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);
1395 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1398 * Allow first faults or private faults to migrate immediately early in
1399 * the lifetime of a task. The magic number 4 is based on waiting for
1400 * two full passes of the "multi-stage node selection" test that is
1403 if ((p->numa_preferred_nid == -1 || p->numa_scan_seq <= 4) &&
1404 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1408 * Multi-stage node selection is used in conjunction with a periodic
1409 * migration fault to build a temporal task<->page relation. By using
1410 * a two-stage filter we remove short/unlikely relations.
1412 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1413 * a task's usage of a particular page (n_p) per total usage of this
1414 * page (n_t) (in a given time-span) to a probability.
1416 * Our periodic faults will sample this probability and getting the
1417 * same result twice in a row, given these samples are fully
1418 * independent, is then given by P(n)^2, provided our sample period
1419 * is sufficiently short compared to the usage pattern.
1421 * This quadric squishes small probabilities, making it less likely we
1422 * act on an unlikely task<->page relation.
1424 if (!cpupid_pid_unset(last_cpupid) &&
1425 cpupid_to_nid(last_cpupid) != dst_nid)
1428 /* Always allow migrate on private faults */
1429 if (cpupid_match_pid(p, last_cpupid))
1432 /* A shared fault, but p->numa_group has not been set up yet. */
1437 * Destination node is much more heavily used than the source
1438 * node? Allow migration.
1440 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1441 ACTIVE_NODE_FRACTION)
1445 * Distribute memory according to CPU & memory use on each node,
1446 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1448 * faults_cpu(dst) 3 faults_cpu(src)
1449 * --------------- * - > ---------------
1450 * faults_mem(dst) 4 faults_mem(src)
1452 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1453 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1456 static unsigned long weighted_cpuload(struct rq *rq);
1457 static unsigned long source_load(int cpu, int type);
1458 static unsigned long target_load(int cpu, int type);
1459 static unsigned long capacity_of(int cpu);
1461 /* Cached statistics for all CPUs within a node */
1465 /* Total compute capacity of CPUs on a node */
1466 unsigned long compute_capacity;
1468 unsigned int nr_running;
1472 * XXX borrowed from update_sg_lb_stats
1474 static void update_numa_stats(struct numa_stats *ns, int nid)
1476 int smt, cpu, cpus = 0;
1477 unsigned long capacity;
1479 memset(ns, 0, sizeof(*ns));
1480 for_each_cpu(cpu, cpumask_of_node(nid)) {
1481 struct rq *rq = cpu_rq(cpu);
1483 ns->nr_running += rq->nr_running;
1484 ns->load += weighted_cpuload(rq);
1485 ns->compute_capacity += capacity_of(cpu);
1491 * If we raced with hotplug and there are no CPUs left in our mask
1492 * the @ns structure is NULL'ed and task_numa_compare() will
1493 * not find this node attractive.
1495 * We'll detect a huge imbalance and bail there.
1500 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1501 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1502 capacity = cpus / smt; /* cores */
1504 capacity = min_t(unsigned, capacity,
1505 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1508 struct task_numa_env {
1509 struct task_struct *p;
1511 int src_cpu, src_nid;
1512 int dst_cpu, dst_nid;
1514 struct numa_stats src_stats, dst_stats;
1519 struct task_struct *best_task;
1524 static void task_numa_assign(struct task_numa_env *env,
1525 struct task_struct *p, long imp)
1527 struct rq *rq = cpu_rq(env->dst_cpu);
1529 /* Bail out if run-queue part of active NUMA balance. */
1530 if (xchg(&rq->numa_migrate_on, 1))
1534 * Clear previous best_cpu/rq numa-migrate flag, since task now
1535 * found a better CPU to move/swap.
1537 if (env->best_cpu != -1) {
1538 rq = cpu_rq(env->best_cpu);
1539 WRITE_ONCE(rq->numa_migrate_on, 0);
1543 put_task_struct(env->best_task);
1548 env->best_imp = imp;
1549 env->best_cpu = env->dst_cpu;
1552 static bool load_too_imbalanced(long src_load, long dst_load,
1553 struct task_numa_env *env)
1556 long orig_src_load, orig_dst_load;
1557 long src_capacity, dst_capacity;
1560 * The load is corrected for the CPU capacity available on each node.
1563 * ------------ vs ---------
1564 * src_capacity dst_capacity
1566 src_capacity = env->src_stats.compute_capacity;
1567 dst_capacity = env->dst_stats.compute_capacity;
1569 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1571 orig_src_load = env->src_stats.load;
1572 orig_dst_load = env->dst_stats.load;
1574 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1576 /* Would this change make things worse? */
1577 return (imb > old_imb);
1581 * Maximum NUMA importance can be 1998 (2*999);
1582 * SMALLIMP @ 30 would be close to 1998/64.
1583 * Used to deter task migration.
1588 * This checks if the overall compute and NUMA accesses of the system would
1589 * be improved if the source tasks was migrated to the target dst_cpu taking
1590 * into account that it might be best if task running on the dst_cpu should
1591 * be exchanged with the source task
1593 static void task_numa_compare(struct task_numa_env *env,
1594 long taskimp, long groupimp, bool maymove)
1596 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1597 struct task_struct *cur;
1598 long src_load, dst_load;
1600 long imp = env->p->numa_group ? groupimp : taskimp;
1602 int dist = env->dist;
1604 if (READ_ONCE(dst_rq->numa_migrate_on))
1608 cur = task_rcu_dereference(&dst_rq->curr);
1609 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1613 * Because we have preemption enabled we can get migrated around and
1614 * end try selecting ourselves (current == env->p) as a swap candidate.
1620 if (maymove && moveimp >= env->best_imp)
1627 * "imp" is the fault differential for the source task between the
1628 * source and destination node. Calculate the total differential for
1629 * the source task and potential destination task. The more negative
1630 * the value is, the more remote accesses that would be expected to
1631 * be incurred if the tasks were swapped.
1633 /* Skip this swap candidate if cannot move to the source cpu */
1634 if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed))
1638 * If dst and source tasks are in the same NUMA group, or not
1639 * in any group then look only at task weights.
1641 if (cur->numa_group == env->p->numa_group) {
1642 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1643 task_weight(cur, env->dst_nid, dist);
1645 * Add some hysteresis to prevent swapping the
1646 * tasks within a group over tiny differences.
1648 if (cur->numa_group)
1652 * Compare the group weights. If a task is all by itself
1653 * (not part of a group), use the task weight instead.
1655 if (cur->numa_group && env->p->numa_group)
1656 imp += group_weight(cur, env->src_nid, dist) -
1657 group_weight(cur, env->dst_nid, dist);
1659 imp += task_weight(cur, env->src_nid, dist) -
1660 task_weight(cur, env->dst_nid, dist);
1663 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1670 * If the NUMA importance is less than SMALLIMP,
1671 * task migration might only result in ping pong
1672 * of tasks and also hurt performance due to cache
1675 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1679 * In the overloaded case, try and keep the load balanced.
1681 load = task_h_load(env->p) - task_h_load(cur);
1685 dst_load = env->dst_stats.load + load;
1686 src_load = env->src_stats.load - load;
1688 if (load_too_imbalanced(src_load, dst_load, env))
1693 * One idle CPU per node is evaluated for a task numa move.
1694 * Call select_idle_sibling to maybe find a better one.
1698 * select_idle_siblings() uses an per-CPU cpumask that
1699 * can be used from IRQ context.
1701 local_irq_disable();
1702 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1707 task_numa_assign(env, cur, imp);
1712 static void task_numa_find_cpu(struct task_numa_env *env,
1713 long taskimp, long groupimp)
1715 long src_load, dst_load, load;
1716 bool maymove = false;
1719 load = task_h_load(env->p);
1720 dst_load = env->dst_stats.load + load;
1721 src_load = env->src_stats.load - load;
1724 * If the improvement from just moving env->p direction is better
1725 * than swapping tasks around, check if a move is possible.
1727 maymove = !load_too_imbalanced(src_load, dst_load, env);
1729 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1730 /* Skip this CPU if the source task cannot migrate */
1731 if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
1735 task_numa_compare(env, taskimp, groupimp, maymove);
1739 static int task_numa_migrate(struct task_struct *p)
1741 struct task_numa_env env = {
1744 .src_cpu = task_cpu(p),
1745 .src_nid = task_node(p),
1747 .imbalance_pct = 112,
1753 struct sched_domain *sd;
1755 unsigned long taskweight, groupweight;
1757 long taskimp, groupimp;
1760 * Pick the lowest SD_NUMA domain, as that would have the smallest
1761 * imbalance and would be the first to start moving tasks about.
1763 * And we want to avoid any moving of tasks about, as that would create
1764 * random movement of tasks -- counter the numa conditions we're trying
1768 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1770 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1774 * Cpusets can break the scheduler domain tree into smaller
1775 * balance domains, some of which do not cross NUMA boundaries.
1776 * Tasks that are "trapped" in such domains cannot be migrated
1777 * elsewhere, so there is no point in (re)trying.
1779 if (unlikely(!sd)) {
1780 sched_setnuma(p, task_node(p));
1784 env.dst_nid = p->numa_preferred_nid;
1785 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1786 taskweight = task_weight(p, env.src_nid, dist);
1787 groupweight = group_weight(p, env.src_nid, dist);
1788 update_numa_stats(&env.src_stats, env.src_nid);
1789 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1790 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1791 update_numa_stats(&env.dst_stats, env.dst_nid);
1793 /* Try to find a spot on the preferred nid. */
1794 task_numa_find_cpu(&env, taskimp, groupimp);
1797 * Look at other nodes in these cases:
1798 * - there is no space available on the preferred_nid
1799 * - the task is part of a numa_group that is interleaved across
1800 * multiple NUMA nodes; in order to better consolidate the group,
1801 * we need to check other locations.
1803 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1804 for_each_online_node(nid) {
1805 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1808 dist = node_distance(env.src_nid, env.dst_nid);
1809 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1811 taskweight = task_weight(p, env.src_nid, dist);
1812 groupweight = group_weight(p, env.src_nid, dist);
1815 /* Only consider nodes where both task and groups benefit */
1816 taskimp = task_weight(p, nid, dist) - taskweight;
1817 groupimp = group_weight(p, nid, dist) - groupweight;
1818 if (taskimp < 0 && groupimp < 0)
1823 update_numa_stats(&env.dst_stats, env.dst_nid);
1824 task_numa_find_cpu(&env, taskimp, groupimp);
1829 * If the task is part of a workload that spans multiple NUMA nodes,
1830 * and is migrating into one of the workload's active nodes, remember
1831 * this node as the task's preferred numa node, so the workload can
1833 * A task that migrated to a second choice node will be better off
1834 * trying for a better one later. Do not set the preferred node here.
1836 if (p->numa_group) {
1837 if (env.best_cpu == -1)
1840 nid = cpu_to_node(env.best_cpu);
1842 if (nid != p->numa_preferred_nid)
1843 sched_setnuma(p, nid);
1846 /* No better CPU than the current one was found. */
1847 if (env.best_cpu == -1)
1850 best_rq = cpu_rq(env.best_cpu);
1851 if (env.best_task == NULL) {
1852 ret = migrate_task_to(p, env.best_cpu);
1853 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1855 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1859 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
1860 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1863 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1864 put_task_struct(env.best_task);
1868 /* Attempt to migrate a task to a CPU on the preferred node. */
1869 static void numa_migrate_preferred(struct task_struct *p)
1871 unsigned long interval = HZ;
1873 /* This task has no NUMA fault statistics yet */
1874 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1877 /* Periodically retry migrating the task to the preferred node */
1878 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1879 p->numa_migrate_retry = jiffies + interval;
1881 /* Success if task is already running on preferred CPU */
1882 if (task_node(p) == p->numa_preferred_nid)
1885 /* Otherwise, try migrate to a CPU on the preferred node */
1886 task_numa_migrate(p);
1890 * Find out how many nodes on the workload is actively running on. Do this by
1891 * tracking the nodes from which NUMA hinting faults are triggered. This can
1892 * be different from the set of nodes where the workload's memory is currently
1895 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1897 unsigned long faults, max_faults = 0;
1898 int nid, active_nodes = 0;
1900 for_each_online_node(nid) {
1901 faults = group_faults_cpu(numa_group, nid);
1902 if (faults > max_faults)
1903 max_faults = faults;
1906 for_each_online_node(nid) {
1907 faults = group_faults_cpu(numa_group, nid);
1908 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1912 numa_group->max_faults_cpu = max_faults;
1913 numa_group->active_nodes = active_nodes;
1917 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1918 * increments. The more local the fault statistics are, the higher the scan
1919 * period will be for the next scan window. If local/(local+remote) ratio is
1920 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1921 * the scan period will decrease. Aim for 70% local accesses.
1923 #define NUMA_PERIOD_SLOTS 10
1924 #define NUMA_PERIOD_THRESHOLD 7
1927 * Increase the scan period (slow down scanning) if the majority of
1928 * our memory is already on our local node, or if the majority of
1929 * the page accesses are shared with other processes.
1930 * Otherwise, decrease the scan period.
1932 static void update_task_scan_period(struct task_struct *p,
1933 unsigned long shared, unsigned long private)
1935 unsigned int period_slot;
1936 int lr_ratio, ps_ratio;
1939 unsigned long remote = p->numa_faults_locality[0];
1940 unsigned long local = p->numa_faults_locality[1];
1943 * If there were no record hinting faults then either the task is
1944 * completely idle or all activity is areas that are not of interest
1945 * to automatic numa balancing. Related to that, if there were failed
1946 * migration then it implies we are migrating too quickly or the local
1947 * node is overloaded. In either case, scan slower
1949 if (local + shared == 0 || p->numa_faults_locality[2]) {
1950 p->numa_scan_period = min(p->numa_scan_period_max,
1951 p->numa_scan_period << 1);
1953 p->mm->numa_next_scan = jiffies +
1954 msecs_to_jiffies(p->numa_scan_period);
1960 * Prepare to scale scan period relative to the current period.
1961 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1962 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1963 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1965 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1966 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1967 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1969 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1971 * Most memory accesses are local. There is no need to
1972 * do fast NUMA scanning, since memory is already local.
1974 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
1977 diff = slot * period_slot;
1978 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
1980 * Most memory accesses are shared with other tasks.
1981 * There is no point in continuing fast NUMA scanning,
1982 * since other tasks may just move the memory elsewhere.
1984 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
1987 diff = slot * period_slot;
1990 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
1991 * yet they are not on the local NUMA node. Speed up
1992 * NUMA scanning to get the memory moved over.
1994 int ratio = max(lr_ratio, ps_ratio);
1995 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1998 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1999 task_scan_min(p), task_scan_max(p));
2000 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2004 * Get the fraction of time the task has been running since the last
2005 * NUMA placement cycle. The scheduler keeps similar statistics, but
2006 * decays those on a 32ms period, which is orders of magnitude off
2007 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2008 * stats only if the task is so new there are no NUMA statistics yet.
2010 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2012 u64 runtime, delta, now;
2013 /* Use the start of this time slice to avoid calculations. */
2014 now = p->se.exec_start;
2015 runtime = p->se.sum_exec_runtime;
2017 if (p->last_task_numa_placement) {
2018 delta = runtime - p->last_sum_exec_runtime;
2019 *period = now - p->last_task_numa_placement;
2021 delta = p->se.avg.load_sum;
2022 *period = LOAD_AVG_MAX;
2025 p->last_sum_exec_runtime = runtime;
2026 p->last_task_numa_placement = now;
2032 * Determine the preferred nid for a task in a numa_group. This needs to
2033 * be done in a way that produces consistent results with group_weight,
2034 * otherwise workloads might not converge.
2036 static int preferred_group_nid(struct task_struct *p, int nid)
2041 /* Direct connections between all NUMA nodes. */
2042 if (sched_numa_topology_type == NUMA_DIRECT)
2046 * On a system with glueless mesh NUMA topology, group_weight
2047 * scores nodes according to the number of NUMA hinting faults on
2048 * both the node itself, and on nearby nodes.
2050 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2051 unsigned long score, max_score = 0;
2052 int node, max_node = nid;
2054 dist = sched_max_numa_distance;
2056 for_each_online_node(node) {
2057 score = group_weight(p, node, dist);
2058 if (score > max_score) {
2067 * Finding the preferred nid in a system with NUMA backplane
2068 * interconnect topology is more involved. The goal is to locate
2069 * tasks from numa_groups near each other in the system, and
2070 * untangle workloads from different sides of the system. This requires
2071 * searching down the hierarchy of node groups, recursively searching
2072 * inside the highest scoring group of nodes. The nodemask tricks
2073 * keep the complexity of the search down.
2075 nodes = node_online_map;
2076 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2077 unsigned long max_faults = 0;
2078 nodemask_t max_group = NODE_MASK_NONE;
2081 /* Are there nodes at this distance from each other? */
2082 if (!find_numa_distance(dist))
2085 for_each_node_mask(a, nodes) {
2086 unsigned long faults = 0;
2087 nodemask_t this_group;
2088 nodes_clear(this_group);
2090 /* Sum group's NUMA faults; includes a==b case. */
2091 for_each_node_mask(b, nodes) {
2092 if (node_distance(a, b) < dist) {
2093 faults += group_faults(p, b);
2094 node_set(b, this_group);
2095 node_clear(b, nodes);
2099 /* Remember the top group. */
2100 if (faults > max_faults) {
2101 max_faults = faults;
2102 max_group = this_group;
2104 * subtle: at the smallest distance there is
2105 * just one node left in each "group", the
2106 * winner is the preferred nid.
2111 /* Next round, evaluate the nodes within max_group. */
2119 static void task_numa_placement(struct task_struct *p)
2121 int seq, nid, max_nid = -1;
2122 unsigned long max_faults = 0;
2123 unsigned long fault_types[2] = { 0, 0 };
2124 unsigned long total_faults;
2125 u64 runtime, period;
2126 spinlock_t *group_lock = NULL;
2129 * The p->mm->numa_scan_seq field gets updated without
2130 * exclusive access. Use READ_ONCE() here to ensure
2131 * that the field is read in a single access:
2133 seq = READ_ONCE(p->mm->numa_scan_seq);
2134 if (p->numa_scan_seq == seq)
2136 p->numa_scan_seq = seq;
2137 p->numa_scan_period_max = task_scan_max(p);
2139 total_faults = p->numa_faults_locality[0] +
2140 p->numa_faults_locality[1];
2141 runtime = numa_get_avg_runtime(p, &period);
2143 /* If the task is part of a group prevent parallel updates to group stats */
2144 if (p->numa_group) {
2145 group_lock = &p->numa_group->lock;
2146 spin_lock_irq(group_lock);
2149 /* Find the node with the highest number of faults */
2150 for_each_online_node(nid) {
2151 /* Keep track of the offsets in numa_faults array */
2152 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2153 unsigned long faults = 0, group_faults = 0;
2156 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2157 long diff, f_diff, f_weight;
2159 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2160 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2161 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2162 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2164 /* Decay existing window, copy faults since last scan */
2165 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2166 fault_types[priv] += p->numa_faults[membuf_idx];
2167 p->numa_faults[membuf_idx] = 0;
2170 * Normalize the faults_from, so all tasks in a group
2171 * count according to CPU use, instead of by the raw
2172 * number of faults. Tasks with little runtime have
2173 * little over-all impact on throughput, and thus their
2174 * faults are less important.
2176 f_weight = div64_u64(runtime << 16, period + 1);
2177 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2179 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2180 p->numa_faults[cpubuf_idx] = 0;
2182 p->numa_faults[mem_idx] += diff;
2183 p->numa_faults[cpu_idx] += f_diff;
2184 faults += p->numa_faults[mem_idx];
2185 p->total_numa_faults += diff;
2186 if (p->numa_group) {
2188 * safe because we can only change our own group
2190 * mem_idx represents the offset for a given
2191 * nid and priv in a specific region because it
2192 * is at the beginning of the numa_faults array.
2194 p->numa_group->faults[mem_idx] += diff;
2195 p->numa_group->faults_cpu[mem_idx] += f_diff;
2196 p->numa_group->total_faults += diff;
2197 group_faults += p->numa_group->faults[mem_idx];
2201 if (!p->numa_group) {
2202 if (faults > max_faults) {
2203 max_faults = faults;
2206 } else if (group_faults > max_faults) {
2207 max_faults = group_faults;
2212 if (p->numa_group) {
2213 numa_group_count_active_nodes(p->numa_group);
2214 spin_unlock_irq(group_lock);
2215 max_nid = preferred_group_nid(p, max_nid);
2219 /* Set the new preferred node */
2220 if (max_nid != p->numa_preferred_nid)
2221 sched_setnuma(p, max_nid);
2224 update_task_scan_period(p, fault_types[0], fault_types[1]);
2227 static inline int get_numa_group(struct numa_group *grp)
2229 return atomic_inc_not_zero(&grp->refcount);
2232 static inline void put_numa_group(struct numa_group *grp)
2234 if (atomic_dec_and_test(&grp->refcount))
2235 kfree_rcu(grp, rcu);
2238 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2241 struct numa_group *grp, *my_grp;
2242 struct task_struct *tsk;
2244 int cpu = cpupid_to_cpu(cpupid);
2247 if (unlikely(!p->numa_group)) {
2248 unsigned int size = sizeof(struct numa_group) +
2249 4*nr_node_ids*sizeof(unsigned long);
2251 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2255 atomic_set(&grp->refcount, 1);
2256 grp->active_nodes = 1;
2257 grp->max_faults_cpu = 0;
2258 spin_lock_init(&grp->lock);
2260 /* Second half of the array tracks nids where faults happen */
2261 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2264 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2265 grp->faults[i] = p->numa_faults[i];
2267 grp->total_faults = p->total_numa_faults;
2270 rcu_assign_pointer(p->numa_group, grp);
2274 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2276 if (!cpupid_match_pid(tsk, cpupid))
2279 grp = rcu_dereference(tsk->numa_group);
2283 my_grp = p->numa_group;
2288 * Only join the other group if its bigger; if we're the bigger group,
2289 * the other task will join us.
2291 if (my_grp->nr_tasks > grp->nr_tasks)
2295 * Tie-break on the grp address.
2297 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2300 /* Always join threads in the same process. */
2301 if (tsk->mm == current->mm)
2304 /* Simple filter to avoid false positives due to PID collisions */
2305 if (flags & TNF_SHARED)
2308 /* Update priv based on whether false sharing was detected */
2311 if (join && !get_numa_group(grp))
2319 BUG_ON(irqs_disabled());
2320 double_lock_irq(&my_grp->lock, &grp->lock);
2322 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2323 my_grp->faults[i] -= p->numa_faults[i];
2324 grp->faults[i] += p->numa_faults[i];
2326 my_grp->total_faults -= p->total_numa_faults;
2327 grp->total_faults += p->total_numa_faults;
2332 spin_unlock(&my_grp->lock);
2333 spin_unlock_irq(&grp->lock);
2335 rcu_assign_pointer(p->numa_group, grp);
2337 put_numa_group(my_grp);
2345 void task_numa_free(struct task_struct *p)
2347 struct numa_group *grp = p->numa_group;
2348 void *numa_faults = p->numa_faults;
2349 unsigned long flags;
2353 spin_lock_irqsave(&grp->lock, flags);
2354 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2355 grp->faults[i] -= p->numa_faults[i];
2356 grp->total_faults -= p->total_numa_faults;
2359 spin_unlock_irqrestore(&grp->lock, flags);
2360 RCU_INIT_POINTER(p->numa_group, NULL);
2361 put_numa_group(grp);
2364 p->numa_faults = NULL;
2369 * Got a PROT_NONE fault for a page on @node.
2371 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2373 struct task_struct *p = current;
2374 bool migrated = flags & TNF_MIGRATED;
2375 int cpu_node = task_node(current);
2376 int local = !!(flags & TNF_FAULT_LOCAL);
2377 struct numa_group *ng;
2380 if (!static_branch_likely(&sched_numa_balancing))
2383 /* for example, ksmd faulting in a user's mm */
2387 /* Allocate buffer to track faults on a per-node basis */
2388 if (unlikely(!p->numa_faults)) {
2389 int size = sizeof(*p->numa_faults) *
2390 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2392 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2393 if (!p->numa_faults)
2396 p->total_numa_faults = 0;
2397 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2401 * First accesses are treated as private, otherwise consider accesses
2402 * to be private if the accessing pid has not changed
2404 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2407 priv = cpupid_match_pid(p, last_cpupid);
2408 if (!priv && !(flags & TNF_NO_GROUP))
2409 task_numa_group(p, last_cpupid, flags, &priv);
2413 * If a workload spans multiple NUMA nodes, a shared fault that
2414 * occurs wholly within the set of nodes that the workload is
2415 * actively using should be counted as local. This allows the
2416 * scan rate to slow down when a workload has settled down.
2419 if (!priv && !local && ng && ng->active_nodes > 1 &&
2420 numa_is_active_node(cpu_node, ng) &&
2421 numa_is_active_node(mem_node, ng))
2425 * Retry task to preferred node migration periodically, in case it
2426 * case it previously failed, or the scheduler moved us.
