p, &period);
2167 /* If the task is part of a group prevent parallel updates to group stats */
2168 if (p->numa_group) {
2169 group_lock = &p->numa_group->lock;
2170 spin_lock_irq(group_lock);
2173 /* Find the node with the highest number of faults */
2174 for_each_online_node(nid) {
2175 /* Keep track of the offsets in numa_faults array */
2176 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2177 unsigned long faults = 0, group_faults = 0;
2180 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2181 long diff, f_diff, f_weight;
2183 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2184 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2185 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2186 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2188 /* Decay existing window, copy faults since last scan */
2189 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2190 fault_types[priv] += p->numa_faults[membuf_idx];
2191 p->numa_faults[membuf_idx] = 0;
2194 * Normalize the faults_from, so all tasks in a group
2195 * count according to CPU use, instead of by the raw
2196 * number of faults. Tasks with little runtime have
2197 * little over-all impact on throughput, and thus their
2198 * faults are less important.
2200 f_weight = div64_u64(runtime << 16, period + 1);
2201 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2203 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2204 p->numa_faults[cpubuf_idx] = 0;
2206 p->numa_faults[mem_idx] += diff;
2207 p->numa_faults[cpu_idx] += f_diff;
2208 faults += p->numa_faults[mem_idx];
2209 p->total_numa_faults += diff;
2210 if (p->numa_group) {
2212 * safe because we can only change our own group
2214 * mem_idx represents the offset for a given
2215 * nid and priv in a specific region because it
2216 * is at the beginning of the numa_faults array.
2218 p->numa_group->faults[mem_idx] += diff;
2219 p->numa_group->faults_cpu[mem_idx] += f_diff;
2220 p->numa_group->total_faults += diff;
2221 group_faults += p->numa_group->faults[mem_idx];
2225 if (faults > max_faults) {
2226 max_faults = faults;
2230 if (group_faults > max_group_faults) {
2231 max_group_faults = group_faults;
2232 max_group_nid = nid;
2236 update_task_scan_period(p, fault_types[0], fault_types[1]);
2238 if (p->numa_group) {
2239 numa_group_count_active_nodes(p->numa_group);
2240 spin_unlock_irq(group_lock);
2241 max_nid = preferred_group_nid(p, max_group_nid);
2245 /* Set the new preferred node */
2246 if (max_nid != p->numa_preferred_nid)
2247 sched_setnuma(p, max_nid);
2249 if (task_node(p) != p->numa_preferred_nid)
2250 numa_migrate_preferred(p);
2254 static inline int get_numa_group(struct numa_group *grp)
2256 return atomic_inc_not_zero(&grp->refcount);
2259 static inline void put_numa_group(struct numa_group *grp)
2261 if (atomic_dec_and_test(&grp->refcount))
2262 kfree_rcu(grp, rcu);
2265 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2268 struct numa_group *grp, *my_grp;
2269 struct task_struct *tsk;
2271 int cpu = cpupid_to_cpu(cpupid);
2274 if (unlikely(!p->numa_group)) {
2275 unsigned int size = sizeof(struct numa_group) +
2276 4*nr_node_ids*sizeof(unsigned long);
2278 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2282 atomic_set(&grp->refcount, 1);
2283 grp->active_nodes = 1;
2284 grp->max_faults_cpu = 0;
2285 spin_lock_init(&grp->lock);
2287 /* Second half of the array tracks nids where faults happen */
2288 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2291 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2292 grp->faults[i] = p->numa_faults[i];
2294 grp->total_faults = p->total_numa_faults;
2297 rcu_assign_pointer(p->numa_group, grp);
2301 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2303 if (!cpupid_match_pid(tsk, cpupid))
2306 grp = rcu_dereference(tsk->numa_group);
2310 my_grp = p->numa_group;
2315 * Only join the other group if its bigger; if we're the bigger group,
2316 * the other task will join us.
2318 if (my_grp->nr_tasks > grp->nr_tasks)
2322 * Tie-break on the grp address.
2324 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2327 /* Always join threads in the same process. */
2328 if (tsk->mm == current->mm)
2331 /* Simple filter to avoid false positives due to PID collisions */
2332 if (flags & TNF_SHARED)
2335 /* Update priv based on whether false sharing was detected */
2338 if (join && !get_numa_group(grp))
2346 BUG_ON(irqs_disabled());
2347 double_lock_irq(&my_grp->lock, &grp->lock);
2349 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2350 my_grp->faults[i] -= p->numa_faults[i];
2351 grp->faults[i] += p->numa_faults[i];
2353 my_grp->total_faults -= p->total_numa_faults;
2354 grp->total_faults += p->total_numa_faults;
2359 spin_unlock(&my_grp->lock);
2360 spin_unlock_irq(&grp->lock);
2362 rcu_assign_pointer(p->numa_group, grp);
2364 put_numa_group(my_grp);
2372 void task_numa_free(struct task_struct *p)
2374 struct numa_group *grp = p->numa_group;
2375 void *numa_faults = p->numa_faults;
2376 unsigned long flags;
2380 spin_lock_irqsave(&grp->lock, flags);
2381 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2382 grp->faults[i] -= p->numa_faults[i];
2383 grp->total_faults -= p->total_numa_faults;
2386 spin_unlock_irqrestore(&grp->lock, flags);
2387 RCU_INIT_POINTER(p->numa_group, NULL);
2388 put_numa_group(grp);
2391 p->numa_faults = NULL;
2396 * Got a PROT_NONE fault for a page on @node.
2398 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2400 struct task_struct *p = current;
2401 bool migrated = flags & TNF_MIGRATED;
2402 int cpu_node = task_node(current);
2403 int local = !!(flags & TNF_FAULT_LOCAL);
2404 struct numa_group *ng;
2407 if (!static_branch_likely(&sched_numa_balancing))
2410 /* for example, ksmd faulting in a user's mm */
2414 /* Allocate buffer to track faults on a per-node basis */
2415 if (unlikely(!p->numa_faults)) {
2416 int size = sizeof(*p->numa_faults) *
2417 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2419 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2420 if (!p->numa_faults)
2423 p->total_numa_faults = 0;
2424 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2428 * First accesses are treated as private, otherwise consider accesses
2429 * to be private if the accessing pid has not changed
2431 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2434 priv = cpupid_match_pid(p, last_cpupid);
2435 if (!priv && !(flags & TNF_NO_GROUP))
2436 task_numa_group(p, last_cpupid, flags, &priv);
2440 * If a workload spans multiple NUMA nodes, a shared fault that
2441 * occurs wholly within the set of nodes that the workload is
2442 * actively using should be counted as local. This allows the
2443 * scan rate to slow down when a workload has settled down.
2446 if (!priv && !local && ng && ng->active_nodes > 1 &&
2447 numa_is_active_node(cpu_node, ng) &&
2448 numa_is_active_node(mem_node, ng))
2451 task_numa_placement(p);
2454 * Retry task to preferred node migration periodically, in case it
2455 * case it previously failed, or the scheduler moved us.
2457 if (time_after(jiffies, p->numa_migrate_retry))
2458 numa_migrate_preferred(p);
2461 p->numa_pages_migrated += pages;
2462 if (flags & TNF_MIGRATE_FAIL)
2463 p->numa_faults_locality[2] += pages;
2465 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2466 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2467 p->numa_faults_locality[local] += pages;
2470 static void reset_ptenuma_scan(struct task_struct *p)
2473 * We only did a read acquisition of the mmap sem, so
2474 * p->mm->numa_scan_seq is written to without exclusive access
2475 * and the update is not guaranteed to be atomic. That's not
2476 * much of an issue though, since this is just used for
2477 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2478 * expensive, to avoid any form of compiler optimizations:
2480 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2481 p->mm->numa_scan_offset = 0;
2485 * The expensive part of numa migration is done from task_work context.
2486 * Triggered from task_tick_numa().
2488 void task_numa_work(struct callback_head *work)
2490 unsigned long migrate, next_scan, now = jiffies;
2491 struct task_struct *p = current;
2492 struct mm_struct *mm = p->mm;
2493 u64 runtime = p->se.sum_exec_runtime;
2494 struct vm_area_struct *vma;
2495 unsigned long start, end;
2496 unsigned long nr_pte_updates = 0;
2497 long pages, virtpages;
2499 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2501 work->next = work; /* protect against double add */
2503 * Who cares about NUMA placement when they're dying.
2505 * NOTE: make sure not to dereference p->mm before this check,
2506 * exit_task_work() happens _after_ exit_mm() so we could be called
2507 * without p->mm even though we still had it when we enqueued this
2510 if (p->flags & PF_EXITING)
2513 if (!mm->numa_next_scan) {
2514 mm->numa_next_scan = now +
2515 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2519 * Enforce maximal scan/migration frequency..
2521 migrate = mm->numa_next_scan;
2522 if (time_before(now, migrate))
2525 if (p->numa_scan_period == 0) {
2526 p->numa_scan_period_max = task_scan_max(p);
2527 p->numa_scan_period = task_scan_start(p);
2530 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2531 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2535 * Delay this task enough that another task of this mm will likely win
2536 * the next time around.
2538 p->node_stamp += 2 * TICK_NSEC;
2540 start = mm->numa_scan_offset;
2541 pages = sysctl_numa_balancing_scan_size;
2542 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2543 virtpages = pages * 8; /* Scan up to this much virtual space */
2548 if (!down_read_trylock(&mm->mmap_sem))
2550 vma = find_vma(mm, start);
2552 reset_ptenuma_scan(p);
2556 for (; vma; vma = vma->vm_next) {
2557 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2558 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2563 * Shared library pages mapped by multiple processes are not
2564 * migrated as it is expected they are cache replicated. Avoid
2565 * hinting faults in read-only file-backed mappings or the vdso
2566 * as migrating the pages will be of marginal benefit.
2569 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2573 * Skip inaccessible VMAs to avoid any confusion between
2574 * PROT_NONE and NUMA hinting ptes
2576 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2580 start = max(start, vma->vm_start);
2581 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2582 end = min(end, vma->vm_end);
2583 nr_pte_updates = change_prot_numa(vma, start, end);
2586 * Try to scan sysctl_numa_balancing_size worth of
2587 * hpages that have at least one present PTE that
2588 * is not already pte-numa. If the VMA contains
2589 * areas that are unused or already full of prot_numa
2590 * PTEs, scan up to virtpages, to skip through those
2594 pages -= (end - start) >> PAGE_SHIFT;
2595 virtpages -= (end - start) >> PAGE_SHIFT;
2598 if (pages <= 0 || virtpages <= 0)
2602 } while (end != vma->vm_end);
2607 * It is possible to reach the end of the VMA list but the last few
2608 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2609 * would find the !migratable VMA on the next scan but not reset the
2610 * scanner to the start so check it now.
2613 mm->numa_scan_offset = start;
2615 reset_ptenuma_scan(p);
2616 up_read(&mm->mmap_sem);
2619 * Make sure tasks use at least 32x as much time to run other code
2620 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2621 * Usually update_task_scan_period slows down scanning enough; on an
2622 * overloaded system we need to limit overhead on a per task basis.
2624 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2625 u64 diff = p->se.sum_exec_runtime - runtime;
2626 p->node_stamp += 32 * diff;
2631 * Drive the periodic memory faults..
2633 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2635 struct callback_head *work = &curr->numa_work;
2639 * We don't care about NUMA placement if we don't have memory.
2641 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2645 * Using runtime rather than walltime has the dual advantage that
2646 * we (mostly) drive the selection from busy threads and that the
2647 * task needs to have done some actual work before we bother with
2650 now = curr->se.sum_exec_runtime;
2651 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2653 if (now > curr->node_stamp + period) {
2654 if (!curr->node_stamp)
2655 curr->numa_scan_period = task_scan_start(curr);
2656 curr->node_stamp += period;
2658 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2659 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2660 task_work_add(curr, work, true);
2666 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2670 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2674 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2678 #endif /* CONFIG_NUMA_BALANCING */
2681 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2683 update_load_add(&cfs_rq->load, se->load.weight);
2684 if (!parent_entity(se))
2685 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2687 if (entity_is_task(se)) {
2688 struct rq *rq = rq_of(cfs_rq);
2690 account_numa_enqueue(rq, task_of(se));
2691 list_add(&se->group_node, &rq->cfs_tasks);
2694 cfs_rq->nr_running++;
2698 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2700 update_load_sub(&cfs_rq->load, se->load.weight);
2701 if (!parent_entity(se))
2702 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2704 if (entity_is_task(se)) {
2705 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2706 list_del_init(&se->group_node);
2709 cfs_rq->nr_running--;
2713 * Signed add and clamp on underflow.
2715 * Explicitly do a load-store to ensure the intermediate value never hits
2716 * memory. This allows lockless observations without ever seeing the negative
2719 #define add_positive(_ptr, _val) do { \
2720 typeof(_ptr) ptr = (_ptr); \
2721 typeof(_val) val = (_val); \
2722 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2726 if (val < 0 && res > var) \
2729 WRITE_ONCE(*ptr, res); \
2733 * Unsigned subtract and clamp on underflow.
2735 * Explicitly do a load-store to ensure the intermediate value never hits
2736 * memory. This allows lockless observations without ever seeing the negative
2739 #define sub_positive(_ptr, _val) do { \
2740 typeof(_ptr) ptr = (_ptr); \
2741 typeof(*ptr) val = (_val); \
2742 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2746 WRITE_ONCE(*ptr, res); \
2751 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2753 cfs_rq->runnable_weight += se->runnable_weight;
2755 cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2756 cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2760 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2762 cfs_rq->runnable_weight -= se->runnable_weight;
2764 sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2765 sub_positive(&cfs_rq->avg.runnable_load_sum,
2766 se_runnable(se) * se->avg.runnable_load_sum);
2770 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2772 cfs_rq->avg.load_avg += se->avg.load_avg;
2773 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2777 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2779 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2780 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2784 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2786 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2788 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2790 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2793 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2794 unsigned long weight, unsigned long runnable)
2797 /* commit outstanding execution time */
2798 if (cfs_rq->curr == se)
2799 update_curr(cfs_rq);
2800 account_entity_dequeue(cfs_rq, se);
2801 dequeue_runnable_load_avg(cfs_rq, se);
2803 dequeue_load_avg(cfs_rq, se);
2805 se->runnable_weight = runnable;
2806 update_load_set(&se->load, weight);
2810 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2812 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2813 se->avg.runnable_load_avg =
2814 div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2818 enqueue_load_avg(cfs_rq, se);
2820 account_entity_enqueue(cfs_rq, se);
2821 enqueue_runnable_load_avg(cfs_rq, se);
2825 void reweight_task(struct task_struct *p, int prio)
2827 struct sched_entity *se = &p->se;
2828 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2829 struct load_weight *load = &se->load;
2830 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2832 reweight_entity(cfs_rq, se, weight, weight);
2833 load->inv_weight = sched_prio_to_wmult[prio];
2836 #ifdef CONFIG_FAIR_GROUP_SCHED
2839 * All this does is approximate the hierarchical proportion which includes that
2840 * global sum we all love to hate.
2842 * That is, the weight of a group entity, is the proportional share of the
2843 * group weight based on the group runqueue weights. That is:
2845 * tg->weight * grq->load.weight
2846 * ge->load.weight = ----------------------------- (1)
2847 * \Sum grq->load.weight
2849 * Now, because computing that sum is prohibitively expensive to compute (been
2850 * there, done that) we approximate it with this average stuff. The average
2851 * moves slower and therefore the approximation is cheaper and more stable.
2853 * So instead of the above, we substitute:
2855 * grq->load.weight -> grq->avg.load_avg (2)
2857 * which yields the following:
2859 * tg->weight * grq->avg.load_avg
2860 * ge->load.weight = ------------------------------ (3)
2863 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2865 * That is shares_avg, and it is right (given the approximation (2)).
2867 * The problem with it is that because the average is slow -- it was designed
2868 * to be exactly that of course -- this leads to transients in boundary
2869 * conditions. In specific, the case where the group was idle and we start the
2870 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2871 * yielding bad latency etc..
2873 * Now, in that special case (1) reduces to:
2875 * tg->weight * grq->load.weight
2876 * ge->load.weight = ----------------------------- = tg->weight (4)
2879 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2881 * So what we do is modify our approximation (3) to approach (4) in the (near)
2886 * tg->weight * grq->load.weight
2887 * --------------------------------------------------- (5)
2888 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2890 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2891 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2894 * tg->weight * grq->load.weight
2895 * ge->load.weight = ----------------------------- (6)
2900 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2901 * max(grq->load.weight, grq->avg.load_avg)
2903 * And that is shares_weight and is icky. In the (near) UP case it approaches
2904 * (4) while in the normal case it approaches (3). It consistently
2905 * overestimates the ge->load.weight and therefore:
2907 * \Sum ge->load.weight >= tg->weight
2911 static long calc_group_shares(struct cfs_rq *cfs_rq)
2913 long tg_weight, tg_shares, load, shares;
2914 struct task_group *tg = cfs_rq->tg;
2916 tg_shares = READ_ONCE(tg->shares);
2918 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
2920 tg_weight = atomic_long_read(&tg->load_avg);
2922 /* Ensure tg_weight >= load */
2923 tg_weight -= cfs_rq->tg_load_avg_contrib;
2926 shares = (tg_shares * load);
2928 shares /= tg_weight;
2931 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2932 * of a group with small tg->shares value. It is a floor value which is
2933 * assigned as a minimum load.weight to the sched_entity representing
2934 * the group on a CPU.
2936 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2937 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2938 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2939 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2942 return clamp_t(long, shares, MIN_SHARES, tg_shares);
2946 * This calculates the effective runnable weight for a group entity based on
2947 * the group entity weight calculated above.
2949 * Because of the above approximation (2), our group entity weight is
2950 * an load_avg based ratio (3). This means that it includes blocked load and
2951 * does not represent the runnable weight.
2953 * Approximate the group entity's runnable weight per ratio from the group
2956 * grq->avg.runnable_load_avg
2957 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
2960 * However, analogous to above, since the avg numbers are slow, this leads to
2961 * transients in the from-idle case. Instead we use:
2963 * ge->runnable_weight = ge->load.weight *
2965 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
2966 * ----------------------------------------------------- (8)
2967 * max(grq->avg.load_avg, grq->load.weight)
2969 * Where these max() serve both to use the 'instant' values to fix the slow
2970 * from-idle and avoid the /0 on to-idle, similar to (6).
2972 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
2974 long runnable, load_avg;
2976 load_avg = max(cfs_rq->avg.load_avg,
2977 scale_load_down(cfs_rq->load.weight));
2979 runnable = max(cfs_rq->avg.runnable_load_avg,
2980 scale_load_down(cfs_rq->runnable_weight));
2984 runnable /= load_avg;
2986 return clamp_t(long, runnable, MIN_SHARES, shares);
2988 #endif /* CONFIG_SMP */
2990 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2993 * Recomputes the group entity based on the current state of its group
2996 static void update_cfs_group(struct sched_entity *se)
2998 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
2999 long shares, runnable;
3004 if (throttled_hierarchy(gcfs_rq))
3008 runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
3010 if (likely(se->load.weight == shares))
3013 shares = calc_group_shares(gcfs_rq);
3014 runnable = calc_group_runnable(gcfs_rq, shares);
3017 reweight_entity(cfs_rq_of(se), se, shares, runnable);
3020 #else /* CONFIG_FAIR_GROUP_SCHED */
3021 static inline void update_cfs_group(struct sched_entity *se)
3024 #endif /* CONFIG_FAIR_GROUP_SCHED */
3026 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3028 struct rq *rq = rq_of(cfs_rq);
3030 if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) {
3032 * There are a few boundary cases this might miss but it should
3033 * get called often enough that that should (hopefully) not be
3036 * It will not get called when we go idle, because the idle
3037 * thread is a different class (!fair), nor will the utilization
3038 * number include things like RT tasks.
3040 * As is, the util number is not freq-invariant (we'd have to
3041 * implement arch_scale_freq_capacity() for that).
3045 cpufreq_update_util(rq, flags);
3050 #ifdef CONFIG_FAIR_GROUP_SCHED
3052 * update_tg_load_avg - update the tg's load avg
3053 * @cfs_rq: the cfs_rq whose avg changed
3054 * @force: update regardless of how small the difference
3056 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3057 * However, because tg->load_avg is a global value there are performance
3060 * In order to avoid having to look at the other cfs_rq's, we use a
3061 * differential update where we store the last value we propagated. This in
3062 * turn allows skipping updates if the differential is 'small'.
3064 * Updating tg's load_avg is necessary before update_cfs_share().
3066 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3068 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3071 * No need to update load_avg for root_task_group as it is not used.
3073 if (cfs_rq->tg == &root_task_group)
3076 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3077 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3078 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3083 * Called within set_task_rq() right before setting a task's CPU. The
3084 * caller only guarantees p->pi_lock is held; no other assumptions,
3085 * including the state of rq->lock, should be made.
3087 void set_task_rq_fair(struct sched_entity *se,
3088 struct cfs_rq *prev, struct cfs_rq *next)
3090 u64 p_last_update_time;
3091 u64 n_last_update_time;
3093 if (!sched_feat(ATTACH_AGE_LOAD))
3097 * We are supposed to update the task to "current" time, then its up to
3098 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3099 * getting what current time is, so simply throw away the out-of-date
3100 * time. This will result in the wakee task is less decayed, but giving
3101 * the wakee more load sounds not bad.
3103 if (!(se->avg.last_update_time && prev))
3106 #ifndef CONFIG_64BIT
3108 u64 p_last_update_time_copy;
3109 u64 n_last_update_time_copy;
3112 p_last_update_time_copy = prev->load_last_update_time_copy;
3113 n_last_update_time_copy = next->load_last_update_time_copy;
3117 p_last_update_time = prev->avg.last_update_time;
3118 n_last_update_time = next->avg.last_update_time;
3120 } while (p_last_update_time != p_last_update_time_copy ||
3121 n_last_update_time != n_last_update_time_copy);
3124 p_last_update_time = prev->avg.last_update_time;
3125 n_last_update_time = next->avg.last_update_time;
3127 __update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se);
3128 se->avg.last_update_time = n_last_update_time;
3133 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3134 * propagate its contribution. The key to this propagation is the invariant
3135 * that for each group:
3137 * ge->avg == grq->avg (1)
3139 * _IFF_ we look at the pure running and runnable sums. Because they
3140 * represent the very same entity, just at different points in the hierarchy.
3142 * Per the above update_tg_cfs_util() is trivial and simply copies the running
3143 * sum over (but still wrong, because the group entity and group rq do not have
3144 * their PELT windows aligned).
3146 * However, update_tg_cfs_runnable() is more complex. So we have:
3148 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3150 * And since, like util, the runnable part should be directly transferable,
3151 * the following would _appear_ to be the straight forward approach:
3153 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3155 * And per (1) we have:
3157 * ge->avg.runnable_avg == grq->avg.runnable_avg
3161 * ge->load.weight * grq->avg.load_avg
3162 * ge->avg.load_avg = ----------------------------------- (4)
3165 * Except that is wrong!
3167 * Because while for entities historical weight is not important and we
3168 * really only care about our future and therefore can consider a pure
3169 * runnable sum, runqueues can NOT do this.
3171 * We specifically want runqueues to have a load_avg that includes
3172 * historical weights. Those represent the blocked load, the load we expect
3173 * to (shortly) return to us. This only works by keeping the weights as
3174 * integral part of the sum. We therefore cannot decompose as per (3).
3176 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3177 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3178 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3179 * runnable section of these tasks overlap (or not). If they were to perfectly
3180 * align the rq as a whole would be runnable 2/3 of the time. If however we
3181 * always have at least 1 runnable task, the rq as a whole is always runnable.
3183 * So we'll have to approximate.. :/
3185 * Given the constraint:
3187 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3189 * We can construct a rule that adds runnable to a rq by assuming minimal
3192 * On removal, we'll assume each task is equally runnable; which yields:
3194 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3196 * XXX: only do this for the part of runnable > running ?
3201 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3203 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3205 /* Nothing to update */
3210 * The relation between sum and avg is:
3212 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3214 * however, the PELT windows are not aligned between grq and gse.
3217 /* Set new sched_entity's utilization */
3218 se->avg.util_avg = gcfs_rq->avg.util_avg;
3219 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3221 /* Update parent cfs_rq utilization */
3222 add_positive(&cfs_rq->avg.util_avg, delta);
3223 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3227 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3229 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3230 unsigned long runnable_load_avg, load_avg;
3231 u64 runnable_load_sum, load_sum = 0;
3237 gcfs_rq->prop_runnable_sum = 0;
3239 if (runnable_sum >= 0) {
3241 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3242 * the CPU is saturated running == runnable.
3244 runnable_sum += se->avg.load_sum;
3245 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3248 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3249 * assuming all tasks are equally runnable.
3251 if (scale_load_down(gcfs_rq->load.weight)) {
3252 load_sum = div_s64(gcfs_rq->avg.load_sum,
3253 scale_load_down(gcfs_rq->load.weight));
3256 /* But make sure to not inflate se's runnable */
3257 runnable_sum = min(se->avg.load_sum, load_sum);
3261 * runnable_sum can't be lower than running_sum
3262 * As running sum is scale with CPU capacity wehreas the runnable sum
3263 * is not we rescale running_sum 1st
3265 running_sum = se->avg.util_sum /
3266 arch_scale_cpu_capacity(NULL, cpu_of(rq_of(cfs_rq)));
3267 runnable_sum = max(runnable_sum, running_sum);
3269 load_sum = (s64)se_weight(se) * runnable_sum;
3270 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3272 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3273 delta_avg = load_avg - se->avg.load_avg;
3275 se->avg.load_sum = runnable_sum;
3276 se->avg.load_avg = load_avg;
3277 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3278 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3280 runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3281 runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3282 delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3283 delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3285 se->avg.runnable_load_sum = runnable_sum;
3286 se->avg.runnable_load_avg = runnable_load_avg;
3289 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3290 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3294 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3296 cfs_rq->propagate = 1;
3297 cfs_rq->prop_runnable_sum += runnable_sum;
3300 /* Update task and its cfs_rq load average */
3301 static inline int propagate_entity_load_avg(struct sched_entity *se)
3303 struct cfs_rq *cfs_rq, *gcfs_rq;
3305 if (entity_is_task(se))
3308 gcfs_rq = group_cfs_rq(se);
3309 if (!gcfs_rq->propagate)
3312 gcfs_rq->propagate = 0;
3314 cfs_rq = cfs_rq_of(se);
3316 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3318 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3319 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3325 * Check if we need to update the load and the utilization of a blocked
3328 static inline bool skip_blocked_update(struct sched_entity *se)
3330 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3333 * If sched_entity still have not zero load or utilization, we have to
3336 if (se->avg.load_avg || se->avg.util_avg)
3340 * If there is a pending propagation, we have to update the load and
3341 * the utilization of the sched_entity:
3343 if (gcfs_rq->propagate)
3347 * Otherwise, the load and the utilization of the sched_entity is
3348 * already zero and there is no pending propagation, so it will be a
3349 * waste of time to try to decay it:
3354 #else /* CONFIG_FAIR_GROUP_SCHED */
3356 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3358 static inline int propagate_entity_load_avg(struct sched_entity *se)
3363 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3365 #endif /* CONFIG_FAIR_GROUP_SCHED */
3368 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3369 * @now: current time, as per cfs_rq_clock_task()
3370 * @cfs_rq: cfs_rq to update
3372 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3373 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3374 * post_init_entity_util_avg().
3376 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3378 * Returns true if the load decayed or we removed load.
3380 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3381 * call update_tg_load_avg() when this function returns true.
3384 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3386 unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3387 struct sched_avg *sa = &cfs_rq->avg;
3390 if (cfs_rq->removed.nr) {
3392 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3394 raw_spin_lock(&cfs_rq->removed.lock);
3395 swap(cfs_rq->removed.util_avg, removed_util);
3396 swap(cfs_rq->removed.load_avg, removed_load);
3397 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3398 cfs_rq->removed.nr = 0;
3399 raw_spin_unlock(&cfs_rq->removed.lock);
3402 sub_positive(&sa->load_avg, r);
3403 sub_positive(&sa->load_sum, r * divider);
3406 sub_positive(&sa->util_avg, r);
3407 sub_positive(&sa->util_sum, r * divider);
3409 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3414 decayed |= __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq);
3416 #ifndef CONFIG_64BIT
3418 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3422 cfs_rq_util_change(cfs_rq, 0);
3428 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3429 * @cfs_rq: cfs_rq to attach to
3430 * @se: sched_entity to attach
3432 * Must call update_cfs_rq_load_avg() before this, since we rely on
3433 * cfs_rq->avg.last_update_time being current.
3435 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3437 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3440 * When we attach the @se to the @cfs_rq, we must align the decay
3441 * window because without that, really weird and wonderful things can
3446 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3447 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3450 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3451 * period_contrib. This isn't strictly correct, but since we're
3452 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3455 se->avg.util_sum = se->avg.util_avg * divider;
3457 se->avg.load_sum = divider;
3458 if (se_weight(se)) {
3460 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3463 se->avg.runnable_load_sum = se->avg.load_sum;
3465 enqueue_load_avg(cfs_rq, se);
3466 cfs_rq->avg.util_avg += se->avg.util_avg;
3467 cfs_rq->avg.util_sum += se->avg.util_sum;
3469 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3471 cfs_rq_util_change(cfs_rq, flags);
3475 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3476 * @cfs_rq: cfs_rq to detach from
3477 * @se: sched_entity to detach
3479 * Must call update_cfs_rq_load_avg() before this, since we rely on
3480 * cfs_rq->avg.last_update_time being current.
3482 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3484 dequeue_load_avg(cfs_rq, se);
3485 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3486 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3488 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3490 cfs_rq_util_change(cfs_rq, 0);
3494 * Optional action to be done while updating the load average
3496 #define UPDATE_TG 0x1
3497 #define SKIP_AGE_LOAD 0x2
3498 #define DO_ATTACH 0x4
3500 /* Update task and its cfs_rq load average */
3501 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3503 u64 now = cfs_rq_clock_task(cfs_rq);
3504 struct rq *rq = rq_of(cfs_rq);
3505 int cpu = cpu_of(rq);
3509 * Track task load average for carrying it to new CPU after migrated, and
3510 * track group sched_entity load average for task_h_load calc in migration
3512 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3513 __update_load_avg_se(now, cpu, cfs_rq, se);
3515 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3516 decayed |= propagate_entity_load_avg(se);
3518 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3521 * DO_ATTACH means we're here from enqueue_entity().
3522 * !last_update_time means we've passed through
3523 * migrate_task_rq_fair() indicating we migrated.
3525 * IOW we're enqueueing a task on a new CPU.
3527 attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
3528 update_tg_load_avg(cfs_rq, 0);
3530 } else if (decayed && (flags & UPDATE_TG))
3531 update_tg_load_avg(cfs_rq, 0);
3534 #ifndef CONFIG_64BIT
3535 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3537 u64 last_update_time_copy;
3538 u64 last_update_time;
3541 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3543 last_update_time = cfs_rq->avg.last_update_time;
3544 } while (last_update_time != last_update_time_copy);
3546 return last_update_time;
3549 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3551 return cfs_rq->avg.last_update_time;
3556 * Synchronize entity load avg of dequeued entity without locking
3559 void sync_entity_load_avg(struct sched_entity *se)
3561 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3562 u64 last_update_time;
3564 last_update_time = cfs_rq_last_update_time(cfs_rq);
3565 __update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se);
3569 * Task first catches up with cfs_rq, and then subtract
3570 * itself from the cfs_rq (task must be off the queue now).
3572 void remove_entity_load_avg(struct sched_entity *se)
3574 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3575 unsigned long flags;
3578 * tasks cannot exit without having gone through wake_up_new_task() ->
3579 * post_init_entity_util_avg() which will have added things to the
3580 * cfs_rq, so we can remove unconditionally.
3582 * Similarly for groups, they will have passed through
3583 * post_init_entity_util_avg() before unregister_sched_fair_group()
3587 sync_entity_load_avg(se);
3589 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3590 ++cfs_rq->removed.nr;
3591 cfs_rq->removed.util_avg += se->avg.util_avg;
3592 cfs_rq->removed.load_avg += se->avg.load_avg;
3593 cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */
3594 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3597 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3599 return cfs_rq->avg.runnable_load_avg;
3602 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3604 return cfs_rq->avg.load_avg;
3607 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3609 static inline unsigned long task_util(struct task_struct *p)
3611 return READ_ONCE(p->se.avg.util_avg);
3614 static inline unsigned long _task_util_est(struct task_struct *p)
3616 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3618 return max(ue.ewma, ue.enqueued);
3621 static inline unsigned long task_util_est(struct task_struct *p)
3623 return max(task_util(p), _task_util_est(p));
3626 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3627 struct task_struct *p)
3629 unsigned int enqueued;
3631 if (!sched_feat(UTIL_EST))
3634 /* Update root cfs_rq's estimated utilization */
3635 enqueued = cfs_rq->avg.util_est.enqueued;
3636 enqueued += (_task_util_est(p) | UTIL_AVG_UNCHANGED);
3637 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3641 * Check if a (signed) value is within a specified (unsigned) margin,
3642 * based on the observation that:
3644 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3646 * NOTE: this only works when value + maring < INT_MAX.
3648 static inline bool within_margin(int value, int margin)
3650 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3654 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3656 long last_ewma_diff;
3659 if (!sched_feat(UTIL_EST))
3662 /* Update root cfs_rq's estimated utilization */
3663 ue.enqueued = cfs_rq->avg.util_est.enqueued;
3664 ue.enqueued -= min_t(unsigned int, ue.enqueued,
3665 (_task_util_est(p) | UTIL_AVG_UNCHANGED));
3666 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3669 * Skip update of task's estimated utilization when the task has not
3670 * yet completed an activation, e.g. being migrated.
3676 * If the PELT values haven't changed since enqueue time,
3677 * skip the util_est update.
3679 ue = p->se.avg.util_est;
3680 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3684 * Skip update of task's estimated utilization when its EWMA is
3685 * already ~1% close to its last activation value.
3687 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3688 last_ewma_diff = ue.enqueued - ue.ewma;
3689 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3693 * Update Task's estimated utilization
3695 * When *p completes an activation we can consolidate another sample
3696 * of the task size. This is done by storing the current PELT value
3697 * as ue.enqueued and by using this value to update the Exponential
3698 * Weighted Moving Average (EWMA):
3700 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
3701 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
3702 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
3703 * = w * ( last_ewma_diff ) + ewma(t-1)
3704 * = w * (last_ewma_diff + ewma(t-1) / w)
3706 * Where 'w' is the weight of new samples, which is configured to be
3707 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3709 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
3710 ue.ewma += last_ewma_diff;
3711 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
3712 WRITE_ONCE(p->se.avg.util_est, ue);
3715 #else /* CONFIG_SMP */
3717 #define UPDATE_TG 0x0
3718 #define SKIP_AGE_LOAD 0x0
3719 #define DO_ATTACH 0x0
3721 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3723 cfs_rq_util_change(cfs_rq, 0);
3726 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3729 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
3731 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3733 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3739 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
3742 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
3745 #endif /* CONFIG_SMP */
3747 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3749 #ifdef CONFIG_SCHED_DEBUG
3750 s64 d = se->vruntime - cfs_rq->min_vruntime;
3755 if (d > 3*sysctl_sched_latency)
3756 schedstat_inc(cfs_rq->nr_spread_over);
3761 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3763 u64 vruntime = cfs_rq->min_vruntime;
3766 * The 'current' period is already promised to the current tasks,
3767 * however the extra weight of the new task will slow them down a
3768 * little, place the new task so that it fits in the slot that
3769 * stays open at the end.
3771 if (initial && sched_feat(START_DEBIT))
3772 vruntime += sched_vslice(cfs_rq, se);
3774 /* sleeps up to a single latency don't count. */
3776 unsigned long thresh = sysctl_sched_latency;
3779 * Halve their sleep time's effect, to allow
3780 * for a gentler effect of sleepers:
3782 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3788 /* ensure we never gain time by being placed backwards. */
3789 se->vruntime = max_vruntime(se->vruntime, vruntime);
3792 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3794 static inline void check_schedstat_required(void)
3796 #ifdef CONFIG_SCHEDSTATS
3797 if (schedstat_enabled())
3800 /* Force schedstat enabled if a dependent tracepoint is active */
3801 if (trace_sched_stat_wait_enabled() ||
3802 trace_sched_stat_sleep_enabled() ||
3803 trace_sched_stat_iowait_enabled() ||
3804 trace_sched_stat_blocked_enabled() ||
3805 trace_sched_stat_runtime_enabled()) {
3806 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3807 "stat_blocked and stat_runtime require the "
3808 "kernel parameter schedstats=enable or "
3809 "kernel.sched_schedstats=1\n");
3820 * update_min_vruntime()
3821 * vruntime -= min_vruntime
3825 * update_min_vruntime()
3826 * vruntime += min_vruntime
3828 * this way the vruntime transition between RQs is done when both
3829 * min_vruntime are up-to-date.
3833 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3834 * vruntime -= min_vruntime
3838 * update_min_vruntime()
3839 * vruntime += min_vruntime
3841 * this way we don't have the most up-to-date min_vruntime on the originating
3842 * CPU and an up-to-date min_vruntime on the destination CPU.
3846 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3848 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3849 bool curr = cfs_rq->curr == se;
3852 * If we're the current task, we must renormalise before calling
3856 se->vruntime += cfs_rq->min_vruntime;
3858 update_curr(cfs_rq);
3861 * Otherwise, renormalise after, such that we're placed at the current
3862 * moment in time, instead of some random moment in the past. Being
3863 * placed in the past could significantly boost this task to the
3864 * fairness detriment of existing tasks.
3866 if (renorm && !curr)
3867 se->vruntime += cfs_rq->min_vruntime;
3870 * When enqueuing a sched_entity, we must:
3871 * - Update loads to have both entity and cfs_rq synced with now.
3872 * - Add its load to cfs_rq->runnable_avg
3873 * - For group_entity, update its weight to reflect the new share of
3875 * - Add its new weight to cfs_rq->load.weight
3877 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
3878 update_cfs_group(se);
3879 enqueue_runnable_load_avg(cfs_rq, se);
3880 account_entity_enqueue(cfs_rq, se);
3882 if (flags & ENQUEUE_WAKEUP)
3883 place_entity(cfs_rq, se, 0);
3885 check_schedstat_required();
3886 update_stats_enqueue(cfs_rq, se, flags);
3887 check_spread(cfs_rq, se);
3889 __enqueue_entity(cfs_rq, se);
3892 if (cfs_rq->nr_running == 1) {
3893 list_add_leaf_cfs_rq(cfs_rq);
3894 check_enqueue_throttle(cfs_rq);
3898 static void __clear_buddies_last(struct sched_entity *se)
3900 for_each_sched_entity(se) {
3901 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3902 if (cfs_rq->last != se)
3905 cfs_rq->last = NULL;
3909 static void __clear_buddies_next(struct sched_entity *se)
3911 for_each_sched_entity(se) {
3912 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3913 if (cfs_rq->next != se)
3916 cfs_rq->next = NULL;
3920 static void __clear_buddies_skip(struct sched_entity *se)
3922 for_each_sched_entity(se) {
3923 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3924 if (cfs_rq->skip != se)
3927 cfs_rq->skip = NULL;
3931 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3933 if (cfs_rq->last == se)
3934 __clear_buddies_last(se);
3936 if (cfs_rq->next == se)
3937 __clear_buddies_next(se);
3939 if (cfs_rq->skip == se)
3940 __clear_buddies_skip(se);
3943 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3946 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3949 * Update run-time statistics of the 'current'.
3951 update_curr(cfs_rq);
3954 * When dequeuing a sched_entity, we must:
3955 * - Update loads to have both entity and cfs_rq synced with now.
3956 * - Substract its load from the cfs_rq->runnable_avg.
3957 * - Substract its previous weight from cfs_rq->load.weight.
3958 * - For group entity, update its weight to reflect the new share
3959 * of its group cfs_rq.
3961 update_load_avg(cfs_rq, se, UPDATE_TG);
3962 dequeue_runnable_load_avg(cfs_rq, se);
3964 update_stats_dequeue(cfs_rq, se, flags);
3966 clear_buddies(cfs_rq, se);
3968 if (se != cfs_rq->curr)
3969 __dequeue_entity(cfs_rq, se);
3971 account_entity_dequeue(cfs_rq, se);
3974 * Normalize after update_curr(); which will also have moved
3975 * min_vruntime if @se is the one holding it back. But before doing
3976 * update_min_vruntime() again, which will discount @se's position and
3977 * can move min_vruntime forward still more.
3979 if (!(flags & DEQUEUE_SLEEP))
3980 se->vruntime -= cfs_rq->min_vruntime;
3982 /* return excess runtime on last dequeue */
3983 return_cfs_rq_runtime(cfs_rq);
3985 update_cfs_group(se);
3988 * Now advance min_vruntime if @se was the entity holding it back,
3989 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
3990 * put back on, and if we advance min_vruntime, we'll be placed back
3991 * further than we started -- ie. we'll be penalized.
3993 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
3994 update_min_vruntime(cfs_rq);
3998 * Preempt the current task with a newly woken task if needed:
4001 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4003 unsigned long ideal_runtime, delta_exec;
4004 struct sched_entity *se;
4007 ideal_runtime = sched_slice(cfs_rq, curr);
4008 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4009 if (delta_exec > ideal_runtime) {
4010 resched_curr(rq_of(cfs_rq));
4012 * The current task ran long enough, ensure it doesn't get
4013 * re-elected due to buddy favours.
4015 clear_buddies(cfs_rq, curr);
4020 * Ensure that a task that missed wakeup preemption by a
4021 * narrow margin doesn't have to wait for a full slice.
4022 * This also mitigates buddy induced latencies under load.
4024 if (delta_exec < sysctl_sched_min_granularity)
4027 se = __pick_first_entity(cfs_rq);
4028 delta = curr->vruntime - se->vruntime;
4033 if (delta > ideal_runtime)
4034 resched_curr(rq_of(cfs_rq));
4038 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4040 /* 'current' is not kept within the tree. */
4043 * Any task has to be enqueued before it get to execute on
4044 * a CPU. So account for the time it spent waiting on the
4047 update_stats_wait_end(cfs_rq, se);
4048 __dequeue_entity(cfs_rq, se);
4049 update_load_avg(cfs_rq, se, UPDATE_TG);
4052 update_stats_curr_start(cfs_rq, se);
4056 * Track our maximum slice length, if the CPU's load is at
4057 * least twice that of our own weight (i.e. dont track it
4058 * when there are only lesser-weight tasks around):
4060 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
4061 schedstat_set(se->statistics.slice_max,
4062 max((u64)schedstat_val(se->statistics.slice_max),
4063 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4066 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4070 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4073 * Pick the next process, keeping these things in mind, in this order:
4074 * 1) keep things fair between processes/task groups
4075 * 2) pick the "next" process, since someone really wants that to run
4076 * 3) pick the "last" process, for cache locality
4077 * 4) do not run the "skip" process, if something else is available
4079 static struct sched_entity *
4080 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4082 struct sched_entity *left = __pick_first_entity(cfs_rq);
4083 struct sched_entity *se;
4086 * If curr is set we have to see if its left of the leftmost entity
4087 * still in the tree, provided there was anything in the tree at all.
4089 if (!left || (curr && entity_before(curr, left)))
4092 se = left; /* ideally we run the leftmost entity */
4095 * Avoid running the skip buddy, if running something else can
4096 * be done without getting too unfair.
4098 if (cfs_rq->skip == se) {
4099 struct sched_entity *second;
4102 second = __pick_first_entity(cfs_rq);
4104 second = __pick_next_entity(se);
4105 if (!second || (curr && entity_before(curr, second)))
4109 if (second && wakeup_preempt_entity(second, left) < 1)
4114 * Prefer last buddy, try to return the CPU to a preempted task.
4116 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4120 * Someone really wants this to run. If it's not unfair, run it.
4122 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4125 clear_buddies(cfs_rq, se);
4130 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4132 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4135 * If still on the runqueue then deactivate_task()
4136 * was not called and update_curr() has to be done:
4139 update_curr(cfs_rq);
4141 /* throttle cfs_rqs exceeding runtime */
4142 check_cfs_rq_runtime(cfs_rq);
4144 check_spread(cfs_rq, prev);
4147 update_stats_wait_start(cfs_rq, prev);
4148 /* Put 'current' back into the tree. */
4149 __enqueue_entity(cfs_rq, prev);
4150 /* in !on_rq case, update occurred at dequeue */
4151 update_load_avg(cfs_rq, prev, 0);
4153 cfs_rq->curr = NULL;
4157 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4160 * Update run-time statistics of the 'current'.
4162 update_curr(cfs_rq);
4165 * Ensure that runnable average is periodically updated.
4167 update_load_avg(cfs_rq, curr, UPDATE_TG);
4168 update_cfs_group(curr);
4170 #ifdef CONFIG_SCHED_HRTICK
4172 * queued ticks are scheduled to match the slice, so don't bother
4173 * validating it and just reschedule.