2428 if (time_after(jiffies, p->numa_migrate_retry)) {
2429 task_numa_placement(p);
2430 numa_migrate_preferred(p);
2434 p->numa_pages_migrated += pages;
2435 if (flags & TNF_MIGRATE_FAIL)
2436 p->numa_faults_locality[2] += pages;
2438 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2439 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2440 p->numa_faults_locality[local] += pages;
2443 static void reset_ptenuma_scan(struct task_struct *p)
2446 * We only did a read acquisition of the mmap sem, so
2447 * p->mm->numa_scan_seq is written to without exclusive access
2448 * and the update is not guaranteed to be atomic. That's not
2449 * much of an issue though, since this is just used for
2450 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2451 * expensive, to avoid any form of compiler optimizations:
2453 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2454 p->mm->numa_scan_offset = 0;
2458 * The expensive part of numa migration is done from task_work context.
2459 * Triggered from task_tick_numa().
2461 void task_numa_work(struct callback_head *work)
2463 unsigned long migrate, next_scan, now = jiffies;
2464 struct task_struct *p = current;
2465 struct mm_struct *mm = p->mm;
2466 u64 runtime = p->se.sum_exec_runtime;
2467 struct vm_area_struct *vma;
2468 unsigned long start, end;
2469 unsigned long nr_pte_updates = 0;
2470 long pages, virtpages;
2472 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2474 work->next = work; /* protect against double add */
2476 * Who cares about NUMA placement when they're dying.
2478 * NOTE: make sure not to dereference p->mm before this check,
2479 * exit_task_work() happens _after_ exit_mm() so we could be called
2480 * without p->mm even though we still had it when we enqueued this
2483 if (p->flags & PF_EXITING)
2486 if (!mm->numa_next_scan) {
2487 mm->numa_next_scan = now +
2488 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2492 * Enforce maximal scan/migration frequency..
2494 migrate = mm->numa_next_scan;
2495 if (time_before(now, migrate))
2498 if (p->numa_scan_period == 0) {
2499 p->numa_scan_period_max = task_scan_max(p);
2500 p->numa_scan_period = task_scan_start(p);
2503 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2504 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2508 * Delay this task enough that another task of this mm will likely win
2509 * the next time around.
2511 p->node_stamp += 2 * TICK_NSEC;
2513 start = mm->numa_scan_offset;
2514 pages = sysctl_numa_balancing_scan_size;
2515 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2516 virtpages = pages * 8; /* Scan up to this much virtual space */
2521 if (!down_read_trylock(&mm->mmap_sem))
2523 vma = find_vma(mm, start);
2525 reset_ptenuma_scan(p);
2529 for (; vma; vma = vma->vm_next) {
2530 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2531 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2536 * Shared library pages mapped by multiple processes are not
2537 * migrated as it is expected they are cache replicated. Avoid
2538 * hinting faults in read-only file-backed mappings or the vdso
2539 * as migrating the pages will be of marginal benefit.
2542 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2546 * Skip inaccessible VMAs to avoid any confusion between
2547 * PROT_NONE and NUMA hinting ptes
2549 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2553 start = max(start, vma->vm_start);
2554 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2555 end = min(end, vma->vm_end);
2556 nr_pte_updates = change_prot_numa(vma, start, end);
2559 * Try to scan sysctl_numa_balancing_size worth of
2560 * hpages that have at least one present PTE that
2561 * is not already pte-numa. If the VMA contains
2562 * areas that are unused or already full of prot_numa
2563 * PTEs, scan up to virtpages, to skip through those
2567 pages -= (end - start) >> PAGE_SHIFT;
2568 virtpages -= (end - start) >> PAGE_SHIFT;
2571 if (pages <= 0 || virtpages <= 0)
2575 } while (end != vma->vm_end);
2580 * It is possible to reach the end of the VMA list but the last few
2581 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2582 * would find the !migratable VMA on the next scan but not reset the
2583 * scanner to the start so check it now.
2586 mm->numa_scan_offset = start;
2588 reset_ptenuma_scan(p);
2589 up_read(&mm->mmap_sem);
2592 * Make sure tasks use at least 32x as much time to run other code
2593 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2594 * Usually update_task_scan_period slows down scanning enough; on an
2595 * overloaded system we need to limit overhead on a per task basis.
2597 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2598 u64 diff = p->se.sum_exec_runtime - runtime;
2599 p->node_stamp += 32 * diff;
2604 * Drive the periodic memory faults..
2606 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2608 struct callback_head *work = &curr->numa_work;
2612 * We don't care about NUMA placement if we don't have memory.
2614 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2618 * Using runtime rather than walltime has the dual advantage that
2619 * we (mostly) drive the selection from busy threads and that the
2620 * task needs to have done some actual work before we bother with
2623 now = curr->se.sum_exec_runtime;
2624 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2626 if (now > curr->node_stamp + period) {
2627 if (!curr->node_stamp)
2628 curr->numa_scan_period = task_scan_start(curr);
2629 curr->node_stamp += period;
2631 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2632 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2633 task_work_add(curr, work, true);
2638 static void update_scan_period(struct task_struct *p, int new_cpu)
2640 int src_nid = cpu_to_node(task_cpu(p));
2641 int dst_nid = cpu_to_node(new_cpu);
2643 if (!static_branch_likely(&sched_numa_balancing))
2646 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2649 if (src_nid == dst_nid)
2653 * Allow resets if faults have been trapped before one scan
2654 * has completed. This is most likely due to a new task that
2655 * is pulled cross-node due to wakeups or load balancing.
2657 if (p->numa_scan_seq) {
2659 * Avoid scan adjustments if moving to the preferred
2660 * node or if the task was not previously running on
2661 * the preferred node.
2663 if (dst_nid == p->numa_preferred_nid ||
2664 (p->numa_preferred_nid != -1 && src_nid != p->numa_preferred_nid))
2668 p->numa_scan_period = task_scan_start(p);
2672 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2676 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2680 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2684 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2688 #endif /* CONFIG_NUMA_BALANCING */
2691 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2693 update_load_add(&cfs_rq->load, se->load.weight);
2694 if (!parent_entity(se))
2695 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2697 if (entity_is_task(se)) {
2698 struct rq *rq = rq_of(cfs_rq);
2700 account_numa_enqueue(rq, task_of(se));
2701 list_add(&se->group_node, &rq->cfs_tasks);
2704 cfs_rq->nr_running++;
2708 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2710 update_load_sub(&cfs_rq->load, se->load.weight);
2711 if (!parent_entity(se))
2712 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2714 if (entity_is_task(se)) {
2715 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2716 list_del_init(&se->group_node);
2719 cfs_rq->nr_running--;
2723 * Signed add and clamp on underflow.
2725 * Explicitly do a load-store to ensure the intermediate value never hits
2726 * memory. This allows lockless observations without ever seeing the negative
2729 #define add_positive(_ptr, _val) do { \
2730 typeof(_ptr) ptr = (_ptr); \
2731 typeof(_val) val = (_val); \
2732 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2736 if (val < 0 && res > var) \
2739 WRITE_ONCE(*ptr, res); \
2743 * Unsigned subtract and clamp on underflow.
2745 * Explicitly do a load-store to ensure the intermediate value never hits
2746 * memory. This allows lockless observations without ever seeing the negative
2749 #define sub_positive(_ptr, _val) do { \
2750 typeof(_ptr) ptr = (_ptr); \
2751 typeof(*ptr) val = (_val); \
2752 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2756 WRITE_ONCE(*ptr, res); \
2761 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2763 cfs_rq->runnable_weight += se->runnable_weight;
2765 cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2766 cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2770 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2772 cfs_rq->runnable_weight -= se->runnable_weight;
2774 sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2775 sub_positive(&cfs_rq->avg.runnable_load_sum,
2776 se_runnable(se) * se->avg.runnable_load_sum);
2780 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2782 cfs_rq->avg.load_avg += se->avg.load_avg;
2783 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2787 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2789 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2790 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2794 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2796 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2798 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2800 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2803 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2804 unsigned long weight, unsigned long runnable)
2807 /* commit outstanding execution time */
2808 if (cfs_rq->curr == se)
2809 update_curr(cfs_rq);
2810 account_entity_dequeue(cfs_rq, se);
2811 dequeue_runnable_load_avg(cfs_rq, se);
2813 dequeue_load_avg(cfs_rq, se);
2815 se->runnable_weight = runnable;
2816 update_load_set(&se->load, weight);
2820 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2822 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2823 se->avg.runnable_load_avg =
2824 div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2828 enqueue_load_avg(cfs_rq, se);
2830 account_entity_enqueue(cfs_rq, se);
2831 enqueue_runnable_load_avg(cfs_rq, se);
2835 void reweight_task(struct task_struct *p, int prio)
2837 struct sched_entity *se = &p->se;
2838 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2839 struct load_weight *load = &se->load;
2840 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2842 reweight_entity(cfs_rq, se, weight, weight);
2843 load->inv_weight = sched_prio_to_wmult[prio];
2846 #ifdef CONFIG_FAIR_GROUP_SCHED
2849 * All this does is approximate the hierarchical proportion which includes that
2850 * global sum we all love to hate.
2852 * That is, the weight of a group entity, is the proportional share of the
2853 * group weight based on the group runqueue weights. That is:
2855 * tg->weight * grq->load.weight
2856 * ge->load.weight = ----------------------------- (1)
2857 * \Sum grq->load.weight
2859 * Now, because computing that sum is prohibitively expensive to compute (been
2860 * there, done that) we approximate it with this average stuff. The average
2861 * moves slower and therefore the approximation is cheaper and more stable.
2863 * So instead of the above, we substitute:
2865 * grq->load.weight -> grq->avg.load_avg (2)
2867 * which yields the following:
2869 * tg->weight * grq->avg.load_avg
2870 * ge->load.weight = ------------------------------ (3)
2873 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2875 * That is shares_avg, and it is right (given the approximation (2)).
2877 * The problem with it is that because the average is slow -- it was designed
2878 * to be exactly that of course -- this leads to transients in boundary
2879 * conditions. In specific, the case where the group was idle and we start the
2880 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2881 * yielding bad latency etc..
2883 * Now, in that special case (1) reduces to:
2885 * tg->weight * grq->load.weight
2886 * ge->load.weight = ----------------------------- = tg->weight (4)
2889 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2891 * So what we do is modify our approximation (3) to approach (4) in the (near)
2896 * tg->weight * grq->load.weight
2897 * --------------------------------------------------- (5)
2898 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2900 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2901 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2904 * tg->weight * grq->load.weight
2905 * ge->load.weight = ----------------------------- (6)
2910 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2911 * max(grq->load.weight, grq->avg.load_avg)
2913 * And that is shares_weight and is icky. In the (near) UP case it approaches
2914 * (4) while in the normal case it approaches (3). It consistently
2915 * overestimates the ge->load.weight and therefore:
2917 * \Sum ge->load.weight >= tg->weight
2921 static long calc_group_shares(struct cfs_rq *cfs_rq)
2923 long tg_weight, tg_shares, load, shares;
2924 struct task_group *tg = cfs_rq->tg;
2926 tg_shares = READ_ONCE(tg->shares);
2928 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
2930 tg_weight = atomic_long_read(&tg->load_avg);
2932 /* Ensure tg_weight >= load */
2933 tg_weight -= cfs_rq->tg_load_avg_contrib;
2936 shares = (tg_shares * load);
2938 shares /= tg_weight;
2941 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2942 * of a group with small tg->shares value. It is a floor value which is
2943 * assigned as a minimum load.weight to the sched_entity representing
2944 * the group on a CPU.
2946 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2947 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2948 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2949 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2952 return clamp_t(long, shares, MIN_SHARES, tg_shares);
2956 * This calculates the effective runnable weight for a group entity based on
2957 * the group entity weight calculated above.
2959 * Because of the above approximation (2), our group entity weight is
2960 * an load_avg based ratio (3). This means that it includes blocked load and
2961 * does not represent the runnable weight.
2963 * Approximate the group entity's runnable weight per ratio from the group
2966 * grq->avg.runnable_load_avg
2967 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
2970 * However, analogous to above, since the avg numbers are slow, this leads to
2971 * transients in the from-idle case. Instead we use:
2973 * ge->runnable_weight = ge->load.weight *
2975 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
2976 * ----------------------------------------------------- (8)
2977 * max(grq->avg.load_avg, grq->load.weight)
2979 * Where these max() serve both to use the 'instant' values to fix the slow
2980 * from-idle and avoid the /0 on to-idle, similar to (6).
2982 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
2984 long runnable, load_avg;
2986 load_avg = max(cfs_rq->avg.load_avg,
2987 scale_load_down(cfs_rq->load.weight));
2989 runnable = max(cfs_rq->avg.runnable_load_avg,
2990 scale_load_down(cfs_rq->runnable_weight));
2994 runnable /= load_avg;
2996 return clamp_t(long, runnable, MIN_SHARES, shares);
2998 #endif /* CONFIG_SMP */
3000 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3003 * Recomputes the group entity based on the current state of its group
3006 static void update_cfs_group(struct sched_entity *se)
3008 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3009 long shares, runnable;
3014 if (throttled_hierarchy(gcfs_rq))
3018 runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
3020 if (likely(se->load.weight == shares))
3023 shares = calc_group_shares(gcfs_rq);
3024 runnable = calc_group_runnable(gcfs_rq, shares);
3027 reweight_entity(cfs_rq_of(se), se, shares, runnable);
3030 #else /* CONFIG_FAIR_GROUP_SCHED */
3031 static inline void update_cfs_group(struct sched_entity *se)
3034 #endif /* CONFIG_FAIR_GROUP_SCHED */
3036 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3038 struct rq *rq = rq_of(cfs_rq);
3040 if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) {
3042 * There are a few boundary cases this might miss but it should
3043 * get called often enough that that should (hopefully) not be
3046 * It will not get called when we go idle, because the idle
3047 * thread is a different class (!fair), nor will the utilization
3048 * number include things like RT tasks.
3050 * As is, the util number is not freq-invariant (we'd have to
3051 * implement arch_scale_freq_capacity() for that).
3055 cpufreq_update_util(rq, flags);
3060 #ifdef CONFIG_FAIR_GROUP_SCHED
3062 * update_tg_load_avg - update the tg's load avg
3063 * @cfs_rq: the cfs_rq whose avg changed
3064 * @force: update regardless of how small the difference
3066 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3067 * However, because tg->load_avg is a global value there are performance
3070 * In order to avoid having to look at the other cfs_rq's, we use a
3071 * differential update where we store the last value we propagated. This in
3072 * turn allows skipping updates if the differential is 'small'.
3074 * Updating tg's load_avg is necessary before update_cfs_share().
3076 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3078 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3081 * No need to update load_avg for root_task_group as it is not used.
3083 if (cfs_rq->tg == &root_task_group)
3086 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3087 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3088 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3093 * Called within set_task_rq() right before setting a task's CPU. The
3094 * caller only guarantees p->pi_lock is held; no other assumptions,
3095 * including the state of rq->lock, should be made.
3097 void set_task_rq_fair(struct sched_entity *se,
3098 struct cfs_rq *prev, struct cfs_rq *next)
3100 u64 p_last_update_time;
3101 u64 n_last_update_time;
3103 if (!sched_feat(ATTACH_AGE_LOAD))
3107 * We are supposed to update the task to "current" time, then its up to
3108 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3109 * getting what current time is, so simply throw away the out-of-date
3110 * time. This will result in the wakee task is less decayed, but giving
3111 * the wakee more load sounds not bad.
3113 if (!(se->avg.last_update_time && prev))
3116 #ifndef CONFIG_64BIT
3118 u64 p_last_update_time_copy;
3119 u64 n_last_update_time_copy;
3122 p_last_update_time_copy = prev->load_last_update_time_copy;
3123 n_last_update_time_copy = next->load_last_update_time_copy;
3127 p_last_update_time = prev->avg.last_update_time;
3128 n_last_update_time = next->avg.last_update_time;
3130 } while (p_last_update_time != p_last_update_time_copy ||
3131 n_last_update_time != n_last_update_time_copy);
3134 p_last_update_time = prev->avg.last_update_time;
3135 n_last_update_time = next->avg.last_update_time;
3137 __update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se);
3138 se->avg.last_update_time = n_last_update_time;
3143 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3144 * propagate its contribution. The key to this propagation is the invariant
3145 * that for each group:
3147 * ge->avg == grq->avg (1)
3149 * _IFF_ we look at the pure running and runnable sums. Because they
3150 * represent the very same entity, just at different points in the hierarchy.
3152 * Per the above update_tg_cfs_util() is trivial and simply copies the running
3153 * sum over (but still wrong, because the group entity and group rq do not have
3154 * their PELT windows aligned).
3156 * However, update_tg_cfs_runnable() is more complex. So we have:
3158 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3160 * And since, like util, the runnable part should be directly transferable,
3161 * the following would _appear_ to be the straight forward approach:
3163 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3165 * And per (1) we have:
3167 * ge->avg.runnable_avg == grq->avg.runnable_avg
3171 * ge->load.weight * grq->avg.load_avg
3172 * ge->avg.load_avg = ----------------------------------- (4)
3175 * Except that is wrong!
3177 * Because while for entities historical weight is not important and we
3178 * really only care about our future and therefore can consider a pure
3179 * runnable sum, runqueues can NOT do this.
3181 * We specifically want runqueues to have a load_avg that includes
3182 * historical weights. Those represent the blocked load, the load we expect
3183 * to (shortly) return to us. This only works by keeping the weights as
3184 * integral part of the sum. We therefore cannot decompose as per (3).
3186 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3187 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3188 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3189 * runnable section of these tasks overlap (or not). If they were to perfectly
3190 * align the rq as a whole would be runnable 2/3 of the time. If however we
3191 * always have at least 1 runnable task, the rq as a whole is always runnable.
3193 * So we'll have to approximate.. :/
3195 * Given the constraint:
3197 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3199 * We can construct a rule that adds runnable to a rq by assuming minimal
3202 * On removal, we'll assume each task is equally runnable; which yields:
3204 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3206 * XXX: only do this for the part of runnable > running ?
3211 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3213 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3215 /* Nothing to update */
3220 * The relation between sum and avg is:
3222 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3224 * however, the PELT windows are not aligned between grq and gse.
3227 /* Set new sched_entity's utilization */
3228 se->avg.util_avg = gcfs_rq->avg.util_avg;
3229 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3231 /* Update parent cfs_rq utilization */
3232 add_positive(&cfs_rq->avg.util_avg, delta);
3233 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3237 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3239 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3240 unsigned long runnable_load_avg, load_avg;
3241 u64 runnable_load_sum, load_sum = 0;
3247 gcfs_rq->prop_runnable_sum = 0;
3249 if (runnable_sum >= 0) {
3251 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3252 * the CPU is saturated running == runnable.
3254 runnable_sum += se->avg.load_sum;
3255 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3258 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3259 * assuming all tasks are equally runnable.
3261 if (scale_load_down(gcfs_rq->load.weight)) {
3262 load_sum = div_s64(gcfs_rq->avg.load_sum,
3263 scale_load_down(gcfs_rq->load.weight));
3266 /* But make sure to not inflate se's runnable */
3267 runnable_sum = min(se->avg.load_sum, load_sum);
3271 * runnable_sum can't be lower than running_sum
3272 * As running sum is scale with CPU capacity wehreas the runnable sum
3273 * is not we rescale running_sum 1st
3275 running_sum = se->avg.util_sum /
3276 arch_scale_cpu_capacity(NULL, cpu_of(rq_of(cfs_rq)));
3277 runnable_sum = max(runnable_sum, running_sum);
3279 load_sum = (s64)se_weight(se) * runnable_sum;
3280 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3282 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3283 delta_avg = load_avg - se->avg.load_avg;
3285 se->avg.load_sum = runnable_sum;
3286 se->avg.load_avg = load_avg;
3287 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3288 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3290 runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3291 runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3292 delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3293 delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3295 se->avg.runnable_load_sum = runnable_sum;
3296 se->avg.runnable_load_avg = runnable_load_avg;
3299 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3300 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3304 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3306 cfs_rq->propagate = 1;
3307 cfs_rq->prop_runnable_sum += runnable_sum;
3310 /* Update task and its cfs_rq load average */
3311 static inline int propagate_entity_load_avg(struct sched_entity *se)
3313 struct cfs_rq *cfs_rq, *gcfs_rq;
3315 if (entity_is_task(se))
3318 gcfs_rq = group_cfs_rq(se);
3319 if (!gcfs_rq->propagate)
3322 gcfs_rq->propagate = 0;
3324 cfs_rq = cfs_rq_of(se);
3326 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3328 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3329 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3335 * Check if we need to update the load and the utilization of a blocked
3338 static inline bool skip_blocked_update(struct sched_entity *se)
3340 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3343 * If sched_entity still have not zero load or utilization, we have to
3346 if (se->avg.load_avg || se->avg.util_avg)
3350 * If there is a pending propagation, we have to update the load and
3351 * the utilization of the sched_entity:
3353 if (gcfs_rq->propagate)
3357 * Otherwise, the load and the utilization of the sched_entity is
3358 * already zero and there is no pending propagation, so it will be a
3359 * waste of time to try to decay it:
3364 #else /* CONFIG_FAIR_GROUP_SCHED */
3366 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3368 static inline int propagate_entity_load_avg(struct sched_entity *se)
3373 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3375 #endif /* CONFIG_FAIR_GROUP_SCHED */
3378 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3379 * @now: current time, as per cfs_rq_clock_task()
3380 * @cfs_rq: cfs_rq to update
3382 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3383 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3384 * post_init_entity_util_avg().
3386 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3388 * Returns true if the load decayed or we removed load.
3390 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3391 * call update_tg_load_avg() when this function returns true.
3394 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3396 unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3397 struct sched_avg *sa = &cfs_rq->avg;
3400 if (cfs_rq->removed.nr) {
3402 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3404 raw_spin_lock(&cfs_rq->removed.lock);
3405 swap(cfs_rq->removed.util_avg, removed_util);
3406 swap(cfs_rq->removed.load_avg, removed_load);
3407 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3408 cfs_rq->removed.nr = 0;
3409 raw_spin_unlock(&cfs_rq->removed.lock);
3412 sub_positive(&sa->load_avg, r);
3413 sub_positive(&sa->load_sum, r * divider);
3416 sub_positive(&sa->util_avg, r);
3417 sub_positive(&sa->util_sum, r * divider);
3419 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3424 decayed |= __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq);
3426 #ifndef CONFIG_64BIT
3428 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3432 cfs_rq_util_change(cfs_rq, 0);
3438 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3439 * @cfs_rq: cfs_rq to attach to
3440 * @se: sched_entity to attach
3441 * @flags: migration hints
3443 * Must call update_cfs_rq_load_avg() before this, since we rely on
3444 * cfs_rq->avg.last_update_time being current.
3446 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3448 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3451 * When we attach the @se to the @cfs_rq, we must align the decay
3452 * window because without that, really weird and wonderful things can
3457 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3458 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3461 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3462 * period_contrib. This isn't strictly correct, but since we're
3463 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3466 se->avg.util_sum = se->avg.util_avg * divider;
3468 se->avg.load_sum = divider;
3469 if (se_weight(se)) {
3471 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3474 se->avg.runnable_load_sum = se->avg.load_sum;
3476 enqueue_load_avg(cfs_rq, se);
3477 cfs_rq->avg.util_avg += se->avg.util_avg;
3478 cfs_rq->avg.util_sum += se->avg.util_sum;
3480 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3482 cfs_rq_util_change(cfs_rq, flags);
3486 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3487 * @cfs_rq: cfs_rq to detach from
3488 * @se: sched_entity to detach
3490 * Must call update_cfs_rq_load_avg() before this, since we rely on
3491 * cfs_rq->avg.last_update_time being current.
3493 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3495 dequeue_load_avg(cfs_rq, se);
3496 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3497 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3499 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3501 cfs_rq_util_change(cfs_rq, 0);
3505 * Optional action to be done while updating the load average
3507 #define UPDATE_TG 0x1
3508 #define SKIP_AGE_LOAD 0x2
3509 #define DO_ATTACH 0x4
3511 /* Update task and its cfs_rq load average */
3512 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3514 u64 now = cfs_rq_clock_task(cfs_rq);
3515 struct rq *rq = rq_of(cfs_rq);
3516 int cpu = cpu_of(rq);
3520 * Track task load average for carrying it to new CPU after migrated, and
3521 * track group sched_entity load average for task_h_load calc in migration
3523 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3524 __update_load_avg_se(now, cpu, cfs_rq, se);
3526 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3527 decayed |= propagate_entity_load_avg(se);
3529 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3532 * DO_ATTACH means we're here from enqueue_entity().
3533 * !last_update_time means we've passed through
3534 * migrate_task_rq_fair() indicating we migrated.
3536 * IOW we're enqueueing a task on a new CPU.
3538 attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
3539 update_tg_load_avg(cfs_rq, 0);
3541 } else if (decayed && (flags & UPDATE_TG))
3542 update_tg_load_avg(cfs_rq, 0);
3545 #ifndef CONFIG_64BIT
3546 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3548 u64 last_update_time_copy;
3549 u64 last_update_time;
3552 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3554 last_update_time = cfs_rq->avg.last_update_time;
3555 } while (last_update_time != last_update_time_copy);
3557 return last_update_time;
3560 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3562 return cfs_rq->avg.last_update_time;
3567 * Synchronize entity load avg of dequeued entity without locking
3570 void sync_entity_load_avg(struct sched_entity *se)
3572 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3573 u64 last_update_time;
3575 last_update_time = cfs_rq_last_update_time(cfs_rq);
3576 __update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se);
3580 * Task first catches up with cfs_rq, and then subtract
3581 * itself from the cfs_rq (task must be off the queue now).
3583 void remove_entity_load_avg(struct sched_entity *se)
3585 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3586 unsigned long flags;
3589 * tasks cannot exit without having gone through wake_up_new_task() ->
3590 * post_init_entity_util_avg() which will have added things to the
3591 * cfs_rq, so we can remove unconditionally.