4176 resched_curr(rq_of(cfs_rq));
4180 * don't let the period tick interfere with the hrtick preemption
4182 if (!sched_feat(DOUBLE_TICK) &&
4183 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4187 if (cfs_rq->nr_running > 1)
4188 check_preempt_tick(cfs_rq, curr);
4192 /**************************************************
4193 * CFS bandwidth control machinery
4196 #ifdef CONFIG_CFS_BANDWIDTH
4198 #ifdef HAVE_JUMP_LABEL
4199 static struct static_key __cfs_bandwidth_used;
4201 static inline bool cfs_bandwidth_used(void)
4203 return static_key_false(&__cfs_bandwidth_used);
4206 void cfs_bandwidth_usage_inc(void)
4208 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4211 void cfs_bandwidth_usage_dec(void)
4213 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4215 #else /* HAVE_JUMP_LABEL */
4216 static bool cfs_bandwidth_used(void)
4221 void cfs_bandwidth_usage_inc(void) {}
4222 void cfs_bandwidth_usage_dec(void) {}
4223 #endif /* HAVE_JUMP_LABEL */
4226 * default period for cfs group bandwidth.
4227 * default: 0.1s, units: nanoseconds
4229 static inline u64 default_cfs_period(void)
4231 return 100000000ULL;
4234 static inline u64 sched_cfs_bandwidth_slice(void)
4236 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4240 * Replenish runtime according to assigned quota and update expiration time.
4241 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4242 * additional synchronization around rq->lock.
4244 * requires cfs_b->lock
4246 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4250 if (cfs_b->quota == RUNTIME_INF)
4253 now = sched_clock_cpu(smp_processor_id());
4254 cfs_b->runtime = cfs_b->quota;
4255 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4256 cfs_b->expires_seq++;
4259 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4261 return &tg->cfs_bandwidth;
4264 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4265 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4267 if (unlikely(cfs_rq->throttle_count))
4268 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4270 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4273 /* returns 0 on failure to allocate runtime */
4274 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4276 struct task_group *tg = cfs_rq->tg;
4277 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4278 u64 amount = 0, min_amount, expires;
4281 /* note: this is a positive sum as runtime_remaining <= 0 */
4282 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4284 raw_spin_lock(&cfs_b->lock);
4285 if (cfs_b->quota == RUNTIME_INF)
4286 amount = min_amount;
4288 start_cfs_bandwidth(cfs_b);
4290 if (cfs_b->runtime > 0) {
4291 amount = min(cfs_b->runtime, min_amount);
4292 cfs_b->runtime -= amount;
4296 expires_seq = cfs_b->expires_seq;
4297 expires = cfs_b->runtime_expires;
4298 raw_spin_unlock(&cfs_b->lock);
4300 cfs_rq->runtime_remaining += amount;
4302 * we may have advanced our local expiration to account for allowed
4303 * spread between our sched_clock and the one on which runtime was
4306 if (cfs_rq->expires_seq != expires_seq) {
4307 cfs_rq->expires_seq = expires_seq;
4308 cfs_rq->runtime_expires = expires;
4311 return cfs_rq->runtime_remaining > 0;
4315 * Note: This depends on the synchronization provided by sched_clock and the
4316 * fact that rq->clock snapshots this value.
4318 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4320 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4322 /* if the deadline is ahead of our clock, nothing to do */
4323 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4326 if (cfs_rq->runtime_remaining < 0)
4330 * If the local deadline has passed we have to consider the
4331 * possibility that our sched_clock is 'fast' and the global deadline
4332 * has not truly expired.
4334 * Fortunately we can check determine whether this the case by checking
4335 * whether the global deadline(cfs_b->expires_seq) has advanced.
4337 if (cfs_rq->expires_seq == cfs_b->expires_seq) {
4338 /* extend local deadline, drift is bounded above by 2 ticks */
4339 cfs_rq->runtime_expires += TICK_NSEC;
4341 /* global deadline is ahead, expiration has passed */
4342 cfs_rq->runtime_remaining = 0;
4346 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4348 /* dock delta_exec before expiring quota (as it could span periods) */
4349 cfs_rq->runtime_remaining -= delta_exec;
4350 expire_cfs_rq_runtime(cfs_rq);
4352 if (likely(cfs_rq->runtime_remaining > 0))
4356 * if we're unable to extend our runtime we resched so that the active
4357 * hierarchy can be throttled
4359 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4360 resched_curr(rq_of(cfs_rq));
4363 static __always_inline
4364 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4366 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4369 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4372 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4374 return cfs_bandwidth_used() && cfs_rq->throttled;
4377 /* check whether cfs_rq, or any parent, is throttled */
4378 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4380 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4384 * Ensure that neither of the group entities corresponding to src_cpu or
4385 * dest_cpu are members of a throttled hierarchy when performing group
4386 * load-balance operations.
4388 static inline int throttled_lb_pair(struct task_group *tg,
4389 int src_cpu, int dest_cpu)
4391 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4393 src_cfs_rq = tg->cfs_rq[src_cpu];
4394 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4396 return throttled_hierarchy(src_cfs_rq) ||
4397 throttled_hierarchy(dest_cfs_rq);
4400 static int tg_unthrottle_up(struct task_group *tg, void *data)
4402 struct rq *rq = data;
4403 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4405 cfs_rq->throttle_count--;
4406 if (!cfs_rq->throttle_count) {
4407 /* adjust cfs_rq_clock_task() */
4408 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4409 cfs_rq->throttled_clock_task;
4415 static int tg_throttle_down(struct task_group *tg, void *data)
4417 struct rq *rq = data;
4418 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4420 /* group is entering throttled state, stop time */
4421 if (!cfs_rq->throttle_count)
4422 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4423 cfs_rq->throttle_count++;
4428 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4430 struct rq *rq = rq_of(cfs_rq);
4431 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4432 struct sched_entity *se;
4433 long task_delta, dequeue = 1;
4436 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4438 /* freeze hierarchy runnable averages while throttled */
4440 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4443 task_delta = cfs_rq->h_nr_running;
4444 for_each_sched_entity(se) {
4445 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4446 /* throttled entity or throttle-on-deactivate */
4451 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4452 qcfs_rq->h_nr_running -= task_delta;
4454 if (qcfs_rq->load.weight)
4459 sub_nr_running(rq, task_delta);
4461 cfs_rq->throttled = 1;
4462 cfs_rq->throttled_clock = rq_clock(rq);
4463 raw_spin_lock(&cfs_b->lock);
4464 empty = list_empty(&cfs_b->throttled_cfs_rq);
4467 * Add to the _head_ of the list, so that an already-started
4468 * distribute_cfs_runtime will not see us
4470 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4473 * If we're the first throttled task, make sure the bandwidth
4477 start_cfs_bandwidth(cfs_b);
4479 raw_spin_unlock(&cfs_b->lock);
4482 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4484 struct rq *rq = rq_of(cfs_rq);
4485 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4486 struct sched_entity *se;
4490 se = cfs_rq->tg->se[cpu_of(rq)];
4492 cfs_rq->throttled = 0;
4494 update_rq_clock(rq);
4496 raw_spin_lock(&cfs_b->lock);
4497 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4498 list_del_rcu(&cfs_rq->throttled_list);
4499 raw_spin_unlock(&cfs_b->lock);
4501 /* update hierarchical throttle state */
4502 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4504 if (!cfs_rq->load.weight)
4507 task_delta = cfs_rq->h_nr_running;
4508 for_each_sched_entity(se) {
4512 cfs_rq = cfs_rq_of(se);
4514 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4515 cfs_rq->h_nr_running += task_delta;
4517 if (cfs_rq_throttled(cfs_rq))
4522 add_nr_running(rq, task_delta);
4524 /* Determine whether we need to wake up potentially idle CPU: */
4525 if (rq->curr == rq->idle && rq->cfs.nr_running)
4529 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4530 u64 remaining, u64 expires)
4532 struct cfs_rq *cfs_rq;
4534 u64 starting_runtime = remaining;
4537 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4539 struct rq *rq = rq_of(cfs_rq);
4543 if (!cfs_rq_throttled(cfs_rq))
4546 runtime = -cfs_rq->runtime_remaining + 1;
4547 if (runtime > remaining)
4548 runtime = remaining;
4549 remaining -= runtime;
4551 cfs_rq->runtime_remaining += runtime;
4552 cfs_rq->runtime_expires = expires;
4554 /* we check whether we're throttled above */
4555 if (cfs_rq->runtime_remaining > 0)
4556 unthrottle_cfs_rq(cfs_rq);
4566 return starting_runtime - remaining;
4570 * Responsible for refilling a task_group's bandwidth and unthrottling its
4571 * cfs_rqs as appropriate. If there has been no activity within the last
4572 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4573 * used to track this state.
4575 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4577 u64 runtime, runtime_expires;
4580 /* no need to continue the timer with no bandwidth constraint */
4581 if (cfs_b->quota == RUNTIME_INF)
4582 goto out_deactivate;
4584 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4585 cfs_b->nr_periods += overrun;
4588 * idle depends on !throttled (for the case of a large deficit), and if
4589 * we're going inactive then everything else can be deferred
4591 if (cfs_b->idle && !throttled)
4592 goto out_deactivate;
4594 __refill_cfs_bandwidth_runtime(cfs_b);
4597 /* mark as potentially idle for the upcoming period */
4602 /* account preceding periods in which throttling occurred */
4603 cfs_b->nr_throttled += overrun;
4605 runtime_expires = cfs_b->runtime_expires;
4608 * This check is repeated as we are holding onto the new bandwidth while
4609 * we unthrottle. This can potentially race with an unthrottled group
4610 * trying to acquire new bandwidth from the global pool. This can result
4611 * in us over-using our runtime if it is all used during this loop, but
4612 * only by limited amounts in that extreme case.
4614 while (throttled && cfs_b->runtime > 0) {
4615 runtime = cfs_b->runtime;
4616 raw_spin_unlock(&cfs_b->lock);
4617 /* we can't nest cfs_b->lock while distributing bandwidth */
4618 runtime = distribute_cfs_runtime(cfs_b, runtime,
4620 raw_spin_lock(&cfs_b->lock);
4622 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4624 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4628 * While we are ensured activity in the period following an
4629 * unthrottle, this also covers the case in which the new bandwidth is
4630 * insufficient to cover the existing bandwidth deficit. (Forcing the
4631 * timer to remain active while there are any throttled entities.)
4641 /* a cfs_rq won't donate quota below this amount */
4642 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4643 /* minimum remaining period time to redistribute slack quota */
4644 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4645 /* how long we wait to gather additional slack before distributing */
4646 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4649 * Are we near the end of the current quota period?
4651 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4652 * hrtimer base being cleared by hrtimer_start. In the case of
4653 * migrate_hrtimers, base is never cleared, so we are fine.
4655 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4657 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4660 /* if the call-back is running a quota refresh is already occurring */
4661 if (hrtimer_callback_running(refresh_timer))
4664 /* is a quota refresh about to occur? */
4665 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4666 if (remaining < min_expire)
4672 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4674 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4676 /* if there's a quota refresh soon don't bother with slack */
4677 if (runtime_refresh_within(cfs_b, min_left))
4680 hrtimer_start(&cfs_b->slack_timer,
4681 ns_to_ktime(cfs_bandwidth_slack_period),
4685 /* we know any runtime found here is valid as update_curr() precedes return */
4686 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4688 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4689 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4691 if (slack_runtime <= 0)
4694 raw_spin_lock(&cfs_b->lock);
4695 if (cfs_b->quota != RUNTIME_INF &&
4696 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4697 cfs_b->runtime += slack_runtime;
4699 /* we are under rq->lock, defer unthrottling using a timer */
4700 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4701 !list_empty(&cfs_b->throttled_cfs_rq))
4702 start_cfs_slack_bandwidth(cfs_b);
4704 raw_spin_unlock(&cfs_b->lock);
4706 /* even if it's not valid for return we don't want to try again */
4707 cfs_rq->runtime_remaining -= slack_runtime;
4710 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4712 if (!cfs_bandwidth_used())
4715 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4718 __return_cfs_rq_runtime(cfs_rq);
4722 * This is done with a timer (instead of inline with bandwidth return) since
4723 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4725 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4727 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4730 /* confirm we're still not at a refresh boundary */
4731 raw_spin_lock(&cfs_b->lock);
4732 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4733 raw_spin_unlock(&cfs_b->lock);
4737 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4738 runtime = cfs_b->runtime;
4740 expires = cfs_b->runtime_expires;
4741 raw_spin_unlock(&cfs_b->lock);
4746 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4748 raw_spin_lock(&cfs_b->lock);
4749 if (expires == cfs_b->runtime_expires)
4750 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4751 raw_spin_unlock(&cfs_b->lock);
4755 * When a group wakes up we want to make sure that its quota is not already
4756 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4757 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4759 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4761 if (!cfs_bandwidth_used())
4764 /* an active group must be handled by the update_curr()->put() path */
4765 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4768 /* ensure the group is not already throttled */
4769 if (cfs_rq_throttled(cfs_rq))
4772 /* update runtime allocation */
4773 account_cfs_rq_runtime(cfs_rq, 0);
4774 if (cfs_rq->runtime_remaining <= 0)
4775 throttle_cfs_rq(cfs_rq);
4778 static void sync_throttle(struct task_group *tg, int cpu)
4780 struct cfs_rq *pcfs_rq, *cfs_rq;
4782 if (!cfs_bandwidth_used())
4788 cfs_rq = tg->cfs_rq[cpu];
4789 pcfs_rq = tg->parent->cfs_rq[cpu];
4791 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4792 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4795 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4796 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4798 if (!cfs_bandwidth_used())
4801 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4805 * it's possible for a throttled entity to be forced into a running
4806 * state (e.g. set_curr_task), in this case we're finished.
4808 if (cfs_rq_throttled(cfs_rq))
4811 throttle_cfs_rq(cfs_rq);
4815 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4817 struct cfs_bandwidth *cfs_b =
4818 container_of(timer, struct cfs_bandwidth, slack_timer);
4820 do_sched_cfs_slack_timer(cfs_b);
4822 return HRTIMER_NORESTART;
4825 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4827 struct cfs_bandwidth *cfs_b =
4828 container_of(timer, struct cfs_bandwidth, period_timer);
4832 raw_spin_lock(&cfs_b->lock);
4834 overrun = hrtimer_forward_now(timer, cfs_b->period);
4838 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4841 cfs_b->period_active = 0;
4842 raw_spin_unlock(&cfs_b->lock);
4844 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4847 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4849 raw_spin_lock_init(&cfs_b->lock);
4851 cfs_b->quota = RUNTIME_INF;
4852 cfs_b->period = ns_to_ktime(default_cfs_period());
4854 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4855 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4856 cfs_b->period_timer.function = sched_cfs_period_timer;
4857 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4858 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4861 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4863 cfs_rq->runtime_enabled = 0;
4864 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4867 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4871 lockdep_assert_held(&cfs_b->lock);
4873 if (cfs_b->period_active)
4876 cfs_b->period_active = 1;
4877 overrun = hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4878 cfs_b->runtime_expires += (overrun + 1) * ktime_to_ns(cfs_b->period);
4879 cfs_b->expires_seq++;
4880 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4883 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4885 /* init_cfs_bandwidth() was not called */
4886 if (!cfs_b->throttled_cfs_rq.next)
4889 hrtimer_cancel(&cfs_b->period_timer);
4890 hrtimer_cancel(&cfs_b->slack_timer);
4894 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
4896 * The race is harmless, since modifying bandwidth settings of unhooked group
4897 * bits doesn't do much.
4900 /* cpu online calback */
4901 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4903 struct task_group *tg;
4905 lockdep_assert_held(&rq->lock);
4908 list_for_each_entry_rcu(tg, &task_groups, list) {
4909 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
4910 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4912 raw_spin_lock(&cfs_b->lock);
4913 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4914 raw_spin_unlock(&cfs_b->lock);
4919 /* cpu offline callback */
4920 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4922 struct task_group *tg;
4924 lockdep_assert_held(&rq->lock);
4927 list_for_each_entry_rcu(tg, &task_groups, list) {
4928 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4930 if (!cfs_rq->runtime_enabled)
4934 * clock_task is not advancing so we just need to make sure
4935 * there's some valid quota amount
4937 cfs_rq->runtime_remaining = 1;
4939 * Offline rq is schedulable till CPU is completely disabled
4940 * in take_cpu_down(), so we prevent new cfs throttling here.
4942 cfs_rq->runtime_enabled = 0;
4944 if (cfs_rq_throttled(cfs_rq))
4945 unthrottle_cfs_rq(cfs_rq);
4950 #else /* CONFIG_CFS_BANDWIDTH */
4951 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4953 return rq_clock_task(rq_of(cfs_rq));
4956 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4957 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4958 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4959 static inline void sync_throttle(struct task_group *tg, int cpu) {}
4960 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4962 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4967 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4972 static inline int throttled_lb_pair(struct task_group *tg,
4973 int src_cpu, int dest_cpu)
4978 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4980 #ifdef CONFIG_FAIR_GROUP_SCHED
4981 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4984 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4988 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4989 static inline void update_runtime_enabled(struct rq *rq) {}
4990 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4992 #endif /* CONFIG_CFS_BANDWIDTH */
4994 /**************************************************
4995 * CFS operations on tasks:
4998 #ifdef CONFIG_SCHED_HRTICK
4999 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5001 struct sched_entity *se = &p->se;
5002 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5004 SCHED_WARN_ON(task_rq(p) != rq);
5006 if (rq->cfs.h_nr_running > 1) {
5007 u64 slice = sched_slice(cfs_rq, se);
5008 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5009 s64 delta = slice - ran;
5016 hrtick_start(rq, delta);
5021 * called from enqueue/dequeue and updates the hrtick when the
5022 * current task is from our class and nr_running is low enough
5025 static void hrtick_update(struct rq *rq)
5027 struct task_struct *curr = rq->curr;
5029 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5032 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5033 hrtick_start_fair(rq, curr);
5035 #else /* !CONFIG_SCHED_HRTICK */
5037 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5041 static inline void hrtick_update(struct rq *rq)
5047 * The enqueue_task method is called before nr_running is
5048 * increased. Here we update the fair scheduling stats and
5049 * then put the task into the rbtree:
5052 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5054 struct cfs_rq *cfs_rq;
5055 struct sched_entity *se = &p->se;
5058 * The code below (indirectly) updates schedutil which looks at
5059 * the cfs_rq utilization to select a frequency.
5060 * Let's add the task's estimated utilization to the cfs_rq's
5061 * estimated utilization, before we update schedutil.
5063 util_est_enqueue(&rq->cfs, p);
5066 * If in_iowait is set, the code below may not trigger any cpufreq
5067 * utilization updates, so do it here explicitly with the IOWAIT flag
5071 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5073 for_each_sched_entity(se) {
5076 cfs_rq = cfs_rq_of(se);
5077 enqueue_entity(cfs_rq, se, flags);
5080 * end evaluation on encountering a throttled cfs_rq
5082 * note: in the case of encountering a throttled cfs_rq we will
5083 * post the final h_nr_running increment below.
5085 if (cfs_rq_throttled(cfs_rq))
5087 cfs_rq->h_nr_running++;
5089 flags = ENQUEUE_WAKEUP;
5092 for_each_sched_entity(se) {
5093 cfs_rq = cfs_rq_of(se);
5094 cfs_rq->h_nr_running++;
5096 if (cfs_rq_throttled(cfs_rq))
5099 update_load_avg(cfs_rq, se, UPDATE_TG);
5100 update_cfs_group(se);
5104 add_nr_running(rq, 1);
5109 static void set_next_buddy(struct sched_entity *se);
5112 * The dequeue_task method is called before nr_running is
5113 * decreased. We remove the task from the rbtree and
5114 * update the fair scheduling stats:
5116 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5118 struct cfs_rq *cfs_rq;
5119 struct sched_entity *se = &p->se;
5120 int task_sleep = flags & DEQUEUE_SLEEP;
5122 for_each_sched_entity(se) {
5123 cfs_rq = cfs_rq_of(se);
5124 dequeue_entity(cfs_rq, se, flags);
5127 * end evaluation on encountering a throttled cfs_rq
5129 * note: in the case of encountering a throttled cfs_rq we will
5130 * post the final h_nr_running decrement below.
5132 if (cfs_rq_throttled(cfs_rq))
5134 cfs_rq->h_nr_running--;
5136 /* Don't dequeue parent if it has other entities besides us */
5137 if (cfs_rq->load.weight) {
5138 /* Avoid re-evaluating load for this entity: */
5139 se = parent_entity(se);
5141 * Bias pick_next to pick a task from this cfs_rq, as
5142 * p is sleeping when it is within its sched_slice.
5144 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5148 flags |= DEQUEUE_SLEEP;
5151 for_each_sched_entity(se) {
5152 cfs_rq = cfs_rq_of(se);
5153 cfs_rq->h_nr_running--;
5155 if (cfs_rq_throttled(cfs_rq))
5158 update_load_avg(cfs_rq, se, UPDATE_TG);
5159 update_cfs_group(se);
5163 sub_nr_running(rq, 1);
5165 util_est_dequeue(&rq->cfs, p, task_sleep);
5171 /* Working cpumask for: load_balance, load_balance_newidle. */
5172 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5173 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5175 #ifdef CONFIG_NO_HZ_COMMON
5177 * per rq 'load' arrray crap; XXX kill this.
5181 * The exact cpuload calculated at every tick would be:
5183 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5185 * If a CPU misses updates for n ticks (as it was idle) and update gets
5186 * called on the n+1-th tick when CPU may be busy, then we have:
5188 * load_n = (1 - 1/2^i)^n * load_0
5189 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
5191 * decay_load_missed() below does efficient calculation of
5193 * load' = (1 - 1/2^i)^n * load
5195 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5196 * This allows us to precompute the above in said factors, thereby allowing the
5197 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5198 * fixed_power_int())
5200 * The calculation is approximated on a 128 point scale.
5202 #define DEGRADE_SHIFT 7
5204 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5205 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5206 { 0, 0, 0, 0, 0, 0, 0, 0 },
5207 { 64, 32, 8, 0, 0, 0, 0, 0 },
5208 { 96, 72, 40, 12, 1, 0, 0, 0 },
5209 { 112, 98, 75, 43, 15, 1, 0, 0 },
5210 { 120, 112, 98, 76, 45, 16, 2, 0 }
5214 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5215 * would be when CPU is idle and so we just decay the old load without
5216 * adding any new load.
5218 static unsigned long
5219 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5223 if (!missed_updates)
5226 if (missed_updates >= degrade_zero_ticks[idx])
5230 return load >> missed_updates;
5232 while (missed_updates) {
5233 if (missed_updates % 2)
5234 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5236 missed_updates >>= 1;
5243 cpumask_var_t idle_cpus_mask;
5245 int has_blocked; /* Idle CPUS has blocked load */
5246 unsigned long next_balance; /* in jiffy units */
5247 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5248 } nohz ____cacheline_aligned;
5250 #endif /* CONFIG_NO_HZ_COMMON */
5253 * __cpu_load_update - update the rq->cpu_load[] statistics
5254 * @this_rq: The rq to update statistics for
5255 * @this_load: The current load
5256 * @pending_updates: The number of missed updates
5258 * Update rq->cpu_load[] statistics. This function is usually called every
5259 * scheduler tick (TICK_NSEC).
5261 * This function computes a decaying average:
5263 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5265 * Because of NOHZ it might not get called on every tick which gives need for
5266 * the @pending_updates argument.
5268 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5269 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5270 * = A * (A * load[i]_n-2 + B) + B
5271 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5272 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5273 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5274 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5275 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5277 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5278 * any change in load would have resulted in the tick being turned back on.
5280 * For regular NOHZ, this reduces to:
5282 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5284 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5287 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5288 unsigned long pending_updates)
5290 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5293 this_rq->nr_load_updates++;
5295 /* Update our load: */
5296 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5297 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5298 unsigned long old_load, new_load;
5300 /* scale is effectively 1 << i now, and >> i divides by scale */
5302 old_load = this_rq->cpu_load[i];
5303 #ifdef CONFIG_NO_HZ_COMMON
5304 old_load = decay_load_missed(old_load, pending_updates - 1, i);
5305 if (tickless_load) {
5306 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5308 * old_load can never be a negative value because a
5309 * decayed tickless_load cannot be greater than the
5310 * original tickless_load.
5312 old_load += tickless_load;
5315 new_load = this_load;
5317 * Round up the averaging division if load is increasing. This
5318 * prevents us from getting stuck on 9 if the load is 10, for
5321 if (new_load > old_load)
5322 new_load += scale - 1;
5324 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5327 sched_avg_update(this_rq);
5330 /* Used instead of source_load when we know the type == 0 */
5331 static unsigned long weighted_cpuload(struct rq *rq)
5333 return cfs_rq_runnable_load_avg(&rq->cfs);
5336 #ifdef CONFIG_NO_HZ_COMMON
5338 * There is no sane way to deal with nohz on smp when using jiffies because the
5339 * CPU doing the jiffies update might drift wrt the CPU doing the jiffy reading
5340 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5342 * Therefore we need to avoid the delta approach from the regular tick when
5343 * possible since that would seriously skew the load calculation. This is why we
5344 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5345 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5346 * loop exit, nohz_idle_balance, nohz full exit...)
5348 * This means we might still be one tick off for nohz periods.
5351 static void cpu_load_update_nohz(struct rq *this_rq,
5352 unsigned long curr_jiffies,
5355 unsigned long pending_updates;
5357 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5358 if (pending_updates) {
5359 this_rq->last_load_update_tick = curr_jiffies;
5361 * In the regular NOHZ case, we were idle, this means load 0.
5362 * In the NOHZ_FULL case, we were non-idle, we should consider
5363 * its weighted load.
5365 cpu_load_update(this_rq, load, pending_updates);
5370 * Called from nohz_idle_balance() to update the load ratings before doing the
5373 static void cpu_load_update_idle(struct rq *this_rq)
5376 * bail if there's load or we're actually up-to-date.
5378 if (weighted_cpuload(this_rq))
5381 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5385 * Record CPU load on nohz entry so we know the tickless load to account
5386 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5387 * than other cpu_load[idx] but it should be fine as cpu_load readers
5388 * shouldn't rely into synchronized cpu_load[*] updates.
5390 void cpu_load_update_nohz_start(void)
5392 struct rq *this_rq = this_rq();
5395 * This is all lockless but should be fine. If weighted_cpuload changes
5396 * concurrently we'll exit nohz. And cpu_load write can race with
5397 * cpu_load_update_idle() but both updater would be writing the same.
5399 this_rq->cpu_load[0] = weighted_cpuload(this_rq);
5403 * Account the tickless load in the end of a nohz frame.
5405 void cpu_load_update_nohz_stop(void)
5407 unsigned long curr_jiffies = READ_ONCE(jiffies);
5408 struct rq *this_rq = this_rq();
5412 if (curr_jiffies == this_rq->last_load_update_tick)
5415 load = weighted_cpuload(this_rq);
5416 rq_lock(this_rq, &rf);
5417 update_rq_clock(this_rq);
5418 cpu_load_update_nohz(this_rq, curr_jiffies, load);
5419 rq_unlock(this_rq, &rf);
5421 #else /* !CONFIG_NO_HZ_COMMON */
5422 static inline void cpu_load_update_nohz(struct rq *this_rq,
5423 unsigned long curr_jiffies,
5424 unsigned long load) { }
5425 #endif /* CONFIG_NO_HZ_COMMON */
5427 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5429 #ifdef CONFIG_NO_HZ_COMMON
5430 /* See the mess around cpu_load_update_nohz(). */
5431 this_rq->last_load_update_tick = READ_ONCE(jiffies);
5433 cpu_load_update(this_rq, load, 1);
5437 * Called from scheduler_tick()
5439 void cpu_load_update_active(struct rq *this_rq)
5441 unsigned long load = weighted_cpuload(this_rq);
5443 if (tick_nohz_tick_stopped())
5444 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5446 cpu_load_update_periodic(this_rq, load);
5450 * Return a low guess at the load of a migration-source CPU weighted
5451 * according to the scheduling class and "nice" value.
5453 * We want to under-estimate the load of migration sources, to
5454 * balance conservatively.
5456 static unsigned long source_load(int cpu, int type)
5458 struct rq *rq = cpu_rq(cpu);
5459 unsigned long total = weighted_cpuload(rq);
5461 if (type == 0 || !sched_feat(LB_BIAS))
5464 return min(rq->cpu_load[type-1], total);
5468 * Return a high guess at the load of a migration-target CPU weighted
5469 * according to the scheduling class and "nice" value.
5471 static unsigned long target_load(int cpu, int type)
5473 struct rq *rq = cpu_rq(cpu);
5474 unsigned long total = weighted_cpuload(rq);
5476 if (type == 0 || !sched_feat(LB_BIAS))
5479 return max(rq->cpu_load[type-1], total);
5482 static unsigned long capacity_of(int cpu)
5484 return cpu_rq(cpu)->cpu_capacity;
5487 static unsigned long capacity_orig_of(int cpu)
5489 return cpu_rq(cpu)->cpu_capacity_orig;
5492 static unsigned long cpu_avg_load_per_task(int cpu)
5494 struct rq *rq = cpu_rq(cpu);
5495 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5496 unsigned long load_avg = weighted_cpuload(rq);
5499 return load_avg / nr_running;
5504 static void record_wakee(struct task_struct *p)
5507 * Only decay a single time; tasks that have less then 1 wakeup per
5508 * jiffy will not have built up many flips.
5510 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5511 current->wakee_flips >>= 1;
5512 current->wakee_flip_decay_ts = jiffies;
5515 if (current->last_wakee != p) {
5516 current->last_wakee = p;
5517 current->wakee_flips++;
5522 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5524 * A waker of many should wake a different task than the one last awakened
5525 * at a frequency roughly N times higher than one of its wakees.
5527 * In order to determine whether we should let the load spread vs consolidating
5528 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5529 * partner, and a factor of lls_size higher frequency in the other.
5531 * With both conditions met, we can be relatively sure that the relationship is
5532 * non-monogamous, with partner count exceeding socket size.
5534 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5535 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5538 static int wake_wide(struct task_struct *p)
5540 unsigned int master = current->wakee_flips;
5541 unsigned int slave = p->wakee_flips;
5542 int factor = this_cpu_read(sd_llc_size);
5545 swap(master, slave);
5546 if (slave < factor || master < slave * factor)
5552 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5553 * soonest. For the purpose of speed we only consider the waking and previous
5556 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5557 * cache-affine and is (or will be) idle.
5559 * wake_affine_weight() - considers the weight to reflect the average
5560 * scheduling latency of the CPUs. This seems to work
5561 * for the overloaded case.
5564 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5567 * If this_cpu is idle, it implies the wakeup is from interrupt
5568 * context. Only allow the move if cache is shared. Otherwise an
5569 * interrupt intensive workload could force all tasks onto one
5570 * node depending on the IO topology or IRQ affinity settings.
5572 * If the prev_cpu is idle and cache affine then avoid a migration.
5573 * There is no guarantee that the cache hot data from an interrupt
5574 * is more important than cache hot data on the prev_cpu and from
5575 * a cpufreq perspective, it's better to have higher utilisation
5578 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5579 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5581 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5584 return nr_cpumask_bits;
5588 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5589 int this_cpu, int prev_cpu, int sync)
5591 s64 this_eff_load, prev_eff_load;
5592 unsigned long task_load;
5594 this_eff_load = target_load(this_cpu, sd->wake_idx);
5597 unsigned long current_load = task_h_load(current);
5599 if (current_load > this_eff_load)
5602 this_eff_load -= current_load;
5605 task_load = task_h_load(p);
5607 this_eff_load += task_load;
5608 if (sched_feat(WA_BIAS))
5609 this_eff_load *= 100;
5610 this_eff_load *= capacity_of(prev_cpu);
5612 prev_eff_load = source_load(prev_cpu, sd->wake_idx);
5613 prev_eff_load -= task_load;
5614 if (sched_feat(WA_BIAS))
5615 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5616 prev_eff_load *= capacity_of(this_cpu);
5619 * If sync, adjust the weight of prev_eff_load such that if
5620 * prev_eff == this_eff that select_idle_sibling() will consider
5621 * stacking the wakee on top of the waker if no other CPU is
5627 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5630 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5631 int this_cpu, int prev_cpu, int sync)
5633 int target = nr_cpumask_bits;
5635 if (sched_feat(WA_IDLE))
5636 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5638 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5639 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5641 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5642 if (target == nr_cpumask_bits)
5645 schedstat_inc(sd->ttwu_move_affine);
5646 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5650 static unsigned long cpu_util_wake(int cpu, struct task_struct *p);
5652 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5654 return max_t(long, capacity_of(cpu) - cpu_util_wake(cpu, p), 0);
5658 * find_idlest_group finds and returns the least busy CPU group within the
5661 * Assumes p is allowed on at least one CPU in sd.
5663 static struct sched_group *
5664 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5665 int this_cpu, int sd_flag)
5667 struct sched_group *idlest = NULL, *group = sd->groups;
5668 struct sched_group *most_spare_sg = NULL;
5669 unsigned long min_runnable_load = ULONG_MAX;
5670 unsigned long this_runnable_load = ULONG_MAX;
5671 unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5672 unsigned long most_spare = 0, this_spare = 0;
5673 int load_idx = sd->forkexec_idx;
5674 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5675 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5676 (sd->imbalance_pct-100) / 100;
5678 if (sd_flag & SD_BALANCE_WAKE)
5679 load_idx = sd->wake_idx;
5682 unsigned long load, avg_load, runnable_load;
5683 unsigned long spare_cap, max_spare_cap;
5687 /* Skip over this group if it has no CPUs allowed */
5688 if (!cpumask_intersects(sched_group_span(group),
5692 local_group = cpumask_test_cpu(this_cpu,
5693 sched_group_span(group));
5696 * Tally up the load of all CPUs in the group and find
5697 * the group containing the CPU with most spare capacity.
5703 for_each_cpu(i, sched_group_span(group)) {
5704 /* Bias balancing toward CPUs of our domain */
5706 load = source_load(i, load_idx);
5708 load = target_load(i, load_idx);
5710 runnable_load += load;
5712 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5714 spare_cap = capacity_spare_wake(i, p);
5716 if (spare_cap > max_spare_cap)
5717 max_spare_cap = spare_cap;
5720 /* Adjust by relative CPU capacity of the group */
5721 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5722 group->sgc->capacity;
5723 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5724 group->sgc->capacity;
5727 this_runnable_load = runnable_load;
5728 this_avg_load = avg_load;
5729 this_spare = max_spare_cap;
5731 if (min_runnable_load > (runnable_load + imbalance)) {
5733 * The runnable load is significantly smaller
5734 * so we can pick this new CPU:
5736 min_runnable_load = runnable_load;
5737 min_avg_load = avg_load;
5739 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5740 (100*min_avg_load > imbalance_scale*avg_load)) {
5742 * The runnable loads are close so take the
5743 * blocked load into account through avg_load:
5745 min_avg_load = avg_load;
5749 if (most_spare < max_spare_cap) {
5750 most_spare = max_spare_cap;
5751 most_spare_sg = group;
5754 } while (group = group->next, group != sd->groups);
5757 * The cross-over point between using spare capacity or least load
5758 * is too conservative for high utilization tasks on partially
5759 * utilized systems if we require spare_capacity > task_util(p),
5760 * so we allow for some task stuffing by using
5761 * spare_capacity > task_util(p)/2.
5763 * Spare capacity can't be used for fork because the utilization has
5764 * not been set yet, we must first select a rq to compute the initial
5767 if (sd_flag & SD_BALANCE_FORK)
5770 if (this_spare > task_util(p) / 2 &&
5771 imbalance_scale*this_spare > 100*most_spare)
5774 if (most_spare > task_util(p) / 2)
5775 return most_spare_sg;
5782 * When comparing groups across NUMA domains, it's possible for the
5783 * local domain to be very lightly loaded relative to the remote
5784 * domains but "imbalance" skews the comparison making remote CPUs
5785 * look much more favourable. When considering cross-domain, add
5786 * imbalance to the runnable load on the remote node and consider
5789 if ((sd->flags & SD_NUMA) &&
5790 min_runnable_load + imbalance >= this_runnable_load)
5793 if (min_runnable_load > (this_runnable_load + imbalance))
5796 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5797 (100*this_avg_load < imbalance_scale*min_avg_load))
5804 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5807 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5809 unsigned long load, min_load = ULONG_MAX;
5810 unsigned int min_exit_latency = UINT_MAX;
5811 u64 latest_idle_timestamp = 0;
5812 int least_loaded_cpu = this_cpu;
5813 int shallowest_idle_cpu = -1;
5816 /* Check if we have any choice: */
5817 if (group->group_weight == 1)
5818 return cpumask_first(sched_group_span(group));
5820 /* Traverse only the allowed CPUs */
5821 for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
5822 if (available_idle_cpu(i)) {
5823 struct rq *rq = cpu_rq(i);
5824 struct cpuidle_state *idle = idle_get_state(rq);
5825 if (idle && idle->exit_latency < min_exit_latency) {
5827 * We give priority to a CPU whose idle state
5828 * has the smallest exit latency irrespective
5829 * of any idle timestamp.
5831 min_exit_latency = idle->exit_latency;
5832 latest_idle_timestamp = rq->idle_stamp;
5833 shallowest_idle_cpu = i;
5834 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5835 rq->idle_stamp > latest_idle_timestamp) {
5837 * If equal or no active idle state, then
5838 * the most recently idled CPU might have
5841 latest_idle_timestamp = rq->idle_stamp;
5842 shallowest_idle_cpu = i;
5844 } else if (shallowest_idle_cpu == -1) {
5845 load = weighted_cpuload(cpu_rq(i));
5846 if (load < min_load) {
5848 least_loaded_cpu = i;
5853 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5856 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5857 int cpu, int prev_cpu, int sd_flag)
5861 if (!cpumask_intersects(sched_domain_span(sd), &p->cpus_allowed))
5865 * We need task's util for capacity_spare_wake, sync it up to prev_cpu's
5868 if (!(sd_flag & SD_BALANCE_FORK))
5869 sync_entity_load_avg(&p->se);
5872 struct sched_group *group;
5873 struct sched_domain *tmp;
5876 if (!(sd->flags & sd_flag)) {
5881 group = find_idlest_group(sd, p, cpu, sd_flag);
5887 new_cpu = find_idlest_group_cpu(group, p, cpu);
5888 if (new_cpu == cpu) {
5889 /* Now try balancing at a lower domain level of 'cpu': */
5894 /* Now try balancing at a lower domain level of 'new_cpu': */
5896 weight = sd->span_weight;
5898 for_each_domain(cpu, tmp) {
5899 if (weight <= tmp->span_weight)
5901 if (tmp->flags & sd_flag)
5909 #ifdef CONFIG_SCHED_SMT
5911 static inline void set_idle_cores(int cpu, int val)
5913 struct sched_domain_shared *sds;
5915 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5917 WRITE_ONCE(sds->has_idle_cores, val);
5920 static inline bool test_idle_cores(int cpu, bool def)
5922 struct sched_domain_shared *sds;
5924 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5926 return READ_ONCE(sds->has_idle_cores);
5932 * Scans the local SMT mask to see if the entire core is idle, and records this
5933 * information in sd_llc_shared->has_idle_cores.
5935 * Since SMT siblings share all cache levels, inspecting this limited remote
5936 * state should be fairly cheap.
5938 void __update_idle_core(struct rq *rq)
5940 int core = cpu_of(rq);
5944 if (test_idle_cores(core, true))
5947 for_each_cpu(cpu, cpu_smt_mask(core)) {
5951 if (!available_idle_cpu(cpu))
5955 set_idle_cores(core, 1);
5961 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5962 * there are no idle cores left in the system; tracked through
5963 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5965 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5967 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5970 if (!static_branch_likely(&sched_smt_present))
5973 if (!test_idle_cores(target, false))
5976 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
5978 for_each_cpu_wrap(core, cpus, target) {
5981 for_each_cpu(cpu, cpu_smt_mask(core)) {
5982 cpumask_clear_cpu(cpu, cpus);
5983 if (!available_idle_cpu(cpu))
5992 * Failed to find an idle core; stop looking for one.
5994 set_idle_cores(target, 0);
6000 * Scan the local SMT mask for idle CPUs.
6002 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6006 if (!static_branch_likely(&sched_smt_present))
6009 for_each_cpu(cpu, cpu_smt_mask(target)) {
6010 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6012 if (available_idle_cpu(cpu))
6019 #else /* CONFIG_SCHED_SMT */
6021 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6026 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6031 #endif /* CONFIG_SCHED_SMT */
6034 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6035 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6036 * average idle time for this rq (as found in rq->avg_idle).
6038 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
6040 struct sched_domain *this_sd;
6041 u64 avg_cost, avg_idle;
6044 int cpu, nr = INT_MAX;
6046 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6051 * Due to large variance we need a large fuzz factor; hackbench in
6052 * particularly is sensitive here.
6054 avg_idle = this_rq()->avg_idle / 512;
6055 avg_cost = this_sd->avg_scan_cost + 1;
6057 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
6060 if (sched_feat(SIS_PROP)) {
6061 u64 span_avg = sd->span_weight * avg_idle;
6062 if (span_avg > 4*avg_cost)
6063 nr = div_u64(span_avg, avg_cost);
6068 time = local_clock();
6070 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
6073 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6075 if (available_idle_cpu(cpu))
6079 time = local_clock() - time;
6080 cost = this_sd->avg_scan_cost;
6081 delta = (s64)(time - cost) / 8;
6082 this_sd->avg_scan_cost += delta;
6088 * Try and locate an idle core/thread in the LLC cache domain.
6090 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6092 struct sched_domain *sd;
6093 int i, recent_used_cpu;
6095 if (available_idle_cpu(target))
6099 * If the previous CPU is cache affine and idle, don't be stupid:
6101 if (prev != target && cpus_share_cache(prev, target) && available_idle_cpu(prev))
6104 /* Check a recently used CPU as a potential idle candidate: */
6105 recent_used_cpu = p->recent_used_cpu;
6106 if (recent_used_cpu != prev &&
6107 recent_used_cpu != target &&
6108 cpus_share_cache(recent_used_cpu, target) &&
6109 available_idle_cpu(recent_used_cpu) &&
6110 cpumask_test_cpu(p->recent_used_cpu, &p->cpus_allowed)) {
6112 * Replace recent_used_cpu with prev as it is a potential
6113 * candidate for the next wake:
6115 p->recent_used_cpu = prev;
6116 return recent_used_cpu;
6119 sd = rcu_dereference(per_cpu(sd_llc, target));
6123 i = select_idle_core(p, sd, target);
6124 if ((unsigned)i < nr_cpumask_bits)
6127 i = select_idle_cpu(p, sd, target);
6128 if ((unsigned)i < nr_cpumask_bits)
6131 i = select_idle_smt(p, sd, target);
6132 if ((unsigned)i < nr_cpumask_bits)
6139 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6140 * @cpu: the CPU to get the utilization of
6142 * The unit of the return value must be the one of capacity so we can compare
6143 * the utilization with the capacity of the CPU that is available for CFS task
6144 * (ie cpu_capacity).
6146 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6147 * recent utilization of currently non-runnable tasks on a CPU. It represents
6148 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6149 * capacity_orig is the cpu_capacity available at the highest frequency
6150 * (arch_scale_freq_capacity()).
6151 * The utilization of a CPU converges towards a sum equal to or less than the
6152 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6153 * the running time on this CPU scaled by capacity_curr.
6155 * The estimated utilization of a CPU is defined to be the maximum between its
6156 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6157 * currently RUNNABLE on that CPU.
6158 * This allows to properly represent the expected utilization of a CPU which
6159 * has just got a big task running since a long sleep period. At the same time
6160 * however it preserves the benefits of the "blocked utilization" in
6161 * describing the potential for other tasks waking up on the same CPU.
6163 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6164 * higher than capacity_orig because of unfortunate rounding in
6165 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6166 * the average stabilizes with the new running time. We need to check that the
6167 * utilization stays within the range of [0..capacity_orig] and cap it if
6168 * necessary. Without utilization capping, a group could be seen as overloaded
6169 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6170 * available capacity. We allow utilization to overshoot capacity_curr (but not
6171 * capacity_orig) as it useful for predicting the capacity required after task
6172 * migrations (scheduler-driven DVFS).
6174 * Return: the (estimated) utilization for the specified CPU
6176 static inline unsigned long cpu_util(int cpu)
6178 struct cfs_rq *cfs_rq;
6181 cfs_rq = &cpu_rq(cpu)->cfs;
6182 util = READ_ONCE(cfs_rq->avg.util_avg);
6184 if (sched_feat(UTIL_EST))
6185 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6187 return min_t(unsigned long, util, capacity_orig_of(cpu));
6191 * cpu_util_wake: Compute CPU utilization with any contributions from
6192 * the waking task p removed.
6194 static unsigned long cpu_util_wake(int cpu, struct task_struct *p)
6196 struct cfs_rq *cfs_rq;
6199 /* Task has no contribution or is new */
6200 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6201 return cpu_util(cpu);
6203 cfs_rq = &cpu_rq(cpu)->cfs;
6204 util = READ_ONCE(cfs_rq->avg.util_avg);
6206 /* Discount task's blocked util from CPU's util */
6207 util -= min_t(unsigned int, util, task_util(p));
6212 * a) if *p is the only task sleeping on this CPU, then:
6213 * cpu_util (== task_util) > util_est (== 0)
6214 * and thus we return:
6215 * cpu_util_wake = (cpu_util - task_util) = 0
6217 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6219 * cpu_util >= task_util
6220 * cpu_util > util_est (== 0)
6221 * and thus we discount *p's blocked utilization to return:
6222 * cpu_util_wake = (cpu_util - task_util) >= 0
6224 * c) if other tasks are RUNNABLE on that CPU and
6225 * util_est > cpu_util
6226 * then we use util_est since it returns a more restrictive
6227 * estimation of the spare capacity on that CPU, by just
6228 * considering the expected utilization of tasks already
6229 * runnable on that CPU.