3593 * Similarly for groups, they will have passed through
3594 * post_init_entity_util_avg() before unregister_sched_fair_group()
3598 sync_entity_load_avg(se);
3600 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3601 ++cfs_rq->removed.nr;
3602 cfs_rq->removed.util_avg += se->avg.util_avg;
3603 cfs_rq->removed.load_avg += se->avg.load_avg;
3604 cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */
3605 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3608 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3610 return cfs_rq->avg.runnable_load_avg;
3613 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3615 return cfs_rq->avg.load_avg;
3618 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3620 static inline unsigned long task_util(struct task_struct *p)
3622 return READ_ONCE(p->se.avg.util_avg);
3625 static inline unsigned long _task_util_est(struct task_struct *p)
3627 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3629 return max(ue.ewma, ue.enqueued);
3632 static inline unsigned long task_util_est(struct task_struct *p)
3634 return max(task_util(p), _task_util_est(p));
3637 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3638 struct task_struct *p)
3640 unsigned int enqueued;
3642 if (!sched_feat(UTIL_EST))
3645 /* Update root cfs_rq's estimated utilization */
3646 enqueued = cfs_rq->avg.util_est.enqueued;
3647 enqueued += (_task_util_est(p) | UTIL_AVG_UNCHANGED);
3648 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3652 * Check if a (signed) value is within a specified (unsigned) margin,
3653 * based on the observation that:
3655 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3657 * NOTE: this only works when value + maring < INT_MAX.
3659 static inline bool within_margin(int value, int margin)
3661 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3665 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3667 long last_ewma_diff;
3670 if (!sched_feat(UTIL_EST))
3673 /* Update root cfs_rq's estimated utilization */
3674 ue.enqueued = cfs_rq->avg.util_est.enqueued;
3675 ue.enqueued -= min_t(unsigned int, ue.enqueued,
3676 (_task_util_est(p) | UTIL_AVG_UNCHANGED));
3677 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3680 * Skip update of task's estimated utilization when the task has not
3681 * yet completed an activation, e.g. being migrated.
3687 * If the PELT values haven't changed since enqueue time,
3688 * skip the util_est update.
3690 ue = p->se.avg.util_est;
3691 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3695 * Skip update of task's estimated utilization when its EWMA is
3696 * already ~1% close to its last activation value.
3698 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3699 last_ewma_diff = ue.enqueued - ue.ewma;
3700 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3704 * Update Task's estimated utilization
3706 * When *p completes an activation we can consolidate another sample
3707 * of the task size. This is done by storing the current PELT value
3708 * as ue.enqueued and by using this value to update the Exponential
3709 * Weighted Moving Average (EWMA):
3711 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
3712 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
3713 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
3714 * = w * ( last_ewma_diff ) + ewma(t-1)
3715 * = w * (last_ewma_diff + ewma(t-1) / w)
3717 * Where 'w' is the weight of new samples, which is configured to be
3718 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3720 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
3721 ue.ewma += last_ewma_diff;
3722 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
3723 WRITE_ONCE(p->se.avg.util_est, ue);
3726 #else /* CONFIG_SMP */
3728 #define UPDATE_TG 0x0
3729 #define SKIP_AGE_LOAD 0x0
3730 #define DO_ATTACH 0x0
3732 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3734 cfs_rq_util_change(cfs_rq, 0);
3737 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3740 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
3742 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3744 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3750 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
3753 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
3756 #endif /* CONFIG_SMP */
3758 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3760 #ifdef CONFIG_SCHED_DEBUG
3761 s64 d = se->vruntime - cfs_rq->min_vruntime;
3766 if (d > 3*sysctl_sched_latency)
3767 schedstat_inc(cfs_rq->nr_spread_over);
3772 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3774 u64 vruntime = cfs_rq->min_vruntime;
3777 * The 'current' period is already promised to the current tasks,
3778 * however the extra weight of the new task will slow them down a
3779 * little, place the new task so that it fits in the slot that
3780 * stays open at the end.
3782 if (initial && sched_feat(START_DEBIT))
3783 vruntime += sched_vslice(cfs_rq, se);
3785 /* sleeps up to a single latency don't count. */
3787 unsigned long thresh = sysctl_sched_latency;
3790 * Halve their sleep time's effect, to allow
3791 * for a gentler effect of sleepers:
3793 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3799 /* ensure we never gain time by being placed backwards. */
3800 se->vruntime = max_vruntime(se->vruntime, vruntime);
3803 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3805 static inline void check_schedstat_required(void)
3807 #ifdef CONFIG_SCHEDSTATS
3808 if (schedstat_enabled())
3811 /* Force schedstat enabled if a dependent tracepoint is active */
3812 if (trace_sched_stat_wait_enabled() ||
3813 trace_sched_stat_sleep_enabled() ||
3814 trace_sched_stat_iowait_enabled() ||
3815 trace_sched_stat_blocked_enabled() ||
3816 trace_sched_stat_runtime_enabled()) {
3817 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3818 "stat_blocked and stat_runtime require the "
3819 "kernel parameter schedstats=enable or "
3820 "kernel.sched_schedstats=1\n");
3831 * update_min_vruntime()
3832 * vruntime -= min_vruntime
3836 * update_min_vruntime()
3837 * vruntime += min_vruntime
3839 * this way the vruntime transition between RQs is done when both
3840 * min_vruntime are up-to-date.
3844 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3845 * vruntime -= min_vruntime
3849 * update_min_vruntime()
3850 * vruntime += min_vruntime
3852 * this way we don't have the most up-to-date min_vruntime on the originating
3853 * CPU and an up-to-date min_vruntime on the destination CPU.
3857 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3859 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3860 bool curr = cfs_rq->curr == se;
3863 * If we're the current task, we must renormalise before calling
3867 se->vruntime += cfs_rq->min_vruntime;
3869 update_curr(cfs_rq);
3872 * Otherwise, renormalise after, such that we're placed at the current
3873 * moment in time, instead of some random moment in the past. Being
3874 * placed in the past could significantly boost this task to the
3875 * fairness detriment of existing tasks.
3877 if (renorm && !curr)
3878 se->vruntime += cfs_rq->min_vruntime;
3881 * When enqueuing a sched_entity, we must:
3882 * - Update loads to have both entity and cfs_rq synced with now.
3883 * - Add its load to cfs_rq->runnable_avg
3884 * - For group_entity, update its weight to reflect the new share of
3886 * - Add its new weight to cfs_rq->load.weight
3888 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
3889 update_cfs_group(se);
3890 enqueue_runnable_load_avg(cfs_rq, se);
3891 account_entity_enqueue(cfs_rq, se);
3893 if (flags & ENQUEUE_WAKEUP)
3894 place_entity(cfs_rq, se, 0);
3896 check_schedstat_required();
3897 update_stats_enqueue(cfs_rq, se, flags);
3898 check_spread(cfs_rq, se);
3900 __enqueue_entity(cfs_rq, se);
3903 if (cfs_rq->nr_running == 1) {
3904 list_add_leaf_cfs_rq(cfs_rq);
3905 check_enqueue_throttle(cfs_rq);
3909 static void __clear_buddies_last(struct sched_entity *se)
3911 for_each_sched_entity(se) {
3912 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3913 if (cfs_rq->last != se)
3916 cfs_rq->last = NULL;
3920 static void __clear_buddies_next(struct sched_entity *se)
3922 for_each_sched_entity(se) {
3923 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3924 if (cfs_rq->next != se)
3927 cfs_rq->next = NULL;
3931 static void __clear_buddies_skip(struct sched_entity *se)
3933 for_each_sched_entity(se) {
3934 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3935 if (cfs_rq->skip != se)
3938 cfs_rq->skip = NULL;
3942 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3944 if (cfs_rq->last == se)
3945 __clear_buddies_last(se);
3947 if (cfs_rq->next == se)
3948 __clear_buddies_next(se);
3950 if (cfs_rq->skip == se)
3951 __clear_buddies_skip(se);
3954 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3957 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3960 * Update run-time statistics of the 'current'.
3962 update_curr(cfs_rq);
3965 * When dequeuing a sched_entity, we must:
3966 * - Update loads to have both entity and cfs_rq synced with now.
3967 * - Substract its load from the cfs_rq->runnable_avg.
3968 * - Substract its previous weight from cfs_rq->load.weight.
3969 * - For group entity, update its weight to reflect the new share
3970 * of its group cfs_rq.
3972 update_load_avg(cfs_rq, se, UPDATE_TG);
3973 dequeue_runnable_load_avg(cfs_rq, se);
3975 update_stats_dequeue(cfs_rq, se, flags);
3977 clear_buddies(cfs_rq, se);
3979 if (se != cfs_rq->curr)
3980 __dequeue_entity(cfs_rq, se);
3982 account_entity_dequeue(cfs_rq, se);
3985 * Normalize after update_curr(); which will also have moved
3986 * min_vruntime if @se is the one holding it back. But before doing
3987 * update_min_vruntime() again, which will discount @se's position and
3988 * can move min_vruntime forward still more.
3990 if (!(flags & DEQUEUE_SLEEP))
3991 se->vruntime -= cfs_rq->min_vruntime;
3993 /* return excess runtime on last dequeue */
3994 return_cfs_rq_runtime(cfs_rq);
3996 update_cfs_group(se);
3999 * Now advance min_vruntime if @se was the entity holding it back,
4000 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4001 * put back on, and if we advance min_vruntime, we'll be placed back
4002 * further than we started -- ie. we'll be penalized.
4004 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4005 update_min_vruntime(cfs_rq);
4009 * Preempt the current task with a newly woken task if needed:
4012 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4014 unsigned long ideal_runtime, delta_exec;
4015 struct sched_entity *se;
4018 ideal_runtime = sched_slice(cfs_rq, curr);
4019 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4020 if (delta_exec > ideal_runtime) {
4021 resched_curr(rq_of(cfs_rq));
4023 * The current task ran long enough, ensure it doesn't get
4024 * re-elected due to buddy favours.
4026 clear_buddies(cfs_rq, curr);
4031 * Ensure that a task that missed wakeup preemption by a
4032 * narrow margin doesn't have to wait for a full slice.
4033 * This also mitigates buddy induced latencies under load.
4035 if (delta_exec < sysctl_sched_min_granularity)
4038 se = __pick_first_entity(cfs_rq);
4039 delta = curr->vruntime - se->vruntime;
4044 if (delta > ideal_runtime)
4045 resched_curr(rq_of(cfs_rq));
4049 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4051 /* 'current' is not kept within the tree. */
4054 * Any task has to be enqueued before it get to execute on
4055 * a CPU. So account for the time it spent waiting on the
4058 update_stats_wait_end(cfs_rq, se);
4059 __dequeue_entity(cfs_rq, se);
4060 update_load_avg(cfs_rq, se, UPDATE_TG);
4063 update_stats_curr_start(cfs_rq, se);
4067 * Track our maximum slice length, if the CPU's load is at
4068 * least twice that of our own weight (i.e. dont track it
4069 * when there are only lesser-weight tasks around):
4071 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
4072 schedstat_set(se->statistics.slice_max,
4073 max((u64)schedstat_val(se->statistics.slice_max),
4074 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4077 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4081 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4084 * Pick the next process, keeping these things in mind, in this order:
4085 * 1) keep things fair between processes/task groups
4086 * 2) pick the "next" process, since someone really wants that to run
4087 * 3) pick the "last" process, for cache locality
4088 * 4) do not run the "skip" process, if something else is available
4090 static struct sched_entity *
4091 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4093 struct sched_entity *left = __pick_first_entity(cfs_rq);
4094 struct sched_entity *se;
4097 * If curr is set we have to see if its left of the leftmost entity
4098 * still in the tree, provided there was anything in the tree at all.
4100 if (!left || (curr && entity_before(curr, left)))
4103 se = left; /* ideally we run the leftmost entity */
4106 * Avoid running the skip buddy, if running something else can
4107 * be done without getting too unfair.
4109 if (cfs_rq->skip == se) {
4110 struct sched_entity *second;
4113 second = __pick_first_entity(cfs_rq);
4115 second = __pick_next_entity(se);
4116 if (!second || (curr && entity_before(curr, second)))
4120 if (second && wakeup_preempt_entity(second, left) < 1)
4125 * Prefer last buddy, try to return the CPU to a preempted task.
4127 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4131 * Someone really wants this to run. If it's not unfair, run it.
4133 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4136 clear_buddies(cfs_rq, se);
4141 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4143 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4146 * If still on the runqueue then deactivate_task()
4147 * was not called and update_curr() has to be done:
4150 update_curr(cfs_rq);
4152 /* throttle cfs_rqs exceeding runtime */
4153 check_cfs_rq_runtime(cfs_rq);
4155 check_spread(cfs_rq, prev);
4158 update_stats_wait_start(cfs_rq, prev);
4159 /* Put 'current' back into the tree. */
4160 __enqueue_entity(cfs_rq, prev);
4161 /* in !on_rq case, update occurred at dequeue */
4162 update_load_avg(cfs_rq, prev, 0);
4164 cfs_rq->curr = NULL;
4168 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4171 * Update run-time statistics of the 'current'.
4173 update_curr(cfs_rq);
4176 * Ensure that runnable average is periodically updated.
4178 update_load_avg(cfs_rq, curr, UPDATE_TG);
4179 update_cfs_group(curr);
4181 #ifdef CONFIG_SCHED_HRTICK
4183 * queued ticks are scheduled to match the slice, so don't bother
4184 * validating it and just reschedule.
4187 resched_curr(rq_of(cfs_rq));
4191 * don't let the period tick interfere with the hrtick preemption
4193 if (!sched_feat(DOUBLE_TICK) &&
4194 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4198 if (cfs_rq->nr_running > 1)
4199 check_preempt_tick(cfs_rq, curr);
4203 /**************************************************
4204 * CFS bandwidth control machinery
4207 #ifdef CONFIG_CFS_BANDWIDTH
4209 #ifdef HAVE_JUMP_LABEL
4210 static struct static_key __cfs_bandwidth_used;
4212 static inline bool cfs_bandwidth_used(void)
4214 return static_key_false(&__cfs_bandwidth_used);
4217 void cfs_bandwidth_usage_inc(void)
4219 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4222 void cfs_bandwidth_usage_dec(void)
4224 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4226 #else /* HAVE_JUMP_LABEL */
4227 static bool cfs_bandwidth_used(void)
4232 void cfs_bandwidth_usage_inc(void) {}
4233 void cfs_bandwidth_usage_dec(void) {}
4234 #endif /* HAVE_JUMP_LABEL */
4237 * default period for cfs group bandwidth.
4238 * default: 0.1s, units: nanoseconds
4240 static inline u64 default_cfs_period(void)
4242 return 100000000ULL;
4245 static inline u64 sched_cfs_bandwidth_slice(void)
4247 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4251 * Replenish runtime according to assigned quota and update expiration time.
4252 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4253 * additional synchronization around rq->lock.
4255 * requires cfs_b->lock
4257 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4261 if (cfs_b->quota == RUNTIME_INF)
4264 now = sched_clock_cpu(smp_processor_id());
4265 cfs_b->runtime = cfs_b->quota;
4266 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4267 cfs_b->expires_seq++;
4270 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4272 return &tg->cfs_bandwidth;
4275 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4276 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4278 if (unlikely(cfs_rq->throttle_count))
4279 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4281 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4284 /* returns 0 on failure to allocate runtime */
4285 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4287 struct task_group *tg = cfs_rq->tg;
4288 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4289 u64 amount = 0, min_amount, expires;
4292 /* note: this is a positive sum as runtime_remaining <= 0 */
4293 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4295 raw_spin_lock(&cfs_b->lock);
4296 if (cfs_b->quota == RUNTIME_INF)
4297 amount = min_amount;
4299 start_cfs_bandwidth(cfs_b);
4301 if (cfs_b->runtime > 0) {
4302 amount = min(cfs_b->runtime, min_amount);
4303 cfs_b->runtime -= amount;
4307 expires_seq = cfs_b->expires_seq;
4308 expires = cfs_b->runtime_expires;
4309 raw_spin_unlock(&cfs_b->lock);
4311 cfs_rq->runtime_remaining += amount;
4313 * we may have advanced our local expiration to account for allowed
4314 * spread between our sched_clock and the one on which runtime was
4317 if (cfs_rq->expires_seq != expires_seq) {
4318 cfs_rq->expires_seq = expires_seq;
4319 cfs_rq->runtime_expires = expires;
4322 return cfs_rq->runtime_remaining > 0;
4326 * Note: This depends on the synchronization provided by sched_clock and the
4327 * fact that rq->clock snapshots this value.
4329 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4331 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4333 /* if the deadline is ahead of our clock, nothing to do */
4334 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4337 if (cfs_rq->runtime_remaining < 0)
4341 * If the local deadline has passed we have to consider the
4342 * possibility that our sched_clock is 'fast' and the global deadline
4343 * has not truly expired.
4345 * Fortunately we can check determine whether this the case by checking
4346 * whether the global deadline(cfs_b->expires_seq) has advanced.
4348 if (cfs_rq->expires_seq == cfs_b->expires_seq) {
4349 /* extend local deadline, drift is bounded above by 2 ticks */
4350 cfs_rq->runtime_expires += TICK_NSEC;
4352 /* global deadline is ahead, expiration has passed */
4353 cfs_rq->runtime_remaining = 0;
4357 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4359 /* dock delta_exec before expiring quota (as it could span periods) */
4360 cfs_rq->runtime_remaining -= delta_exec;
4361 expire_cfs_rq_runtime(cfs_rq);
4363 if (likely(cfs_rq->runtime_remaining > 0))
4367 * if we're unable to extend our runtime we resched so that the active
4368 * hierarchy can be throttled
4370 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4371 resched_curr(rq_of(cfs_rq));
4374 static __always_inline
4375 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4377 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4380 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4383 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4385 return cfs_bandwidth_used() && cfs_rq->throttled;
4388 /* check whether cfs_rq, or any parent, is throttled */
4389 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4391 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4395 * Ensure that neither of the group entities corresponding to src_cpu or
4396 * dest_cpu are members of a throttled hierarchy when performing group
4397 * load-balance operations.
4399 static inline int throttled_lb_pair(struct task_group *tg,
4400 int src_cpu, int dest_cpu)
4402 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4404 src_cfs_rq = tg->cfs_rq[src_cpu];
4405 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4407 return throttled_hierarchy(src_cfs_rq) ||
4408 throttled_hierarchy(dest_cfs_rq);
4411 static int tg_unthrottle_up(struct task_group *tg, void *data)
4413 struct rq *rq = data;
4414 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4416 cfs_rq->throttle_count--;
4417 if (!cfs_rq->throttle_count) {
4418 /* adjust cfs_rq_clock_task() */
4419 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4420 cfs_rq->throttled_clock_task;
4426 static int tg_throttle_down(struct task_group *tg, void *data)
4428 struct rq *rq = data;
4429 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4431 /* group is entering throttled state, stop time */
4432 if (!cfs_rq->throttle_count)
4433 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4434 cfs_rq->throttle_count++;
4439 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4441 struct rq *rq = rq_of(cfs_rq);
4442 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4443 struct sched_entity *se;
4444 long task_delta, dequeue = 1;
4447 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4449 /* freeze hierarchy runnable averages while throttled */
4451 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4454 task_delta = cfs_rq->h_nr_running;
4455 for_each_sched_entity(se) {
4456 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4457 /* throttled entity or throttle-on-deactivate */
4462 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4463 qcfs_rq->h_nr_running -= task_delta;
4465 if (qcfs_rq->load.weight)
4470 sub_nr_running(rq, task_delta);
4472 cfs_rq->throttled = 1;
4473 cfs_rq->throttled_clock = rq_clock(rq);
4474 raw_spin_lock(&cfs_b->lock);
4475 empty = list_empty(&cfs_b->throttled_cfs_rq);
4478 * Add to the _head_ of the list, so that an already-started
4479 * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
4480 * not running add to the tail so that later runqueues don't get starved.
4482 if (cfs_b->distribute_running)
4483 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4485 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4488 * If we're the first throttled task, make sure the bandwidth
4492 start_cfs_bandwidth(cfs_b);
4494 raw_spin_unlock(&cfs_b->lock);
4497 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4499 struct rq *rq = rq_of(cfs_rq);
4500 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4501 struct sched_entity *se;
4505 se = cfs_rq->tg->se[cpu_of(rq)];
4507 cfs_rq->throttled = 0;
4509 update_rq_clock(rq);
4511 raw_spin_lock(&cfs_b->lock);
4512 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4513 list_del_rcu(&cfs_rq->throttled_list);
4514 raw_spin_unlock(&cfs_b->lock);
4516 /* update hierarchical throttle state */
4517 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4519 if (!cfs_rq->load.weight)
4522 task_delta = cfs_rq->h_nr_running;
4523 for_each_sched_entity(se) {
4527 cfs_rq = cfs_rq_of(se);
4529 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4530 cfs_rq->h_nr_running += task_delta;
4532 if (cfs_rq_throttled(cfs_rq))
4537 add_nr_running(rq, task_delta);
4539 /* Determine whether we need to wake up potentially idle CPU: */
4540 if (rq->curr == rq->idle && rq->cfs.nr_running)
4544 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4545 u64 remaining, u64 expires)
4547 struct cfs_rq *cfs_rq;
4549 u64 starting_runtime = remaining;
4552 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4554 struct rq *rq = rq_of(cfs_rq);
4558 if (!cfs_rq_throttled(cfs_rq))
4561 runtime = -cfs_rq->runtime_remaining + 1;
4562 if (runtime > remaining)
4563 runtime = remaining;
4564 remaining -= runtime;
4566 cfs_rq->runtime_remaining += runtime;
4567 cfs_rq->runtime_expires = expires;
4569 /* we check whether we're throttled above */
4570 if (cfs_rq->runtime_remaining > 0)
4571 unthrottle_cfs_rq(cfs_rq);
4581 return starting_runtime - remaining;
4585 * Responsible for refilling a task_group's bandwidth and unthrottling its
4586 * cfs_rqs as appropriate. If there has been no activity within the last
4587 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4588 * used to track this state.
4590 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4592 u64 runtime, runtime_expires;
4595 /* no need to continue the timer with no bandwidth constraint */
4596 if (cfs_b->quota == RUNTIME_INF)
4597 goto out_deactivate;
4599 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4600 cfs_b->nr_periods += overrun;
4603 * idle depends on !throttled (for the case of a large deficit), and if
4604 * we're going inactive then everything else can be deferred
4606 if (cfs_b->idle && !throttled)
4607 goto out_deactivate;
4609 __refill_cfs_bandwidth_runtime(cfs_b);
4612 /* mark as potentially idle for the upcoming period */
4617 /* account preceding periods in which throttling occurred */
4618 cfs_b->nr_throttled += overrun;
4620 runtime_expires = cfs_b->runtime_expires;
4623 * This check is repeated as we are holding onto the new bandwidth while
4624 * we unthrottle. This can potentially race with an unthrottled group
4625 * trying to acquire new bandwidth from the global pool. This can result
4626 * in us over-using our runtime if it is all used during this loop, but
4627 * only by limited amounts in that extreme case.
4629 while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
4630 runtime = cfs_b->runtime;
4631 cfs_b->distribute_running = 1;
4632 raw_spin_unlock(&cfs_b->lock);
4633 /* we can't nest cfs_b->lock while distributing bandwidth */
4634 runtime = distribute_cfs_runtime(cfs_b, runtime,
4636 raw_spin_lock(&cfs_b->lock);
4638 cfs_b->distribute_running = 0;
4639 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4641 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4645 * While we are ensured activity in the period following an
4646 * unthrottle, this also covers the case in which the new bandwidth is
4647 * insufficient to cover the existing bandwidth deficit. (Forcing the
4648 * timer to remain active while there are any throttled entities.)
4658 /* a cfs_rq won't donate quota below this amount */
4659 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4660 /* minimum remaining period time to redistribute slack quota */
4661 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4662 /* how long we wait to gather additional slack before distributing */
4663 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4666 * Are we near the end of the current quota period?
4668 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4669 * hrtimer base being cleared by hrtimer_start. In the case of
4670 * migrate_hrtimers, base is never cleared, so we are fine.