6231 * Cases a) and b) are covered by the above code, while case c) is
6232 * covered by the following code when estimated utilization is
6235 if (sched_feat(UTIL_EST))
6236 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6239 * Utilization (estimated) can exceed the CPU capacity, thus let's
6240 * clamp to the maximum CPU capacity to ensure consistency with
6241 * the cpu_util call.
6243 return min_t(unsigned long, util, capacity_orig_of(cpu));
6247 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6248 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6250 * In that case WAKE_AFFINE doesn't make sense and we'll let
6251 * BALANCE_WAKE sort things out.
6253 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6255 long min_cap, max_cap;
6257 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6258 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6260 /* Minimum capacity is close to max, no need to abort wake_affine */
6261 if (max_cap - min_cap < max_cap >> 3)
6264 /* Bring task utilization in sync with prev_cpu */
6265 sync_entity_load_avg(&p->se);
6267 return min_cap * 1024 < task_util(p) * capacity_margin;
6271 * select_task_rq_fair: Select target runqueue for the waking task in domains
6272 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6273 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6275 * Balances load by selecting the idlest CPU in the idlest group, or under
6276 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6278 * Returns the target CPU number.
6280 * preempt must be disabled.
6283 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6285 struct sched_domain *tmp, *sd = NULL;
6286 int cpu = smp_processor_id();
6287 int new_cpu = prev_cpu;
6288 int want_affine = 0;
6289 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6291 if (sd_flag & SD_BALANCE_WAKE) {
6293 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
6294 && cpumask_test_cpu(cpu, &p->cpus_allowed);
6298 for_each_domain(cpu, tmp) {
6299 if (!(tmp->flags & SD_LOAD_BALANCE))
6303 * If both 'cpu' and 'prev_cpu' are part of this domain,
6304 * cpu is a valid SD_WAKE_AFFINE target.
6306 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6307 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6308 if (cpu != prev_cpu)
6309 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6311 sd = NULL; /* Prefer wake_affine over balance flags */
6315 if (tmp->flags & sd_flag)
6317 else if (!want_affine)
6323 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6324 } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6327 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6330 current->recent_used_cpu = cpu;
6337 static void detach_entity_cfs_rq(struct sched_entity *se);
6340 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6341 * cfs_rq_of(p) references at time of call are still valid and identify the
6342 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6344 static void migrate_task_rq_fair(struct task_struct *p)
6347 * As blocked tasks retain absolute vruntime the migration needs to
6348 * deal with this by subtracting the old and adding the new
6349 * min_vruntime -- the latter is done by enqueue_entity() when placing
6350 * the task on the new runqueue.
6352 if (p->state == TASK_WAKING) {
6353 struct sched_entity *se = &p->se;
6354 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6357 #ifndef CONFIG_64BIT
6358 u64 min_vruntime_copy;
6361 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6363 min_vruntime = cfs_rq->min_vruntime;
6364 } while (min_vruntime != min_vruntime_copy);
6366 min_vruntime = cfs_rq->min_vruntime;
6369 se->vruntime -= min_vruntime;
6372 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6374 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6375 * rq->lock and can modify state directly.
6377 lockdep_assert_held(&task_rq(p)->lock);
6378 detach_entity_cfs_rq(&p->se);
6382 * We are supposed to update the task to "current" time, then
6383 * its up to date and ready to go to new CPU/cfs_rq. But we
6384 * have difficulty in getting what current time is, so simply
6385 * throw away the out-of-date time. This will result in the
6386 * wakee task is less decayed, but giving the wakee more load
6389 remove_entity_load_avg(&p->se);
6392 /* Tell new CPU we are migrated */
6393 p->se.avg.last_update_time = 0;
6395 /* We have migrated, no longer consider this task hot */
6396 p->se.exec_start = 0;
6399 static void task_dead_fair(struct task_struct *p)
6401 remove_entity_load_avg(&p->se);
6403 #endif /* CONFIG_SMP */
6405 static unsigned long wakeup_gran(struct sched_entity *se)
6407 unsigned long gran = sysctl_sched_wakeup_granularity;
6410 * Since its curr running now, convert the gran from real-time
6411 * to virtual-time in his units.
6413 * By using 'se' instead of 'curr' we penalize light tasks, so
6414 * they get preempted easier. That is, if 'se' < 'curr' then
6415 * the resulting gran will be larger, therefore penalizing the
6416 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6417 * be smaller, again penalizing the lighter task.
6419 * This is especially important for buddies when the leftmost
6420 * task is higher priority than the buddy.
6422 return calc_delta_fair(gran, se);
6426 * Should 'se' preempt 'curr'.
6440 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6442 s64 gran, vdiff = curr->vruntime - se->vruntime;
6447 gran = wakeup_gran(se);
6454 static void set_last_buddy(struct sched_entity *se)
6456 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6459 for_each_sched_entity(se) {
6460 if (SCHED_WARN_ON(!se->on_rq))
6462 cfs_rq_of(se)->last = se;
6466 static void set_next_buddy(struct sched_entity *se)
6468 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6471 for_each_sched_entity(se) {
6472 if (SCHED_WARN_ON(!se->on_rq))
6474 cfs_rq_of(se)->next = se;
6478 static void set_skip_buddy(struct sched_entity *se)
6480 for_each_sched_entity(se)
6481 cfs_rq_of(se)->skip = se;
6485 * Preempt the current task with a newly woken task if needed:
6487 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6489 struct task_struct *curr = rq->curr;
6490 struct sched_entity *se = &curr->se, *pse = &p->se;
6491 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6492 int scale = cfs_rq->nr_running >= sched_nr_latency;
6493 int next_buddy_marked = 0;
6495 if (unlikely(se == pse))
6499 * This is possible from callers such as attach_tasks(), in which we
6500 * unconditionally check_prempt_curr() after an enqueue (which may have
6501 * lead to a throttle). This both saves work and prevents false
6502 * next-buddy nomination below.
6504 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6507 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6508 set_next_buddy(pse);
6509 next_buddy_marked = 1;
6513 * We can come here with TIF_NEED_RESCHED already set from new task
6516 * Note: this also catches the edge-case of curr being in a throttled
6517 * group (e.g. via set_curr_task), since update_curr() (in the
6518 * enqueue of curr) will have resulted in resched being set. This
6519 * prevents us from potentially nominating it as a false LAST_BUDDY
6522 if (test_tsk_need_resched(curr))
6525 /* Idle tasks are by definition preempted by non-idle tasks. */
6526 if (unlikely(curr->policy == SCHED_IDLE) &&
6527 likely(p->policy != SCHED_IDLE))
6531 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6532 * is driven by the tick):
6534 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6537 find_matching_se(&se, &pse);
6538 update_curr(cfs_rq_of(se));
6540 if (wakeup_preempt_entity(se, pse) == 1) {
6542 * Bias pick_next to pick the sched entity that is
6543 * triggering this preemption.
6545 if (!next_buddy_marked)
6546 set_next_buddy(pse);
6555 * Only set the backward buddy when the current task is still
6556 * on the rq. This can happen when a wakeup gets interleaved
6557 * with schedule on the ->pre_schedule() or idle_balance()
6558 * point, either of which can * drop the rq lock.
6560 * Also, during early boot the idle thread is in the fair class,
6561 * for obvious reasons its a bad idea to schedule back to it.
6563 if (unlikely(!se->on_rq || curr == rq->idle))
6566 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6570 static struct task_struct *
6571 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6573 struct cfs_rq *cfs_rq = &rq->cfs;
6574 struct sched_entity *se;
6575 struct task_struct *p;
6579 if (!cfs_rq->nr_running)
6582 #ifdef CONFIG_FAIR_GROUP_SCHED
6583 if (prev->sched_class != &fair_sched_class)
6587 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6588 * likely that a next task is from the same cgroup as the current.
6590 * Therefore attempt to avoid putting and setting the entire cgroup
6591 * hierarchy, only change the part that actually changes.
6595 struct sched_entity *curr = cfs_rq->curr;
6598 * Since we got here without doing put_prev_entity() we also
6599 * have to consider cfs_rq->curr. If it is still a runnable
6600 * entity, update_curr() will update its vruntime, otherwise
6601 * forget we've ever seen it.
6605 update_curr(cfs_rq);
6610 * This call to check_cfs_rq_runtime() will do the
6611 * throttle and dequeue its entity in the parent(s).
6612 * Therefore the nr_running test will indeed
6615 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6618 if (!cfs_rq->nr_running)
6625 se = pick_next_entity(cfs_rq, curr);
6626 cfs_rq = group_cfs_rq(se);
6632 * Since we haven't yet done put_prev_entity and if the selected task
6633 * is a different task than we started out with, try and touch the
6634 * least amount of cfs_rqs.
6637 struct sched_entity *pse = &prev->se;
6639 while (!(cfs_rq = is_same_group(se, pse))) {
6640 int se_depth = se->depth;
6641 int pse_depth = pse->depth;
6643 if (se_depth <= pse_depth) {
6644 put_prev_entity(cfs_rq_of(pse), pse);
6645 pse = parent_entity(pse);
6647 if (se_depth >= pse_depth) {
6648 set_next_entity(cfs_rq_of(se), se);
6649 se = parent_entity(se);
6653 put_prev_entity(cfs_rq, pse);
6654 set_next_entity(cfs_rq, se);
6661 put_prev_task(rq, prev);
6664 se = pick_next_entity(cfs_rq, NULL);
6665 set_next_entity(cfs_rq, se);
6666 cfs_rq = group_cfs_rq(se);
6671 done: __maybe_unused;
6674 * Move the next running task to the front of
6675 * the list, so our cfs_tasks list becomes MRU
6678 list_move(&p->se.group_node, &rq->cfs_tasks);
6681 if (hrtick_enabled(rq))
6682 hrtick_start_fair(rq, p);
6687 new_tasks = idle_balance(rq, rf);
6690 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6691 * possible for any higher priority task to appear. In that case we
6692 * must re-start the pick_next_entity() loop.
6704 * Account for a descheduled task:
6706 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6708 struct sched_entity *se = &prev->se;
6709 struct cfs_rq *cfs_rq;
6711 for_each_sched_entity(se) {
6712 cfs_rq = cfs_rq_of(se);
6713 put_prev_entity(cfs_rq, se);
6718 * sched_yield() is very simple
6720 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6722 static void yield_task_fair(struct rq *rq)
6724 struct task_struct *curr = rq->curr;
6725 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6726 struct sched_entity *se = &curr->se;
6729 * Are we the only task in the tree?
6731 if (unlikely(rq->nr_running == 1))
6734 clear_buddies(cfs_rq, se);
6736 if (curr->policy != SCHED_BATCH) {
6737 update_rq_clock(rq);
6739 * Update run-time statistics of the 'current'.
6741 update_curr(cfs_rq);
6743 * Tell update_rq_clock() that we've just updated,
6744 * so we don't do microscopic update in schedule()
6745 * and double the fastpath cost.
6747 rq_clock_skip_update(rq);
6753 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6755 struct sched_entity *se = &p->se;
6757 /* throttled hierarchies are not runnable */
6758 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6761 /* Tell the scheduler that we'd really like pse to run next. */
6764 yield_task_fair(rq);
6770 /**************************************************
6771 * Fair scheduling class load-balancing methods.
6775 * The purpose of load-balancing is to achieve the same basic fairness the
6776 * per-CPU scheduler provides, namely provide a proportional amount of compute
6777 * time to each task. This is expressed in the following equation:
6779 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6781 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
6782 * W_i,0 is defined as:
6784 * W_i,0 = \Sum_j w_i,j (2)
6786 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
6787 * is derived from the nice value as per sched_prio_to_weight[].
6789 * The weight average is an exponential decay average of the instantaneous
6792 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6794 * C_i is the compute capacity of CPU i, typically it is the
6795 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6796 * can also include other factors [XXX].
6798 * To achieve this balance we define a measure of imbalance which follows
6799 * directly from (1):
6801 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6803 * We them move tasks around to minimize the imbalance. In the continuous
6804 * function space it is obvious this converges, in the discrete case we get
6805 * a few fun cases generally called infeasible weight scenarios.
6808 * - infeasible weights;
6809 * - local vs global optima in the discrete case. ]
6814 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6815 * for all i,j solution, we create a tree of CPUs that follows the hardware
6816 * topology where each level pairs two lower groups (or better). This results
6817 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
6818 * tree to only the first of the previous level and we decrease the frequency
6819 * of load-balance at each level inv. proportional to the number of CPUs in
6825 * \Sum { --- * --- * 2^i } = O(n) (5)
6827 * `- size of each group
6828 * | | `- number of CPUs doing load-balance
6830 * `- sum over all levels
6832 * Coupled with a limit on how many tasks we can migrate every balance pass,
6833 * this makes (5) the runtime complexity of the balancer.
6835 * An important property here is that each CPU is still (indirectly) connected
6836 * to every other CPU in at most O(log n) steps:
6838 * The adjacency matrix of the resulting graph is given by:
6841 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6844 * And you'll find that:
6846 * A^(log_2 n)_i,j != 0 for all i,j (7)
6848 * Showing there's indeed a path between every CPU in at most O(log n) steps.
6849 * The task movement gives a factor of O(m), giving a convergence complexity
6852 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6857 * In order to avoid CPUs going idle while there's still work to do, new idle
6858 * balancing is more aggressive and has the newly idle CPU iterate up the domain
6859 * tree itself instead of relying on other CPUs to bring it work.
6861 * This adds some complexity to both (5) and (8) but it reduces the total idle
6869 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6872 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6877 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6879 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
6881 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6884 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6885 * rewrite all of this once again.]
6888 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6890 enum fbq_type { regular, remote, all };
6892 #define LBF_ALL_PINNED 0x01
6893 #define LBF_NEED_BREAK 0x02
6894 #define LBF_DST_PINNED 0x04
6895 #define LBF_SOME_PINNED 0x08
6896 #define LBF_NOHZ_STATS 0x10
6897 #define LBF_NOHZ_AGAIN 0x20
6900 struct sched_domain *sd;
6908 struct cpumask *dst_grpmask;
6910 enum cpu_idle_type idle;
6912 /* The set of CPUs under consideration for load-balancing */
6913 struct cpumask *cpus;
6918 unsigned int loop_break;
6919 unsigned int loop_max;
6921 enum fbq_type fbq_type;
6922 struct list_head tasks;
6926 * Is this task likely cache-hot:
6928 static int task_hot(struct task_struct *p, struct lb_env *env)
6932 lockdep_assert_held(&env->src_rq->lock);
6934 if (p->sched_class != &fair_sched_class)
6937 if (unlikely(p->policy == SCHED_IDLE))
6941 * Buddy candidates are cache hot:
6943 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6944 (&p->se == cfs_rq_of(&p->se)->next ||
6945 &p->se == cfs_rq_of(&p->se)->last))
6948 if (sysctl_sched_migration_cost == -1)
6950 if (sysctl_sched_migration_cost == 0)
6953 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6955 return delta < (s64)sysctl_sched_migration_cost;
6958 #ifdef CONFIG_NUMA_BALANCING
6960 * Returns 1, if task migration degrades locality
6961 * Returns 0, if task migration improves locality i.e migration preferred.
6962 * Returns -1, if task migration is not affected by locality.
6964 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6966 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6967 unsigned long src_faults, dst_faults;
6968 int src_nid, dst_nid;
6970 if (!static_branch_likely(&sched_numa_balancing))
6973 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6976 src_nid = cpu_to_node(env->src_cpu);
6977 dst_nid = cpu_to_node(env->dst_cpu);
6979 if (src_nid == dst_nid)
6982 /* Migrating away from the preferred node is always bad. */
6983 if (src_nid == p->numa_preferred_nid) {
6984 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6990 /* Encourage migration to the preferred node. */
6991 if (dst_nid == p->numa_preferred_nid)
6994 /* Leaving a core idle is often worse than degrading locality. */
6995 if (env->idle != CPU_NOT_IDLE)
6999 src_faults = group_faults(p, src_nid);
7000 dst_faults = group_faults(p, dst_nid);
7002 src_faults = task_faults(p, src_nid);
7003 dst_faults = task_faults(p, dst_nid);
7006 return dst_faults < src_faults;
7010 static inline int migrate_degrades_locality(struct task_struct *p,
7018 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7021 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7025 lockdep_assert_held(&env->src_rq->lock);
7028 * We do not migrate tasks that are:
7029 * 1) throttled_lb_pair, or
7030 * 2) cannot be migrated to this CPU due to cpus_allowed, or
7031 * 3) running (obviously), or
7032 * 4) are cache-hot on their current CPU.
7034 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7037 if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) {
7040 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7042 env->flags |= LBF_SOME_PINNED;
7045 * Remember if this task can be migrated to any other CPU in
7046 * our sched_group. We may want to revisit it if we couldn't
7047 * meet load balance goals by pulling other tasks on src_cpu.
7049 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7050 * already computed one in current iteration.
7052 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7055 /* Prevent to re-select dst_cpu via env's CPUs: */
7056 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7057 if (cpumask_test_cpu(cpu, &p->cpus_allowed)) {
7058 env->flags |= LBF_DST_PINNED;
7059 env->new_dst_cpu = cpu;
7067 /* Record that we found atleast one task that could run on dst_cpu */
7068 env->flags &= ~LBF_ALL_PINNED;
7070 if (task_running(env->src_rq, p)) {
7071 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7076 * Aggressive migration if:
7077 * 1) destination numa is preferred
7078 * 2) task is cache cold, or
7079 * 3) too many balance attempts have failed.
7081 tsk_cache_hot = migrate_degrades_locality(p, env);
7082 if (tsk_cache_hot == -1)
7083 tsk_cache_hot = task_hot(p, env);
7085 if (tsk_cache_hot <= 0 ||
7086 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7087 if (tsk_cache_hot == 1) {
7088 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7089 schedstat_inc(p->se.statistics.nr_forced_migrations);
7094 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7099 * detach_task() -- detach the task for the migration specified in env
7101 static void detach_task(struct task_struct *p, struct lb_env *env)
7103 lockdep_assert_held(&env->src_rq->lock);
7105 p->on_rq = TASK_ON_RQ_MIGRATING;
7106 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7107 set_task_cpu(p, env->dst_cpu);
7111 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7112 * part of active balancing operations within "domain".
7114 * Returns a task if successful and NULL otherwise.
7116 static struct task_struct *detach_one_task(struct lb_env *env)
7118 struct task_struct *p;
7120 lockdep_assert_held(&env->src_rq->lock);
7122 list_for_each_entry_reverse(p,
7123 &env->src_rq->cfs_tasks, se.group_node) {
7124 if (!can_migrate_task(p, env))
7127 detach_task(p, env);
7130 * Right now, this is only the second place where
7131 * lb_gained[env->idle] is updated (other is detach_tasks)
7132 * so we can safely collect stats here rather than
7133 * inside detach_tasks().
7135 schedstat_inc(env->sd->lb_gained[env->idle]);
7141 static const unsigned int sched_nr_migrate_break = 32;
7144 * detach_tasks() -- tries to detach up to imbalance weighted load from
7145 * busiest_rq, as part of a balancing operation within domain "sd".
7147 * Returns number of detached tasks if successful and 0 otherwise.
7149 static int detach_tasks(struct lb_env *env)
7151 struct list_head *tasks = &env->src_rq->cfs_tasks;
7152 struct task_struct *p;
7156 lockdep_assert_held(&env->src_rq->lock);
7158 if (env->imbalance <= 0)
7161 while (!list_empty(tasks)) {
7163 * We don't want to steal all, otherwise we may be treated likewise,
7164 * which could at worst lead to a livelock crash.
7166 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7169 p = list_last_entry(tasks, struct task_struct, se.group_node);
7172 /* We've more or less seen every task there is, call it quits */
7173 if (env->loop > env->loop_max)
7176 /* take a breather every nr_migrate tasks */
7177 if (env->loop > env->loop_break) {
7178 env->loop_break += sched_nr_migrate_break;
7179 env->flags |= LBF_NEED_BREAK;
7183 if (!can_migrate_task(p, env))
7186 load = task_h_load(p);
7188 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7191 if ((load / 2) > env->imbalance)
7194 detach_task(p, env);
7195 list_add(&p->se.group_node, &env->tasks);
7198 env->imbalance -= load;
7200 #ifdef CONFIG_PREEMPT
7202 * NEWIDLE balancing is a source of latency, so preemptible
7203 * kernels will stop after the first task is detached to minimize
7204 * the critical section.
7206 if (env->idle == CPU_NEWLY_IDLE)
7211 * We only want to steal up to the prescribed amount of
7214 if (env->imbalance <= 0)
7219 list_move(&p->se.group_node, tasks);
7223 * Right now, this is one of only two places we collect this stat
7224 * so we can safely collect detach_one_task() stats here rather
7225 * than inside detach_one_task().
7227 schedstat_add(env->sd->lb_gained[env->idle], detached);
7233 * attach_task() -- attach the task detached by detach_task() to its new rq.
7235 static void attach_task(struct rq *rq, struct task_struct *p)
7237 lockdep_assert_held(&rq->lock);
7239 BUG_ON(task_rq(p) != rq);
7240 activate_task(rq, p, ENQUEUE_NOCLOCK);
7241 p->on_rq = TASK_ON_RQ_QUEUED;
7242 check_preempt_curr(rq, p, 0);
7246 * attach_one_task() -- attaches the task returned from detach_one_task() to
7249 static void attach_one_task(struct rq *rq, struct task_struct *p)
7254 update_rq_clock(rq);
7260 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7263 static void attach_tasks(struct lb_env *env)
7265 struct list_head *tasks = &env->tasks;
7266 struct task_struct *p;
7269 rq_lock(env->dst_rq, &rf);
7270 update_rq_clock(env->dst_rq);
7272 while (!list_empty(tasks)) {
7273 p = list_first_entry(tasks, struct task_struct, se.group_node);
7274 list_del_init(&p->se.group_node);
7276 attach_task(env->dst_rq, p);
7279 rq_unlock(env->dst_rq, &rf);
7282 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7284 if (cfs_rq->avg.load_avg)
7287 if (cfs_rq->avg.util_avg)
7293 static inline bool others_rqs_have_blocked(struct rq *rq)
7295 if (READ_ONCE(rq->avg_rt.util_avg))
7298 if (READ_ONCE(rq->avg_dl.util_avg))
7304 #ifdef CONFIG_FAIR_GROUP_SCHED
7306 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7308 if (cfs_rq->load.weight)
7311 if (cfs_rq->avg.load_sum)
7314 if (cfs_rq->avg.util_sum)
7317 if (cfs_rq->avg.runnable_load_sum)
7323 static void update_blocked_averages(int cpu)
7325 struct rq *rq = cpu_rq(cpu);
7326 struct cfs_rq *cfs_rq, *pos;
7330 rq_lock_irqsave(rq, &rf);
7331 update_rq_clock(rq);
7334 * Iterates the task_group tree in a bottom up fashion, see
7335 * list_add_leaf_cfs_rq() for details.