4672 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4674 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4677 /* if the call-back is running a quota refresh is already occurring */
4678 if (hrtimer_callback_running(refresh_timer))
4681 /* is a quota refresh about to occur? */
4682 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4683 if (remaining < min_expire)
4689 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4691 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4693 /* if there's a quota refresh soon don't bother with slack */
4694 if (runtime_refresh_within(cfs_b, min_left))
4697 hrtimer_start(&cfs_b->slack_timer,
4698 ns_to_ktime(cfs_bandwidth_slack_period),
4702 /* we know any runtime found here is valid as update_curr() precedes return */
4703 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4705 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4706 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4708 if (slack_runtime <= 0)
4711 raw_spin_lock(&cfs_b->lock);
4712 if (cfs_b->quota != RUNTIME_INF &&
4713 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4714 cfs_b->runtime += slack_runtime;
4716 /* we are under rq->lock, defer unthrottling using a timer */
4717 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4718 !list_empty(&cfs_b->throttled_cfs_rq))
4719 start_cfs_slack_bandwidth(cfs_b);
4721 raw_spin_unlock(&cfs_b->lock);
4723 /* even if it's not valid for return we don't want to try again */
4724 cfs_rq->runtime_remaining -= slack_runtime;
4727 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4729 if (!cfs_bandwidth_used())
4732 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4735 __return_cfs_rq_runtime(cfs_rq);
4739 * This is done with a timer (instead of inline with bandwidth return) since
4740 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4742 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4744 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4747 /* confirm we're still not at a refresh boundary */
4748 raw_spin_lock(&cfs_b->lock);
4749 if (cfs_b->distribute_running) {
4750 raw_spin_unlock(&cfs_b->lock);
4754 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4755 raw_spin_unlock(&cfs_b->lock);
4759 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4760 runtime = cfs_b->runtime;
4762 expires = cfs_b->runtime_expires;
4764 cfs_b->distribute_running = 1;
4766 raw_spin_unlock(&cfs_b->lock);
4771 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4773 raw_spin_lock(&cfs_b->lock);
4774 if (expires == cfs_b->runtime_expires)
4775 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4776 cfs_b->distribute_running = 0;
4777 raw_spin_unlock(&cfs_b->lock);
4781 * When a group wakes up we want to make sure that its quota is not already
4782 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4783 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4785 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4787 if (!cfs_bandwidth_used())
4790 /* an active group must be handled by the update_curr()->put() path */
4791 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4794 /* ensure the group is not already throttled */
4795 if (cfs_rq_throttled(cfs_rq))
4798 /* update runtime allocation */
4799 account_cfs_rq_runtime(cfs_rq, 0);
4800 if (cfs_rq->runtime_remaining <= 0)
4801 throttle_cfs_rq(cfs_rq);
4804 static void sync_throttle(struct task_group *tg, int cpu)
4806 struct cfs_rq *pcfs_rq, *cfs_rq;
4808 if (!cfs_bandwidth_used())
4814 cfs_rq = tg->cfs_rq[cpu];
4815 pcfs_rq = tg->parent->cfs_rq[cpu];
4817 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4818 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4821 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4822 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4824 if (!cfs_bandwidth_used())
4827 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4831 * it's possible for a throttled entity to be forced into a running
4832 * state (e.g. set_curr_task), in this case we're finished.
4834 if (cfs_rq_throttled(cfs_rq))
4837 throttle_cfs_rq(cfs_rq);
4841 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4843 struct cfs_bandwidth *cfs_b =
4844 container_of(timer, struct cfs_bandwidth, slack_timer);
4846 do_sched_cfs_slack_timer(cfs_b);
4848 return HRTIMER_NORESTART;
4851 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4853 struct cfs_bandwidth *cfs_b =
4854 container_of(timer, struct cfs_bandwidth, period_timer);
4858 raw_spin_lock(&cfs_b->lock);
4860 overrun = hrtimer_forward_now(timer, cfs_b->period);
4864 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4867 cfs_b->period_active = 0;
4868 raw_spin_unlock(&cfs_b->lock);
4870 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4873 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4875 raw_spin_lock_init(&cfs_b->lock);
4877 cfs_b->quota = RUNTIME_INF;
4878 cfs_b->period = ns_to_ktime(default_cfs_period());
4880 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4881 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4882 cfs_b->period_timer.function = sched_cfs_period_timer;
4883 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4884 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4885 cfs_b->distribute_running = 0;
4888 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4890 cfs_rq->runtime_enabled = 0;
4891 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4894 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4898 lockdep_assert_held(&cfs_b->lock);
4900 if (cfs_b->period_active)
4903 cfs_b->period_active = 1;
4904 overrun = hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4905 cfs_b->runtime_expires += (overrun + 1) * ktime_to_ns(cfs_b->period);
4906 cfs_b->expires_seq++;
4907 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4910 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4912 /* init_cfs_bandwidth() was not called */
4913 if (!cfs_b->throttled_cfs_rq.next)
4916 hrtimer_cancel(&cfs_b->period_timer);
4917 hrtimer_cancel(&cfs_b->slack_timer);
4921 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
4923 * The race is harmless, since modifying bandwidth settings of unhooked group
4924 * bits doesn't do much.
4927 /* cpu online calback */
4928 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4930 struct task_group *tg;
4932 lockdep_assert_held(&rq->lock);
4935 list_for_each_entry_rcu(tg, &task_groups, list) {
4936 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
4937 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4939 raw_spin_lock(&cfs_b->lock);
4940 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4941 raw_spin_unlock(&cfs_b->lock);
4946 /* cpu offline callback */
4947 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4949 struct task_group *tg;
4951 lockdep_assert_held(&rq->lock);
4954 list_for_each_entry_rcu(tg, &task_groups, list) {
4955 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4957 if (!cfs_rq->runtime_enabled)
4961 * clock_task is not advancing so we just need to make sure
4962 * there's some valid quota amount
4964 cfs_rq->runtime_remaining = 1;
4966 * Offline rq is schedulable till CPU is completely disabled
4967 * in take_cpu_down(), so we prevent new cfs throttling here.
4969 cfs_rq->runtime_enabled = 0;
4971 if (cfs_rq_throttled(cfs_rq))
4972 unthrottle_cfs_rq(cfs_rq);
4977 #else /* CONFIG_CFS_BANDWIDTH */
4978 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4980 return rq_clock_task(rq_of(cfs_rq));
4983 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4984 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4985 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4986 static inline void sync_throttle(struct task_group *tg, int cpu) {}
4987 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4989 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4994 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4999 static inline int throttled_lb_pair(struct task_group *tg,
5000 int src_cpu, int dest_cpu)
5005 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5007 #ifdef CONFIG_FAIR_GROUP_SCHED
5008 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5011 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5015 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5016 static inline void update_runtime_enabled(struct rq *rq) {}
5017 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5019 #endif /* CONFIG_CFS_BANDWIDTH */
5021 /**************************************************
5022 * CFS operations on tasks:
5025 #ifdef CONFIG_SCHED_HRTICK
5026 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5028 struct sched_entity *se = &p->se;
5029 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5031 SCHED_WARN_ON(task_rq(p) != rq);
5033 if (rq->cfs.h_nr_running > 1) {
5034 u64 slice = sched_slice(cfs_rq, se);
5035 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5036 s64 delta = slice - ran;
5043 hrtick_start(rq, delta);
5048 * called from enqueue/dequeue and updates the hrtick when the
5049 * current task is from our class and nr_running is low enough
5052 static void hrtick_update(struct rq *rq)
5054 struct task_struct *curr = rq->curr;
5056 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5059 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5060 hrtick_start_fair(rq, curr);
5062 #else /* !CONFIG_SCHED_HRTICK */
5064 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5068 static inline void hrtick_update(struct rq *rq)
5074 * The enqueue_task method is called before nr_running is
5075 * increased. Here we update the fair scheduling stats and
5076 * then put the task into the rbtree:
5079 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5081 struct cfs_rq *cfs_rq;
5082 struct sched_entity *se = &p->se;
5085 * The code below (indirectly) updates schedutil which looks at
5086 * the cfs_rq utilization to select a frequency.
5087 * Let's add the task's estimated utilization to the cfs_rq's
5088 * estimated utilization, before we update schedutil.
5090 util_est_enqueue(&rq->cfs, p);
5093 * If in_iowait is set, the code below may not trigger any cpufreq
5094 * utilization updates, so do it here explicitly with the IOWAIT flag
5098 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5100 for_each_sched_entity(se) {
5103 cfs_rq = cfs_rq_of(se);
5104 enqueue_entity(cfs_rq, se, flags);
5107 * end evaluation on encountering a throttled cfs_rq
5109 * note: in the case of encountering a throttled cfs_rq we will
5110 * post the final h_nr_running increment below.
5112 if (cfs_rq_throttled(cfs_rq))
5114 cfs_rq->h_nr_running++;
5116 flags = ENQUEUE_WAKEUP;
5119 for_each_sched_entity(se) {
5120 cfs_rq = cfs_rq_of(se);
5121 cfs_rq->h_nr_running++;
5123 if (cfs_rq_throttled(cfs_rq))
5126 update_load_avg(cfs_rq, se, UPDATE_TG);
5127 update_cfs_group(se);
5131 add_nr_running(rq, 1);
5136 static void set_next_buddy(struct sched_entity *se);
5139 * The dequeue_task method is called before nr_running is
5140 * decreased. We remove the task from the rbtree and
5141 * update the fair scheduling stats:
5143 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5145 struct cfs_rq *cfs_rq;
5146 struct sched_entity *se = &p->se;
5147 int task_sleep = flags & DEQUEUE_SLEEP;
5149 for_each_sched_entity(se) {
5150 cfs_rq = cfs_rq_of(se);
5151 dequeue_entity(cfs_rq, se, flags);
5154 * end evaluation on encountering a throttled cfs_rq
5156 * note: in the case of encountering a throttled cfs_rq we will
5157 * post the final h_nr_running decrement below.
5159 if (cfs_rq_throttled(cfs_rq))
5161 cfs_rq->h_nr_running--;
5163 /* Don't dequeue parent if it has other entities besides us */
5164 if (cfs_rq->load.weight) {
5165 /* Avoid re-evaluating load for this entity: */
5166 se = parent_entity(se);
5168 * Bias pick_next to pick a task from this cfs_rq, as
5169 * p is sleeping when it is within its sched_slice.
5171 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5175 flags |= DEQUEUE_SLEEP;
5178 for_each_sched_entity(se) {
5179 cfs_rq = cfs_rq_of(se);
5180 cfs_rq->h_nr_running--;
5182 if (cfs_rq_throttled(cfs_rq))
5185 update_load_avg(cfs_rq, se, UPDATE_TG);
5186 update_cfs_group(se);
5190 sub_nr_running(rq, 1);
5192 util_est_dequeue(&rq->cfs, p, task_sleep);
5198 /* Working cpumask for: load_balance, load_balance_newidle. */
5199 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5200 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5202 #ifdef CONFIG_NO_HZ_COMMON
5204 * per rq 'load' arrray crap; XXX kill this.
5208 * The exact cpuload calculated at every tick would be:
5210 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5212 * If a CPU misses updates for n ticks (as it was idle) and update gets
5213 * called on the n+1-th tick when CPU may be busy, then we have:
5215 * load_n = (1 - 1/2^i)^n * load_0
5216 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
5218 * decay_load_missed() below does efficient calculation of
5220 * load' = (1 - 1/2^i)^n * load
5222 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5223 * This allows us to precompute the above in said factors, thereby allowing the
5224 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5225 * fixed_power_int())
5227 * The calculation is approximated on a 128 point scale.
5229 #define DEGRADE_SHIFT 7
5231 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5232 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5233 { 0, 0, 0, 0, 0, 0, 0, 0 },
5234 { 64, 32, 8, 0, 0, 0, 0, 0 },
5235 { 96, 72, 40, 12, 1, 0, 0, 0 },
5236 { 112, 98, 75, 43, 15, 1, 0, 0 },
5237 { 120, 112, 98, 76, 45, 16, 2, 0 }
5241 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5242 * would be when CPU is idle and so we just decay the old load without
5243 * adding any new load.
5245 static unsigned long
5246 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5250 if (!missed_updates)
5253 if (missed_updates >= degrade_zero_ticks[idx])
5257 return load >> missed_updates;
5259 while (missed_updates) {
5260 if (missed_updates % 2)
5261 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5263 missed_updates >>= 1;
5270 cpumask_var_t idle_cpus_mask;
5272 int has_blocked; /* Idle CPUS has blocked load */
5273 unsigned long next_balance; /* in jiffy units */
5274 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5275 } nohz ____cacheline_aligned;
5277 #endif /* CONFIG_NO_HZ_COMMON */
5280 * __cpu_load_update - update the rq->cpu_load[] statistics
5281 * @this_rq: The rq to update statistics for
5282 * @this_load: The current load
5283 * @pending_updates: The number of missed updates
5285 * Update rq->cpu_load[] statistics. This function is usually called every
5286 * scheduler tick (TICK_NSEC).
5288 * This function computes a decaying average:
5290 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5292 * Because of NOHZ it might not get called on every tick which gives need for
5293 * the @pending_updates argument.
5295 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5296 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5297 * = A * (A * load[i]_n-2 + B) + B
5298 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5299 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5300 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5301 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5302 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5304 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5305 * any change in load would have resulted in the tick being turned back on.
5307 * For regular NOHZ, this reduces to:
5309 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5311 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5314 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5315 unsigned long pending_updates)
5317 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5320 this_rq->nr_load_updates++;
5322 /* Update our load: */
5323 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5324 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5325 unsigned long old_load, new_load;
5327 /* scale is effectively 1 << i now, and >> i divides by scale */
5329 old_load = this_rq->cpu_load[i];
5330 #ifdef CONFIG_NO_HZ_COMMON
5331 old_load = decay_load_missed(old_load, pending_updates - 1, i);
5332 if (tickless_load) {
5333 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5335 * old_load can never be a negative value because a
5336 * decayed tickless_load cannot be greater than the
5337 * original tickless_load.
5339 old_load += tickless_load;
5342 new_load = this_load;
5344 * Round up the averaging division if load is increasing. This
5345 * prevents us from getting stuck on 9 if the load is 10, for
5348 if (new_load > old_load)
5349 new_load += scale - 1;
5351 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5355 /* Used instead of source_load when we know the type == 0 */
5356 static unsigned long weighted_cpuload(struct rq *rq)
5358 return cfs_rq_runnable_load_avg(&rq->cfs);
5361 #ifdef CONFIG_NO_HZ_COMMON
5363 * There is no sane way to deal with nohz on smp when using jiffies because the
5364 * CPU doing the jiffies update might drift wrt the CPU doing the jiffy reading
5365 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5367 * Therefore we need to avoid the delta approach from the regular tick when
5368 * possible since that would seriously skew the load calculation. This is why we
5369 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5370 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5371 * loop exit, nohz_idle_balance, nohz full exit...)
5373 * This means we might still be one tick off for nohz periods.
5376 static void cpu_load_update_nohz(struct rq *this_rq,
5377 unsigned long curr_jiffies,
5380 unsigned long pending_updates;
5382 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5383 if (pending_updates) {
5384 this_rq->last_load_update_tick = curr_jiffies;
5386 * In the regular NOHZ case, we were idle, this means load 0.
5387 * In the NOHZ_FULL case, we were non-idle, we should consider
5388 * its weighted load.
5390 cpu_load_update(this_rq, load, pending_updates);
5395 * Called from nohz_idle_balance() to update the load ratings before doing the
5398 static void cpu_load_update_idle(struct rq *this_rq)
5401 * bail if there's load or we're actually up-to-date.
5403 if (weighted_cpuload(this_rq))
5406 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5410 * Record CPU load on nohz entry so we know the tickless load to account
5411 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5412 * than other cpu_load[idx] but it should be fine as cpu_load readers
5413 * shouldn't rely into synchronized cpu_load[*] updates.
5415 void cpu_load_update_nohz_start(void)
5417 struct rq *this_rq = this_rq();
5420 * This is all lockless but should be fine. If weighted_cpuload changes
5421 * concurrently we'll exit nohz. And cpu_load write can race with
5422 * cpu_load_update_idle() but both updater would be writing the same.
5424 this_rq->cpu_load[0] = weighted_cpuload(this_rq);
5428 * Account the tickless load in the end of a nohz frame.
5430 void cpu_load_update_nohz_stop(void)
5432 unsigned long curr_jiffies = READ_ONCE(jiffies);
5433 struct rq *this_rq = this_rq();
5437 if (curr_jiffies == this_rq->last_load_update_tick)
5440 load = weighted_cpuload(this_rq);
5441 rq_lock(this_rq, &rf);
5442 update_rq_clock(this_rq);
5443 cpu_load_update_nohz(this_rq, curr_jiffies, load);
5444 rq_unlock(this_rq, &rf);
5446 #else /* !CONFIG_NO_HZ_COMMON */
5447 static inline void cpu_load_update_nohz(struct rq *this_rq,
5448 unsigned long curr_jiffies,
5449 unsigned long load) { }
5450 #endif /* CONFIG_NO_HZ_COMMON */
5452 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5454 #ifdef CONFIG_NO_HZ_COMMON
5455 /* See the mess around cpu_load_update_nohz(). */
5456 this_rq->last_load_update_tick = READ_ONCE(jiffies);
5458 cpu_load_update(this_rq, load, 1);
5462 * Called from scheduler_tick()
5464 void cpu_load_update_active(struct rq *this_rq)
5466 unsigned long load = weighted_cpuload(this_rq);
5468 if (tick_nohz_tick_stopped())
5469 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5471 cpu_load_update_periodic(this_rq, load);
5475 * Return a low guess at the load of a migration-source CPU weighted
5476 * according to the scheduling class and "nice" value.
5478 * We want to under-estimate the load of migration sources, to
5479 * balance conservatively.
5481 static unsigned long source_load(int cpu, int type)
5483 struct rq *rq = cpu_rq(cpu);
5484 unsigned long total = weighted_cpuload(rq);
5486 if (type == 0 || !sched_feat(LB_BIAS))
5489 return min(rq->cpu_load[type-1], total);
5493 * Return a high guess at the load of a migration-target CPU weighted
5494 * according to the scheduling class and "nice" value.
5496 static unsigned long target_load(int cpu, int type)
5498 struct rq *rq = cpu_rq(cpu);
5499 unsigned long total = weighted_cpuload(rq);
5501 if (type == 0 || !sched_feat(LB_BIAS))
5504 return max(rq->cpu_load[type-1], total);
5507 static unsigned long capacity_of(int cpu)
5509 return cpu_rq(cpu)->cpu_capacity;
5512 static unsigned long capacity_orig_of(int cpu)
5514 return cpu_rq(cpu)->cpu_capacity_orig;
5517 static unsigned long cpu_avg_load_per_task(int cpu)
5519 struct rq *rq = cpu_rq(cpu);
5520 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5521 unsigned long load_avg = weighted_cpuload(rq);
5524 return load_avg / nr_running;
5529 static void record_wakee(struct task_struct *p)
5532 * Only decay a single time; tasks that have less then 1 wakeup per
5533 * jiffy will not have built up many flips.
5535 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5536 current->wakee_flips >>= 1;
5537 current->wakee_flip_decay_ts = jiffies;
5540 if (current->last_wakee != p) {
5541 current->last_wakee = p;
5542 current->wakee_flips++;
5547 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5549 * A waker of many should wake a different task than the one last awakened
5550 * at a frequency roughly N times higher than one of its wakees.
5552 * In order to determine whether we should let the load spread vs consolidating
5553 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5554 * partner, and a factor of lls_size higher frequency in the other.
5556 * With both conditions met, we can be relatively sure that the relationship is
5557 * non-monogamous, with partner count exceeding socket size.
5559 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5560 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5563 static int wake_wide(struct task_struct *p)
5565 unsigned int master = current->wakee_flips;
5566 unsigned int slave = p->wakee_flips;
5567 int factor = this_cpu_read(sd_llc_size);
5570 swap(master, slave);
5571 if (slave < factor || master < slave * factor)
5577 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5578 * soonest. For the purpose of speed we only consider the waking and previous
5581 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5582 * cache-affine and is (or will be) idle.
5584 * wake_affine_weight() - considers the weight to reflect the average
5585 * scheduling latency of the CPUs. This seems to work
5586 * for the overloaded case.
5589 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5592 * If this_cpu is idle, it implies the wakeup is from interrupt
5593 * context. Only allow the move if cache is shared. Otherwise an
5594 * interrupt intensive workload could force all tasks onto one
5595 * node depending on the IO topology or IRQ affinity settings.
5597 * If the prev_cpu is idle and cache affine then avoid a migration.
5598 * There is no guarantee that the cache hot data from an interrupt
5599 * is more important than cache hot data on the prev_cpu and from
5600 * a cpufreq perspective, it's better to have higher utilisation
5603 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5604 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5606 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5609 return nr_cpumask_bits;
5613 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5614 int this_cpu, int prev_cpu, int sync)
5616 s64 this_eff_load, prev_eff_load;
5617 unsigned long task_load;
5619 this_eff_load = target_load(this_cpu, sd->wake_idx);
5622 unsigned long current_load = task_h_load(current);
5624 if (current_load > this_eff_load)
5627 this_eff_load -= current_load;
5630 task_load = task_h_load(p);
5632 this_eff_load += task_load;
5633 if (sched_feat(WA_BIAS))
5634 this_eff_load *= 100;
5635 this_eff_load *= capacity_of(prev_cpu);
5637 prev_eff_load = source_load(prev_cpu, sd->wake_idx);
5638 prev_eff_load -= task_load;
5639 if (sched_feat(WA_BIAS))
5640 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5641 prev_eff_load *= capacity_of(this_cpu);
5644 * If sync, adjust the weight of prev_eff_load such that if
5645 * prev_eff == this_eff that select_idle_sibling() will consider
5646 * stacking the wakee on top of the waker if no other CPU is
5652 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5655 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5656 int this_cpu, int prev_cpu, int sync)
5658 int target = nr_cpumask_bits;
5660 if (sched_feat(WA_IDLE))
5661 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5663 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5664 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5666 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5667 if (target == nr_cpumask_bits)
5670 schedstat_inc(sd->ttwu_move_affine);
5671 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5675 static unsigned long cpu_util_without(int cpu, struct task_struct *p);
5677 static unsigned long capacity_spare_without(int cpu, struct task_struct *p)
5679 return max_t(long, capacity_of(cpu) - cpu_util_without(cpu, p), 0);
5683 * find_idlest_group finds and returns the least busy CPU group within the
5686 * Assumes p is allowed on at least one CPU in sd.
5688 static struct sched_group *
5689 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5690 int this_cpu, int sd_flag)
5692 struct sched_group *idlest = NULL, *group = sd->groups;
5693 struct sched_group *most_spare_sg = NULL;
5694 unsigned long min_runnable_load = ULONG_MAX;
5695 unsigned long this_runnable_load = ULONG_MAX;
5696 unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5697 unsigned long most_spare = 0, this_spare = 0;
5698 int load_idx = sd->forkexec_idx;
5699 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5700 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5701 (sd->imbalance_pct-100) / 100;
5703 if (sd_flag & SD_BALANCE_WAKE)
5704 load_idx = sd->wake_idx;
5707 unsigned long load, avg_load, runnable_load;
5708 unsigned long spare_cap, max_spare_cap;
5712 /* Skip over this group if it has no CPUs allowed */
5713 if (!cpumask_intersects(sched_group_span(group),
5717 local_group = cpumask_test_cpu(this_cpu,
5718 sched_group_span(group));
5721 * Tally up the load of all CPUs in the group and find
5722 * the group containing the CPU with most spare capacity.
5728 for_each_cpu(i, sched_group_span(group)) {
5729 /* Bias balancing toward CPUs of our domain */
5731 load = source_load(i, load_idx);
5733 load = target_load(i, load_idx);
5735 runnable_load += load;
5737 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5739 spare_cap = capacity_spare_without(i, p);
5741 if (spare_cap > max_spare_cap)
5742 max_spare_cap = spare_cap;
5745 /* Adjust by relative CPU capacity of the group */
5746 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5747 group->sgc->capacity;
5748 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5749 group->sgc->capacity;
5752 this_runnable_load = runnable_load;
5753 this_avg_load = avg_load;
5754 this_spare = max_spare_cap;
5756 if (min_runnable_load > (runnable_load + imbalance)) {
5758 * The runnable load is significantly smaller
5759 * so we can pick this new CPU:
5761 min_runnable_load = runnable_load;
5762 min_avg_load = avg_load;
5764 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5765 (100*min_avg_load > imbalance_scale*avg_load)) {
5767 * The runnable loads are close so take the
5768 * blocked load into account through avg_load:
5770 min_avg_load = avg_load;
5774 if (most_spare < max_spare_cap) {
5775 most_spare = max_spare_cap;
5776 most_spare_sg = group;
5779 } while (group = group->next, group != sd->groups);
5782 * The cross-over point between using spare capacity or least load
5783 * is too conservative for high utilization tasks on partially
5784 * utilized systems if we require spare_capacity > task_util(p),
5785 * so we allow for some task stuffing by using
5786 * spare_capacity > task_util(p)/2.
5788 * Spare capacity can't be used for fork because the utilization has
5789 * not been set yet, we must first select a rq to compute the initial
5792 if (sd_flag & SD_BALANCE_FORK)
5795 if (this_spare > task_util(p) / 2 &&
5796 imbalance_scale*this_spare > 100*most_spare)
5799 if (most_spare > task_util(p) / 2)
5800 return most_spare_sg;
5807 * When comparing groups across NUMA domains, it's possible for the
5808 * local domain to be very lightly loaded relative to the remote
5809 * domains but "imbalance" skews the comparison making remote CPUs
5810 * look much more favourable. When considering cross-domain, add
5811 * imbalance to the runnable load on the remote node and consider
5814 if ((sd->flags & SD_NUMA) &&
5815 min_runnable_load + imbalance >= this_runnable_load)
5818 if (min_runnable_load > (this_runnable_load + imbalance))
5821 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5822 (100*this_avg_load < imbalance_scale*min_avg_load))
5829 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5832 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5834 unsigned long load, min_load = ULONG_MAX;
5835 unsigned int min_exit_latency = UINT_MAX;
5836 u64 latest_idle_timestamp = 0;
5837 int least_loaded_cpu = this_cpu;
5838 int shallowest_idle_cpu = -1;
5841 /* Check if we have any choice: */
5842 if (group->group_weight == 1)
5843 return cpumask_first(sched_group_span(group));
5845 /* Traverse only the allowed CPUs */
5846 for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
5847 if (available_idle_cpu(i)) {
5848 struct rq *rq = cpu_rq(i);
5849 struct cpuidle_state *idle = idle_get_state(rq);
5850 if (idle && idle->exit_latency < min_exit_latency) {
5852 * We give priority to a CPU whose idle state
5853 * has the smallest exit latency irrespective
5854 * of any idle timestamp.