7337 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7338 struct sched_entity *se;
7340 /* throttled entities do not contribute to load */
7341 if (throttled_hierarchy(cfs_rq))
7344 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
7345 update_tg_load_avg(cfs_rq, 0);
7347 /* Propagate pending load changes to the parent, if any: */
7348 se = cfs_rq->tg->se[cpu];
7349 if (se && !skip_blocked_update(se))
7350 update_load_avg(cfs_rq_of(se), se, 0);
7353 * There can be a lot of idle CPU cgroups. Don't let fully
7354 * decayed cfs_rqs linger on the list.
7356 if (cfs_rq_is_decayed(cfs_rq))
7357 list_del_leaf_cfs_rq(cfs_rq);
7359 /* Don't need periodic decay once load/util_avg are null */
7360 if (cfs_rq_has_blocked(cfs_rq))
7363 update_rt_rq_load_avg(rq_clock_task(rq), rq, 0);
7364 update_dl_rq_load_avg(rq_clock_task(rq), rq, 0);
7365 /* Don't need periodic decay once load/util_avg are null */
7366 if (others_rqs_have_blocked(rq))
7369 #ifdef CONFIG_NO_HZ_COMMON
7370 rq->last_blocked_load_update_tick = jiffies;
7372 rq->has_blocked_load = 0;
7374 rq_unlock_irqrestore(rq, &rf);
7378 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7379 * This needs to be done in a top-down fashion because the load of a child
7380 * group is a fraction of its parents load.
7382 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7384 struct rq *rq = rq_of(cfs_rq);
7385 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7386 unsigned long now = jiffies;
7389 if (cfs_rq->last_h_load_update == now)
7392 cfs_rq->h_load_next = NULL;
7393 for_each_sched_entity(se) {
7394 cfs_rq = cfs_rq_of(se);
7395 cfs_rq->h_load_next = se;
7396 if (cfs_rq->last_h_load_update == now)
7401 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7402 cfs_rq->last_h_load_update = now;
7405 while ((se = cfs_rq->h_load_next) != NULL) {
7406 load = cfs_rq->h_load;
7407 load = div64_ul(load * se->avg.load_avg,
7408 cfs_rq_load_avg(cfs_rq) + 1);
7409 cfs_rq = group_cfs_rq(se);
7410 cfs_rq->h_load = load;
7411 cfs_rq->last_h_load_update = now;
7415 static unsigned long task_h_load(struct task_struct *p)
7417 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7419 update_cfs_rq_h_load(cfs_rq);
7420 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7421 cfs_rq_load_avg(cfs_rq) + 1);
7424 static inline void update_blocked_averages(int cpu)
7426 struct rq *rq = cpu_rq(cpu);
7427 struct cfs_rq *cfs_rq = &rq->cfs;
7430 rq_lock_irqsave(rq, &rf);
7431 update_rq_clock(rq);
7432 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
7433 update_rt_rq_load_avg(rq_clock_task(rq), rq, 0);
7434 update_dl_rq_load_avg(rq_clock_task(rq), rq, 0);
7435 #ifdef CONFIG_NO_HZ_COMMON
7436 rq->last_blocked_load_update_tick = jiffies;
7437 if (!cfs_rq_has_blocked(cfs_rq) && !others_rqs_have_blocked(rq))
7438 rq->has_blocked_load = 0;
7440 rq_unlock_irqrestore(rq, &rf);
7443 static unsigned long task_h_load(struct task_struct *p)
7445 return p->se.avg.load_avg;
7449 /********** Helpers for find_busiest_group ************************/
7458 * sg_lb_stats - stats of a sched_group required for load_balancing
7460 struct sg_lb_stats {
7461 unsigned long avg_load; /*Avg load across the CPUs of the group */
7462 unsigned long group_load; /* Total load over the CPUs of the group */
7463 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7464 unsigned long load_per_task;
7465 unsigned long group_capacity;
7466 unsigned long group_util; /* Total utilization of the group */
7467 unsigned int sum_nr_running; /* Nr tasks running in the group */
7468 unsigned int idle_cpus;
7469 unsigned int group_weight;
7470 enum group_type group_type;
7471 int group_no_capacity;
7472 #ifdef CONFIG_NUMA_BALANCING
7473 unsigned int nr_numa_running;
7474 unsigned int nr_preferred_running;
7479 * sd_lb_stats - Structure to store the statistics of a sched_domain
7480 * during load balancing.
7482 struct sd_lb_stats {
7483 struct sched_group *busiest; /* Busiest group in this sd */
7484 struct sched_group *local; /* Local group in this sd */
7485 unsigned long total_running;
7486 unsigned long total_load; /* Total load of all groups in sd */
7487 unsigned long total_capacity; /* Total capacity of all groups in sd */
7488 unsigned long avg_load; /* Average load across all groups in sd */
7490 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7491 struct sg_lb_stats local_stat; /* Statistics of the local group */
7494 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7497 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7498 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7499 * We must however clear busiest_stat::avg_load because
7500 * update_sd_pick_busiest() reads this before assignment.
7502 *sds = (struct sd_lb_stats){
7505 .total_running = 0UL,
7507 .total_capacity = 0UL,
7510 .sum_nr_running = 0,
7511 .group_type = group_other,
7517 * get_sd_load_idx - Obtain the load index for a given sched domain.
7518 * @sd: The sched_domain whose load_idx is to be obtained.
7519 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7521 * Return: The load index.
7523 static inline int get_sd_load_idx(struct sched_domain *sd,
7524 enum cpu_idle_type idle)
7530 load_idx = sd->busy_idx;
7533 case CPU_NEWLY_IDLE:
7534 load_idx = sd->newidle_idx;
7537 load_idx = sd->idle_idx;
7544 static unsigned long scale_rt_capacity(int cpu)
7546 struct rq *rq = cpu_rq(cpu);
7547 u64 total, used, age_stamp, avg;
7551 * Since we're reading these variables without serialization make sure
7552 * we read them once before doing sanity checks on them.
7554 age_stamp = READ_ONCE(rq->age_stamp);
7555 avg = READ_ONCE(rq->rt_avg);
7556 delta = __rq_clock_broken(rq) - age_stamp;
7558 if (unlikely(delta < 0))
7561 total = sched_avg_period() + delta;
7563 used = div_u64(avg, total);
7565 if (likely(used < SCHED_CAPACITY_SCALE))
7566 return SCHED_CAPACITY_SCALE - used;
7571 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7573 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7574 struct sched_group *sdg = sd->groups;
7576 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7578 capacity *= scale_rt_capacity(cpu);
7579 capacity >>= SCHED_CAPACITY_SHIFT;
7584 cpu_rq(cpu)->cpu_capacity = capacity;
7585 sdg->sgc->capacity = capacity;
7586 sdg->sgc->min_capacity = capacity;
7589 void update_group_capacity(struct sched_domain *sd, int cpu)
7591 struct sched_domain *child = sd->child;
7592 struct sched_group *group, *sdg = sd->groups;
7593 unsigned long capacity, min_capacity;
7594 unsigned long interval;
7596 interval = msecs_to_jiffies(sd->balance_interval);
7597 interval = clamp(interval, 1UL, max_load_balance_interval);
7598 sdg->sgc->next_update = jiffies + interval;
7601 update_cpu_capacity(sd, cpu);
7606 min_capacity = ULONG_MAX;
7608 if (child->flags & SD_OVERLAP) {
7610 * SD_OVERLAP domains cannot assume that child groups
7611 * span the current group.
7614 for_each_cpu(cpu, sched_group_span(sdg)) {
7615 struct sched_group_capacity *sgc;
7616 struct rq *rq = cpu_rq(cpu);
7619 * build_sched_domains() -> init_sched_groups_capacity()
7620 * gets here before we've attached the domains to the
7623 * Use capacity_of(), which is set irrespective of domains
7624 * in update_cpu_capacity().
7626 * This avoids capacity from being 0 and
7627 * causing divide-by-zero issues on boot.
7629 if (unlikely(!rq->sd)) {
7630 capacity += capacity_of(cpu);
7632 sgc = rq->sd->groups->sgc;
7633 capacity += sgc->capacity;
7636 min_capacity = min(capacity, min_capacity);
7640 * !SD_OVERLAP domains can assume that child groups
7641 * span the current group.
7644 group = child->groups;
7646 struct sched_group_capacity *sgc = group->sgc;
7648 capacity += sgc->capacity;
7649 min_capacity = min(sgc->min_capacity, min_capacity);
7650 group = group->next;
7651 } while (group != child->groups);
7654 sdg->sgc->capacity = capacity;
7655 sdg->sgc->min_capacity = min_capacity;
7659 * Check whether the capacity of the rq has been noticeably reduced by side
7660 * activity. The imbalance_pct is used for the threshold.
7661 * Return true is the capacity is reduced
7664 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7666 return ((rq->cpu_capacity * sd->imbalance_pct) <
7667 (rq->cpu_capacity_orig * 100));
7671 * Group imbalance indicates (and tries to solve) the problem where balancing
7672 * groups is inadequate due to ->cpus_allowed constraints.
7674 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
7675 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
7678 * { 0 1 2 3 } { 4 5 6 7 }
7681 * If we were to balance group-wise we'd place two tasks in the first group and
7682 * two tasks in the second group. Clearly this is undesired as it will overload
7683 * cpu 3 and leave one of the CPUs in the second group unused.
7685 * The current solution to this issue is detecting the skew in the first group
7686 * by noticing the lower domain failed to reach balance and had difficulty
7687 * moving tasks due to affinity constraints.
7689 * When this is so detected; this group becomes a candidate for busiest; see
7690 * update_sd_pick_busiest(). And calculate_imbalance() and
7691 * find_busiest_group() avoid some of the usual balance conditions to allow it
7692 * to create an effective group imbalance.
7694 * This is a somewhat tricky proposition since the next run might not find the
7695 * group imbalance and decide the groups need to be balanced again. A most
7696 * subtle and fragile situation.
7699 static inline int sg_imbalanced(struct sched_group *group)
7701 return group->sgc->imbalance;
7705 * group_has_capacity returns true if the group has spare capacity that could
7706 * be used by some tasks.
7707 * We consider that a group has spare capacity if the * number of task is
7708 * smaller than the number of CPUs or if the utilization is lower than the
7709 * available capacity for CFS tasks.
7710 * For the latter, we use a threshold to stabilize the state, to take into
7711 * account the variance of the tasks' load and to return true if the available
7712 * capacity in meaningful for the load balancer.
7713 * As an example, an available capacity of 1% can appear but it doesn't make
7714 * any benefit for the load balance.
7717 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7719 if (sgs->sum_nr_running < sgs->group_weight)
7722 if ((sgs->group_capacity * 100) >
7723 (sgs->group_util * env->sd->imbalance_pct))
7730 * group_is_overloaded returns true if the group has more tasks than it can
7732 * group_is_overloaded is not equals to !group_has_capacity because a group
7733 * with the exact right number of tasks, has no more spare capacity but is not
7734 * overloaded so both group_has_capacity and group_is_overloaded return
7738 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7740 if (sgs->sum_nr_running <= sgs->group_weight)
7743 if ((sgs->group_capacity * 100) <
7744 (sgs->group_util * env->sd->imbalance_pct))
7751 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7752 * per-CPU capacity than sched_group ref.
7755 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7757 return sg->sgc->min_capacity * capacity_margin <
7758 ref->sgc->min_capacity * 1024;
7762 group_type group_classify(struct sched_group *group,
7763 struct sg_lb_stats *sgs)
7765 if (sgs->group_no_capacity)
7766 return group_overloaded;
7768 if (sg_imbalanced(group))
7769 return group_imbalanced;
7774 static bool update_nohz_stats(struct rq *rq, bool force)
7776 #ifdef CONFIG_NO_HZ_COMMON
7777 unsigned int cpu = rq->cpu;
7779 if (!rq->has_blocked_load)
7782 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
7785 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
7788 update_blocked_averages(cpu);
7790 return rq->has_blocked_load;
7797 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7798 * @env: The load balancing environment.
7799 * @group: sched_group whose statistics are to be updated.
7800 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7801 * @local_group: Does group contain this_cpu.
7802 * @sgs: variable to hold the statistics for this group.
7803 * @overload: Indicate more than one runnable task for any CPU.
7805 static inline void update_sg_lb_stats(struct lb_env *env,
7806 struct sched_group *group, int load_idx,
7807 int local_group, struct sg_lb_stats *sgs,
7813 memset(sgs, 0, sizeof(*sgs));
7815 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
7816 struct rq *rq = cpu_rq(i);
7818 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
7819 env->flags |= LBF_NOHZ_AGAIN;
7821 /* Bias balancing toward CPUs of our domain: */
7823 load = target_load(i, load_idx);
7825 load = source_load(i, load_idx);
7827 sgs->group_load += load;
7828 sgs->group_util += cpu_util(i);
7829 sgs->sum_nr_running += rq->cfs.h_nr_running;
7831 nr_running = rq->nr_running;
7835 #ifdef CONFIG_NUMA_BALANCING
7836 sgs->nr_numa_running += rq->nr_numa_running;
7837 sgs->nr_preferred_running += rq->nr_preferred_running;
7839 sgs->sum_weighted_load += weighted_cpuload(rq);
7841 * No need to call idle_cpu() if nr_running is not 0
7843 if (!nr_running && idle_cpu(i))
7847 /* Adjust by relative CPU capacity of the group */
7848 sgs->group_capacity = group->sgc->capacity;
7849 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7851 if (sgs->sum_nr_running)
7852 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7854 sgs->group_weight = group->group_weight;
7856 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7857 sgs->group_type = group_classify(group, sgs);
7861 * update_sd_pick_busiest - return 1 on busiest group
7862 * @env: The load balancing environment.
7863 * @sds: sched_domain statistics
7864 * @sg: sched_group candidate to be checked for being the busiest
7865 * @sgs: sched_group statistics
7867 * Determine if @sg is a busier group than the previously selected
7870 * Return: %true if @sg is a busier group than the previously selected
7871 * busiest group. %false otherwise.
7873 static bool update_sd_pick_busiest(struct lb_env *env,
7874 struct sd_lb_stats *sds,
7875 struct sched_group *sg,
7876 struct sg_lb_stats *sgs)
7878 struct sg_lb_stats *busiest = &sds->busiest_stat;
7880 if (sgs->group_type > busiest->group_type)
7883 if (sgs->group_type < busiest->group_type)
7886 if (sgs->avg_load <= busiest->avg_load)
7889 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7893 * Candidate sg has no more than one task per CPU and
7894 * has higher per-CPU capacity. Migrating tasks to less
7895 * capable CPUs may harm throughput. Maximize throughput,
7896 * power/energy consequences are not considered.
7898 if (sgs->sum_nr_running <= sgs->group_weight &&
7899 group_smaller_cpu_capacity(sds->local, sg))
7903 /* This is the busiest node in its class. */
7904 if (!(env->sd->flags & SD_ASYM_PACKING))
7907 /* No ASYM_PACKING if target CPU is already busy */
7908 if (env->idle == CPU_NOT_IDLE)
7911 * ASYM_PACKING needs to move all the work to the highest
7912 * prority CPUs in the group, therefore mark all groups
7913 * of lower priority than ourself as busy.
7915 if (sgs->sum_nr_running &&
7916 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
7920 /* Prefer to move from lowest priority CPU's work */
7921 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
7922 sg->asym_prefer_cpu))
7929 #ifdef CONFIG_NUMA_BALANCING
7930 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7932 if (sgs->sum_nr_running > sgs->nr_numa_running)
7934 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7939 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7941 if (rq->nr_running > rq->nr_numa_running)
7943 if (rq->nr_running > rq->nr_preferred_running)
7948 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7953 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7957 #endif /* CONFIG_NUMA_BALANCING */
7960 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7961 * @env: The load balancing environment.
7962 * @sds: variable to hold the statistics for this sched_domain.
7964 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7966 struct sched_domain *child = env->sd->child;
7967 struct sched_group *sg = env->sd->groups;
7968 struct sg_lb_stats *local = &sds->local_stat;
7969 struct sg_lb_stats tmp_sgs;
7970 int load_idx, prefer_sibling = 0;
7971 bool overload = false;
7973 if (child && child->flags & SD_PREFER_SIBLING)
7976 #ifdef CONFIG_NO_HZ_COMMON
7977 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
7978 env->flags |= LBF_NOHZ_STATS;
7981 load_idx = get_sd_load_idx(env->sd, env->idle);
7984 struct sg_lb_stats *sgs = &tmp_sgs;
7987 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
7992 if (env->idle != CPU_NEWLY_IDLE ||
7993 time_after_eq(jiffies, sg->sgc->next_update))
7994 update_group_capacity(env->sd, env->dst_cpu);
7997 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
8004 * In case the child domain prefers tasks go to siblings
8005 * first, lower the sg capacity so that we'll try
8006 * and move all the excess tasks away. We lower the capacity
8007 * of a group only if the local group has the capacity to fit
8008 * these excess tasks. The extra check prevents the case where
8009 * you always pull from the heaviest group when it is already
8010 * under-utilized (possible with a large weight task outweighs
8011 * the tasks on the system).
8013 if (prefer_sibling && sds->local &&
8014 group_has_capacity(env, local) &&
8015 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
8016 sgs->group_no_capacity = 1;
8017 sgs->group_type = group_classify(sg, sgs);
8020 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8022 sds->busiest_stat = *sgs;
8026 /* Now, start updating sd_lb_stats */
8027 sds->total_running += sgs->sum_nr_running;
8028 sds->total_load += sgs->group_load;
8029 sds->total_capacity += sgs->group_capacity;
8032 } while (sg != env->sd->groups);
8034 #ifdef CONFIG_NO_HZ_COMMON
8035 if ((env->flags & LBF_NOHZ_AGAIN) &&
8036 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8038 WRITE_ONCE(nohz.next_blocked,
8039 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8043 if (env->sd->flags & SD_NUMA)
8044 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8046 if (!env->sd->parent) {
8047 /* update overload indicator if we are at root domain */
8048 if (env->dst_rq->rd->overload != overload)
8049 env->dst_rq->rd->overload = overload;
8054 * check_asym_packing - Check to see if the group is packed into the
8057 * This is primarily intended to used at the sibling level. Some
8058 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8059 * case of POWER7, it can move to lower SMT modes only when higher
8060 * threads are idle. When in lower SMT modes, the threads will
8061 * perform better since they share less core resources. Hence when we
8062 * have idle threads, we want them to be the higher ones.
8064 * This packing function is run on idle threads. It checks to see if
8065 * the busiest CPU in this domain (core in the P7 case) has a higher
8066 * CPU number than the packing function is being run on. Here we are
8067 * assuming lower CPU number will be equivalent to lower a SMT thread
8070 * Return: 1 when packing is required and a task should be moved to
8071 * this CPU. The amount of the imbalance is returned in env->imbalance.
8073 * @env: The load balancing environment.
8074 * @sds: Statistics of the sched_domain which is to be packed
8076 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8080 if (!(env->sd->flags & SD_ASYM_PACKING))
8083 if (env->idle == CPU_NOT_IDLE)
8089 busiest_cpu = sds->busiest->asym_prefer_cpu;
8090 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
8093 env->imbalance = DIV_ROUND_CLOSEST(
8094 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
8095 SCHED_CAPACITY_SCALE);
8101 * fix_small_imbalance - Calculate the minor imbalance that exists
8102 * amongst the groups of a sched_domain, during
8104 * @env: The load balancing environment.
8105 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8108 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8110 unsigned long tmp, capa_now = 0, capa_move = 0;
8111 unsigned int imbn = 2;
8112 unsigned long scaled_busy_load_per_task;
8113 struct sg_lb_stats *local, *busiest;
8115 local = &sds->local_stat;
8116 busiest = &sds->busiest_stat;
8118 if (!local->sum_nr_running)
8119 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8120 else if (busiest->load_per_task > local->load_per_task)
8123 scaled_busy_load_per_task =
8124 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8125 busiest->group_capacity;
8127 if (busiest->avg_load + scaled_busy_load_per_task >=
8128 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8129 env->imbalance = busiest->load_per_task;
8134 * OK, we don't have enough imbalance to justify moving tasks,
8135 * however we may be able to increase total CPU capacity used by
8139 capa_now += busiest->group_capacity *
8140 min(busiest->load_per_task, busiest->avg_load);
8141 capa_now += local->group_capacity *
8142 min(local->load_per_task, local->avg_load);
8143 capa_now /= SCHED_CAPACITY_SCALE;
8145 /* Amount of load we'd subtract */
8146 if (busiest->avg_load > scaled_busy_load_per_task) {
8147 capa_move += busiest->group_capacity *
8148 min(busiest->load_per_task,
8149 busiest->avg_load - scaled_busy_load_per_task);
8152 /* Amount of load we'd add */
8153 if (busiest->avg_load * busiest->group_capacity <
8154 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8155 tmp = (busiest->avg_load * busiest->group_capacity) /
8156 local->group_capacity;
8158 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8159 local->group_capacity;
8161 capa_move += local->group_capacity *
8162 min(local->load_per_task, local->avg_load + tmp);
8163 capa_move /= SCHED_CAPACITY_SCALE;
8165 /* Move if we gain throughput */
8166 if (capa_move > capa_now)
8167 env->imbalance = busiest->load_per_task;
8171 * calculate_imbalance - Calculate the amount of imbalance present within the
8172 * groups of a given sched_domain during load balance.
8173 * @env: load balance environment
8174 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8176 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8178 unsigned long max_pull, load_above_capacity = ~0UL;
8179 struct sg_lb_stats *local, *busiest;
8181 local = &sds->local_stat;
8182 busiest = &sds->busiest_stat;
8184 if (busiest->group_type == group_imbalanced) {
8186 * In the group_imb case we cannot rely on group-wide averages
8187 * to ensure CPU-load equilibrium, look at wider averages. XXX
8189 busiest->load_per_task =
8190 min(busiest->load_per_task, sds->avg_load);
8194 * Avg load of busiest sg can be less and avg load of local sg can
8195 * be greater than avg load across all sgs of sd because avg load
8196 * factors in sg capacity and sgs with smaller group_type are
8197 * skipped when updating the busiest sg:
8199 if (busiest->avg_load <= sds->avg_load ||
8200 local->avg_load >= sds->avg_load) {
8202 return fix_small_imbalance(env, sds);
8206 * If there aren't any idle CPUs, avoid creating some.
8208 if (busiest->group_type == group_overloaded &&
8209 local->group_type == group_overloaded) {
8210 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
8211 if (load_above_capacity > busiest->group_capacity) {
8212 load_above_capacity -= busiest->group_capacity;
8213 load_above_capacity *= scale_load_down(NICE_0_LOAD);
8214 load_above_capacity /= busiest->group_capacity;
8216 load_above_capacity = ~0UL;
8220 * We're trying to get all the CPUs to the average_load, so we don't
8221 * want to push ourselves above the average load, nor do we wish to
8222 * reduce the max loaded CPU below the average load. At the same time,
8223 * we also don't want to reduce the group load below the group
8224 * capacity. Thus we look for the minimum possible imbalance.
8226 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8228 /* How much load to actually move to equalise the imbalance */
8229 env->imbalance = min(
8230 max_pull * busiest->group_capacity,
8231 (sds->avg_load - local->avg_load) * local->group_capacity
8232 ) / SCHED_CAPACITY_SCALE;
8235 * if *imbalance is less than the average load per runnable task
8236 * there is no guarantee that any tasks will be moved so we'll have
8237 * a think about bumping its value to force at least one task to be
8240 if (env->imbalance < busiest->load_per_task)
8241 return fix_small_imbalance(env, sds);
8244 /******* find_busiest_group() helpers end here *********************/
8247 * find_busiest_group - Returns the busiest group within the sched_domain
8248 * if there is an imbalance.
8250 * Also calculates the amount of weighted load which should be moved
8251 * to restore balance.
8253 * @env: The load balancing environment.
8255 * Return: - The busiest group if imbalance exists.
8257 static struct sched_group *find_busiest_group(struct lb_env *env)
8259 struct sg_lb_stats *local, *busiest;
8260 struct sd_lb_stats sds;
8262 init_sd_lb_stats(&sds);
8265 * Compute the various statistics relavent for load balancing at
8268 update_sd_lb_stats(env, &sds);
8269 local = &sds.local_stat;
8270 busiest = &sds.busiest_stat;
8272 /* ASYM feature bypasses nice load balance check */
8273 if (check_asym_packing(env, &sds))
8276 /* There is no busy sibling group to pull tasks from */
8277 if (!sds.busiest || busiest->sum_nr_running == 0)
8280 /* XXX broken for overlapping NUMA groups */
8281 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8282 / sds.total_capacity;
8285 * If the busiest group is imbalanced the below checks don't
8286 * work because they assume all things are equal, which typically
8287 * isn't true due to cpus_allowed constraints and the like.
8289 if (busiest->group_type == group_imbalanced)
8293 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8294 * capacities from resulting in underutilization due to avg_load.
8296 if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8297 busiest->group_no_capacity)
8301 * If the local group is busier than the selected busiest group
8302 * don't try and pull any tasks.
8304 if (local->avg_load >= busiest->avg_load)
8308 * Don't pull any tasks if this group is already above the domain
8311 if (local->avg_load >= sds.avg_load)
8314 if (env->idle == CPU_IDLE) {
8316 * This CPU is idle. If the busiest group is not overloaded
8317 * and there is no imbalance between this and busiest group
8318 * wrt idle CPUs, it is balanced. The imbalance becomes
8319 * significant if the diff is greater than 1 otherwise we
8320 * might end up to just move the imbalance on another group
8322 if ((busiest->group_type != group_overloaded) &&
8323 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8327 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8328 * imbalance_pct to be conservative.
8330 if (100 * busiest->avg_load <=
8331 env->sd->imbalance_pct * local->avg_load)
8336 /* Looks like there is an imbalance. Compute it */
8337 calculate_imbalance(env, &sds);
8346 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
8348 static struct rq *find_busiest_queue(struct lb_env *env,
8349 struct sched_group *group)
8351 struct rq *busiest = NULL, *rq;
8352 unsigned long busiest_load = 0, busiest_capacity = 1;
8355 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8356 unsigned long capacity, wl;
8360 rt = fbq_classify_rq(rq);
8363 * We classify groups/runqueues into three groups:
8364 * - regular: there are !numa tasks
8365 * - remote: there are numa tasks that run on the 'wrong' node
8366 * - all: there is no distinction
8368 * In order to avoid migrating ideally placed numa tasks,
8369 * ignore those when there's better options.
8371 * If we ignore the actual busiest queue to migrate another
8372 * task, the next balance pass can still reduce the busiest
8373 * queue by moving tasks around inside the node.
8375 * If we cannot move enough load due to this classification
8376 * the next pass will adjust the group classification and
8377 * allow migration of more tasks.
8379 * Both cases only affect the total convergence complexity.
8381 if (rt > env->fbq_type)
8384 capacity = capacity_of(i);
8386 wl = weighted_cpuload(rq);
8389 * When comparing with imbalance, use weighted_cpuload()
8390 * which is not scaled with the CPU capacity.
8393 if (rq->nr_running == 1 && wl > env->imbalance &&
8394 !check_cpu_capacity(rq, env->sd))
8398 * For the load comparisons with the other CPU's, consider
8399 * the weighted_cpuload() scaled with the CPU capacity, so
8400 * that the load can be moved away from the CPU that is
8401 * potentially running at a lower capacity.
8403 * Thus we're looking for max(wl_i / capacity_i), crosswise
8404 * multiplication to rid ourselves of the division works out
8405 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8406 * our previous maximum.
8408 if (wl * busiest_capacity > busiest_load * capacity) {
8410 busiest_capacity = capacity;
8419 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8420 * so long as it is large enough.
8422 #define MAX_PINNED_INTERVAL 512
8424 static int need_active_balance(struct lb_env *env)
8426 struct sched_domain *sd = env->sd;
8428 if (env->idle == CPU_NEWLY_IDLE) {
8431 * ASYM_PACKING needs to force migrate tasks from busy but
8432 * lower priority CPUs in order to pack all tasks in the
8433 * highest priority CPUs.
8435 if ((sd->flags & SD_ASYM_PACKING) &&
8436 sched_asym_prefer(env->dst_cpu, env->src_cpu))
8441 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8442 * It's worth migrating the task if the src_cpu's capacity is reduced
8443 * because of other sched_class or IRQs if more capacity stays
8444 * available on dst_cpu.
8446 if ((env->idle != CPU_NOT_IDLE) &&
8447 (env->src_rq->cfs.h_nr_running == 1)) {
8448 if ((check_cpu_capacity(env->src_rq, sd)) &&
8449 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8453 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8456 static int active_load_balance_cpu_stop(void *data);
8458 static int should_we_balance(struct lb_env *env)
8460 struct sched_group *sg = env->sd->groups;
8461 int cpu, balance_cpu = -1;
8464 * Ensure the balancing environment is consistent; can happen
8465 * when the softirq triggers 'during' hotplug.
8467 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8471 * In the newly idle case, we will allow all the CPUs
8472 * to do the newly idle load balance.
8474 if (env->idle == CPU_NEWLY_IDLE)
8477 /* Try to find first idle CPU */
8478 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8486 if (balance_cpu == -1)
8487 balance_cpu = group_balance_cpu(sg);
8490 * First idle CPU or the first CPU(busiest) in this sched group
8491 * is eligible for doing load balancing at this and above domains.
8493 return balance_cpu == env->dst_cpu;
8497 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8498 * tasks if there is an imbalance.
8500 static int load_balance(int this_cpu, struct rq *this_rq,
8501 struct sched_domain *sd, enum cpu_idle_type idle,
8502 int *continue_balancing)
8504 int ld_moved, cur_ld_moved, active_balance = 0;
8505 struct sched_domain *sd_parent = sd->parent;
8506 struct sched_group *group;
8509 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8511 struct lb_env env = {
8513 .dst_cpu = this_cpu,
8515 .dst_grpmask = sched_group_span(sd->groups),
8517 .loop_break = sched_nr_migrate_break,
8520 .tasks = LIST_HEAD_INIT(env.tasks),
8523 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8525 schedstat_inc(sd->lb_count[idle]);
8528 if (!should_we_balance(&env)) {
8529 *continue_balancing = 0;
8533 group = find_busiest_group(&env);
8535 schedstat_inc(sd->lb_nobusyg[idle]);
8539 busiest = find_busiest_queue(&env, group);
8541 schedstat_inc(sd->lb_nobusyq[idle]);
8545 BUG_ON(busiest == env.dst_rq);
8547 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8549 env.src_cpu = busiest->cpu;
8550 env.src_rq = busiest;
8553 if (busiest->nr_running > 1) {
8555 * Attempt to move tasks. If find_busiest_group has found
8556 * an imbalance but busiest->nr_running <= 1, the group is
8557 * still unbalanced. ld_moved simply stays zero, so it is
8558 * correctly treated as an imbalance.
8560 env.flags |= LBF_ALL_PINNED;
8561 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8564 rq_lock_irqsave(busiest, &rf);
8565 update_rq_clock(busiest);
8568 * cur_ld_moved - load moved in current iteration
8569 * ld_moved - cumulative load moved across iterations
8571 cur_ld_moved = detach_tasks(&env);
8574 * We've detached some tasks from busiest_rq. Every
8575 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8576 * unlock busiest->lock, and we are able to be sure
8577 * that nobody can manipulate the tasks in parallel.
8578 * See task_rq_lock() family for the details.
8581 rq_unlock(busiest, &rf);
8585 ld_moved += cur_ld_moved;
8588 local_irq_restore(rf.flags);
8590 if (env.flags & LBF_NEED_BREAK) {
8591 env.flags &= ~LBF_NEED_BREAK;
8596 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8597 * us and move them to an alternate dst_cpu in our sched_group
8598 * where they can run. The upper limit on how many times we
8599 * iterate on same src_cpu is dependent on number of CPUs in our
8602 * This changes load balance semantics a bit on who can move
8603 * load to a given_cpu. In addition to the given_cpu itself
8604 * (or a ilb_cpu acting on its behalf where given_cpu is
8605 * nohz-idle), we now have balance_cpu in a position to move
8606 * load to given_cpu. In rare situations, this may cause
8607 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8608 * _independently_ and at _same_ time to move some load to
8609 * given_cpu) causing exceess load to be moved to given_cpu.
8610 * This however should not happen so much in practice and
8611 * moreover subsequent load balance cycles should correct the
8612 * excess load moved.
8614 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8616 /* Prevent to re-select dst_cpu via env's CPUs */
8617 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8619 env.dst_rq = cpu_rq(env.new_dst_cpu);
8620 env.dst_cpu = env.new_dst_cpu;
8621 env.flags &= ~LBF_DST_PINNED;
8623 env.loop_break = sched_nr_migrate_break;
8626 * Go back to "more_balance" rather than "redo" since we
8627 * need to continue with same src_cpu.
8633 * We failed to reach balance because of affinity.
8636 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8638 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8639 *group_imbalance = 1;
8642 /* All tasks on this runqueue were pinned by CPU affinity */
8643 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8644 cpumask_clear_cpu(cpu_of(busiest), cpus);
8646 * Attempting to continue load balancing at the current
8647 * sched_domain level only makes sense if there are
8648 * active CPUs remaining as possible busiest CPUs to
8649 * pull load from which are not contained within the
8650 * destination group that is receiving any migrated
8653 if (!cpumask_subset(cpus, env.dst_grpmask)) {
8655 env.loop_break = sched_nr_migrate_break;
8658 goto out_all_pinned;
8663 schedstat_inc(sd->lb_failed[idle]);
8665 * Increment the failure counter only on periodic balance.
8666 * We do not want newidle balance, which can be very
8667 * frequent, pollute the failure counter causing
8668 * excessive cache_hot migrations and active balances.
8670 if (idle != CPU_NEWLY_IDLE)
8671 sd->nr_balance_failed++;
8673 if (need_active_balance(&env)) {
8674 unsigned long flags;
8676 raw_spin_lock_irqsave(&busiest->lock, flags);
8679 * Don't kick the active_load_balance_cpu_stop,
8680 * if the curr task on busiest CPU can't be
8681 * moved to this_cpu:
8683 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
8684 raw_spin_unlock_irqrestore(&busiest->lock,
8686 env.flags |= LBF_ALL_PINNED;
8687 goto out_one_pinned;
8691 * ->active_balance synchronizes accesses to
8692 * ->active_balance_work. Once set, it's cleared
8693 * only after active load balance is finished.
8695 if (!busiest->active_balance) {
8696 busiest->active_balance = 1;
8697 busiest->push_cpu = this_cpu;
8700 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8702 if (active_balance) {
8703 stop_one_cpu_nowait(cpu_of(busiest),
8704 active_load_balance_cpu_stop, busiest,
8705 &busiest->active_balance_work);
8708 /* We've kicked active balancing, force task migration. */
8709 sd->nr_balance_failed = sd->cache_nice_tries+1;
8712 sd->nr_balance_failed = 0;
8714 if (likely(!active_balance)) {
8715 /* We were unbalanced, so reset the balancing interval */
8716 sd->balance_interval = sd->min_interval;
8719 * If we've begun active balancing, start to back off. This
8720 * case may not be covered by the all_pinned logic if there
8721 * is only 1 task on the busy runqueue (because we don't call
8724 if (sd->balance_interval < sd->max_interval)
8725 sd->balance_interval *= 2;
8732 * We reach balance although we may have faced some affinity
8733 * constraints. Clear the imbalance flag if it was set.
8736 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8738 if (*group_imbalance)
8739 *group_imbalance = 0;
8744 * We reach balance because all tasks are pinned at this level so
8745 * we can't migrate them. Let the imbalance flag set so parent level
8746 * can try to migrate them.
8748 schedstat_inc(sd->lb_balanced[idle]);
8750 sd->nr_balance_failed = 0;
8753 /* tune up the balancing interval */
8754 if (((env.flags & LBF_ALL_PINNED) &&
8755 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8756 (sd->balance_interval < sd->max_interval))
8757 sd->balance_interval *= 2;
8764 static inline unsigned long
8765 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8767 unsigned long interval = sd->balance_interval;
8770 interval *= sd->busy_factor;
8772 /* scale ms to jiffies */
8773 interval = msecs_to_jiffies(interval);
8774 interval = clamp(interval, 1UL, max_load_balance_interval);
8780 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
8782 unsigned long interval, next;
8784 /* used by idle balance, so cpu_busy = 0 */
8785 interval = get_sd_balance_interval(sd, 0);
8786 next = sd->last_balance + interval;
8788 if (time_after(*next_balance, next))
8789 *next_balance = next;
8793 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
8794 * running tasks off the busiest CPU onto idle CPUs. It requires at
8795 * least 1 task to be running on each physical CPU where possible, and
8796 * avoids physical / logical imbalances.
8798 static int active_load_balance_cpu_stop(void *data)
8800 struct rq *busiest_rq = data;
8801 int busiest_cpu = cpu_of(busiest_rq);
8802 int target_cpu = busiest_rq->push_cpu;
8803 struct rq *target_rq = cpu_rq(target_cpu);
8804 struct sched_domain *sd;
8805 struct task_struct *p = NULL;
8808 rq_lock_irq(busiest_rq, &rf);
8810 * Between queueing the stop-work and running it is a hole in which
8811 * CPUs can become inactive. We should not move tasks from or to
8814 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
8817 /* Make sure the requested CPU hasn't gone down in the meantime: */
8818 if (unlikely(busiest_cpu != smp_processor_id() ||
8819 !busiest_rq->active_balance))
8822 /* Is there any task to move? */
8823 if (busiest_rq->nr_running <= 1)
8827 * This condition is "impossible", if it occurs
8828 * we need to fix it. Originally reported by
8829 * Bjorn Helgaas on a 128-CPU setup.
8831 BUG_ON(busiest_rq == target_rq);
8833 /* Search for an sd spanning us and the target CPU. */
8835 for_each_domain(target_cpu, sd) {
8836 if ((sd->flags & SD_LOAD_BALANCE) &&
8837 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8842 struct lb_env env = {
8844 .dst_cpu = target_cpu,
8845 .dst_rq = target_rq,
8846 .src_cpu = busiest_rq->cpu,
8847 .src_rq = busiest_rq,
8850 * can_migrate_task() doesn't need to compute new_dst_cpu
8851 * for active balancing. Since we have CPU_IDLE, but no
8852 * @dst_grpmask we need to make that test go away with lying
8855 .flags = LBF_DST_PINNED,
8858 schedstat_inc(sd->alb_count);
8859 update_rq_clock(busiest_rq);
8861 p = detach_one_task(&env);
8863 schedstat_inc(sd->alb_pushed);
8864 /* Active balancing done, reset the failure counter. */
8865 sd->nr_balance_failed = 0;
8867 schedstat_inc(sd->alb_failed);
8872 busiest_rq->active_balance = 0;
8873 rq_unlock(busiest_rq, &rf);
8876 attach_one_task(target_rq, p);
8883 static DEFINE_SPINLOCK(balancing);
8886 * Scale the max load_balance interval with the number of CPUs in the system.
8887 * This trades load-balance latency on larger machines for less cross talk.
8889 void update_max_interval(void)
8891 max_load_balance_interval = HZ*num_online_cpus()/10;
8895 * It checks each scheduling domain to see if it is due to be balanced,
8896 * and initiates a balancing operation if so.
8898 * Balancing parameters are set up in init_sched_domains.
8900 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8902 int continue_balancing = 1;
8904 unsigned long interval;
8905 struct sched_domain *sd;
8906 /* Earliest time when we have to do rebalance again */
8907 unsigned long next_balance = jiffies + 60*HZ;
8908 int update_next_balance = 0;
8909 int need_serialize, need_decay = 0;
8913 for_each_domain(cpu, sd) {
8915 * Decay the newidle max times here because this is a regular
8916 * visit to all the domains. Decay ~1% per second.
8918 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8919 sd->max_newidle_lb_cost =
8920 (sd->max_newidle_lb_cost * 253) / 256;
8921 sd->next_decay_max_lb_cost = jiffies + HZ;
8924 max_cost += sd->max_newidle_lb_cost;
8926 if (!(sd->flags & SD_LOAD_BALANCE))
8930 * Stop the load balance at this level. There is another
8931 * CPU in our sched group which is doing load balancing more
8934 if (!continue_balancing) {
8940 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8942 need_serialize = sd->flags & SD_SERIALIZE;
8943 if (need_serialize) {
8944 if (!spin_trylock(&balancing))
8948 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8949 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8951 * The LBF_DST_PINNED logic could have changed
8952 * env->dst_cpu, so we can't know our idle
8953 * state even if we migrated tasks. Update it.
8955 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8957 sd->last_balance = jiffies;
8958 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8961 spin_unlock(&balancing);
8963 if (time_after(next_balance, sd->last_balance + interval)) {
8964 next_balance = sd->last_balance + interval;
8965 update_next_balance = 1;
8970 * Ensure the rq-wide value also decays but keep it at a
8971 * reasonable floor to avoid funnies with rq->avg_idle.
8973 rq->max_idle_balance_cost =
8974 max((u64)sysctl_sched_migration_cost, max_cost);
8979 * next_balance will be updated only when there is a need.
8980 * When the cpu is attached to null domain for ex, it will not be
8983 if (likely(update_next_balance)) {
8984 rq->next_balance = next_balance;
8986 #ifdef CONFIG_NO_HZ_COMMON
8988 * If this CPU has been elected to perform the nohz idle
8989 * balance. Other idle CPUs have already rebalanced with
8990 * nohz_idle_balance() and nohz.next_balance has been
8991 * updated accordingly. This CPU is now running the idle load
8992 * balance for itself and we need to update the
8993 * nohz.next_balance accordingly.
8995 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8996 nohz.next_balance = rq->next_balance;
9001 static inline int on_null_domain(struct rq *rq)
9003 return unlikely(!rcu_dereference_sched(rq->sd));
9006 #ifdef CONFIG_NO_HZ_COMMON
9008 * idle load balancing details
9009 * - When one of the busy CPUs notice that there may be an idle rebalancing
9010 * needed, they will kick the idle load balancer, which then does idle
9011 * load balancing for all the idle CPUs.
9014 static inline int find_new_ilb(void)
9016 int ilb = cpumask_first(nohz.idle_cpus_mask);
9018 if (ilb < nr_cpu_ids && idle_cpu(ilb))
9025 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
9026 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
9027 * CPU (if there is one).
9029 static void kick_ilb(unsigned int flags)
9033 nohz.next_balance++;
9035 ilb_cpu = find_new_ilb();
9037 if (ilb_cpu >= nr_cpu_ids)
9040 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
9041 if (flags & NOHZ_KICK_MASK)
9045 * Use smp_send_reschedule() instead of resched_cpu().
9046 * This way we generate a sched IPI on the target CPU which
9047 * is idle. And the softirq performing nohz idle load balance
9048 * will be run before returning from the IPI.
9050 smp_send_reschedule(ilb_cpu);
9054 * Current heuristic for kicking the idle load balancer in the presence
9055 * of an idle cpu in the system.
9056 * - This rq has more than one task.
9057 * - This rq has at least one CFS task and the capacity of the CPU is
9058 * significantly reduced because of RT tasks or IRQs.
9059 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9060 * multiple busy cpu.
9061 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9062 * domain span are idle.
9064 static void nohz_balancer_kick(struct rq *rq)
9066 unsigned long now = jiffies;
9067 struct sched_domain_shared *sds;
9068 struct sched_domain *sd;
9069 int nr_busy, i, cpu = rq->cpu;
9070 unsigned int flags = 0;
9072 if (unlikely(rq->idle_balance))
9076 * We may be recently in ticked or tickless idle mode. At the first
9077 * busy tick after returning from idle, we will update the busy stats.
9079 nohz_balance_exit_idle(rq);
9082 * None are in tickless mode and hence no need for NOHZ idle load
9085 if (likely(!atomic_read(&nohz.nr_cpus)))
9088 if (READ_ONCE(nohz.has_blocked) &&
9089 time_after(now, READ_ONCE(nohz.next_blocked)))
9090 flags = NOHZ_STATS_KICK;
9092 if (time_before(now, nohz.next_balance))
9095 if (rq->nr_running >= 2) {
9096 flags = NOHZ_KICK_MASK;
9101 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9104 * XXX: write a coherent comment on why we do this.
9105 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
9107 nr_busy = atomic_read(&sds->nr_busy_cpus);
9109 flags = NOHZ_KICK_MASK;
9115 sd = rcu_dereference(rq->sd);
9117 if ((rq->cfs.h_nr_running >= 1) &&
9118 check_cpu_capacity(rq, sd)) {
9119 flags = NOHZ_KICK_MASK;
9124 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9126 for_each_cpu(i, sched_domain_span(sd)) {
9128 !cpumask_test_cpu(i, nohz.idle_cpus_mask))
9131 if (sched_asym_prefer(i, cpu)) {
9132 flags = NOHZ_KICK_MASK;
9144 static void set_cpu_sd_state_busy(int cpu)
9146 struct sched_domain *sd;
9149 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9151 if (!sd || !sd->nohz_idle)
9155 atomic_inc(&sd->shared->nr_busy_cpus);
9160 void nohz_balance_exit_idle(struct rq *rq)
9162 SCHED_WARN_ON(rq != this_rq());
9164 if (likely(!rq->nohz_tick_stopped))
9167 rq->nohz_tick_stopped = 0;
9168 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
9169 atomic_dec(&nohz.nr_cpus);
9171 set_cpu_sd_state_busy(rq->cpu);
9174 static void set_cpu_sd_state_idle(int cpu)
9176 struct sched_domain *sd;
9179 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9181 if (!sd || sd->nohz_idle)
9185 atomic_dec(&sd->shared->nr_busy_cpus);
9191 * This routine will record that the CPU is going idle with tick stopped.