5856 min_exit_latency = idle->exit_latency;
5857 latest_idle_timestamp = rq->idle_stamp;
5858 shallowest_idle_cpu = i;
5859 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5860 rq->idle_stamp > latest_idle_timestamp) {
5862 * If equal or no active idle state, then
5863 * the most recently idled CPU might have
5866 latest_idle_timestamp = rq->idle_stamp;
5867 shallowest_idle_cpu = i;
5869 } else if (shallowest_idle_cpu == -1) {
5870 load = weighted_cpuload(cpu_rq(i));
5871 if (load < min_load) {
5873 least_loaded_cpu = i;
5878 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5881 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5882 int cpu, int prev_cpu, int sd_flag)
5886 if (!cpumask_intersects(sched_domain_span(sd), &p->cpus_allowed))
5890 * We need task's util for capacity_spare_without, sync it up to
5891 * prev_cpu's last_update_time.
5893 if (!(sd_flag & SD_BALANCE_FORK))
5894 sync_entity_load_avg(&p->se);
5897 struct sched_group *group;
5898 struct sched_domain *tmp;
5901 if (!(sd->flags & sd_flag)) {
5906 group = find_idlest_group(sd, p, cpu, sd_flag);
5912 new_cpu = find_idlest_group_cpu(group, p, cpu);
5913 if (new_cpu == cpu) {
5914 /* Now try balancing at a lower domain level of 'cpu': */
5919 /* Now try balancing at a lower domain level of 'new_cpu': */
5921 weight = sd->span_weight;
5923 for_each_domain(cpu, tmp) {
5924 if (weight <= tmp->span_weight)
5926 if (tmp->flags & sd_flag)
5934 #ifdef CONFIG_SCHED_SMT
5935 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5937 static inline void set_idle_cores(int cpu, int val)
5939 struct sched_domain_shared *sds;
5941 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5943 WRITE_ONCE(sds->has_idle_cores, val);
5946 static inline bool test_idle_cores(int cpu, bool def)
5948 struct sched_domain_shared *sds;
5950 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5952 return READ_ONCE(sds->has_idle_cores);
5958 * Scans the local SMT mask to see if the entire core is idle, and records this
5959 * information in sd_llc_shared->has_idle_cores.
5961 * Since SMT siblings share all cache levels, inspecting this limited remote
5962 * state should be fairly cheap.
5964 void __update_idle_core(struct rq *rq)
5966 int core = cpu_of(rq);
5970 if (test_idle_cores(core, true))
5973 for_each_cpu(cpu, cpu_smt_mask(core)) {
5977 if (!available_idle_cpu(cpu))
5981 set_idle_cores(core, 1);
5987 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5988 * there are no idle cores left in the system; tracked through
5989 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5991 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5993 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5996 if (!static_branch_likely(&sched_smt_present))
5999 if (!test_idle_cores(target, false))
6002 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
6004 for_each_cpu_wrap(core, cpus, target) {
6007 for_each_cpu(cpu, cpu_smt_mask(core)) {
6008 cpumask_clear_cpu(cpu, cpus);
6009 if (!available_idle_cpu(cpu))
6018 * Failed to find an idle core; stop looking for one.
6020 set_idle_cores(target, 0);
6026 * Scan the local SMT mask for idle CPUs.
6028 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6032 if (!static_branch_likely(&sched_smt_present))
6035 for_each_cpu(cpu, cpu_smt_mask(target)) {
6036 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6038 if (available_idle_cpu(cpu))
6045 #else /* CONFIG_SCHED_SMT */
6047 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6052 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6057 #endif /* CONFIG_SCHED_SMT */
6060 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6061 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6062 * average idle time for this rq (as found in rq->avg_idle).
6064 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
6066 struct sched_domain *this_sd;
6067 u64 avg_cost, avg_idle;
6070 int cpu, nr = INT_MAX;
6072 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6077 * Due to large variance we need a large fuzz factor; hackbench in
6078 * particularly is sensitive here.
6080 avg_idle = this_rq()->avg_idle / 512;
6081 avg_cost = this_sd->avg_scan_cost + 1;
6083 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
6086 if (sched_feat(SIS_PROP)) {
6087 u64 span_avg = sd->span_weight * avg_idle;
6088 if (span_avg > 4*avg_cost)
6089 nr = div_u64(span_avg, avg_cost);
6094 time = local_clock();
6096 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
6099 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6101 if (available_idle_cpu(cpu))
6105 time = local_clock() - time;
6106 cost = this_sd->avg_scan_cost;
6107 delta = (s64)(time - cost) / 8;
6108 this_sd->avg_scan_cost += delta;
6114 * Try and locate an idle core/thread in the LLC cache domain.
6116 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6118 struct sched_domain *sd;
6119 int i, recent_used_cpu;
6121 if (available_idle_cpu(target))
6125 * If the previous CPU is cache affine and idle, don't be stupid:
6127 if (prev != target && cpus_share_cache(prev, target) && available_idle_cpu(prev))
6130 /* Check a recently used CPU as a potential idle candidate: */
6131 recent_used_cpu = p->recent_used_cpu;
6132 if (recent_used_cpu != prev &&
6133 recent_used_cpu != target &&
6134 cpus_share_cache(recent_used_cpu, target) &&
6135 available_idle_cpu(recent_used_cpu) &&
6136 cpumask_test_cpu(p->recent_used_cpu, &p->cpus_allowed)) {
6138 * Replace recent_used_cpu with prev as it is a potential
6139 * candidate for the next wake:
6141 p->recent_used_cpu = prev;
6142 return recent_used_cpu;
6145 sd = rcu_dereference(per_cpu(sd_llc, target));
6149 i = select_idle_core(p, sd, target);
6150 if ((unsigned)i < nr_cpumask_bits)
6153 i = select_idle_cpu(p, sd, target);
6154 if ((unsigned)i < nr_cpumask_bits)
6157 i = select_idle_smt(p, sd, target);
6158 if ((unsigned)i < nr_cpumask_bits)
6165 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6166 * @cpu: the CPU to get the utilization of
6168 * The unit of the return value must be the one of capacity so we can compare
6169 * the utilization with the capacity of the CPU that is available for CFS task
6170 * (ie cpu_capacity).
6172 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6173 * recent utilization of currently non-runnable tasks on a CPU. It represents
6174 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6175 * capacity_orig is the cpu_capacity available at the highest frequency
6176 * (arch_scale_freq_capacity()).
6177 * The utilization of a CPU converges towards a sum equal to or less than the
6178 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6179 * the running time on this CPU scaled by capacity_curr.
6181 * The estimated utilization of a CPU is defined to be the maximum between its
6182 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6183 * currently RUNNABLE on that CPU.
6184 * This allows to properly represent the expected utilization of a CPU which
6185 * has just got a big task running since a long sleep period. At the same time
6186 * however it preserves the benefits of the "blocked utilization" in
6187 * describing the potential for other tasks waking up on the same CPU.
6189 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6190 * higher than capacity_orig because of unfortunate rounding in
6191 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6192 * the average stabilizes with the new running time. We need to check that the
6193 * utilization stays within the range of [0..capacity_orig] and cap it if
6194 * necessary. Without utilization capping, a group could be seen as overloaded
6195 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6196 * available capacity. We allow utilization to overshoot capacity_curr (but not
6197 * capacity_orig) as it useful for predicting the capacity required after task
6198 * migrations (scheduler-driven DVFS).
6200 * Return: the (estimated) utilization for the specified CPU
6202 static inline unsigned long cpu_util(int cpu)
6204 struct cfs_rq *cfs_rq;
6207 cfs_rq = &cpu_rq(cpu)->cfs;
6208 util = READ_ONCE(cfs_rq->avg.util_avg);
6210 if (sched_feat(UTIL_EST))
6211 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6213 return min_t(unsigned long, util, capacity_orig_of(cpu));
6217 * cpu_util_without: compute cpu utilization without any contributions from *p
6218 * @cpu: the CPU which utilization is requested
6219 * @p: the task which utilization should be discounted
6221 * The utilization of a CPU is defined by the utilization of tasks currently
6222 * enqueued on that CPU as well as tasks which are currently sleeping after an
6223 * execution on that CPU.
6225 * This method returns the utilization of the specified CPU by discounting the
6226 * utilization of the specified task, whenever the task is currently
6227 * contributing to the CPU utilization.
6229 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6231 struct cfs_rq *cfs_rq;
6234 /* Task has no contribution or is new */
6235 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6236 return cpu_util(cpu);
6238 cfs_rq = &cpu_rq(cpu)->cfs;
6239 util = READ_ONCE(cfs_rq->avg.util_avg);
6241 /* Discount task's util from CPU's util */
6242 util -= min_t(unsigned int, util, task_util(p));
6247 * a) if *p is the only task sleeping on this CPU, then:
6248 * cpu_util (== task_util) > util_est (== 0)
6249 * and thus we return:
6250 * cpu_util_without = (cpu_util - task_util) = 0
6252 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6254 * cpu_util >= task_util
6255 * cpu_util > util_est (== 0)
6256 * and thus we discount *p's blocked utilization to return:
6257 * cpu_util_without = (cpu_util - task_util) >= 0
6259 * c) if other tasks are RUNNABLE on that CPU and
6260 * util_est > cpu_util
6261 * then we use util_est since it returns a more restrictive
6262 * estimation of the spare capacity on that CPU, by just
6263 * considering the expected utilization of tasks already
6264 * runnable on that CPU.
6266 * Cases a) and b) are covered by the above code, while case c) is
6267 * covered by the following code when estimated utilization is
6270 if (sched_feat(UTIL_EST)) {
6271 unsigned int estimated =
6272 READ_ONCE(cfs_rq->avg.util_est.enqueued);
6275 * Despite the following checks we still have a small window
6276 * for a possible race, when an execl's select_task_rq_fair()
6277 * races with LB's detach_task():
6280 * p->on_rq = TASK_ON_RQ_MIGRATING;
6281 * ---------------------------------- A
6282 * deactivate_task() \
6283 * dequeue_task() + RaceTime
6284 * util_est_dequeue() /
6285 * ---------------------------------- B
6287 * The additional check on "current == p" it's required to
6288 * properly fix the execl regression and it helps in further
6289 * reducing the chances for the above race.
6291 if (unlikely(task_on_rq_queued(p) || current == p)) {
6292 estimated -= min_t(unsigned int, estimated,
6293 (_task_util_est(p) | UTIL_AVG_UNCHANGED));
6295 util = max(util, estimated);
6299 * Utilization (estimated) can exceed the CPU capacity, thus let's
6300 * clamp to the maximum CPU capacity to ensure consistency with
6301 * the cpu_util call.
6303 return min_t(unsigned long, util, capacity_orig_of(cpu));
6307 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6308 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6310 * In that case WAKE_AFFINE doesn't make sense and we'll let
6311 * BALANCE_WAKE sort things out.
6313 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6315 long min_cap, max_cap;
6317 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6318 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6320 /* Minimum capacity is close to max, no need to abort wake_affine */
6321 if (max_cap - min_cap < max_cap >> 3)
6324 /* Bring task utilization in sync with prev_cpu */
6325 sync_entity_load_avg(&p->se);
6327 return min_cap * 1024 < task_util(p) * capacity_margin;
6331 * select_task_rq_fair: Select target runqueue for the waking task in domains
6332 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6333 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6335 * Balances load by selecting the idlest CPU in the idlest group, or under
6336 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6338 * Returns the target CPU number.
6340 * preempt must be disabled.
6343 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6345 struct sched_domain *tmp, *sd = NULL;
6346 int cpu = smp_processor_id();
6347 int new_cpu = prev_cpu;
6348 int want_affine = 0;
6349 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6351 if (sd_flag & SD_BALANCE_WAKE) {
6353 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
6354 && cpumask_test_cpu(cpu, &p->cpus_allowed);
6358 for_each_domain(cpu, tmp) {
6359 if (!(tmp->flags & SD_LOAD_BALANCE))
6363 * If both 'cpu' and 'prev_cpu' are part of this domain,
6364 * cpu is a valid SD_WAKE_AFFINE target.
6366 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6367 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6368 if (cpu != prev_cpu)
6369 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6371 sd = NULL; /* Prefer wake_affine over balance flags */
6375 if (tmp->flags & sd_flag)
6377 else if (!want_affine)
6383 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6384 } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6387 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6390 current->recent_used_cpu = cpu;
6397 static void detach_entity_cfs_rq(struct sched_entity *se);
6400 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6401 * cfs_rq_of(p) references at time of call are still valid and identify the
6402 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6404 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6407 * As blocked tasks retain absolute vruntime the migration needs to
6408 * deal with this by subtracting the old and adding the new
6409 * min_vruntime -- the latter is done by enqueue_entity() when placing
6410 * the task on the new runqueue.
6412 if (p->state == TASK_WAKING) {
6413 struct sched_entity *se = &p->se;
6414 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6417 #ifndef CONFIG_64BIT
6418 u64 min_vruntime_copy;
6421 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6423 min_vruntime = cfs_rq->min_vruntime;
6424 } while (min_vruntime != min_vruntime_copy);
6426 min_vruntime = cfs_rq->min_vruntime;
6429 se->vruntime -= min_vruntime;
6432 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6434 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6435 * rq->lock and can modify state directly.
6437 lockdep_assert_held(&task_rq(p)->lock);
6438 detach_entity_cfs_rq(&p->se);
6442 * We are supposed to update the task to "current" time, then
6443 * its up to date and ready to go to new CPU/cfs_rq. But we
6444 * have difficulty in getting what current time is, so simply
6445 * throw away the out-of-date time. This will result in the
6446 * wakee task is less decayed, but giving the wakee more load
6449 remove_entity_load_avg(&p->se);
6452 /* Tell new CPU we are migrated */
6453 p->se.avg.last_update_time = 0;
6455 /* We have migrated, no longer consider this task hot */
6456 p->se.exec_start = 0;
6458 update_scan_period(p, new_cpu);
6461 static void task_dead_fair(struct task_struct *p)
6463 remove_entity_load_avg(&p->se);
6465 #endif /* CONFIG_SMP */
6467 static unsigned long wakeup_gran(struct sched_entity *se)
6469 unsigned long gran = sysctl_sched_wakeup_granularity;
6472 * Since its curr running now, convert the gran from real-time
6473 * to virtual-time in his units.
6475 * By using 'se' instead of 'curr' we penalize light tasks, so
6476 * they get preempted easier. That is, if 'se' < 'curr' then
6477 * the resulting gran will be larger, therefore penalizing the
6478 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6479 * be smaller, again penalizing the lighter task.
6481 * This is especially important for buddies when the leftmost
6482 * task is higher priority than the buddy.
6484 return calc_delta_fair(gran, se);
6488 * Should 'se' preempt 'curr'.
6502 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6504 s64 gran, vdiff = curr->vruntime - se->vruntime;
6509 gran = wakeup_gran(se);
6516 static void set_last_buddy(struct sched_entity *se)
6518 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6521 for_each_sched_entity(se) {
6522 if (SCHED_WARN_ON(!se->on_rq))
6524 cfs_rq_of(se)->last = se;
6528 static void set_next_buddy(struct sched_entity *se)
6530 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6533 for_each_sched_entity(se) {
6534 if (SCHED_WARN_ON(!se->on_rq))
6536 cfs_rq_of(se)->next = se;
6540 static void set_skip_buddy(struct sched_entity *se)
6542 for_each_sched_entity(se)
6543 cfs_rq_of(se)->skip = se;
6547 * Preempt the current task with a newly woken task if needed:
6549 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6551 struct task_struct *curr = rq->curr;
6552 struct sched_entity *se = &curr->se, *pse = &p->se;
6553 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6554 int scale = cfs_rq->nr_running >= sched_nr_latency;
6555 int next_buddy_marked = 0;
6557 if (unlikely(se == pse))
6561 * This is possible from callers such as attach_tasks(), in which we
6562 * unconditionally check_prempt_curr() after an enqueue (which may have
6563 * lead to a throttle). This both saves work and prevents false
6564 * next-buddy nomination below.
6566 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6569 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6570 set_next_buddy(pse);
6571 next_buddy_marked = 1;
6575 * We can come here with TIF_NEED_RESCHED already set from new task
6578 * Note: this also catches the edge-case of curr being in a throttled
6579 * group (e.g. via set_curr_task), since update_curr() (in the
6580 * enqueue of curr) will have resulted in resched being set. This
6581 * prevents us from potentially nominating it as a false LAST_BUDDY
6584 if (test_tsk_need_resched(curr))
6587 /* Idle tasks are by definition preempted by non-idle tasks. */
6588 if (unlikely(curr->policy == SCHED_IDLE) &&
6589 likely(p->policy != SCHED_IDLE))
6593 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6594 * is driven by the tick):
6596 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6599 find_matching_se(&se, &pse);
6600 update_curr(cfs_rq_of(se));
6602 if (wakeup_preempt_entity(se, pse) == 1) {
6604 * Bias pick_next to pick the sched entity that is
6605 * triggering this preemption.
6607 if (!next_buddy_marked)
6608 set_next_buddy(pse);
6617 * Only set the backward buddy when the current task is still
6618 * on the rq. This can happen when a wakeup gets interleaved
6619 * with schedule on the ->pre_schedule() or idle_balance()
6620 * point, either of which can * drop the rq lock.
6622 * Also, during early boot the idle thread is in the fair class,
6623 * for obvious reasons its a bad idea to schedule back to it.
6625 if (unlikely(!se->on_rq || curr == rq->idle))
6628 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6632 static struct task_struct *
6633 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6635 struct cfs_rq *cfs_rq = &rq->cfs;
6636 struct sched_entity *se;
6637 struct task_struct *p;
6641 if (!cfs_rq->nr_running)
6644 #ifdef CONFIG_FAIR_GROUP_SCHED
6645 if (prev->sched_class != &fair_sched_class)
6649 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6650 * likely that a next task is from the same cgroup as the current.
6652 * Therefore attempt to avoid putting and setting the entire cgroup
6653 * hierarchy, only change the part that actually changes.
6657 struct sched_entity *curr = cfs_rq->curr;
6660 * Since we got here without doing put_prev_entity() we also
6661 * have to consider cfs_rq->curr. If it is still a runnable
6662 * entity, update_curr() will update its vruntime, otherwise
6663 * forget we've ever seen it.
6667 update_curr(cfs_rq);
6672 * This call to check_cfs_rq_runtime() will do the
6673 * throttle and dequeue its entity in the parent(s).
6674 * Therefore the nr_running test will indeed
6677 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6680 if (!cfs_rq->nr_running)
6687 se = pick_next_entity(cfs_rq, curr);
6688 cfs_rq = group_cfs_rq(se);
6694 * Since we haven't yet done put_prev_entity and if the selected task
6695 * is a different task than we started out with, try and touch the
6696 * least amount of cfs_rqs.
6699 struct sched_entity *pse = &prev->se;
6701 while (!(cfs_rq = is_same_group(se, pse))) {
6702 int se_depth = se->depth;
6703 int pse_depth = pse->depth;
6705 if (se_depth <= pse_depth) {
6706 put_prev_entity(cfs_rq_of(pse), pse);
6707 pse = parent_entity(pse);
6709 if (se_depth >= pse_depth) {
6710 set_next_entity(cfs_rq_of(se), se);
6711 se = parent_entity(se);
6715 put_prev_entity(cfs_rq, pse);
6716 set_next_entity(cfs_rq, se);
6723 put_prev_task(rq, prev);
6726 se = pick_next_entity(cfs_rq, NULL);
6727 set_next_entity(cfs_rq, se);
6728 cfs_rq = group_cfs_rq(se);
6733 done: __maybe_unused;
6736 * Move the next running task to the front of
6737 * the list, so our cfs_tasks list becomes MRU
6740 list_move(&p->se.group_node, &rq->cfs_tasks);
6743 if (hrtick_enabled(rq))
6744 hrtick_start_fair(rq, p);
6749 new_tasks = idle_balance(rq, rf);
6752 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6753 * possible for any higher priority task to appear. In that case we
6754 * must re-start the pick_next_entity() loop.
6766 * Account for a descheduled task:
6768 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6770 struct sched_entity *se = &prev->se;
6771 struct cfs_rq *cfs_rq;
6773 for_each_sched_entity(se) {
6774 cfs_rq = cfs_rq_of(se);
6775 put_prev_entity(cfs_rq, se);
6780 * sched_yield() is very simple
6782 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6784 static void yield_task_fair(struct rq *rq)
6786 struct task_struct *curr = rq->curr;
6787 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6788 struct sched_entity *se = &curr->se;
6791 * Are we the only task in the tree?
6793 if (unlikely(rq->nr_running == 1))
6796 clear_buddies(cfs_rq, se);
6798 if (curr->policy != SCHED_BATCH) {
6799 update_rq_clock(rq);
6801 * Update run-time statistics of the 'current'.
6803 update_curr(cfs_rq);
6805 * Tell update_rq_clock() that we've just updated,
6806 * so we don't do microscopic update in schedule()
6807 * and double the fastpath cost.
6809 rq_clock_skip_update(rq);
6815 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6817 struct sched_entity *se = &p->se;
6819 /* throttled hierarchies are not runnable */
6820 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6823 /* Tell the scheduler that we'd really like pse to run next. */
6826 yield_task_fair(rq);
6832 /**************************************************
6833 * Fair scheduling class load-balancing methods.
6837 * The purpose of load-balancing is to achieve the same basic fairness the
6838 * per-CPU scheduler provides, namely provide a proportional amount of compute
6839 * time to each task. This is expressed in the following equation:
6841 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6843 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
6844 * W_i,0 is defined as:
6846 * W_i,0 = \Sum_j w_i,j (2)
6848 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
6849 * is derived from the nice value as per sched_prio_to_weight[].
6851 * The weight average is an exponential decay average of the instantaneous
6854 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6856 * C_i is the compute capacity of CPU i, typically it is the
6857 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6858 * can also include other factors [XXX].
6860 * To achieve this balance we define a measure of imbalance which follows
6861 * directly from (1):
6863 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6865 * We them move tasks around to minimize the imbalance. In the continuous
6866 * function space it is obvious this converges, in the discrete case we get
6867 * a few fun cases generally called infeasible weight scenarios.
6870 * - infeasible weights;
6871 * - local vs global optima in the discrete case. ]
6876 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6877 * for all i,j solution, we create a tree of CPUs that follows the hardware
6878 * topology where each level pairs two lower groups (or better). This results
6879 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
6880 * tree to only the first of the previous level and we decrease the frequency
6881 * of load-balance at each level inv. proportional to the number of CPUs in
6887 * \Sum { --- * --- * 2^i } = O(n) (5)
6889 * `- size of each group
6890 * | | `- number of CPUs doing load-balance
6892 * `- sum over all levels
6894 * Coupled with a limit on how many tasks we can migrate every balance pass,
6895 * this makes (5) the runtime complexity of the balancer.
6897 * An important property here is that each CPU is still (indirectly) connected
6898 * to every other CPU in at most O(log n) steps:
6900 * The adjacency matrix of the resulting graph is given by:
6903 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6906 * And you'll find that:
6908 * A^(log_2 n)_i,j != 0 for all i,j (7)
6910 * Showing there's indeed a path between every CPU in at most O(log n) steps.
6911 * The task movement gives a factor of O(m), giving a convergence complexity
6914 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6919 * In order to avoid CPUs going idle while there's still work to do, new idle
6920 * balancing is more aggressive and has the newly idle CPU iterate up the domain
6921 * tree itself instead of relying on other CPUs to bring it work.
6923 * This adds some complexity to both (5) and (8) but it reduces the total idle
6931 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6934 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6939 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6941 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
6943 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6946 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6947 * rewrite all of this once again.]
6950 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6952 enum fbq_type { regular, remote, all };
6954 #define LBF_ALL_PINNED 0x01
6955 #define LBF_NEED_BREAK 0x02
6956 #define LBF_DST_PINNED 0x04
6957 #define LBF_SOME_PINNED 0x08
6958 #define LBF_NOHZ_STATS 0x10
6959 #define LBF_NOHZ_AGAIN 0x20
6962 struct sched_domain *sd;
6970 struct cpumask *dst_grpmask;
6972 enum cpu_idle_type idle;
6974 /* The set of CPUs under consideration for load-balancing */
6975 struct cpumask *cpus;
6980 unsigned int loop_break;
6981 unsigned int loop_max;
6983 enum fbq_type fbq_type;
6984 struct list_head tasks;
6988 * Is this task likely cache-hot:
6990 static int task_hot(struct task_struct *p, struct lb_env *env)
6994 lockdep_assert_held(&env->src_rq->lock);
6996 if (p->sched_class != &fair_sched_class)
6999 if (unlikely(p->policy == SCHED_IDLE))
7003 * Buddy candidates are cache hot:
7005 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7006 (&p->se == cfs_rq_of(&p->se)->next ||
7007 &p->se == cfs_rq_of(&p->se)->last))
7010 if (sysctl_sched_migration_cost == -1)
7012 if (sysctl_sched_migration_cost == 0)
7015 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7017 return delta < (s64)sysctl_sched_migration_cost;
7020 #ifdef CONFIG_NUMA_BALANCING
7022 * Returns 1, if task migration degrades locality
7023 * Returns 0, if task migration improves locality i.e migration preferred.
7024 * Returns -1, if task migration is not affected by locality.
7026 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7028 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7029 unsigned long src_weight, dst_weight;
7030 int src_nid, dst_nid, dist;
7032 if (!static_branch_likely(&sched_numa_balancing))
7035 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7038 src_nid = cpu_to_node(env->src_cpu);
7039 dst_nid = cpu_to_node(env->dst_cpu);
7041 if (src_nid == dst_nid)
7044 /* Migrating away from the preferred node is always bad. */
7045 if (src_nid == p->numa_preferred_nid) {
7046 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7052 /* Encourage migration to the preferred node. */
7053 if (dst_nid == p->numa_preferred_nid)
7056 /* Leaving a core idle is often worse than degrading locality. */
7057 if (env->idle == CPU_IDLE)
7060 dist = node_distance(src_nid, dst_nid);
7062 src_weight = group_weight(p, src_nid, dist);
7063 dst_weight = group_weight(p, dst_nid, dist);
7065 src_weight = task_weight(p, src_nid, dist);
7066 dst_weight = task_weight(p, dst_nid, dist);
7069 return dst_weight < src_weight;
7073 static inline int migrate_degrades_locality(struct task_struct *p,
7081 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7084 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7088 lockdep_assert_held(&env->src_rq->lock);
7091 * We do not migrate tasks that are:
7092 * 1) throttled_lb_pair, or
7093 * 2) cannot be migrated to this CPU due to cpus_allowed, or
7094 * 3) running (obviously), or
7095 * 4) are cache-hot on their current CPU.
7097 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7100 if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) {
7103 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7105 env->flags |= LBF_SOME_PINNED;
7108 * Remember if this task can be migrated to any other CPU in
7109 * our sched_group. We may want to revisit it if we couldn't
7110 * meet load balance goals by pulling other tasks on src_cpu.
7112 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7113 * already computed one in current iteration.
7115 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7118 /* Prevent to re-select dst_cpu via env's CPUs: */
7119 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7120 if (cpumask_test_cpu(cpu, &p->cpus_allowed)) {
7121 env->flags |= LBF_DST_PINNED;
7122 env->new_dst_cpu = cpu;
7130 /* Record that we found atleast one task that could run on dst_cpu */
7131 env->flags &= ~LBF_ALL_PINNED;
7133 if (task_running(env->src_rq, p)) {
7134 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7139 * Aggressive migration if:
7140 * 1) destination numa is preferred
7141 * 2) task is cache cold, or
7142 * 3) too many balance attempts have failed.
7144 tsk_cache_hot = migrate_degrades_locality(p, env);
7145 if (tsk_cache_hot == -1)
7146 tsk_cache_hot = task_hot(p, env);
7148 if (tsk_cache_hot <= 0 ||
7149 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7150 if (tsk_cache_hot == 1) {
7151 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7152 schedstat_inc(p->se.statistics.nr_forced_migrations);
7157 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7162 * detach_task() -- detach the task for the migration specified in env
7164 static void detach_task(struct task_struct *p, struct lb_env *env)
7166 lockdep_assert_held(&env->src_rq->lock);
7168 p->on_rq = TASK_ON_RQ_MIGRATING;
7169 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7170 set_task_cpu(p, env->dst_cpu);
7174 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7175 * part of active balancing operations within "domain".
7177 * Returns a task if successful and NULL otherwise.
7179 static struct task_struct *detach_one_task(struct lb_env *env)
7181 struct task_struct *p;
7183 lockdep_assert_held(&env->src_rq->lock);
7185 list_for_each_entry_reverse(p,
7186 &env->src_rq->cfs_tasks, se.group_node) {
7187 if (!can_migrate_task(p, env))
7190 detach_task(p, env);
7193 * Right now, this is only the second place where
7194 * lb_gained[env->idle] is updated (other is detach_tasks)
7195 * so we can safely collect stats here rather than
7196 * inside detach_tasks().
7198 schedstat_inc(env->sd->lb_gained[env->idle]);
7204 static const unsigned int sched_nr_migrate_break = 32;
7207 * detach_tasks() -- tries to detach up to imbalance weighted load from
7208 * busiest_rq, as part of a balancing operation within domain "sd".
7210 * Returns number of detached tasks if successful and 0 otherwise.
7212 static int detach_tasks(struct lb_env *env)
7214 struct list_head *tasks = &env->src_rq->cfs_tasks;
7215 struct task_struct *p;
7219 lockdep_assert_held(&env->src_rq->lock);
7221 if (env->imbalance <= 0)
7224 while (!list_empty(tasks)) {
7226 * We don't want to steal all, otherwise we may be treated likewise,
7227 * which could at worst lead to a livelock crash.
7229 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7232 p = list_last_entry(tasks, struct task_struct, se.group_node);
7235 /* We've more or less seen every task there is, call it quits */
7236 if (env->loop > env->loop_max)
7239 /* take a breather every nr_migrate tasks */
7240 if (env->loop > env->loop_break) {
7241 env->loop_break += sched_nr_migrate_break;
7242 env->flags |= LBF_NEED_BREAK;
7246 if (!can_migrate_task(p, env))
7249 load = task_h_load(p);
7251 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7254 if ((load / 2) > env->imbalance)
7257 detach_task(p, env);
7258 list_add(&p->se.group_node, &env->tasks);
7261 env->imbalance -= load;
7263 #ifdef CONFIG_PREEMPT
7265 * NEWIDLE balancing is a source of latency, so preemptible
7266 * kernels will stop after the first task is detached to minimize
7267 * the critical section.
7269 if (env->idle == CPU_NEWLY_IDLE)
7274 * We only want to steal up to the prescribed amount of
7277 if (env->imbalance <= 0)
7282 list_move(&p->se.group_node, tasks);
7286 * Right now, this is one of only two places we collect this stat
7287 * so we can safely collect detach_one_task() stats here rather
7288 * than inside detach_one_task().
7290 schedstat_add(env->sd->lb_gained[env->idle], detached);
7296 * attach_task() -- attach the task detached by detach_task() to its new rq.
7298 static void attach_task(struct rq *rq, struct task_struct *p)
7300 lockdep_assert_held(&rq->lock);
7302 BUG_ON(task_rq(p) != rq);
7303 activate_task(rq, p, ENQUEUE_NOCLOCK);
7304 p->on_rq = TASK_ON_RQ_QUEUED;
7305 check_preempt_curr(rq, p, 0);
7309 * attach_one_task() -- attaches the task returned from detach_one_task() to
7312 static void attach_one_task(struct rq *rq, struct task_struct *p)
7317 update_rq_clock(rq);
7323 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7326 static void attach_tasks(struct lb_env *env)
7328 struct list_head *tasks = &env->tasks;
7329 struct task_struct *p;
7332 rq_lock(env->dst_rq, &rf);
7333 update_rq_clock(env->dst_rq);
7335 while (!list_empty(tasks)) {
7336 p = list_first_entry(tasks, struct task_struct, se.group_node);
7337 list_del_init(&p->se.group_node);
7339 attach_task(env->dst_rq, p);
7342 rq_unlock(env->dst_rq, &rf);
7345 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7347 if (cfs_rq->avg.load_avg)
7350 if (cfs_rq->avg.util_avg)
7356 static inline bool others_have_blocked(struct rq *rq)
7358 if (READ_ONCE(rq->avg_rt.util_avg))
7361 if (READ_ONCE(rq->avg_dl.util_avg))
7364 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7365 if (READ_ONCE(rq->avg_irq.util_avg))
7372 #ifdef CONFIG_FAIR_GROUP_SCHED
7374 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7376 if (cfs_rq->load.weight)
7379 if (cfs_rq->avg.load_sum)
7382 if (cfs_rq->avg.util_sum)
7385 if (cfs_rq->avg.runnable_load_sum)
7391 static void update_blocked_averages(int cpu)
7393 struct rq *rq = cpu_rq(cpu);
7394 struct cfs_rq *cfs_rq, *pos;
7395 const struct sched_class *curr_class;
7399 rq_lock_irqsave(rq, &rf);
7400 update_rq_clock(rq);
7403 * Iterates the task_group tree in a bottom up fashion, see
7404 * list_add_leaf_cfs_rq() for details.
7406 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7407 struct sched_entity *se;
7409 /* throttled entities do not contribute to load */
7410 if (throttled_hierarchy(cfs_rq))
7413 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
7414 update_tg_load_avg(cfs_rq, 0);
7416 /* Propagate pending load changes to the parent, if any: */
7417 se = cfs_rq->tg->se[cpu];
7418 if (se && !skip_blocked_update(se))
7419 update_load_avg(cfs_rq_of(se), se, 0);
7422 * There can be a lot of idle CPU cgroups. Don't let fully
7423 * decayed cfs_rqs linger on the list.
7425 if (cfs_rq_is_decayed(cfs_rq))
7426 list_del_leaf_cfs_rq(cfs_rq);
7428 /* Don't need periodic decay once load/util_avg are null */
7429 if (cfs_rq_has_blocked(cfs_rq))
7433 curr_class = rq->curr->sched_class;
7434 update_rt_rq_load_avg(rq_clock_task(rq), rq, curr_class == &rt_sched_class);
7435 update_dl_rq_load_avg(rq_clock_task(rq), rq, curr_class == &dl_sched_class);
7436 update_irq_load_avg(rq, 0);
7437 /* Don't need periodic decay once load/util_avg are null */
7438 if (others_have_blocked(rq))
7441 #ifdef CONFIG_NO_HZ_COMMON
7442 rq->last_blocked_load_update_tick = jiffies;
7444 rq->has_blocked_load = 0;
7446 rq_unlock_irqrestore(rq, &rf);
7450 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7451 * This needs to be done in a top-down fashion because the load of a child
7452 * group is a fraction of its parents load.
7454 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7456 struct rq *rq = rq_of(cfs_rq);
7457 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7458 unsigned long now = jiffies;
7461 if (cfs_rq->last_h_load_update == now)
7464 cfs_rq->h_load_next = NULL;
7465 for_each_sched_entity(se) {
7466 cfs_rq = cfs_rq_of(se);
7467 cfs_rq->h_load_next = se;
7468 if (cfs_rq->last_h_load_update == now)
7473 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7474 cfs_rq->last_h_load_update = now;
7477 while ((se = cfs_rq->h_load_next) != NULL) {
7478 load = cfs_rq->h_load;
7479 load = div64_ul(load * se->avg.load_avg,
7480 cfs_rq_load_avg(cfs_rq) + 1);
7481 cfs_rq = group_cfs_rq(se);
7482 cfs_rq->h_load = load;
7483 cfs_rq->last_h_load_update = now;
7487 static unsigned long task_h_load(struct task_struct *p)
7489 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7491 update_cfs_rq_h_load(cfs_rq);
7492 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7493 cfs_rq_load_avg(cfs_rq) + 1);
7496 static inline void update_blocked_averages(int cpu)
7498 struct rq *rq = cpu_rq(cpu);
7499 struct cfs_rq *cfs_rq = &rq->cfs;
7500 const struct sched_class *curr_class;
7503 rq_lock_irqsave(rq, &rf);
7504 update_rq_clock(rq);
7505 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
7507 curr_class = rq->curr->sched_class;
7508 update_rt_rq_load_avg(rq_clock_task(rq), rq, curr_class == &rt_sched_class);
7509 update_dl_rq_load_avg(rq_clock_task(rq), rq, curr_class == &dl_sched_class);
7510 update_irq_load_avg(rq, 0);
7511 #ifdef CONFIG_NO_HZ_COMMON
7512 rq->last_blocked_load_update_tick = jiffies;
7513 if (!cfs_rq_has_blocked(cfs_rq) && !others_have_blocked(rq))
7514 rq->has_blocked_load = 0;
7516 rq_unlock_irqrestore(rq, &rf);
7519 static unsigned long task_h_load(struct task_struct *p)
7521 return p->se.avg.load_avg;
7525 /********** Helpers for find_busiest_group ************************/
7534 * sg_lb_stats - stats of a sched_group required for load_balancing
7536 struct sg_lb_stats {
7537 unsigned long avg_load; /*Avg load across the CPUs of the group */
7538 unsigned long group_load; /* Total load over the CPUs of the group */
7539 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7540 unsigned long load_per_task;
7541 unsigned long group_capacity;
7542 unsigned long group_util; /* Total utilization of the group */
7543 unsigned int sum_nr_running; /* Nr tasks running in the group */
7544 unsigned int idle_cpus;
7545 unsigned int group_weight;
7546 enum group_type group_type;
7547 int group_no_capacity;
7548 #ifdef CONFIG_NUMA_BALANCING
7549 unsigned int nr_numa_running;
7550 unsigned int nr_preferred_running;
7555 * sd_lb_stats - Structure to store the statistics of a sched_domain
7556 * during load balancing.
7558 struct sd_lb_stats {
7559 struct sched_group *busiest; /* Busiest group in this sd */
7560 struct sched_group *local; /* Local group in this sd */
7561 unsigned long total_running;
7562 unsigned long total_load; /* Total load of all groups in sd */
7563 unsigned long total_capacity; /* Total capacity of all groups in sd */
7564 unsigned long avg_load; /* Average load across all groups in sd */
7566 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7567 struct sg_lb_stats local_stat; /* Statistics of the local group */
7570 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7573 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7574 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7575 * We must however clear busiest_stat::avg_load because
7576 * update_sd_pick_busiest() reads this before assignment.
7578 *sds = (struct sd_lb_stats){
7581 .total_running = 0UL,
7583 .total_capacity = 0UL,
7586 .sum_nr_running = 0,
7587 .group_type = group_other,
7593 * get_sd_load_idx - Obtain the load index for a given sched domain.
7594 * @sd: The sched_domain whose load_idx is to be obtained.
7595 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7597 * Return: The load index.
7599 static inline int get_sd_load_idx(struct sched_domain *sd,
7600 enum cpu_idle_type idle)
7606 load_idx = sd->busy_idx;
7609 case CPU_NEWLY_IDLE:
7610 load_idx = sd->newidle_idx;
7613 load_idx = sd->idle_idx;
7620 static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu)
7622 struct rq *rq = cpu_rq(cpu);
7623 unsigned long max = arch_scale_cpu_capacity(sd, cpu);
7624 unsigned long used, free;
7627 irq = cpu_util_irq(rq);
7629 if (unlikely(irq >= max))
7632 used = READ_ONCE(rq->avg_rt.util_avg);
7633 used += READ_ONCE(rq->avg_dl.util_avg);
7635 if (unlikely(used >= max))
7640 return scale_irq_capacity(free, irq, max);
7643 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7645 unsigned long capacity = scale_rt_capacity(sd, cpu);
7646 struct sched_group *sdg = sd->groups;
7648 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(sd, cpu);
7653 cpu_rq(cpu)->cpu_capacity = capacity;
7654 sdg->sgc->capacity = capacity;
7655 sdg->sgc->min_capacity = capacity;
7658 void update_group_capacity(struct sched_domain *sd, int cpu)
7660 struct sched_domain *child = sd->child;
7661 struct sched_group *group, *sdg = sd->groups;
7662 unsigned long capacity, min_capacity;
7663 unsigned long interval;
7665 interval = msecs_to_jiffies(sd->balance_interval);
7666 interval = clamp(interval, 1UL, max_load_balance_interval);
7667 sdg->sgc->next_update = jiffies + interval;
7670 update_cpu_capacity(sd, cpu);
7675 min_capacity = ULONG_MAX;
7677 if (child->flags & SD_OVERLAP) {
7679 * SD_OVERLAP domains cannot assume that child groups
7680 * span the current group.
7683 for_each_cpu(cpu, sched_group_span(sdg)) {
7684 struct sched_group_capacity *sgc;
7685 struct rq *rq = cpu_rq(cpu);
7688 * build_sched_domains() -> init_sched_groups_capacity()
7689 * gets here before we've attached the domains to the
7692 * Use capacity_of(), which is set irrespective of domains
7693 * in update_cpu_capacity().
7695 * This avoids capacity from being 0 and
7696 * causing divide-by-zero issues on boot.
7698 if (unlikely(!rq->sd)) {
7699 capacity += capacity_of(cpu);
7701 sgc = rq->sd->groups->sgc;
7702 capacity += sgc->capacity;
7705 min_capacity = min(capacity, min_capacity);
7709 * !SD_OVERLAP domains can assume that child groups
7710 * span the current group.
7713 group = child->groups;
7715 struct sched_group_capacity *sgc = group->sgc;
7717 capacity += sgc->capacity;
7718 min_capacity = min(sgc->min_capacity, min_capacity);
7719 group = group->next;
7720 } while (group != child->groups);
7723 sdg->sgc->capacity = capacity;
7724 sdg->sgc->min_capacity = min_capacity;
7728 * Check whether the capacity of the rq has been noticeably reduced by side
7729 * activity. The imbalance_pct is used for the threshold.
7730 * Return true is the capacity is reduced
7733 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7735 return ((rq->cpu_capacity * sd->imbalance_pct) <
7736 (rq->cpu_capacity_orig * 100));
7740 * Group imbalance indicates (and tries to solve) the problem where balancing
7741 * groups is inadequate due to ->cpus_allowed constraints.
7743 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
7744 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
7747 * { 0 1 2 3 } { 4 5 6 7 }
7750 * If we were to balance group-wise we'd place two tasks in the first group and
7751 * two tasks in the second group. Clearly this is undesired as it will overload
7752 * cpu 3 and leave one of the CPUs in the second group unused.
7754 * The current solution to this issue is detecting the skew in the first group
7755 * by noticing the lower domain failed to reach balance and had difficulty
7756 * moving tasks due to affinity constraints.
7758 * When this is so detected; this group becomes a candidate for busiest; see
7759 * update_sd_pick_busiest(). And calculate_imbalance() and
7760 * find_busiest_group() avoid some of the usual balance conditions to allow it
7761 * to create an effective group imbalance.
7763 * This is a somewhat tricky proposition since the next run might not find the
7764 * group imbalance and decide the groups need to be balanced again. A most
7765 * subtle and fragile situation.
7768 static inline int sg_imbalanced(struct sched_group *group)
7770 return group->sgc->imbalance;
7774 * group_has_capacity returns true if the group has spare capacity that could
7775 * be used by some tasks.
7776 * We consider that a group has spare capacity if the * number of task is
7777 * smaller than the number of CPUs or if the utilization is lower than the
7778 * available capacity for CFS tasks.
7779 * For the latter, we use a threshold to stabilize the state, to take into
7780 * account the variance of the tasks' load and to return true if the available
7781 * capacity in meaningful for the load balancer.
7782 * As an example, an available capacity of 1% can appear but it doesn't make
7783 * any benefit for the load balance.
7786 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7788 if (sgs->sum_nr_running < sgs->group_weight)
7791 if ((sgs->group_capacity * 100) >
7792 (sgs->group_util * env->sd->imbalance_pct))
7799 * group_is_overloaded returns true if the group has more tasks than it can
7801 * group_is_overloaded is not equals to !group_has_capacity because a group
7802 * with the exact right number of tasks, has no more spare capacity but is not
7803 * overloaded so both group_has_capacity and group_is_overloaded return
7807 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7809 if (sgs->sum_nr_running <= sgs->group_weight)
7812 if ((sgs->group_capacity * 100) <
7813 (sgs->group_util * env->sd->imbalance_pct))
7820 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7821 * per-CPU capacity than sched_group ref.
7824 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7826 return sg->sgc->min_capacity * capacity_margin <
7827 ref->sgc->min_capacity * 1024;
7831 group_type group_classify(struct sched_group *group,
7832 struct sg_lb_stats *sgs)
7834 if (sgs->group_no_capacity)
7835 return group_overloaded;
7837 if (sg_imbalanced(group))
7838 return group_imbalanced;
7843 static bool update_nohz_stats(struct rq *rq, bool force)
7845 #ifdef CONFIG_NO_HZ_COMMON
7846 unsigned int cpu = rq->cpu;
7848 if (!rq->has_blocked_load)
7851 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
7854 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
7857 update_blocked_averages(cpu);
7859 return rq->has_blocked_load;
7866 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7867 * @env: The load balancing environment.
7868 * @group: sched_group whose statistics are to be updated.
7869 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7870 * @local_group: Does group contain this_cpu.
7871 * @sgs: variable to hold the statistics for this group.
7872 * @overload: Indicate more than one runnable task for any CPU.
7874 static inline void update_sg_lb_stats(struct lb_env *env,
7875 struct sched_group *group, int load_idx,
7876 int local_group, struct sg_lb_stats *sgs,
7882 memset(sgs, 0, sizeof(*sgs));
7884 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
7885 struct rq *rq = cpu_rq(i);
7887 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
7888 env->flags |= LBF_NOHZ_AGAIN;
7890 /* Bias balancing toward CPUs of our domain: */
7892 load = target_load(i, load_idx);
7894 load = source_load(i, load_idx);
7896 sgs->group_load += load;
7897 sgs->group_util += cpu_util(i);
7898 sgs->sum_nr_running += rq->cfs.h_nr_running;
7900 nr_running = rq->nr_running;
7904 #ifdef CONFIG_NUMA_BALANCING
7905 sgs->nr_numa_running += rq->nr_numa_running;
7906 sgs->nr_preferred_running += rq->nr_preferred_running;
7908 sgs->sum_weighted_load += weighted_cpuload(rq);
7910 * No need to call idle_cpu() if nr_running is not 0
7912 if (!nr_running && idle_cpu(i))
7916 /* Adjust by relative CPU capacity of the group */
7917 sgs->group_capacity = group->sgc->capacity;
7918 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7920 if (sgs->sum_nr_running)
7921 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7923 sgs->group_weight = group->group_weight;
7925 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7926 sgs->group_type = group_classify(group, sgs);
7930 * update_sd_pick_busiest - return 1 on busiest group
7931 * @env: The load balancing environment.
7932 * @sds: sched_domain statistics
7933 * @sg: sched_group candidate to be checked for being the busiest
7934 * @sgs: sched_group statistics
7936 * Determine if @sg is a busier group than the previously selected
7939 * Return: %true if @sg is a busier group than the previously selected
7940 * busiest group. %false otherwise.
7942 static bool update_sd_pick_busiest(struct lb_env *env,
7943 struct sd_lb_stats *sds,
7944 struct sched_group *sg,
7945 struct sg_lb_stats *sgs)
7947 struct sg_lb_stats *busiest = &sds->busiest_stat;
7949 if (sgs->group_type > busiest->group_type)
7952 if (sgs->group_type < busiest->group_type)
7955 if (sgs->avg_load <= busiest->avg_load)
7958 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7962 * Candidate sg has no more than one task per CPU and
7963 * has higher per-CPU capacity. Migrating tasks to less
7964 * capable CPUs may harm throughput. Maximize throughput,
7965 * power/energy consequences are not considered.
7967 if (sgs->sum_nr_running <= sgs->group_weight &&
7968 group_smaller_cpu_capacity(sds->local, sg))
7972 /* This is the busiest node in its class. */
7973 if (!(env->sd->flags & SD_ASYM_PACKING))
7976 /* No ASYM_PACKING if target CPU is already busy */
7977 if (env->idle == CPU_NOT_IDLE)
7980 * ASYM_PACKING needs to move all the work to the highest
7981 * prority CPUs in the group, therefore mark all groups
7982 * of lower priority than ourself as busy.
7984 if (sgs->sum_nr_running &&
7985 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
7989 /* Prefer to move from lowest priority CPU's work */
7990 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
7991 sg->asym_prefer_cpu))
7998 #ifdef CONFIG_NUMA_BALANCING
7999 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8001 if (sgs->sum_nr_running > sgs->nr_numa_running)
8003 if (sgs->sum_nr_running > sgs->nr_preferred_running)
8008 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8010 if (rq->nr_running > rq->nr_numa_running)
8012 if (rq->nr_running > rq->nr_preferred_running)
8017 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8022 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8026 #endif /* CONFIG_NUMA_BALANCING */
8029 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8030 * @env: The load balancing environment.
8031 * @sds: variable to hold the statistics for this sched_domain.
8033 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8035 struct sched_domain *child = env->sd->child;
8036 struct sched_group *sg = env->sd->groups;
8037 struct sg_lb_stats *local = &sds->local_stat;
8038 struct sg_lb_stats tmp_sgs;
8039 int load_idx, prefer_sibling = 0;
8040 bool overload = false;
8042 if (child && child->flags & SD_PREFER_SIBLING)
8045 #ifdef CONFIG_NO_HZ_COMMON
8046 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
8047 env->flags |= LBF_NOHZ_STATS;
8050 load_idx = get_sd_load_idx(env->sd, env->idle);
8053 struct sg_lb_stats *sgs = &tmp_sgs;
8056 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
8061 if (env->idle != CPU_NEWLY_IDLE ||
8062 time_after_eq(jiffies, sg->sgc->next_update))
8063 update_group_capacity(env->sd, env->dst_cpu);
8066 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
8073 * In case the child domain prefers tasks go to siblings
8074 * first, lower the sg capacity so that we'll try
8075 * and move all the excess tasks away. We lower the capacity
8076 * of a group only if the local group has the capacity to fit
8077 * these excess tasks. The extra check prevents the case where
8078 * you always pull from the heaviest group when it is already
8079 * under-utilized (possible with a large weight task outweighs
8080 * the tasks on the system).