9192 * This info will be used in performing idle load balancing in the future.
9194 void nohz_balance_enter_idle(int cpu)
9196 struct rq *rq = cpu_rq(cpu);
9198 SCHED_WARN_ON(cpu != smp_processor_id());
9200 /* If this CPU is going down, then nothing needs to be done: */
9201 if (!cpu_active(cpu))
9204 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9205 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9209 * Can be set safely without rq->lock held
9210 * If a clear happens, it will have evaluated last additions because
9211 * rq->lock is held during the check and the clear
9213 rq->has_blocked_load = 1;
9216 * The tick is still stopped but load could have been added in the
9217 * meantime. We set the nohz.has_blocked flag to trig a check of the
9218 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
9219 * of nohz.has_blocked can only happen after checking the new load
9221 if (rq->nohz_tick_stopped)
9224 /* If we're a completely isolated CPU, we don't play: */
9225 if (on_null_domain(rq))
9228 rq->nohz_tick_stopped = 1;
9230 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9231 atomic_inc(&nohz.nr_cpus);
9234 * Ensures that if nohz_idle_balance() fails to observe our
9235 * @idle_cpus_mask store, it must observe the @has_blocked
9238 smp_mb__after_atomic();
9240 set_cpu_sd_state_idle(cpu);
9244 * Each time a cpu enter idle, we assume that it has blocked load and
9245 * enable the periodic update of the load of idle cpus
9247 WRITE_ONCE(nohz.has_blocked, 1);
9251 * Internal function that runs load balance for all idle cpus. The load balance
9252 * can be a simple update of blocked load or a complete load balance with
9253 * tasks movement depending of flags.
9254 * The function returns false if the loop has stopped before running
9255 * through all idle CPUs.
9257 static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
9258 enum cpu_idle_type idle)
9260 /* Earliest time when we have to do rebalance again */
9261 unsigned long now = jiffies;
9262 unsigned long next_balance = now + 60*HZ;
9263 bool has_blocked_load = false;
9264 int update_next_balance = 0;
9265 int this_cpu = this_rq->cpu;
9270 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
9273 * We assume there will be no idle load after this update and clear
9274 * the has_blocked flag. If a cpu enters idle in the mean time, it will
9275 * set the has_blocked flag and trig another update of idle load.
9276 * Because a cpu that becomes idle, is added to idle_cpus_mask before
9277 * setting the flag, we are sure to not clear the state and not
9278 * check the load of an idle cpu.
9280 WRITE_ONCE(nohz.has_blocked, 0);
9283 * Ensures that if we miss the CPU, we must see the has_blocked
9284 * store from nohz_balance_enter_idle().
9288 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9289 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9293 * If this CPU gets work to do, stop the load balancing
9294 * work being done for other CPUs. Next load
9295 * balancing owner will pick it up.
9297 if (need_resched()) {
9298 has_blocked_load = true;
9302 rq = cpu_rq(balance_cpu);
9304 has_blocked_load |= update_nohz_stats(rq, true);
9307 * If time for next balance is due,
9310 if (time_after_eq(jiffies, rq->next_balance)) {
9313 rq_lock_irqsave(rq, &rf);
9314 update_rq_clock(rq);
9315 cpu_load_update_idle(rq);
9316 rq_unlock_irqrestore(rq, &rf);
9318 if (flags & NOHZ_BALANCE_KICK)
9319 rebalance_domains(rq, CPU_IDLE);
9322 if (time_after(next_balance, rq->next_balance)) {
9323 next_balance = rq->next_balance;
9324 update_next_balance = 1;
9328 /* Newly idle CPU doesn't need an update */
9329 if (idle != CPU_NEWLY_IDLE) {
9330 update_blocked_averages(this_cpu);
9331 has_blocked_load |= this_rq->has_blocked_load;
9334 if (flags & NOHZ_BALANCE_KICK)
9335 rebalance_domains(this_rq, CPU_IDLE);
9337 WRITE_ONCE(nohz.next_blocked,
9338 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
9340 /* The full idle balance loop has been done */
9344 /* There is still blocked load, enable periodic update */
9345 if (has_blocked_load)
9346 WRITE_ONCE(nohz.has_blocked, 1);
9349 * next_balance will be updated only when there is a need.
9350 * When the CPU is attached to null domain for ex, it will not be
9353 if (likely(update_next_balance))
9354 nohz.next_balance = next_balance;
9360 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9361 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9363 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9365 int this_cpu = this_rq->cpu;
9368 if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
9371 if (idle != CPU_IDLE) {
9372 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9377 * barrier, pairs with nohz_balance_enter_idle(), ensures ...
9379 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9380 if (!(flags & NOHZ_KICK_MASK))
9383 _nohz_idle_balance(this_rq, flags, idle);
9388 static void nohz_newidle_balance(struct rq *this_rq)
9390 int this_cpu = this_rq->cpu;
9393 * This CPU doesn't want to be disturbed by scheduler
9396 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
9399 /* Will wake up very soon. No time for doing anything else*/
9400 if (this_rq->avg_idle < sysctl_sched_migration_cost)
9403 /* Don't need to update blocked load of idle CPUs*/
9404 if (!READ_ONCE(nohz.has_blocked) ||
9405 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
9408 raw_spin_unlock(&this_rq->lock);
9410 * This CPU is going to be idle and blocked load of idle CPUs
9411 * need to be updated. Run the ilb locally as it is a good
9412 * candidate for ilb instead of waking up another idle CPU.
9413 * Kick an normal ilb if we failed to do the update.
9415 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
9416 kick_ilb(NOHZ_STATS_KICK);
9417 raw_spin_lock(&this_rq->lock);
9420 #else /* !CONFIG_NO_HZ_COMMON */
9421 static inline void nohz_balancer_kick(struct rq *rq) { }
9423 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9428 static inline void nohz_newidle_balance(struct rq *this_rq) { }
9429 #endif /* CONFIG_NO_HZ_COMMON */
9432 * idle_balance is called by schedule() if this_cpu is about to become
9433 * idle. Attempts to pull tasks from other CPUs.
9435 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
9437 unsigned long next_balance = jiffies + HZ;
9438 int this_cpu = this_rq->cpu;
9439 struct sched_domain *sd;
9440 int pulled_task = 0;
9444 * We must set idle_stamp _before_ calling idle_balance(), such that we
9445 * measure the duration of idle_balance() as idle time.
9447 this_rq->idle_stamp = rq_clock(this_rq);
9450 * Do not pull tasks towards !active CPUs...
9452 if (!cpu_active(this_cpu))
9456 * This is OK, because current is on_cpu, which avoids it being picked
9457 * for load-balance and preemption/IRQs are still disabled avoiding
9458 * further scheduler activity on it and we're being very careful to
9459 * re-start the picking loop.
9461 rq_unpin_lock(this_rq, rf);
9463 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
9464 !this_rq->rd->overload) {
9467 sd = rcu_dereference_check_sched_domain(this_rq->sd);
9469 update_next_balance(sd, &next_balance);
9472 nohz_newidle_balance(this_rq);
9477 raw_spin_unlock(&this_rq->lock);
9479 update_blocked_averages(this_cpu);
9481 for_each_domain(this_cpu, sd) {
9482 int continue_balancing = 1;
9483 u64 t0, domain_cost;
9485 if (!(sd->flags & SD_LOAD_BALANCE))
9488 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
9489 update_next_balance(sd, &next_balance);
9493 if (sd->flags & SD_BALANCE_NEWIDLE) {
9494 t0 = sched_clock_cpu(this_cpu);
9496 pulled_task = load_balance(this_cpu, this_rq,
9498 &continue_balancing);
9500 domain_cost = sched_clock_cpu(this_cpu) - t0;
9501 if (domain_cost > sd->max_newidle_lb_cost)
9502 sd->max_newidle_lb_cost = domain_cost;
9504 curr_cost += domain_cost;
9507 update_next_balance(sd, &next_balance);
9510 * Stop searching for tasks to pull if there are
9511 * now runnable tasks on this rq.
9513 if (pulled_task || this_rq->nr_running > 0)
9518 raw_spin_lock(&this_rq->lock);
9520 if (curr_cost > this_rq->max_idle_balance_cost)
9521 this_rq->max_idle_balance_cost = curr_cost;
9525 * While browsing the domains, we released the rq lock, a task could
9526 * have been enqueued in the meantime. Since we're not going idle,
9527 * pretend we pulled a task.
9529 if (this_rq->cfs.h_nr_running && !pulled_task)
9532 /* Move the next balance forward */
9533 if (time_after(this_rq->next_balance, next_balance))
9534 this_rq->next_balance = next_balance;
9536 /* Is there a task of a high priority class? */
9537 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
9541 this_rq->idle_stamp = 0;
9543 rq_repin_lock(this_rq, rf);
9549 * run_rebalance_domains is triggered when needed from the scheduler tick.
9550 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9552 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9554 struct rq *this_rq = this_rq();
9555 enum cpu_idle_type idle = this_rq->idle_balance ?
9556 CPU_IDLE : CPU_NOT_IDLE;
9559 * If this CPU has a pending nohz_balance_kick, then do the
9560 * balancing on behalf of the other idle CPUs whose ticks are
9561 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9562 * give the idle CPUs a chance to load balance. Else we may
9563 * load balance only within the local sched_domain hierarchy
9564 * and abort nohz_idle_balance altogether if we pull some load.
9566 if (nohz_idle_balance(this_rq, idle))
9569 /* normal load balance */
9570 update_blocked_averages(this_rq->cpu);
9571 rebalance_domains(this_rq, idle);
9575 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9577 void trigger_load_balance(struct rq *rq)
9579 /* Don't need to rebalance while attached to NULL domain */
9580 if (unlikely(on_null_domain(rq)))
9583 if (time_after_eq(jiffies, rq->next_balance))
9584 raise_softirq(SCHED_SOFTIRQ);
9586 nohz_balancer_kick(rq);
9589 static void rq_online_fair(struct rq *rq)
9593 update_runtime_enabled(rq);
9596 static void rq_offline_fair(struct rq *rq)
9600 /* Ensure any throttled groups are reachable by pick_next_task */
9601 unthrottle_offline_cfs_rqs(rq);
9604 #endif /* CONFIG_SMP */
9607 * scheduler tick hitting a task of our scheduling class.
9609 * NOTE: This function can be called remotely by the tick offload that
9610 * goes along full dynticks. Therefore no local assumption can be made
9611 * and everything must be accessed through the @rq and @curr passed in
9614 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9616 struct cfs_rq *cfs_rq;
9617 struct sched_entity *se = &curr->se;
9619 for_each_sched_entity(se) {
9620 cfs_rq = cfs_rq_of(se);
9621 entity_tick(cfs_rq, se, queued);
9624 if (static_branch_unlikely(&sched_numa_balancing))
9625 task_tick_numa(rq, curr);
9629 * called on fork with the child task as argument from the parent's context
9630 * - child not yet on the tasklist
9631 * - preemption disabled
9633 static void task_fork_fair(struct task_struct *p)
9635 struct cfs_rq *cfs_rq;
9636 struct sched_entity *se = &p->se, *curr;
9637 struct rq *rq = this_rq();
9641 update_rq_clock(rq);
9643 cfs_rq = task_cfs_rq(current);
9644 curr = cfs_rq->curr;
9646 update_curr(cfs_rq);
9647 se->vruntime = curr->vruntime;
9649 place_entity(cfs_rq, se, 1);
9651 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9653 * Upon rescheduling, sched_class::put_prev_task() will place
9654 * 'current' within the tree based on its new key value.
9656 swap(curr->vruntime, se->vruntime);
9660 se->vruntime -= cfs_rq->min_vruntime;
9665 * Priority of the task has changed. Check to see if we preempt
9669 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9671 if (!task_on_rq_queued(p))
9675 * Reschedule if we are currently running on this runqueue and
9676 * our priority decreased, or if we are not currently running on
9677 * this runqueue and our priority is higher than the current's
9679 if (rq->curr == p) {
9680 if (p->prio > oldprio)
9683 check_preempt_curr(rq, p, 0);
9686 static inline bool vruntime_normalized(struct task_struct *p)
9688 struct sched_entity *se = &p->se;
9691 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9692 * the dequeue_entity(.flags=0) will already have normalized the
9699 * When !on_rq, vruntime of the task has usually NOT been normalized.
9700 * But there are some cases where it has already been normalized:
9702 * - A forked child which is waiting for being woken up by
9703 * wake_up_new_task().
9704 * - A task which has been woken up by try_to_wake_up() and
9705 * waiting for actually being woken up by sched_ttwu_pending().
9707 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9713 #ifdef CONFIG_FAIR_GROUP_SCHED
9715 * Propagate the changes of the sched_entity across the tg tree to make it
9716 * visible to the root
9718 static void propagate_entity_cfs_rq(struct sched_entity *se)
9720 struct cfs_rq *cfs_rq;
9722 /* Start to propagate at parent */
9725 for_each_sched_entity(se) {
9726 cfs_rq = cfs_rq_of(se);
9728 if (cfs_rq_throttled(cfs_rq))
9731 update_load_avg(cfs_rq, se, UPDATE_TG);
9735 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
9738 static void detach_entity_cfs_rq(struct sched_entity *se)
9740 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9742 /* Catch up with the cfs_rq and remove our load when we leave */
9743 update_load_avg(cfs_rq, se, 0);
9744 detach_entity_load_avg(cfs_rq, se);
9745 update_tg_load_avg(cfs_rq, false);
9746 propagate_entity_cfs_rq(se);
9749 static void attach_entity_cfs_rq(struct sched_entity *se)
9751 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9753 #ifdef CONFIG_FAIR_GROUP_SCHED
9755 * Since the real-depth could have been changed (only FAIR
9756 * class maintain depth value), reset depth properly.
9758 se->depth = se->parent ? se->parent->depth + 1 : 0;
9761 /* Synchronize entity with its cfs_rq */
9762 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
9763 attach_entity_load_avg(cfs_rq, se, 0);
9764 update_tg_load_avg(cfs_rq, false);
9765 propagate_entity_cfs_rq(se);
9768 static void detach_task_cfs_rq(struct task_struct *p)
9770 struct sched_entity *se = &p->se;
9771 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9773 if (!vruntime_normalized(p)) {
9775 * Fix up our vruntime so that the current sleep doesn't
9776 * cause 'unlimited' sleep bonus.
9778 place_entity(cfs_rq, se, 0);
9779 se->vruntime -= cfs_rq->min_vruntime;
9782 detach_entity_cfs_rq(se);
9785 static void attach_task_cfs_rq(struct task_struct *p)
9787 struct sched_entity *se = &p->se;
9788 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9790 attach_entity_cfs_rq(se);
9792 if (!vruntime_normalized(p))
9793 se->vruntime += cfs_rq->min_vruntime;
9796 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9798 detach_task_cfs_rq(p);
9801 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9803 attach_task_cfs_rq(p);
9805 if (task_on_rq_queued(p)) {
9807 * We were most likely switched from sched_rt, so
9808 * kick off the schedule if running, otherwise just see
9809 * if we can still preempt the current task.
9814 check_preempt_curr(rq, p, 0);
9818 /* Account for a task changing its policy or group.
9820 * This routine is mostly called to set cfs_rq->curr field when a task
9821 * migrates between groups/classes.
9823 static void set_curr_task_fair(struct rq *rq)
9825 struct sched_entity *se = &rq->curr->se;
9827 for_each_sched_entity(se) {
9828 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9830 set_next_entity(cfs_rq, se);
9831 /* ensure bandwidth has been allocated on our new cfs_rq */
9832 account_cfs_rq_runtime(cfs_rq, 0);
9836 void init_cfs_rq(struct cfs_rq *cfs_rq)
9838 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
9839 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9840 #ifndef CONFIG_64BIT
9841 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9844 raw_spin_lock_init(&cfs_rq->removed.lock);
9848 #ifdef CONFIG_FAIR_GROUP_SCHED
9849 static void task_set_group_fair(struct task_struct *p)
9851 struct sched_entity *se = &p->se;
9853 set_task_rq(p, task_cpu(p));
9854 se->depth = se->parent ? se->parent->depth + 1 : 0;
9857 static void task_move_group_fair(struct task_struct *p)
9859 detach_task_cfs_rq(p);
9860 set_task_rq(p, task_cpu(p));
9863 /* Tell se's cfs_rq has been changed -- migrated */
9864 p->se.avg.last_update_time = 0;
9866 attach_task_cfs_rq(p);
9869 static void task_change_group_fair(struct task_struct *p, int type)
9872 case TASK_SET_GROUP:
9873 task_set_group_fair(p);
9876 case TASK_MOVE_GROUP:
9877 task_move_group_fair(p);
9882 void free_fair_sched_group(struct task_group *tg)
9886 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9888 for_each_possible_cpu(i) {
9890 kfree(tg->cfs_rq[i]);
9899 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9901 struct sched_entity *se;
9902 struct cfs_rq *cfs_rq;
9905 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
9908 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
9912 tg->shares = NICE_0_LOAD;
9914 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9916 for_each_possible_cpu(i) {
9917 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9918 GFP_KERNEL, cpu_to_node(i));
9922 se = kzalloc_node(sizeof(struct sched_entity),
9923 GFP_KERNEL, cpu_to_node(i));
9927 init_cfs_rq(cfs_rq);
9928 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9929 init_entity_runnable_average(se);
9940 void online_fair_sched_group(struct task_group *tg)
9942 struct sched_entity *se;
9946 for_each_possible_cpu(i) {
9950 raw_spin_lock_irq(&rq->lock);
9951 update_rq_clock(rq);
9952 attach_entity_cfs_rq(se);
9953 sync_throttle(tg, i);
9954 raw_spin_unlock_irq(&rq->lock);
9958 void unregister_fair_sched_group(struct task_group *tg)
9960 unsigned long flags;
9964 for_each_possible_cpu(cpu) {
9966 remove_entity_load_avg(tg->se[cpu]);
9969 * Only empty task groups can be destroyed; so we can speculatively
9970 * check on_list without danger of it being re-added.
9972 if (!tg->cfs_rq[cpu]->on_list)
9977 raw_spin_lock_irqsave(&rq->lock, flags);
9978 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9979 raw_spin_unlock_irqrestore(&rq->lock, flags);
9983 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9984 struct sched_entity *se, int cpu,
9985 struct sched_entity *parent)
9987 struct rq *rq = cpu_rq(cpu);
9991 init_cfs_rq_runtime(cfs_rq);
9993 tg->cfs_rq[cpu] = cfs_rq;
9996 /* se could be NULL for root_task_group */
10001 se->cfs_rq = &rq->cfs;
10004 se->cfs_rq = parent->my_q;
10005 se->depth = parent->depth + 1;
10009 /* guarantee group entities always have weight */
10010 update_load_set(&se->load, NICE_0_LOAD);
10011 se->parent = parent;
10014 static DEFINE_MUTEX(shares_mutex);
10016 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10021 * We can't change the weight of the root cgroup.
10026 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10028 mutex_lock(&shares_mutex);
10029 if (tg->shares == shares)
10032 tg->shares = shares;
10033 for_each_possible_cpu(i) {
10034 struct rq *rq = cpu_rq(i);
10035 struct sched_entity *se = tg->se[i];
10036 struct rq_flags rf;
10038 /* Propagate contribution to hierarchy */
10039 rq_lock_irqsave(rq, &rf);
10040 update_rq_clock(rq);
10041 for_each_sched_entity(se) {
10042 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
10043 update_cfs_group(se);
10045 rq_unlock_irqrestore(rq, &rf);
10049 mutex_unlock(&shares_mutex);
10052 #else /* CONFIG_FAIR_GROUP_SCHED */
10054 void free_fair_sched_group(struct task_group *tg) { }
10056 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10061 void online_fair_sched_group(struct task_group *tg) { }
10063 void unregister_fair_sched_group(struct task_group *tg) { }
10065 #endif /* CONFIG_FAIR_GROUP_SCHED */
10068 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10070 struct sched_entity *se = &task->se;
10071 unsigned int rr_interval = 0;
10074 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10077 if (rq->cfs.load.weight)
10078 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10080 return rr_interval;
10084 * All the scheduling class methods:
10086 const struct sched_class fair_sched_class = {
10087 .next = &idle_sched_class,
10088 .enqueue_task = enqueue_task_fair,
10089 .dequeue_task = dequeue_task_fair,
10090 .yield_task = yield_task_fair,
10091 .yield_to_task = yield_to_task_fair,
10093 .check_preempt_curr = check_preempt_wakeup,
10095 .pick_next_task = pick_next_task_fair,
10096 .put_prev_task = put_prev_task_fair,
10099 .select_task_rq = select_task_rq_fair,
10100 .migrate_task_rq = migrate_task_rq_fair,
10102 .rq_online = rq_online_fair,
10103 .rq_offline = rq_offline_fair,
10105 .task_dead = task_dead_fair,
10106 .set_cpus_allowed = set_cpus_allowed_common,
10109 .set_curr_task = set_curr_task_fair,
10110 .task_tick = task_tick_fair,
10111 .task_fork = task_fork_fair,
10113 .prio_changed = prio_changed_fair,
10114 .switched_from = switched_from_fair,
10115 .switched_to = switched_to_fair,
10117 .get_rr_interval = get_rr_interval_fair,
10119 .update_curr = update_curr_fair,
10121 #ifdef CONFIG_FAIR_GROUP_SCHED
10122 .task_change_group = task_change_group_fair,
10126 #ifdef CONFIG_SCHED_DEBUG
10127 void print_cfs_stats(struct seq_file *m, int cpu)
10129 struct cfs_rq *cfs_rq, *pos;
10132 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
10133 print_cfs_rq(m, cpu, cfs_rq);
10137 #ifdef CONFIG_NUMA_BALANCING
10138 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10141 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10143 for_each_online_node(node) {
10144 if (p->numa_faults) {
10145 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10146 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10148 if (p->numa_group) {
10149 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
10150 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
10152 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10155 #endif /* CONFIG_NUMA_BALANCING */
10156 #endif /* CONFIG_SCHED_DEBUG */
10158 __init void init_sched_fair_class(void)
10161 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10163 #ifdef CONFIG_NO_HZ_COMMON
10164 nohz.next_balance = jiffies;
10165 nohz.next_blocked = jiffies;
10166 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);