8082 if (prefer_sibling && sds->local &&
8083 group_has_capacity(env, local) &&
8084 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
8085 sgs->group_no_capacity = 1;
8086 sgs->group_type = group_classify(sg, sgs);
8089 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8091 sds->busiest_stat = *sgs;
8095 /* Now, start updating sd_lb_stats */
8096 sds->total_running += sgs->sum_nr_running;
8097 sds->total_load += sgs->group_load;
8098 sds->total_capacity += sgs->group_capacity;
8101 } while (sg != env->sd->groups);
8103 #ifdef CONFIG_NO_HZ_COMMON
8104 if ((env->flags & LBF_NOHZ_AGAIN) &&
8105 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8107 WRITE_ONCE(nohz.next_blocked,
8108 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8112 if (env->sd->flags & SD_NUMA)
8113 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8115 if (!env->sd->parent) {
8116 /* update overload indicator if we are at root domain */
8117 if (env->dst_rq->rd->overload != overload)
8118 env->dst_rq->rd->overload = overload;
8123 * check_asym_packing - Check to see if the group is packed into the
8126 * This is primarily intended to used at the sibling level. Some
8127 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8128 * case of POWER7, it can move to lower SMT modes only when higher
8129 * threads are idle. When in lower SMT modes, the threads will
8130 * perform better since they share less core resources. Hence when we
8131 * have idle threads, we want them to be the higher ones.
8133 * This packing function is run on idle threads. It checks to see if
8134 * the busiest CPU in this domain (core in the P7 case) has a higher
8135 * CPU number than the packing function is being run on. Here we are
8136 * assuming lower CPU number will be equivalent to lower a SMT thread
8139 * Return: 1 when packing is required and a task should be moved to
8140 * this CPU. The amount of the imbalance is returned in env->imbalance.
8142 * @env: The load balancing environment.
8143 * @sds: Statistics of the sched_domain which is to be packed
8145 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8149 if (!(env->sd->flags & SD_ASYM_PACKING))
8152 if (env->idle == CPU_NOT_IDLE)
8158 busiest_cpu = sds->busiest->asym_prefer_cpu;
8159 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
8162 env->imbalance = DIV_ROUND_CLOSEST(
8163 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
8164 SCHED_CAPACITY_SCALE);
8170 * fix_small_imbalance - Calculate the minor imbalance that exists
8171 * amongst the groups of a sched_domain, during
8173 * @env: The load balancing environment.
8174 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8177 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8179 unsigned long tmp, capa_now = 0, capa_move = 0;
8180 unsigned int imbn = 2;
8181 unsigned long scaled_busy_load_per_task;
8182 struct sg_lb_stats *local, *busiest;
8184 local = &sds->local_stat;
8185 busiest = &sds->busiest_stat;
8187 if (!local->sum_nr_running)
8188 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8189 else if (busiest->load_per_task > local->load_per_task)
8192 scaled_busy_load_per_task =
8193 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8194 busiest->group_capacity;
8196 if (busiest->avg_load + scaled_busy_load_per_task >=
8197 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8198 env->imbalance = busiest->load_per_task;
8203 * OK, we don't have enough imbalance to justify moving tasks,
8204 * however we may be able to increase total CPU capacity used by
8208 capa_now += busiest->group_capacity *
8209 min(busiest->load_per_task, busiest->avg_load);
8210 capa_now += local->group_capacity *
8211 min(local->load_per_task, local->avg_load);
8212 capa_now /= SCHED_CAPACITY_SCALE;
8214 /* Amount of load we'd subtract */
8215 if (busiest->avg_load > scaled_busy_load_per_task) {
8216 capa_move += busiest->group_capacity *
8217 min(busiest->load_per_task,
8218 busiest->avg_load - scaled_busy_load_per_task);
8221 /* Amount of load we'd add */
8222 if (busiest->avg_load * busiest->group_capacity <
8223 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8224 tmp = (busiest->avg_load * busiest->group_capacity) /
8225 local->group_capacity;
8227 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8228 local->group_capacity;
8230 capa_move += local->group_capacity *
8231 min(local->load_per_task, local->avg_load + tmp);
8232 capa_move /= SCHED_CAPACITY_SCALE;
8234 /* Move if we gain throughput */
8235 if (capa_move > capa_now)
8236 env->imbalance = busiest->load_per_task;
8240 * calculate_imbalance - Calculate the amount of imbalance present within the
8241 * groups of a given sched_domain during load balance.
8242 * @env: load balance environment
8243 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8245 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8247 unsigned long max_pull, load_above_capacity = ~0UL;
8248 struct sg_lb_stats *local, *busiest;
8250 local = &sds->local_stat;
8251 busiest = &sds->busiest_stat;
8253 if (busiest->group_type == group_imbalanced) {
8255 * In the group_imb case we cannot rely on group-wide averages
8256 * to ensure CPU-load equilibrium, look at wider averages. XXX
8258 busiest->load_per_task =
8259 min(busiest->load_per_task, sds->avg_load);
8263 * Avg load of busiest sg can be less and avg load of local sg can
8264 * be greater than avg load across all sgs of sd because avg load
8265 * factors in sg capacity and sgs with smaller group_type are
8266 * skipped when updating the busiest sg:
8268 if (busiest->avg_load <= sds->avg_load ||
8269 local->avg_load >= sds->avg_load) {
8271 return fix_small_imbalance(env, sds);
8275 * If there aren't any idle CPUs, avoid creating some.
8277 if (busiest->group_type == group_overloaded &&
8278 local->group_type == group_overloaded) {
8279 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
8280 if (load_above_capacity > busiest->group_capacity) {
8281 load_above_capacity -= busiest->group_capacity;
8282 load_above_capacity *= scale_load_down(NICE_0_LOAD);
8283 load_above_capacity /= busiest->group_capacity;
8285 load_above_capacity = ~0UL;
8289 * We're trying to get all the CPUs to the average_load, so we don't
8290 * want to push ourselves above the average load, nor do we wish to
8291 * reduce the max loaded CPU below the average load. At the same time,
8292 * we also don't want to reduce the group load below the group
8293 * capacity. Thus we look for the minimum possible imbalance.
8295 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8297 /* How much load to actually move to equalise the imbalance */
8298 env->imbalance = min(
8299 max_pull * busiest->group_capacity,
8300 (sds->avg_load - local->avg_load) * local->group_capacity
8301 ) / SCHED_CAPACITY_SCALE;
8304 * if *imbalance is less than the average load per runnable task
8305 * there is no guarantee that any tasks will be moved so we'll have
8306 * a think about bumping its value to force at least one task to be
8309 if (env->imbalance < busiest->load_per_task)
8310 return fix_small_imbalance(env, sds);
8313 /******* find_busiest_group() helpers end here *********************/
8316 * find_busiest_group - Returns the busiest group within the sched_domain
8317 * if there is an imbalance.
8319 * Also calculates the amount of weighted load which should be moved
8320 * to restore balance.
8322 * @env: The load balancing environment.
8324 * Return: - The busiest group if imbalance exists.
8326 static struct sched_group *find_busiest_group(struct lb_env *env)
8328 struct sg_lb_stats *local, *busiest;
8329 struct sd_lb_stats sds;
8331 init_sd_lb_stats(&sds);
8334 * Compute the various statistics relavent for load balancing at
8337 update_sd_lb_stats(env, &sds);
8338 local = &sds.local_stat;
8339 busiest = &sds.busiest_stat;
8341 /* ASYM feature bypasses nice load balance check */
8342 if (check_asym_packing(env, &sds))
8345 /* There is no busy sibling group to pull tasks from */
8346 if (!sds.busiest || busiest->sum_nr_running == 0)
8349 /* XXX broken for overlapping NUMA groups */
8350 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8351 / sds.total_capacity;
8354 * If the busiest group is imbalanced the below checks don't
8355 * work because they assume all things are equal, which typically
8356 * isn't true due to cpus_allowed constraints and the like.
8358 if (busiest->group_type == group_imbalanced)
8362 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8363 * capacities from resulting in underutilization due to avg_load.
8365 if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8366 busiest->group_no_capacity)
8370 * If the local group is busier than the selected busiest group
8371 * don't try and pull any tasks.
8373 if (local->avg_load >= busiest->avg_load)
8377 * Don't pull any tasks if this group is already above the domain
8380 if (local->avg_load >= sds.avg_load)
8383 if (env->idle == CPU_IDLE) {
8385 * This CPU is idle. If the busiest group is not overloaded
8386 * and there is no imbalance between this and busiest group
8387 * wrt idle CPUs, it is balanced. The imbalance becomes
8388 * significant if the diff is greater than 1 otherwise we
8389 * might end up to just move the imbalance on another group
8391 if ((busiest->group_type != group_overloaded) &&
8392 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8396 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8397 * imbalance_pct to be conservative.
8399 if (100 * busiest->avg_load <=
8400 env->sd->imbalance_pct * local->avg_load)
8405 /* Looks like there is an imbalance. Compute it */
8406 calculate_imbalance(env, &sds);
8407 return env->imbalance ? sds.busiest : NULL;
8415 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
8417 static struct rq *find_busiest_queue(struct lb_env *env,
8418 struct sched_group *group)
8420 struct rq *busiest = NULL, *rq;
8421 unsigned long busiest_load = 0, busiest_capacity = 1;
8424 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8425 unsigned long capacity, wl;
8429 rt = fbq_classify_rq(rq);
8432 * We classify groups/runqueues into three groups:
8433 * - regular: there are !numa tasks
8434 * - remote: there are numa tasks that run on the 'wrong' node
8435 * - all: there is no distinction
8437 * In order to avoid migrating ideally placed numa tasks,
8438 * ignore those when there's better options.
8440 * If we ignore the actual busiest queue to migrate another
8441 * task, the next balance pass can still reduce the busiest
8442 * queue by moving tasks around inside the node.
8444 * If we cannot move enough load due to this classification
8445 * the next pass will adjust the group classification and
8446 * allow migration of more tasks.
8448 * Both cases only affect the total convergence complexity.
8450 if (rt > env->fbq_type)
8453 capacity = capacity_of(i);
8455 wl = weighted_cpuload(rq);
8458 * When comparing with imbalance, use weighted_cpuload()
8459 * which is not scaled with the CPU capacity.
8462 if (rq->nr_running == 1 && wl > env->imbalance &&
8463 !check_cpu_capacity(rq, env->sd))
8467 * For the load comparisons with the other CPU's, consider
8468 * the weighted_cpuload() scaled with the CPU capacity, so
8469 * that the load can be moved away from the CPU that is
8470 * potentially running at a lower capacity.
8472 * Thus we're looking for max(wl_i / capacity_i), crosswise
8473 * multiplication to rid ourselves of the division works out
8474 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8475 * our previous maximum.
8477 if (wl * busiest_capacity > busiest_load * capacity) {
8479 busiest_capacity = capacity;
8488 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8489 * so long as it is large enough.
8491 #define MAX_PINNED_INTERVAL 512
8493 static int need_active_balance(struct lb_env *env)
8495 struct sched_domain *sd = env->sd;
8497 if (env->idle == CPU_NEWLY_IDLE) {
8500 * ASYM_PACKING needs to force migrate tasks from busy but
8501 * lower priority CPUs in order to pack all tasks in the
8502 * highest priority CPUs.
8504 if ((sd->flags & SD_ASYM_PACKING) &&
8505 sched_asym_prefer(env->dst_cpu, env->src_cpu))
8510 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8511 * It's worth migrating the task if the src_cpu's capacity is reduced
8512 * because of other sched_class or IRQs if more capacity stays
8513 * available on dst_cpu.
8515 if ((env->idle != CPU_NOT_IDLE) &&
8516 (env->src_rq->cfs.h_nr_running == 1)) {
8517 if ((check_cpu_capacity(env->src_rq, sd)) &&
8518 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8522 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8525 static int active_load_balance_cpu_stop(void *data);
8527 static int should_we_balance(struct lb_env *env)
8529 struct sched_group *sg = env->sd->groups;
8530 int cpu, balance_cpu = -1;
8533 * Ensure the balancing environment is consistent; can happen
8534 * when the softirq triggers 'during' hotplug.
8536 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8540 * In the newly idle case, we will allow all the CPUs
8541 * to do the newly idle load balance.
8543 if (env->idle == CPU_NEWLY_IDLE)
8546 /* Try to find first idle CPU */
8547 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8555 if (balance_cpu == -1)
8556 balance_cpu = group_balance_cpu(sg);
8559 * First idle CPU or the first CPU(busiest) in this sched group
8560 * is eligible for doing load balancing at this and above domains.
8562 return balance_cpu == env->dst_cpu;
8566 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8567 * tasks if there is an imbalance.
8569 static int load_balance(int this_cpu, struct rq *this_rq,
8570 struct sched_domain *sd, enum cpu_idle_type idle,
8571 int *continue_balancing)
8573 int ld_moved, cur_ld_moved, active_balance = 0;
8574 struct sched_domain *sd_parent = sd->parent;
8575 struct sched_group *group;
8578 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8580 struct lb_env env = {
8582 .dst_cpu = this_cpu,
8584 .dst_grpmask = sched_group_span(sd->groups),
8586 .loop_break = sched_nr_migrate_break,
8589 .tasks = LIST_HEAD_INIT(env.tasks),
8592 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8594 schedstat_inc(sd->lb_count[idle]);
8597 if (!should_we_balance(&env)) {
8598 *continue_balancing = 0;
8602 group = find_busiest_group(&env);
8604 schedstat_inc(sd->lb_nobusyg[idle]);
8608 busiest = find_busiest_queue(&env, group);
8610 schedstat_inc(sd->lb_nobusyq[idle]);
8614 BUG_ON(busiest == env.dst_rq);
8616 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8618 env.src_cpu = busiest->cpu;
8619 env.src_rq = busiest;
8622 if (busiest->nr_running > 1) {
8624 * Attempt to move tasks. If find_busiest_group has found
8625 * an imbalance but busiest->nr_running <= 1, the group is
8626 * still unbalanced. ld_moved simply stays zero, so it is
8627 * correctly treated as an imbalance.
8629 env.flags |= LBF_ALL_PINNED;
8630 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8633 rq_lock_irqsave(busiest, &rf);
8634 update_rq_clock(busiest);
8637 * cur_ld_moved - load moved in current iteration
8638 * ld_moved - cumulative load moved across iterations
8640 cur_ld_moved = detach_tasks(&env);
8643 * We've detached some tasks from busiest_rq. Every
8644 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8645 * unlock busiest->lock, and we are able to be sure
8646 * that nobody can manipulate the tasks in parallel.
8647 * See task_rq_lock() family for the details.
8650 rq_unlock(busiest, &rf);
8654 ld_moved += cur_ld_moved;
8657 local_irq_restore(rf.flags);
8659 if (env.flags & LBF_NEED_BREAK) {
8660 env.flags &= ~LBF_NEED_BREAK;
8665 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8666 * us and move them to an alternate dst_cpu in our sched_group
8667 * where they can run. The upper limit on how many times we
8668 * iterate on same src_cpu is dependent on number of CPUs in our
8671 * This changes load balance semantics a bit on who can move
8672 * load to a given_cpu. In addition to the given_cpu itself
8673 * (or a ilb_cpu acting on its behalf where given_cpu is
8674 * nohz-idle), we now have balance_cpu in a position to move
8675 * load to given_cpu. In rare situations, this may cause
8676 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8677 * _independently_ and at _same_ time to move some load to
8678 * given_cpu) causing exceess load to be moved to given_cpu.
8679 * This however should not happen so much in practice and
8680 * moreover subsequent load balance cycles should correct the
8681 * excess load moved.
8683 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8685 /* Prevent to re-select dst_cpu via env's CPUs */
8686 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8688 env.dst_rq = cpu_rq(env.new_dst_cpu);
8689 env.dst_cpu = env.new_dst_cpu;
8690 env.flags &= ~LBF_DST_PINNED;
8692 env.loop_break = sched_nr_migrate_break;
8695 * Go back to "more_balance" rather than "redo" since we
8696 * need to continue with same src_cpu.
8702 * We failed to reach balance because of affinity.
8705 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8707 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8708 *group_imbalance = 1;
8711 /* All tasks on this runqueue were pinned by CPU affinity */
8712 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8713 cpumask_clear_cpu(cpu_of(busiest), cpus);
8715 * Attempting to continue load balancing at the current
8716 * sched_domain level only makes sense if there are
8717 * active CPUs remaining as possible busiest CPUs to
8718 * pull load from which are not contained within the
8719 * destination group that is receiving any migrated
8722 if (!cpumask_subset(cpus, env.dst_grpmask)) {
8724 env.loop_break = sched_nr_migrate_break;
8727 goto out_all_pinned;
8732 schedstat_inc(sd->lb_failed[idle]);
8734 * Increment the failure counter only on periodic balance.
8735 * We do not want newidle balance, which can be very
8736 * frequent, pollute the failure counter causing
8737 * excessive cache_hot migrations and active balances.
8739 if (idle != CPU_NEWLY_IDLE)
8740 sd->nr_balance_failed++;
8742 if (need_active_balance(&env)) {
8743 unsigned long flags;
8745 raw_spin_lock_irqsave(&busiest->lock, flags);
8748 * Don't kick the active_load_balance_cpu_stop,
8749 * if the curr task on busiest CPU can't be
8750 * moved to this_cpu:
8752 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
8753 raw_spin_unlock_irqrestore(&busiest->lock,
8755 env.flags |= LBF_ALL_PINNED;
8756 goto out_one_pinned;
8760 * ->active_balance synchronizes accesses to
8761 * ->active_balance_work. Once set, it's cleared
8762 * only after active load balance is finished.
8764 if (!busiest->active_balance) {
8765 busiest->active_balance = 1;
8766 busiest->push_cpu = this_cpu;
8769 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8771 if (active_balance) {
8772 stop_one_cpu_nowait(cpu_of(busiest),
8773 active_load_balance_cpu_stop, busiest,
8774 &busiest->active_balance_work);
8777 /* We've kicked active balancing, force task migration. */
8778 sd->nr_balance_failed = sd->cache_nice_tries+1;
8781 sd->nr_balance_failed = 0;
8783 if (likely(!active_balance)) {
8784 /* We were unbalanced, so reset the balancing interval */
8785 sd->balance_interval = sd->min_interval;
8788 * If we've begun active balancing, start to back off. This
8789 * case may not be covered by the all_pinned logic if there
8790 * is only 1 task on the busy runqueue (because we don't call
8793 if (sd->balance_interval < sd->max_interval)
8794 sd->balance_interval *= 2;
8801 * We reach balance although we may have faced some affinity
8802 * constraints. Clear the imbalance flag if it was set.
8805 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8807 if (*group_imbalance)
8808 *group_imbalance = 0;
8813 * We reach balance because all tasks are pinned at this level so
8814 * we can't migrate them. Let the imbalance flag set so parent level
8815 * can try to migrate them.
8817 schedstat_inc(sd->lb_balanced[idle]);
8819 sd->nr_balance_failed = 0;
8822 /* tune up the balancing interval */
8823 if (((env.flags & LBF_ALL_PINNED) &&
8824 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8825 (sd->balance_interval < sd->max_interval))
8826 sd->balance_interval *= 2;
8833 static inline unsigned long
8834 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8836 unsigned long interval = sd->balance_interval;
8839 interval *= sd->busy_factor;
8841 /* scale ms to jiffies */
8842 interval = msecs_to_jiffies(interval);
8843 interval = clamp(interval, 1UL, max_load_balance_interval);
8849 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
8851 unsigned long interval, next;
8853 /* used by idle balance, so cpu_busy = 0 */
8854 interval = get_sd_balance_interval(sd, 0);
8855 next = sd->last_balance + interval;
8857 if (time_after(*next_balance, next))
8858 *next_balance = next;
8862 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
8863 * running tasks off the busiest CPU onto idle CPUs. It requires at
8864 * least 1 task to be running on each physical CPU where possible, and
8865 * avoids physical / logical imbalances.
8867 static int active_load_balance_cpu_stop(void *data)
8869 struct rq *busiest_rq = data;
8870 int busiest_cpu = cpu_of(busiest_rq);
8871 int target_cpu = busiest_rq->push_cpu;
8872 struct rq *target_rq = cpu_rq(target_cpu);
8873 struct sched_domain *sd;
8874 struct task_struct *p = NULL;
8877 rq_lock_irq(busiest_rq, &rf);
8879 * Between queueing the stop-work and running it is a hole in which
8880 * CPUs can become inactive. We should not move tasks from or to
8883 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
8886 /* Make sure the requested CPU hasn't gone down in the meantime: */
8887 if (unlikely(busiest_cpu != smp_processor_id() ||
8888 !busiest_rq->active_balance))
8891 /* Is there any task to move? */
8892 if (busiest_rq->nr_running <= 1)
8896 * This condition is "impossible", if it occurs
8897 * we need to fix it. Originally reported by
8898 * Bjorn Helgaas on a 128-CPU setup.
8900 BUG_ON(busiest_rq == target_rq);
8902 /* Search for an sd spanning us and the target CPU. */
8904 for_each_domain(target_cpu, sd) {
8905 if ((sd->flags & SD_LOAD_BALANCE) &&
8906 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8911 struct lb_env env = {
8913 .dst_cpu = target_cpu,
8914 .dst_rq = target_rq,
8915 .src_cpu = busiest_rq->cpu,
8916 .src_rq = busiest_rq,
8919 * can_migrate_task() doesn't need to compute new_dst_cpu
8920 * for active balancing. Since we have CPU_IDLE, but no
8921 * @dst_grpmask we need to make that test go away with lying
8924 .flags = LBF_DST_PINNED,
8927 schedstat_inc(sd->alb_count);
8928 update_rq_clock(busiest_rq);
8930 p = detach_one_task(&env);
8932 schedstat_inc(sd->alb_pushed);
8933 /* Active balancing done, reset the failure counter. */
8934 sd->nr_balance_failed = 0;
8936 schedstat_inc(sd->alb_failed);
8941 busiest_rq->active_balance = 0;
8942 rq_unlock(busiest_rq, &rf);
8945 attach_one_task(target_rq, p);
8952 static DEFINE_SPINLOCK(balancing);
8955 * Scale the max load_balance interval with the number of CPUs in the system.
8956 * This trades load-balance latency on larger machines for less cross talk.
8958 void update_max_interval(void)
8960 max_load_balance_interval = HZ*num_online_cpus()/10;
8964 * It checks each scheduling domain to see if it is due to be balanced,
8965 * and initiates a balancing operation if so.
8967 * Balancing parameters are set up in init_sched_domains.
8969 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8971 int continue_balancing = 1;
8973 unsigned long interval;
8974 struct sched_domain *sd;
8975 /* Earliest time when we have to do rebalance again */
8976 unsigned long next_balance = jiffies + 60*HZ;
8977 int update_next_balance = 0;
8978 int need_serialize, need_decay = 0;
8982 for_each_domain(cpu, sd) {
8984 * Decay the newidle max times here because this is a regular
8985 * visit to all the domains. Decay ~1% per second.
8987 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8988 sd->max_newidle_lb_cost =
8989 (sd->max_newidle_lb_cost * 253) / 256;
8990 sd->next_decay_max_lb_cost = jiffies + HZ;
8993 max_cost += sd->max_newidle_lb_cost;
8995 if (!(sd->flags & SD_LOAD_BALANCE))
8999 * Stop the load balance at this level. There is another
9000 * CPU in our sched group which is doing load balancing more
9003 if (!continue_balancing) {
9009 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9011 need_serialize = sd->flags & SD_SERIALIZE;
9012 if (need_serialize) {
9013 if (!spin_trylock(&balancing))
9017 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9018 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9020 * The LBF_DST_PINNED logic could have changed
9021 * env->dst_cpu, so we can't know our idle
9022 * state even if we migrated tasks. Update it.
9024 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9026 sd->last_balance = jiffies;
9027 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9030 spin_unlock(&balancing);
9032 if (time_after(next_balance, sd->last_balance + interval)) {
9033 next_balance = sd->last_balance + interval;
9034 update_next_balance = 1;
9039 * Ensure the rq-wide value also decays but keep it at a
9040 * reasonable floor to avoid funnies with rq->avg_idle.
9042 rq->max_idle_balance_cost =
9043 max((u64)sysctl_sched_migration_cost, max_cost);
9048 * next_balance will be updated only when there is a need.
9049 * When the cpu is attached to null domain for ex, it will not be
9052 if (likely(update_next_balance)) {
9053 rq->next_balance = next_balance;
9055 #ifdef CONFIG_NO_HZ_COMMON
9057 * If this CPU has been elected to perform the nohz idle
9058 * balance. Other idle CPUs have already rebalanced with
9059 * nohz_idle_balance() and nohz.next_balance has been
9060 * updated accordingly. This CPU is now running the idle load
9061 * balance for itself and we need to update the
9062 * nohz.next_balance accordingly.
9064 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9065 nohz.next_balance = rq->next_balance;
9070 static inline int on_null_domain(struct rq *rq)
9072 return unlikely(!rcu_dereference_sched(rq->sd));
9075 #ifdef CONFIG_NO_HZ_COMMON
9077 * idle load balancing details
9078 * - When one of the busy CPUs notice that there may be an idle rebalancing
9079 * needed, they will kick the idle load balancer, which then does idle
9080 * load balancing for all the idle CPUs.
9083 static inline int find_new_ilb(void)
9085 int ilb = cpumask_first(nohz.idle_cpus_mask);
9087 if (ilb < nr_cpu_ids && idle_cpu(ilb))
9094 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
9095 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
9096 * CPU (if there is one).
9098 static void kick_ilb(unsigned int flags)
9102 nohz.next_balance++;
9104 ilb_cpu = find_new_ilb();
9106 if (ilb_cpu >= nr_cpu_ids)
9109 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
9110 if (flags & NOHZ_KICK_MASK)
9114 * Use smp_send_reschedule() instead of resched_cpu().
9115 * This way we generate a sched IPI on the target CPU which
9116 * is idle. And the softirq performing nohz idle load balance
9117 * will be run before returning from the IPI.
9119 smp_send_reschedule(ilb_cpu);
9123 * Current heuristic for kicking the idle load balancer in the presence
9124 * of an idle cpu in the system.
9125 * - This rq has more than one task.
9126 * - This rq has at least one CFS task and the capacity of the CPU is
9127 * significantly reduced because of RT tasks or IRQs.
9128 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9129 * multiple busy cpu.
9130 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9131 * domain span are idle.
9133 static void nohz_balancer_kick(struct rq *rq)
9135 unsigned long now = jiffies;
9136 struct sched_domain_shared *sds;
9137 struct sched_domain *sd;
9138 int nr_busy, i, cpu = rq->cpu;
9139 unsigned int flags = 0;
9141 if (unlikely(rq->idle_balance))
9145 * We may be recently in ticked or tickless idle mode. At the first
9146 * busy tick after returning from idle, we will update the busy stats.
9148 nohz_balance_exit_idle(rq);
9151 * None are in tickless mode and hence no need for NOHZ idle load
9154 if (likely(!atomic_read(&nohz.nr_cpus)))
9157 if (READ_ONCE(nohz.has_blocked) &&
9158 time_after(now, READ_ONCE(nohz.next_blocked)))
9159 flags = NOHZ_STATS_KICK;
9161 if (time_before(now, nohz.next_balance))
9164 if (rq->nr_running >= 2) {
9165 flags = NOHZ_KICK_MASK;
9170 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9173 * XXX: write a coherent comment on why we do this.
9174 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
9176 nr_busy = atomic_read(&sds->nr_busy_cpus);
9178 flags = NOHZ_KICK_MASK;
9184 sd = rcu_dereference(rq->sd);
9186 if ((rq->cfs.h_nr_running >= 1) &&
9187 check_cpu_capacity(rq, sd)) {
9188 flags = NOHZ_KICK_MASK;
9193 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9195 for_each_cpu(i, sched_domain_span(sd)) {
9197 !cpumask_test_cpu(i, nohz.idle_cpus_mask))
9200 if (sched_asym_prefer(i, cpu)) {
9201 flags = NOHZ_KICK_MASK;
9213 static void set_cpu_sd_state_busy(int cpu)
9215 struct sched_domain *sd;
9218 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9220 if (!sd || !sd->nohz_idle)
9224 atomic_inc(&sd->shared->nr_busy_cpus);
9229 void nohz_balance_exit_idle(struct rq *rq)
9231 SCHED_WARN_ON(rq != this_rq());
9233 if (likely(!rq->nohz_tick_stopped))
9236 rq->nohz_tick_stopped = 0;
9237 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
9238 atomic_dec(&nohz.nr_cpus);
9240 set_cpu_sd_state_busy(rq->cpu);
9243 static void set_cpu_sd_state_idle(int cpu)
9245 struct sched_domain *sd;
9248 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9250 if (!sd || sd->nohz_idle)
9254 atomic_dec(&sd->shared->nr_busy_cpus);
9260 * This routine will record that the CPU is going idle with tick stopped.
9261 * This info will be used in performing idle load balancing in the future.
9263 void nohz_balance_enter_idle(int cpu)
9265 struct rq *rq = cpu_rq(cpu);
9267 SCHED_WARN_ON(cpu != smp_processor_id());
9269 /* If this CPU is going down, then nothing needs to be done: */
9270 if (!cpu_active(cpu))
9273 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9274 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9278 * Can be set safely without rq->lock held
9279 * If a clear happens, it will have evaluated last additions because
9280 * rq->lock is held during the check and the clear
9282 rq->has_blocked_load = 1;
9285 * The tick is still stopped but load could have been added in the
9286 * meantime. We set the nohz.has_blocked flag to trig a check of the
9287 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
9288 * of nohz.has_blocked can only happen after checking the new load
9290 if (rq->nohz_tick_stopped)
9293 /* If we're a completely isolated CPU, we don't play: */
9294 if (on_null_domain(rq))
9297 rq->nohz_tick_stopped = 1;
9299 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9300 atomic_inc(&nohz.nr_cpus);
9303 * Ensures that if nohz_idle_balance() fails to observe our
9304 * @idle_cpus_mask store, it must observe the @has_blocked
9307 smp_mb__after_atomic();
9309 set_cpu_sd_state_idle(cpu);
9313 * Each time a cpu enter idle, we assume that it has blocked load and
9314 * enable the periodic update of the load of idle cpus
9316 WRITE_ONCE(nohz.has_blocked, 1);
9320 * Internal function that runs load balance for all idle cpus. The load balance
9321 * can be a simple update of blocked load or a complete load balance with
9322 * tasks movement depending of flags.
9323 * The function returns false if the loop has stopped before running
9324 * through all idle CPUs.
9326 static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
9327 enum cpu_idle_type idle)
9329 /* Earliest time when we have to do rebalance again */
9330 unsigned long now = jiffies;
9331 unsigned long next_balance = now + 60*HZ;
9332 bool has_blocked_load = false;
9333 int update_next_balance = 0;
9334 int this_cpu = this_rq->cpu;
9339 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
9342 * We assume there will be no idle load after this update and clear
9343 * the has_blocked flag. If a cpu enters idle in the mean time, it will
9344 * set the has_blocked flag and trig another update of idle load.
9345 * Because a cpu that becomes idle, is added to idle_cpus_mask before
9346 * setting the flag, we are sure to not clear the state and not
9347 * check the load of an idle cpu.
9349 WRITE_ONCE(nohz.has_blocked, 0);
9352 * Ensures that if we miss the CPU, we must see the has_blocked
9353 * store from nohz_balance_enter_idle().
9357 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9358 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9362 * If this CPU gets work to do, stop the load balancing
9363 * work being done for other CPUs. Next load
9364 * balancing owner will pick it up.
9366 if (need_resched()) {
9367 has_blocked_load = true;
9371 rq = cpu_rq(balance_cpu);
9373 has_blocked_load |= update_nohz_stats(rq, true);
9376 * If time for next balance is due,
9379 if (time_after_eq(jiffies, rq->next_balance)) {
9382 rq_lock_irqsave(rq, &rf);
9383 update_rq_clock(rq);
9384 cpu_load_update_idle(rq);
9385 rq_unlock_irqrestore(rq, &rf);
9387 if (flags & NOHZ_BALANCE_KICK)
9388 rebalance_domains(rq, CPU_IDLE);
9391 if (time_after(next_balance, rq->next_balance)) {
9392 next_balance = rq->next_balance;
9393 update_next_balance = 1;
9397 /* Newly idle CPU doesn't need an update */
9398 if (idle != CPU_NEWLY_IDLE) {
9399 update_blocked_averages(this_cpu);
9400 has_blocked_load |= this_rq->has_blocked_load;
9403 if (flags & NOHZ_BALANCE_KICK)
9404 rebalance_domains(this_rq, CPU_IDLE);
9406 WRITE_ONCE(nohz.next_blocked,
9407 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
9409 /* The full idle balance loop has been done */
9413 /* There is still blocked load, enable periodic update */
9414 if (has_blocked_load)
9415 WRITE_ONCE(nohz.has_blocked, 1);
9418 * next_balance will be updated only when there is a need.
9419 * When the CPU is attached to null domain for ex, it will not be
9422 if (likely(update_next_balance))
9423 nohz.next_balance = next_balance;
9429 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9430 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9432 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9434 int this_cpu = this_rq->cpu;
9437 if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
9440 if (idle != CPU_IDLE) {
9441 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9446 * barrier, pairs with nohz_balance_enter_idle(), ensures ...
9448 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9449 if (!(flags & NOHZ_KICK_MASK))
9452 _nohz_idle_balance(this_rq, flags, idle);
9457 static void nohz_newidle_balance(struct rq *this_rq)
9459 int this_cpu = this_rq->cpu;
9462 * This CPU doesn't want to be disturbed by scheduler
9465 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
9468 /* Will wake up very soon. No time for doing anything else*/
9469 if (this_rq->avg_idle < sysctl_sched_migration_cost)
9472 /* Don't need to update blocked load of idle CPUs*/
9473 if (!READ_ONCE(nohz.has_blocked) ||
9474 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
9477 raw_spin_unlock(&this_rq->lock);
9479 * This CPU is going to be idle and blocked load of idle CPUs
9480 * need to be updated. Run the ilb locally as it is a good
9481 * candidate for ilb instead of waking up another idle CPU.
9482 * Kick an normal ilb if we failed to do the update.
9484 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
9485 kick_ilb(NOHZ_STATS_KICK);
9486 raw_spin_lock(&this_rq->lock);
9489 #else /* !CONFIG_NO_HZ_COMMON */
9490 static inline void nohz_balancer_kick(struct rq *rq) { }
9492 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9497 static inline void nohz_newidle_balance(struct rq *this_rq) { }
9498 #endif /* CONFIG_NO_HZ_COMMON */
9501 * idle_balance is called by schedule() if this_cpu is about to become
9502 * idle. Attempts to pull tasks from other CPUs.
9504 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
9506 unsigned long next_balance = jiffies + HZ;
9507 int this_cpu = this_rq->cpu;
9508 struct sched_domain *sd;
9509 int pulled_task = 0;
9513 * We must set idle_stamp _before_ calling idle_balance(), such that we
9514 * measure the duration of idle_balance() as idle time.
9516 this_rq->idle_stamp = rq_clock(this_rq);
9519 * Do not pull tasks towards !active CPUs...
9521 if (!cpu_active(this_cpu))
9525 * This is OK, because current is on_cpu, which avoids it being picked
9526 * for load-balance and preemption/IRQs are still disabled avoiding
9527 * further scheduler activity on it and we're being very careful to
9528 * re-start the picking loop.
9530 rq_unpin_lock(this_rq, rf);
9532 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
9533 !this_rq->rd->overload) {
9536 sd = rcu_dereference_check_sched_domain(this_rq->sd);
9538 update_next_balance(sd, &next_balance);
9541 nohz_newidle_balance(this_rq);
9546 raw_spin_unlock(&this_rq->lock);
9548 update_blocked_averages(this_cpu);
9550 for_each_domain(this_cpu, sd) {
9551 int continue_balancing = 1;
9552 u64 t0, domain_cost;
9554 if (!(sd->flags & SD_LOAD_BALANCE))
9557 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
9558 update_next_balance(sd, &next_balance);
9562 if (sd->flags & SD_BALANCE_NEWIDLE) {
9563 t0 = sched_clock_cpu(this_cpu);
9565 pulled_task = load_balance(this_cpu, this_rq,
9567 &continue_balancing);
9569 domain_cost = sched_clock_cpu(this_cpu) - t0;
9570 if (domain_cost > sd->max_newidle_lb_cost)
9571 sd->max_newidle_lb_cost = domain_cost;
9573 curr_cost += domain_cost;
9576 update_next_balance(sd, &next_balance);
9579 * Stop searching for tasks to pull if there are
9580 * now runnable tasks on this rq.
9582 if (pulled_task || this_rq->nr_running > 0)
9587 raw_spin_lock(&this_rq->lock);
9589 if (curr_cost > this_rq->max_idle_balance_cost)
9590 this_rq->max_idle_balance_cost = curr_cost;
9594 * While browsing the domains, we released the rq lock, a task could
9595 * have been enqueued in the meantime. Since we're not going idle,
9596 * pretend we pulled a task.
9598 if (this_rq->cfs.h_nr_running && !pulled_task)
9601 /* Move the next balance forward */
9602 if (time_after(this_rq->next_balance, next_balance))
9603 this_rq->next_balance = next_balance;
9605 /* Is there a task of a high priority class? */
9606 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
9610 this_rq->idle_stamp = 0;
9612 rq_repin_lock(this_rq, rf);
9618 * run_rebalance_domains is triggered when needed from the scheduler tick.
9619 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9621 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9623 struct rq *this_rq = this_rq();
9624 enum cpu_idle_type idle = this_rq->idle_balance ?
9625 CPU_IDLE : CPU_NOT_IDLE;
9628 * If this CPU has a pending nohz_balance_kick, then do the
9629 * balancing on behalf of the other idle CPUs whose ticks are
9630 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9631 * give the idle CPUs a chance to load balance. Else we may
9632 * load balance only within the local sched_domain hierarchy
9633 * and abort nohz_idle_balance altogether if we pull some load.
9635 if (nohz_idle_balance(this_rq, idle))
9638 /* normal load balance */
9639 update_blocked_averages(this_rq->cpu);
9640 rebalance_domains(this_rq, idle);
9644 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9646 void trigger_load_balance(struct rq *rq)
9648 /* Don't need to rebalance while attached to NULL domain */
9649 if (unlikely(on_null_domain(rq)))
9652 if (time_after_eq(jiffies, rq->next_balance))
9653 raise_softirq(SCHED_SOFTIRQ);
9655 nohz_balancer_kick(rq);
9658 static void rq_online_fair(struct rq *rq)
9662 update_runtime_enabled(rq);
9665 static void rq_offline_fair(struct rq *rq)
9669 /* Ensure any throttled groups are reachable by pick_next_task */
9670 unthrottle_offline_cfs_rqs(rq);
9673 #endif /* CONFIG_SMP */
9676 * scheduler tick hitting a task of our scheduling class.
9678 * NOTE: This function can be called remotely by the tick offload that
9679 * goes along full dynticks. Therefore no local assumption can be made
9680 * and everything must be accessed through the @rq and @curr passed in
9683 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9685 struct cfs_rq *cfs_rq;
9686 struct sched_entity *se = &curr->se;
9688 for_each_sched_entity(se) {
9689 cfs_rq = cfs_rq_of(se);
9690 entity_tick(cfs_rq, se, queued);
9693 if (static_branch_unlikely(&sched_numa_balancing))
9694 task_tick_numa(rq, curr);
9698 * called on fork with the child task as argument from the parent's context
9699 * - child not yet on the tasklist
9700 * - preemption disabled
9702 static void task_fork_fair(struct task_struct *p)
9704 struct cfs_rq *cfs_rq;
9705 struct sched_entity *se = &p->se, *curr;
9706 struct rq *rq = this_rq();
9710 update_rq_clock(rq);
9712 cfs_rq = task_cfs_rq(current);
9713 curr = cfs_rq->curr;
9715 update_curr(cfs_rq);
9716 se->vruntime = curr->vruntime;
9718 place_entity(cfs_rq, se, 1);
9720 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9722 * Upon rescheduling, sched_class::put_prev_task() will place
9723 * 'current' within the tree based on its new key value.
9725 swap(curr->vruntime, se->vruntime);
9729 se->vruntime -= cfs_rq->min_vruntime;
9734 * Priority of the task has changed. Check to see if we preempt
9738 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9740 if (!task_on_rq_queued(p))
9744 * Reschedule if we are currently running on this runqueue and
9745 * our priority decreased, or if we are not currently running on
9746 * this runqueue and our priority is higher than the current's
9748 if (rq->curr == p) {
9749 if (p->prio > oldprio)
9752 check_preempt_curr(rq, p, 0);
9755 static inline bool vruntime_normalized(struct task_struct *p)
9757 struct sched_entity *se = &p->se;
9760 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9761 * the dequeue_entity(.flags=0) will already have normalized the
9768 * When !on_rq, vruntime of the task has usually NOT been normalized.
9769 * But there are some cases where it has already been normalized:
9771 * - A forked child which is waiting for being woken up by
9772 * wake_up_new_task().
9773 * - A task which has been woken up by try_to_wake_up() and
9774 * waiting for actually being woken up by sched_ttwu_pending().
9776 if (!se->sum_exec_runtime ||
9777 (p->state == TASK_WAKING && p->sched_remote_wakeup))
9783 #ifdef CONFIG_FAIR_GROUP_SCHED
9785 * Propagate the changes of the sched_entity across the tg tree to make it
9786 * visible to the root
9788 static void propagate_entity_cfs_rq(struct sched_entity *se)
9790 struct cfs_rq *cfs_rq;
9792 /* Start to propagate at parent */
9795 for_each_sched_entity(se) {
9796 cfs_rq = cfs_rq_of(se);
9798 if (cfs_rq_throttled(cfs_rq))
9801 update_load_avg(cfs_rq, se, UPDATE_TG);
9805 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
9808 static void detach_entity_cfs_rq(struct sched_entity *se)
9810 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9812 /* Catch up with the cfs_rq and remove our load when we leave */
9813 update_load_avg(cfs_rq, se, 0);
9814 detach_entity_load_avg(cfs_rq, se);
9815 update_tg_load_avg(cfs_rq, false);
9816 propagate_entity_cfs_rq(se);
9819 static void attach_entity_cfs_rq(struct sched_entity *se)
9821 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9823 #ifdef CONFIG_FAIR_GROUP_SCHED
9825 * Since the real-depth could have been changed (only FAIR
9826 * class maintain depth value), reset depth properly.
9828 se->depth = se->parent ? se->parent->depth + 1 : 0;
9831 /* Synchronize entity with its cfs_rq */
9832 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
9833 attach_entity_load_avg(cfs_rq, se, 0);
9834 update_tg_load_avg(cfs_rq, false);
9835 propagate_entity_cfs_rq(se);
9838 static void detach_task_cfs_rq(struct task_struct *p)
9840 struct sched_entity *se = &p->se;
9841 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9843 if (!vruntime_normalized(p)) {
9845 * Fix up our vruntime so that the current sleep doesn't
9846 * cause 'unlimited' sleep bonus.
9848 place_entity(cfs_rq, se, 0);
9849 se->vruntime -= cfs_rq->min_vruntime;
9852 detach_entity_cfs_rq(se);
9855 static void attach_task_cfs_rq(struct task_struct *p)
9857 struct sched_entity *se = &p->se;
9858 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9860 attach_entity_cfs_rq(se);
9862 if (!vruntime_normalized(p))
9863 se->vruntime += cfs_rq->min_vruntime;
9866 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9868 detach_task_cfs_rq(p);
9871 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9873 attach_task_cfs_rq(p);
9875 if (task_on_rq_queued(p)) {
9877 * We were most likely switched from sched_rt, so
9878 * kick off the schedule if running, otherwise just see
9879 * if we can still preempt the current task.
9884 check_preempt_curr(rq, p, 0);
9888 /* Account for a task changing its policy or group.
9890 * This routine is mostly called to set cfs_rq->curr field when a task
9891 * migrates between groups/classes.
9893 static void set_curr_task_fair(struct rq *rq)
9895 struct sched_entity *se = &rq->curr->se;
9897 for_each_sched_entity(se) {
9898 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9900 set_next_entity(cfs_rq, se);
9901 /* ensure bandwidth has been allocated on our new cfs_rq */
9902 account_cfs_rq_runtime(cfs_rq, 0);
9906 void init_cfs_rq(struct cfs_rq *cfs_rq)
9908 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
9909 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9910 #ifndef CONFIG_64BIT
9911 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9914 raw_spin_lock_init(&cfs_rq->removed.lock);
9918 #ifdef CONFIG_FAIR_GROUP_SCHED
9919 static void task_set_group_fair(struct task_struct *p)
9921 struct sched_entity *se = &p->se;
9923 set_task_rq(p, task_cpu(p));
9924 se->depth = se->parent ? se->parent->depth + 1 : 0;
9927 static void task_move_group_fair(struct task_struct *p)
9929 detach_task_cfs_rq(p);
9930 set_task_rq(p, task_cpu(p));
9933 /* Tell se's cfs_rq has been changed -- migrated */
9934 p->se.avg.last_update_time = 0;
9936 attach_task_cfs_rq(p);
9939 static void task_change_group_fair(struct task_struct *p, int type)
9942 case TASK_SET_GROUP:
9943 task_set_group_fair(p);
9946 case TASK_MOVE_GROUP:
9947 task_move_group_fair(p);
9952 void free_fair_sched_group(struct task_group *tg)
9956 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9958 for_each_possible_cpu(i) {
9960 kfree(tg->cfs_rq[i]);
9969 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9971 struct sched_entity *se;
9972 struct cfs_rq *cfs_rq;
9975 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
9978 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
9982 tg->shares = NICE_0_LOAD;
9984 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9986 for_each_possible_cpu(i) {
9987 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9988 GFP_KERNEL, cpu_to_node(i));
9992 se = kzalloc_node(sizeof(struct sched_entity),
9993 GFP_KERNEL, cpu_to_node(i));
9997 init_cfs_rq(cfs_rq);
9998 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9999 init_entity_runnable_average(se);
10010 void online_fair_sched_group(struct task_group *tg)
10012 struct sched_entity *se;
10016 for_each_possible_cpu(i) {
10020 raw_spin_lock_irq(&rq->lock);
10021 update_rq_clock(rq);
10022 attach_entity_cfs_rq(se);
10023 sync_throttle(tg, i);
10024 raw_spin_unlock_irq(&rq->lock);
10028 void unregister_fair_sched_group(struct task_group *tg)
10030 unsigned long flags;
10034 for_each_possible_cpu(cpu) {
10036 remove_entity_load_avg(tg->se[cpu]);
10039 * Only empty task groups can be destroyed; so we can speculatively
10040 * check on_list without danger of it being re-added.
10042 if (!tg->cfs_rq[cpu]->on_list)
10047 raw_spin_lock_irqsave(&rq->lock, flags);
10048 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
10049 raw_spin_unlock_irqrestore(&rq->lock, flags);
10053 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
10054 struct sched_entity *se, int cpu,
10055 struct sched_entity *parent)
10057 struct rq *rq = cpu_rq(cpu);
10061 init_cfs_rq_runtime(cfs_rq);
10063 tg->cfs_rq[cpu] = cfs_rq;
10066 /* se could be NULL for root_task_group */
10071 se->cfs_rq = &rq->cfs;
10074 se->cfs_rq = parent->my_q;
10075 se->depth = parent->depth + 1;
10079 /* guarantee group entities always have weight */
10080 update_load_set(&se->load, NICE_0_LOAD);
10081 se->parent = parent;
10084 static DEFINE_MUTEX(shares_mutex);
10086 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10091 * We can't change the weight of the root cgroup.
10096 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10098 mutex_lock(&shares_mutex);
10099 if (tg->shares == shares)
10102 tg->shares = shares;
10103 for_each_possible_cpu(i) {
10104 struct rq *rq = cpu_rq(i);
10105 struct sched_entity *se = tg->se[i];
10106 struct rq_flags rf;
10108 /* Propagate contribution to hierarchy */
10109 rq_lock_irqsave(rq, &rf);
10110 update_rq_clock(rq);
10111 for_each_sched_entity(se) {
10112 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
10113 update_cfs_group(se);
10115 rq_unlock_irqrestore(rq, &rf);
10119 mutex_unlock(&shares_mutex);
10122 #else /* CONFIG_FAIR_GROUP_SCHED */
10124 void free_fair_sched_group(struct task_group *tg) { }
10126 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10131 void online_fair_sched_group(struct task_group *tg) { }
10133 void unregister_fair_sched_group(struct task_group *tg) { }
10135 #endif /* CONFIG_FAIR_GROUP_SCHED */
10138 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10140 struct sched_entity *se = &task->se;
10141 unsigned int rr_interval = 0;
10144 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10147 if (rq->cfs.load.weight)
10148 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10150 return rr_interval;
10154 * All the scheduling class methods:
10156 const struct sched_class fair_sched_class = {
10157 .next = &idle_sched_class,
10158 .enqueue_task = enqueue_task_fair,
10159 .dequeue_task = dequeue_task_fair,
10160 .yield_task = yield_task_fair,
10161 .yield_to_task = yield_to_task_fair,
10163 .check_preempt_curr = check_preempt_wakeup,
10165 .pick_next_task = pick_next_task_fair,
10166 .put_prev_task = put_prev_task_fair,
10169 .select_task_rq = select_task_rq_fair,
10170 .migrate_task_rq = migrate_task_rq_fair,
10172 .rq_online = rq_online_fair,
10173 .rq_offline = rq_offline_fair,
10175 .task_dead = task_dead_fair,
10176 .set_cpus_allowed = set_cpus_allowed_common,
10179 .set_curr_task = set_curr_task_fair,
10180 .task_tick = task_tick_fair,
10181 .task_fork = task_fork_fair,
10183 .prio_changed = prio_changed_fair,
10184 .switched_from = switched_from_fair,
10185 .switched_to = switched_to_fair,
10187 .get_rr_interval = get_rr_interval_fair,
10189 .update_curr = update_curr_fair,
10191 #ifdef CONFIG_FAIR_GROUP_SCHED
10192 .task_change_group = task_change_group_fair,
10196 #ifdef CONFIG_SCHED_DEBUG
10197 void print_cfs_stats(struct seq_file *m, int cpu)
10199 struct cfs_rq *cfs_rq, *pos;
10202 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
10203 print_cfs_rq(m, cpu, cfs_rq);
10207 #ifdef CONFIG_NUMA_BALANCING
10208 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10211 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10213 for_each_online_node(node) {
10214 if (p->numa_faults) {
10215 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10216 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10218 if (p->numa_group) {
10219 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
10220 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
10222 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10225 #endif /* CONFIG_NUMA_BALANCING */
10226 #endif /* CONFIG_SCHED_DEBUG */
10228 __init void init_sched_fair_class(void)
10231 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10233 #ifdef CONFIG_NO_HZ_COMMON
10234 nohz.next_balance = jiffies;
10235 nohz.next_blocked = jiffies;
10236 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);